Detection of Foot Contact Using Inertial Measurement Units in Sports Movements: A Systematic Review

Margherita Mendicino; José Miguel Palha de Araújo dos Santos; Pietro Margheriti; Stefano Zaffagnini; Stefano Di Paolo

doi:10.3390/app151810250

,

and

¹

2nd Orthopaedic and Traumatological Clinic, IRCCS Rizzoli Orthopaedic Institute, Via G Pupilli 1, 40136 Bologna, Italy

²

Faculdade de Engenharia da Universidade do Porto (FEUP), University of Porto, 4200-465 Porto, Portugal

³

Department of Biomedical and Neuromotor Sciences, University of Bologna, 40123 Bologna, Italy

^*

Author to whom correspondence should be addressed.

Appl. Sci.2025, 15(18), 10250;https://doi.org/10.3390/app151810250

This article belongs to the Special Issue Sports Biomechanics and Injury Prevention

Version Notes

Order Reprints

Featured Application

In a scientific context that strongly advocates on-field real-world analysis of sport biomechanical data, the present work systematically presents the current applications of foot contact detection algorithms and critically reviews the room for their improvement in the context of injury prevention.

Abstract

Inertial Measurement Units (IMUs) offer promising alternatives to traditional motion capture systems, especially in real-world sports scenarios. Accurate foot contact detection (FCD) is crucial for biomechanical analysis, and since on-the-field force plates are unsuitable, IMU-based FCD algorithms have been extensively investigated. However, sports activities leading to musculoskeletal injuries are multidirectional and high-dynamics in nature and FCD algorithms, which have mostly been studied in gait analysis, might sensibly worsen performance. This systematic review (PROSPERO, ID: CRD420251010584) aimed to evaluate IMU-based FCD algorithms applied to high-dynamics sports tasks, identifying strengths, limitations, and areas for improvement. A multi-database search was conducted until May 2025. Studies were included if they applied IMU-based FCD algorithms in high-dynamic movements. In total, 37 studies evaluating 71 FCD algorithms were included. Most papers focused on running, with only 3 on cut manoeuvres. Almost all studies involved healthy individuals only, and foot linear acceleration was the most inspected FCD metric. FCD algorithms demonstrated high accuracy, though speed variation impacted performance in 23/37 studies. This review highlights the lack of validated IMU-based FCD algorithms for high-dynamic sports movements and emphasizes the need for improved methods to advance sports biomechanics testing in injury prevention.

Keywords:

IMU; algorithm; on-field biomechanics; ACL; football; injury prevention

1. Introduction

High-dynamic movements are responsible for the most severe sports-related ligament, muscle, and tendon injuries [,,,]. Cutting maneuvers, sudden decelerations, and landing from jumps are common to several team sports []. These movements are typically executed at high velocity and characterized by rapid changes in speed and direction often in unpredictable conditions, thus involve significant force generation [,].

Recent studies highlighted that injuries to ligaments, e.g., anterior cruciate ligament (ACL) [] muscles, e.g., hamstrings [] and tendons, e.g., Achilles’ tendon [], occur during such movements. Thus, a deeper understanding of athletes’ motion during these sports gestures has been advocated to improve injury mitigation strategies []. For instance, biomechanical analysis of cutting manoeuvres has been adopted to depict key risk factors and inform injury prevention protocols for ACL rupture [,,].

The forces leading to such injuries are the highest at the impact on the ground. The foot contact window (or stance phase) corresponds to the moment when the greatest loads act on the whole kinetic chain. Such loads becomes particularly relevant in the presence of suboptimal body postures: e.g., an anterior foot strike combined with excessive internal tibial rotation has been shown to significantly increase ACL strain []. These altered loading patterns usually generate altered kinematics and kinetics, such as high knee abduction moments, increased knee valgus angles, and trunk misalignment [,], that could be mitigated through optimized prevention and rehabilitation strategies []. In this scenario, improving foot contact detection (FCD)—defined as the identification of the window in which the foot makes contact with the ground—could play a crucial role in biomechanical analysis of high-intensity sports movements.

Key foot contact events used in FCD include initial contact (IC), the instant when any part of the foot first establishes contact with the ground, and toe-off (TO), the moment when the foot loses contact with the ground. The total contact time, or foot contact window, defined as the duration between IC and TO, enables precise segmentation of the ground contact phase. The precise identification of such temporal markers allow to properly isolate key joint kinematics and kinetics [,]. Additionally, they serve as indirect indicators of biomechanical control: for instance, prolonged or asymmetrical contact times, or delayed IC, have been associated with increased joint loading and risky preparatory mechanisms, both considered typical injury risk patterns []. Altered foot contact has been associated with neuromuscular imbalances, such as trunk or pelvic instability, often observed at IC during high-dynamic tasks []. Accurate detection of foot contact events is therefore essential to identify high-risk movement patterns and to support effective injury prevention strategies. FCD is typically performed using gold standard laboratory motion capture systems, employing cameras and force platforms or instrumented treadmills []. Despite being adequate for the assessment of daily life motion (gait, running, sit-to-stand), laboratory equipment is inherently restricted to indoor settings and limitedly suitable for sports applications [].

Inertial Measurement Units (IMUs) emerged as an alternative to inspect sports biomechanics in ecological environment. Several validation studies demonstrated that IMU systems provide accurate kinematics for on-field sport applications [,,]. Due to the absence of force plates, IMUs require dedicated algorithms to identify the foot contact window. Over the past decades, several FCD algorithms have been developed to test patients and athletes in the wild. In linear movements, e.g., gait and running analysis, IMU-based FCD algorithms have demonstrated high reliability compared to gold-standard (e.g., r > 0.93 in gait and in running) [,,]. IMU-based FCD algorithms predominantly employ movement-specific thresholds for velocity and acceleration at different body levels (foot, shins, sternum, etc.) []. However, technical challenges may limit the transferability of FCD algorithms for linear movements to high-dynamic sports tasks, such as increased signal variability, step frequency, unpredictable movement directions []. Accordingly, IMU-based FCD algorithms validated in the context of gait analysis may not perform equally in high-dynamic movements. FCD algorithms specifically designed for sports applications are advisable, and their performance should be specifically evaluated in such conditions.

Despite the recent growth of IMUs in on-field sport testing, there is lack of knowledge regarding FCD algorithms in sport movements, as highlighted in recent state-of-the-art, systematic, and scoping reviews [,,]. The accurate detection of foot contact window would play a key role in the adoption of IMUs in sports biomechanics testing, especially when related to the identification of athletes at risk of sustaining musculoskeletal injuries [,,].

Therefore, the aim of the present systematic review was to explore different algorithms for FCD through IMUs in high-dynamic sports movements. The ultimate goal was to critically identify current algorithms’ strengths and limitations and provide insights into methodological improvements towards the information of sports injury prevention programmes.

2. Materials and Methods

2.1. Search Strategy

This systematic review was performed following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (Figure 1).

Figure 1. Flow chart describing the identification, screening, eligibility assessment, and inclusion process of the present systematic review, following the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA). A total of 328 records were identified (322 from databases and 6 from grey literature). After title and abstract screening, 275 studies were excluded, and a further 16 studies were excluded after full-text assessment. Reasons for exclusion are detailed within the diagram. In total 37 studies met all predefined inclusion criteria and were included in the systematic review.

A comprehensive literature search was conducted across four electronic databases (PubMed, Scopus, Embase, and Cochrane Library) between 5 May and 15 May 2025. These databases were selected because they collectively cover a broad range of biomedical, engineering, and sports science literature, ensuring that relevant interdisciplinary studies were captured. The search covered the full time span available for each database up to 15 May 2025, without applying restrictions on publication year, ensuring broad coverage of literature. Only articles published in English were considered. Conference abstracts, editorials and letters were excluded to provide sufficient methodological detail for quality assessment and replication. The search strategy was developed using Boolean logic and adapted as necessary for the syntax of each database. The final combined search string was as follows:

(“identification” OR “identify” OR “detection” OR “detect” OR “determine” OR “determination” OR “validation” OR “capturing” OR “develop” OR “validate”) AND (“method” OR “algorithm” OR “methods” OR “algorithms” OR “approach”) AND (“run” OR “running” OR “jog” OR “jogging” OR “sprint” OR “sprinting” OR “cut” OR “cutting” OR “high dynamic” OR “change of direction” OR “deceleration” OR “multidirectional” OR “sport” OR “football” OR “soccer” OR “frisbee” OR “basketball” OR “rugby” OR “skating”) AND (“IMU” OR “inertial measurement” OR “wearable sensor” OR “inertial sensor” OR “accelerometer” OR “wearable system”) AND (“foot contact” OR “stance phase” OR “stride” OR “ground foot contact” OR “ground contact” OR “foot-strike” OR “event” OR “initial contact” OR “step”).

All titles and abstracts were independently screened by two authors (J.M.S., P.M.) for appropriate inclusion. Conflicts regarding study inclusion were resolved by a third reviewer (M.M.) and, when necessary, a fourth (S.D.P.). The full texts of the studies were evaluated when eligibility could not be assessed from the title and abstract. Moreover, the reference section of the eligible articles was screened for other possible eligible studies. Grey literature sources were also systematically searched and assessed for eligibility to minimize publication bias and to ensure inclusion of potentially relevant technical reports (n = 6).

This systematic review was prospectively submitted to the PROSPERO international prospective register of systematic reviews (ID: CRD420251010584) in May 2025, just prior to the beginning of the literature search.

2.2. Inclusion and Exclusion Criteria

Inclusion and exclusion criteria are detailed in Table 1. The inclusion criteria covered articles reporting a fully automatic method for FCD using wearable technology (i.e., minimal or no manual detection and user interaction required), articles analysing high-dynamic movements (namely, movements performed during sports activities), articles using IMUs, and articles focusing on FCD. For this review, high-dynamic tasks were defined as “movements with rapid directional or speed changes and high biomechanical demand (e.g., cutting, sprinting, deceleration)” []. Studies were excluded if they employed semi-automatic or manual methods for identifying initial contact and toe-off events, as these approaches rely on human intervention, limit scalability, and do not meet the requirement for fully automatic detection of the foot contact window. Additionally, articles that identified only one of the two key events (initial contact or toe-off) were excluded, because both events are required to define the complete foot contact window, which is the central variable of interest of this review. Review articles and studies with topics not related to FCD (e.g., fall assessments, investigation in non-human species, etc.), were also excluded from the analysis, to maintain focus on sport-specific applications of wearable technology.

Table 1. Inclusion and exclusion criteria used in the article screening process.

2.3. Quality Assessment

The PROBAST (Prediction model Risk Of Bias ASsessment Tool) and GRADE (Grading of Recommendations, Assessment, Development and Evaluation) approaches were used to evaluate the methodological quality and certainty of evidence for each included study. PROBAST was selected because it is specifically designed for evaluating risk of bias and applicability in studies developing or validating prediction models, which is directly relevant to FCD methods. PROBAST evaluates risk of bias across four domains: participants, predictors, outcome, and analysis, and categorizes each study as low, moderate, high, or unclear risk []. GRADE was employed to complement PROBAST, as it provides a transparent and structured framework for rating the overall certainty of evidence across studies, enabling more robust conclusions regarding the strength of recommendations. GRADE assesses certainty of evidence based on study design, risk of bias, inconsistency, indirectness, imprecision, and publication bias, classifying studies into four levels: high, moderate, low, or very low certainty []. Each included study reporting a validation procedure for FCD was assessed using both tools. Assessments were conducted independently by two reviewers (S.D.P., M.M.) and any discrepancies were resolved through discussion and consensus, following standard procedures recommended by the Cochrane Handbook to enhance assessment reliability and minimize subjective bias []. To enhance transparency in the quality appraisal, inter-rater agreement was quantified. The percentage of agreement between the two authors was calculated for both PROBAST and GRADE scales, as the proportion of articles for assigned the same rating. Agreement through Cohen’s kappa was also computed to account for agreement beyond chance.

2.4. Data Extraction

Data extraction from the included studies was performed jointly by two authors (J.M.S. and P.M.) using a standardized Excel (v16, Microsoft, Redmond, WA, USA) form to ensure consistency across studies. The collected information included aim of the study, movement task and speed, target population, IMU system adopted and its sampling frequency, number of IMUs and their placement, algorithms explained, ground-truth system and its sampling frequency, results, conclusions and considerations. The definitions of each extracted data (e.g., algorithm names) and the assessment metrics (e.g., dispersion units) were presented as reported by the authors of each retrieved paper to maximize fidelity. A description of the algorithm steps was also provided, with the level of detail varying according to the depth of information provided in each article.

3. Results

After the initial search, 328 articles were considered for eligibility. According to the inclusion and exclusion criteria, 275 articles were excluded following title and abstract screening, and an additional 16 were excluded after full-text analysis (Figure 1), resulting in a total of 291 excluded articles. Consequently, 37 articles were included in this review.

The information extracted from each of the selected articles were summarised in Table 2 according to chronological date of publication. A focus on the specific FCD algorithms characteristics was provided in Table 3. All findings were graphically summarized in Figure 2, to provide a clear and immediate overview of the results obtained from the present systematic review. The figure highlights common practices as well as gaps in the literature, which are further discussed in the Discussion section.

Figure 2. Graphical summary of the systematic review findings on IMU-based foot contact detection algorithms in high-dynamic movements. The key information collected for each study included in the systematic review (movement task analysed, speed, gender, ground truth system adopted, IMU system brand, number of IMUs, IMU placement, and metrics adopted in the foot contact detection algorithms) were summarized in percentages (%) and ranges. Left panel: Most studies investigated running tasks (78%), with reported speeds ranging from 1.9 to 9 m/s; nearly half (49%) reported an effect of velocity on algorithm performance. Male participants were more frequently included (67%), while 38% of studies reported a balanced gender distribution. For ground-truth validation, the majority relied on force-sensing systems, mainly insoles (20%). Right panel: Most studies employed a single IMU (51%), placed mainly on the trunk, foot, or shank, with sampling frequencies between 50 Hz and 2400 Hz. Xsens Technologies was the most frequently adopted IMU system (used in 14% of studies). Algorithmic approaches varied, with 65% focusing on linear acceleration metrics, and only 41% providing a detailed description of the algorithm.

3.1. Movement Tasks and Speed

Overall, 28 studies investigated only running tasks [,,,,,,,,,,,,,,,,,,,,,,,,,,,], 3 focused on sprinting tasks [,,], 2 investigated both running tasks and change of direction [,], 3 focused on skating sprints [,,], and only 1 examined only change of direction tasks [].

Twelve out of the 37 articles did not provide any information on the approaching speeds reached by subjects during the movement tasks [,,,,,,,,,,,]. Among the remaining articles (25/37), the reported speed varied considerably, from speeds as low as 1.9 m/s up to 9 m/s. Among the only three articles investigating change of direction tasks, two did not report the cut angle [,], while the other one analysed both 90° and 180° cuts [].

3.2. Target Population & Sample Size

The majority of the included articles investigated healthy individuals, as explicitly stated by the authors. However, 7 out of 37 studies did not provide specific information regarding participants’ health status [,,,,,,]. Only one study employed a case–control design, comparing healthy children with those affected by Duchenne muscular dystrophy []. In terms of sample size, about two-third of the articles (24/37) included less than twenty participants [,,,,,,,,,,,,,,,,,,,,,,,]. In 21/37 articles, data were also presented as step samples [,,,,,,,,,,,,,,,,,,,,], ranging from 144 steps to 90,537 steps. In this context, a step is defined as the window between the initial contact and toe-off events of the same foot, representing one complete foot contact. The remaining studies (16/37) instead, reported the total number of trials [,,,,,,,,,,] or did not provide any specification regarding the sample size [,,,,]. Gender distribution was not reported in 3 out of 37 included studies [,,]. Among the studies that reported gender, the total sample included 466 males (67%) and 232 females (33%); 14/37 studies presented a gender distribution within a 40–60% range [,,,,,,,,,,,,,].

3.3. IMU System & Placement

The most used IMU system was the Xsens Technologies system (Movella Inc., Henderson, NV, USA), which appeared in 5/37 articles [,,,,] followed by Casio Computer Co. (Shibuya-ku, Tokyo, Japan) used in 3/37 studies [,,] and ADXRS652 from Analog Devices (Norwood, MA, USA) [,,] also used three times. Physilog (Gait Up SA, Lausanne, Switzerland) [,], Shimmer3 (Shimmer Inc., Dublin, Ireland) [,] and MPU-9150 (InvenSense, Inc., San Jose, CA, USA) [,] were all used in 2/37 studies. Sample frequency ranged from 50 Hz to 2400 Hz. Regarding IMU placement, 19/37 articles adopted a single-sensor setup [,,,,,,,,,,,,,,,,,,], 11/37 adopted a 2-sensors setup [,,,,,,,,,,], and 7/37 adopted a multiple sensors setup (3–8 IMUs) [,,,,,,]. IMUs were placed mainly on the foot (16/37), on the trunk (17/37), and on the shank (10/37).

3.4. Algorithm Structure

Among the 37 included articles, a total of 71 algorithms were described, since some described more than one algorithm for FCD. Three algorithms chose different metrics for IC and TO detection. Most of the algorithms (46/71) focused solely on linear acceleration metrics. In particular, 25/71 used foot acceleration, 8/71 used ankle acceleration, 7/71 used pelvis linear acceleration, 3/71 used shank linear acceleration and 3/71 used trunk acceleration. Instead, 25/71 algorithms employed different metrics for FCD, primarily angular velocity-based metrics, such as foot velocity (10/24), shank velocity (5/24), ankle velocity (2/24), pelvis velocity (2/24), trunk velocity (1/24), but also foot angular acceleration (2/24), pelvis linear velocity (1/24), foot inclination angle (1/24) and knee flexion angle (1/24).

Regarding the level of detail provided, 19/71 provided only a general overview of the method used and inspected variables [,,,,,,,]. About the remaining articles, 23/71 algorithms included only a brief descriptions of the method(s) [,,,,,,,,,,,,,,,], while 29/71 provided very detailed explanations, with step-by-step descriptions of the presented algorithm(s) [,,,,,,,,,,,,]. All algorithms’ details are disclosed in Table 3.

3.5. Ground-Truth Data

To obtain ground-truth data for algorithms validation, a variety of systems were employed across the reviewed studies. Seven articles used multiple reference systems [,,,,,,]. Among the 37 articles, 22 relied on force-sensing systems, including insole force sensors (n = 9) [,,,,,,,,], embedded force platforms (n = 7) [,,,,,,], and instrumented treadmills (n = 6) [,,,,,]. The remaining studies adopted alternative methods for collecting ground-truth data: 8 articles used consumer-grade cameras [,,,,,,,], 6 employed high-speed camera systems [,,,,,], and 5 utilized photoelectric bars [,,,,]. Additionally, 2 studies used different algorithms (either embedded in smartphones [], sourced from literature [] or using Kinovea software (v0.9.5, 2019) []) and 1 employed an infrared camera [].

3.6. Main Results & Effect of Speed

Different validation metrics were adopted across the reviewed articles. The most commonly used was the offset in milliseconds from ground-truth, followed by Mean Absolute Error (MAE), Interclass Correlation Coefficient (ICC), Mean Absolute Deviation (MAD), Root Mean Square Error (RMSE) and Pearson’s Correlation Coefficient (r). Each study defined its specific threshold for interpreting algorithm performance. According to each threshold, all FCD algorithms showed low error, high correlation coefficients, and low RMSE. Not all articles presented the FCD results of both IC and TO.

To facilitate a direct comparison between studies reporting similar metrics, all results have been systematically organized into five separate tables (Table 4, Table 5, Table 6, Table 7 and Table 8), each corresponding to a specific validation metric. In particular, studies were grouped based on MAE, MAD, RMSE, systematic error measures (offset, bias, Limits of Agreement), and reliability/agreement measures (ICC, Pearson’s r, accuracy, Mean Absolute Percentage Error).

Across the reviewed studies, algorithms consistently showed better performance in detecting IC compared to TO, as reflected by lower error values across multiple metrics including MAE, MAD, and RMSE. Derived parameters such as contact time (CT) and flight time (FT), which depend on both IC and TO, exhibited higher RMSE values, compared to discrete event detection. Task complexity influenced algorithm accuracy, with more variable movement signals generally associated with increased errors. Moreover, performance tended to decrease at higher movement speeds, as observed consistently across error measures and reliability metrics. Specifically, an effect of speed on algorithms’ performance was highlighted in 18/37 articles [,,,,,,,,,,,,,,,,,], with most articles describing a decrease in overall FCD algorithm performance at increasing speeds. It should be noted, however, that few articles reported increased outcomes at higher speeds.

Several algorithms demonstrated near-zero bias and narrow limits of agreement in systematic error assessments. Reliability and agreement measures were generally very high (often above 0.9) for IC and ground contact time (GCT), confirming robust consistency across methods. Notably, among the best-performing algorithms, Blauberger’s method [] maintained strong accuracy even under sprinting conditions, highlighting its robustness.

3.7. Machine Learning

Only 4 out of 71 algorithms adopted a machine learning or deep learning approach for FCD [,,,]. The remaining relied on heuristic, rule-based models.

Donahue & Hahn implemented a Beta Process Auto-Regressive Hidden Markov Model (BP-AR-HMM), achieving >97% accuracy and outperforming the heuristic approach (95%), though increasing computational complexity []. This was the only model capable of directly estimating the IC and TO events, form the input feature of foot angular velocity. The model was trained using the entire dataset without a traditional splitting strategy, leveraging the reversible jump Markov Chain Monte Carlo (MCMC) method to isolate and extract features, eliminating the need for traditional cross-validation procedures.

In a subsequent paper, the same author applied a bidirectional Long Short-Term Memory network (BD-LSTM), which used foot and pelvis accelerations and angular velocities as input features. The BD-LSTM was trained and evaluated using a Leave-One-Out Cross-Validation (LOOCV) strategy to assess generalization across participants. The model achieved RMSE between 14 and 59 ms in gait event detection, with reduced performance at higher speeds [].

Similarly, Bach et al., proposed a machine learning model that achieved high accuracy, with mean absolute errors of approximately 21–29 ms []. Their approach employed a reservoir computer trained to predict vertical ground reaction forces from shank accelerometer data. The reservoir computer was configured with 1000 nodes, a spectral radius of 0.5, an input scaling range between 0.1 and 0.5, a leak rate of 1, a damping factor of 0.5, and a small amount of noise added during training to minimize overfitting. The model was trained using two data splitting strategies: (1) a 50% training, 25% validation, and 25% test split with 100 repetitions using random sampling, and (2) a Leave-M-out cross-validation.

Lucot et al. developed a deep learning step counting algorithm based on a Recurrent Neural Network (RNN) architecture using Long Short-Term Memory (LSTM) layers. The model reached a mean absolute percentage error below 1% (MAPE = 0.75) for step count []. The input data were accelerometer and gyroscope signals. For training and evaluation, a 7-fold cross-validation was used. The hyperparameters tested during model optimization included: number of LSTM units (16–128, nominal 64), LSTM layers (1–3, nominal 2), dropout rate (0.1–0.5, nominal 0.4), input window size (25–250 ms, nominal 50 ms), activation function (tanh), learning rate (10⁻²–10⁻⁴, nominal 10⁻³), batch size (64–2048, nominal 1024), and optional dense layer configuration.

For clarity and completeness, model type, input feature(s), output and data splitting strategies are also specified in Table 2 for these four algorithms.

3.8. Quality Assessment

Most algorithms (30/37) demonstrated a low risk of bias according to PROBAST, particularly in studies employing foot- or shank-mounted inertial measurement units validated against force plates or 3D motion capture systems. Only 2/37 (older or early-phase) studies were rated as high risk due to limited validation data or unclear outcome definitions. Regarding GRADE, a substantial portion of studies (20/37) reached moderate certainty of evidence, reflecting consistent but sometimes limited generalizability or applicability to real-world scenarios. This mainly included studies providing good algorithm performance but limited ecological testing or restricted cohorts. High-certainty evidence was assigned to studies with robust validation methods and strong agreement with gold-standard systems (10/37). In contrast, low-certainty ratings (7/37) were typically due to small sample sizes, narrow participant characteristics, or limited test conditions. These assessments indicate that while many methods are promising and well-validated under controlled conditions, only a small number have been validated in diverse or ecological scenarios (Table A1). The level of agreement between the authors who independently assessed the studies was 81.1% for PROBAST (Cohen’s kappa = 0.59) and 75.7% for GRADE (Cohen’s kappa = 0.58), indicating moderate inter-rater reliability.

Table 2. Summary of articles on “IMU-based foot contact detection algorithms”. The articles included in the systematic review were presented according to different data extracted: aim of the study, movement task(s) inspected, movement speed, target population, sample size, IMU System (brand), sampling frequency, number of IMUs adopted, IMU placement, algorithm details, ground truth system adopted for validation, study results, study conclusions and considerations. IMU—inertial measurement unit; M—male; F—female; Fs—sample frequency; IC—initial contact; TO—toe off; CT—contact time; ST—stride time; FT -flight time; SL—stride length; GCT—ground-contact time; COD—change of direction; RMSE—root mean square error; MAD—mean absolute deviation; MAE—mean absolute error; MAPE—mean absolute percentage error; IQR—interquartile range; LoA—Limits of Agreement; SD—standard deviation; SE—standard error; CI—confidence interval; RAME—relative absolute mean error; TEE—typical error of the estimate; VT—vertical; AP—anteroposterior; ML—mediolateral.

Reference	Year	Aim	Movement Task, Speed	Target Population, Sample	IMU System, Frequency	nº of IMUs, Placement	Algorithms	Ground Truth, Frequency	Results	Conclusions & Considerations
Purcell []	2006	Develop and validate an accelerometer-based method for estimating foot-ground contact time (GCT) during the acceleration phase of maximal sprinting and during constant speed running	(1) Steady-state running (2) Maximal sprint (self-selected speeds)	n = 6 (Gender not reported) Healthy subjects Samples 114 trials	Analog Devices (Norwood, MA, USA) Fs = 250 Hz	Right shank n = 1 pair of triaxial accelerometers	IC: Resultant and ML Shank Accelerations TO: ML and AP Shank Accelerations	Piezoelectric force plate (Kistler, Winterthur, Switzerland) Fs = 1000 Hz	MAE, r for contact time (ms) (1) Steady-state running MAE_jog = 0 ± 12 ms MAE_run = −2 ± 3 ms MAE_sprint = −1 ± 1 ms r = [0.892; 0.997] (2) Acceleration phase of maximal sprinting MAE_step1 = −8 ± 9 ms MAE_step3 = −2 ± 5 ms MAE_step5 = 0 ± 1 ms r = [0.951; 0.991]	Conclusions Body-mounted accelerometers provide close estimation of foot-GCT during running Considerations Best estimates achieved at higher speeds, which were also associated with highest accelerations
Lee J. []	2010	Determine the agreement between an inertial sensor and a standard method for measuring running gait, analyzing stride, step, and stance durations at increasing velocity	Running low (2.8 to 3.3 m/s) medium (3.6 to 4.2 m/s) high (4.4 to 5.3 m/s)	n = 10 (6 M; 4 F) Healthy elite runners Samples 504 steps	MiniTraqua Version 1 (Australian Institute of Sport, ACT, Australia) Fs = 100 Hz	Sacrum n = 1	IC and TO: Sacrum Anterior–posterior Acceleration	Infrared camera, (Qualisys Medical AB, Gothenburg, Sweden) Fs = 500 Hz	Stride, Step, Stance at different speeds: CI 95% = [−24; 23] ms SE = 0.4 ms–0.9 ms r >= 0.90 (most measures) r = 0.78 (medium velocity step) r = 0.76 (high velocity step)	Conclusions Inertial sensors are suitable for measuring stride, step, and stance duration Considerations No statistically significant change in systems’ agreement at different velocities.
Bergamini []	2012	Identify consistent features in signals from a single trunk-mounted IMU to detect foot-strike and foot-off instants, and to estimate stance and stride duration during the maintenance phase of sprint running	Sprint running Speeds: 5.7 to 7.4 m/s for amateurs 9.7 to 10.8 m/s for elite athletes	(A) n = 5 (3 M; 2 F) Amateurs Samples 30 strides (E) n = 6 (3 M; 3 F) Elite athletes Samples 36 strides	FreeSense (Sensorize, Rome, Italy) Fs = 200 Hz	Lower back trunk (L1) n = 1	IC and TO: Second Derivative of Trunk Angular Velocity	(A) 6 force platforms (Kistler, Winterthur, Switzerland) Fs = 200 Hz (E) High speed video recording Exilim EX-F1 (Casio Computer Co., Ltd., Tokyo, Japan) Fs = 300 FPS	MAE = 0.005 s with a 95% LoA < 0.025 s for both (A) and (E) between IMU and relative reference system	Conclusions The IMU is suitable for reliably estimating stance and stride duration during sprint running Considerations The peak of angular velocity proved to be consistent across levels of expertise, athletes, and trials
Sinclair []	2013	Determine if specific gait events could be accurately and consistently identified using an accelerometer mounted to the distal tibia, by comparing the predicted events to those detected using force platform data	Overground running (4.0 m/s ± 5%)	n = 16 (11 M; 5 F) Healthy volunteers Samples 160 trials	ACL 300 (Biometrics, Newport, UK) Fs = 1000 Hz	Antero-medial aspect of the distal tibia (8 cm above the medial malleolus) n = 1	IC and TO: Tibial VT Acceleration	Piezoelectric force platform, Model 9281CA (Kistler, Winterthur, Switzerland) Fs = 1000 Hz	Average error (mean, CI): IC = 1.68 [−2.94; 6.25] TO = −3.59 [−5.4; 1.78] Absolute error (mean, CI): IC = 5.46 [1.89; 9.03] TO = 5.00 [3.49; 8.53]	Conclusions Shank-mounted accelerometers can be used to accurately and reliably detect gait events Considerations The paper highlights the influence of speed as an area for future research.
Fortune []	2014	Validate step counts and cadence calculations from tri-axial accelerometer data during dynamic activity	Walking and jogging in a straight line	n = 12 (3 M; 9 F) Healthy subjects Samples 105 trials	Custom-built Activity Monitoring System (AMS) incorporating tri-axial accelerometer Fs = 100 Hz	waist, right thigh, and ankles (left, right). n = 4	IC: Right ankle’s AP Acceleration TO: Not reported	Video recording with a handheld camera Fs = 60 Hz	Inter-rater reliability IQR = 92% (8%) ICC(A,1) > 0.97 (walking and jogging) Median sensitivity walking, IQR = 91% (5%) jogging, IQR = 97% (6%)	Conclusions The methods are suitable for step counting and cadence measurement using the specified accelerometer placements in a free-living environment. Considerations The adaptive threshold method is robust across a wide range of gait velocities.
Harrison & Whelan []	2015	Develop a method for identifying peak impact accelerations in the anterior–posterior axis using IMUs during running and compare this with initial contact via force plates	Running (50% of maximal effort)	n = 7 (3 M; 4 F) Sprinters Samples 380 steps	Delsys Trigno (Delsys, Natick MA, USA) Fs = 148.15 Hz	Anterior tibia n = 1	IC and TO: Tibial AP Acceleration	AMTI force plate (Advanced Mechanical Technology, Inc., Watertown, MA, USA) Fs = 1000 Hz	MAD = [−0.017; 0.015] s ICC > 0.99 R > 0.99	Conclusions A single accelerometer on the anterior tibia provides an effective and reliable technique to identify key events in the running gait cycle Considerations The level of agreement between force plate and accelerometer methods increased with running speed
Gindre []	2015	Assessing the reliability and validity of the Myotest accelerometer-based system for measuring running stride kinematics	Running 60 m (12, 15, 18 and 21 km/h)	n = 20 (20 M; 0 F) Habitual runners Samples 160 trials	Myotest (Myotest SA, Sion, Switzerland) Fs = 500 Hz	Waist (on a belt) n = 1	IC and TO: Pelvis VT Acceleration	(1) Optojump (OJ) photoelectric system (Microgate, Bolzano, Italy) Fs = 1000 Hz (2) Video recording from 2 Casio High Speed EXILIM EX-FH25 cameras (Casio Computer Co., Ltd., Tokyo, Japan) Fs = 300 Hz	Contact time (ms) (1) IMU vs. Optojump 12 k/h: Δ(%) = 38 15 k/h: Δ(%) = 35 18 k/h: Δ(%) = 35 21 k/h: Δ(%) = 36 (2) IMU vs. cameras 12 k/h: Δ(%) = 34 15 k/h: Δ(%) = 31 18 k/h: Δ(%) = 32 21 k/h: Δ(%) = 33	Conclusions The device is reliable for measuring contact time, aerial time, and step frequency during running Considerations Further refinement is needed to improve robustness against movement artefacts and non-vertical motions, which can affect measurement quality
Lee H. []	2015	Propose a novel algorithm for robust FCD that functions irrespective of step mode and device pose in real smartphone usage environments	Walking (1.6 to 2.1 steps/s), Running (2.2 to 3.5 steps/s) Free-walking (chosen speed)	n = 30 (15 M; 15 F) Healthy adults and adolescents Samples 300 steps per subject, per task	Smartphones accelerometer Fs = 50 Hz	7 device poses: texting, calling, pocket, swinging, handbag, backpack, arm-band n = 1	IC and TO: Resultant Acceleration	2 commercial algorithms embedded in smartphones and 1 algorithm from the literature	Overall average accuracy Walking: 99.6% Running: 99.4% Free-walking: 99.3%	Conclusions The algorithm achieves high accuracy and low power consumption across various step modes and device poses, outperforming existing methods Considerations The algorithm does not consider device pose changes when the user is static, which could lead to false FCDs
Ammann []	2016	Validate ground contact time detection through a IMU worn on the shoelaces during running	Running (3.0 to 9.0 m/s)	n = 12 (7 M; 5 F) Healthy high-level running athletes Samples 144 steps	MPU-9150 (InvenSense Inc., San Jose, CA, USA) Fs = 1000 Hz	Foot (left, right) n = 2	IC and TO: Foot Acceleration	High speed camera system MarathonPro Videal AG (Niederönz, Switzerland) Fs = 1000 Hz	Ground Contact Time: Estimation −1.3 ± 6.1% ms At maximal speed ICC = 0.808 CI 95% = [0.653; 0.894] Overall ICC = 0.984 CI 95% = [0.977; 0.989]	Conclusions The algorithm seemed to be consistently valid across running speeds Considerations TO signal is less pronounced than IC signal, particularly at higher speeds, resulting in shorter CT
Stetter []	2016	Present and validate a new accelerometer-based approach for automated identification of ice hockey skating strides and a method to detect ice contact and swing phases	30 m skating sprints (time = [4.11; 5.18] s)	n = 6 (Gender not reported) Ice hockey players Samples 30 trials	Analog Devices Inc. (Norwood, MA) Fs = 2400 Hz	Right skate at the center of the chassis n = 1	IC and TO: Foot VT Acceleration Note: TO = BO, Blade-Off	Insole pressure measurement system (Pedar-X, novel GmbH, Munich, Germany) Fs = 90 Hz	Contact time Bland-Altmann LoA = [−0.0187; 00185] s Bias = −0.000093 s Stride time Bland-Altmann LoA = [−0.0168; 00152] s Bias = −0.000827 s	Conclusions The method is valid for in-field ice hockey testing of contact and stride times during forward skating Considerations The lower sampling rate and data interpolation of the reference system are limitations
Schmidt []	2016	Develop and validate a wireless sensor network-based device for detecting and monitoring stance durations during sprinting	Maximal sprints	n = 12 (10 M; 2 F) Track and field athletes Samples 380 steps	MPU-9150 (InvenSense Inc., San Jose, CA, USA) Fs = 1000 Hz	Ankle (left, right) n = 2	IC: Ankle VT Acceleration and Angular Velocity TO: Ankle VT Acceleration	Optojump (OJ) photoelectric system (Microgate, Bolzano, Italy) Fs = 1000 Hz	Detection rate: 95.7% MAD = 4.3 ms Systematic error: −2.5 ± 4.8 ms 95% LoA = [−11.8; 6.8]	Conclusions The IMU system provides reliable and accurate measurements of stance durations, making it a suitable tool for sprint diagnostics Considerations Extreme deviations in single steps and systematic errors among individuals highlight limitations, suggesting a need for individually adapted algorithms for even higher accuracy
Khandelwal & Wickström []	2017	Introduce the MAREA gait database and assess the impact of different real-world scenarios on the performance of state-of-the-art accelerometer-based gait event detection algorithms	(1) Indoor walk and run on treadmill (4 to 8 km/h) (2) Outdoor walk and run (self-selected speed)	n = 20 (12 M; 8 F) Healthy adult subjects Samples 105 trials	Shimmer3 (Shimmer Inc., Dublin, Ireland) Fs = 128 Hz	Waist, left wrist and ankles (left, right) n = 4	(1) A_JR IC and TO: Resultant Ankle Acceleration (2) A_RT IC and TO: VT and AP Ankle Acceleration (3) A_RS IC and TO: VT and AP Ankle Acceleration (4) A_MA IC and TO: Resultant Ankle Acceleration (5) A_AS IC and TO: VT and ML Ankle Acceleration (6) A_SK IC and TO: Resultant Ankle Acceleration	Piezo-electric force sensitive resistors fixed at the extreme ends of the sole of instrumented shoes Fs = 128 Hz	(1) Steady walking in a controlled indoor environment median F1 score IC = 0.98, TO = 0.94 (2) Outdoor walk and run median F1 score IC = 0.82, TO = 0.53 All evaluated algorithms displayed better performance for detecting IC compared to TO in different scenarios.	Conclusions Gait event detection methods, while performing well in controlled indoor settings, demonstrate decreased performance in real-world and high-dynamic scenarios Considerations Algorithms from (1) to (5) are sampling frequency and speed dependent (better performance at lower speeds). Algorithm (6), specifically designed for varying speeds, is more robust but still slightly declines in highly dynamic scenarios
Brahms []	2018	Test the validity of a single foot-mounted, 9-degree of freedom IMU to estimate stride length (SL) for running	Running (2.71 to 4.36 m/s)	n = 11 (7 M; 4 F) Healthy volunteers Samples 331 steps	Xsens Technologies (Movella Inc., Henderson, NV, USA) Fs = 100 Hz	Right foot n = 1	Stance: Foot Resultant Angular Velocity & Foot Resultant Acceleration IC and TO: Not reported	Optical motion capture system, (Vicon, Oxford, UK) Fs = 100 Hz	SL: ICC = 0.955 r = 0.961	Conclusions The IMU-based method accurately determines SLs at typical running speeds Considerations Accuracy may vary between runners with stable stance phase versus those with rotational motion at stance phase
Falbriard []	2018	Assess the performance of different kinematic features measured by foot-worn inertial sensors for detecting running gait temporal events	Running (2.8 m/s to 5.5 m/s)	n = 41 (28 M; 13 F) Healthy volunteers Samples: training set 4836 steps validation set 12092 steps	Physilog 4 (Gait Up, Lausanne, Switzerland) Fs = 500 Hz	Foot (left, right) n = 2	IC and TO: - Foot Pitch Angular Velocity - Foot pitch Angular Acceleration - Foot Pitch Angular Jerk (1st Derivative of Angular Acceleration) - Foot Roll Angular Velocity - Foot Angular Velocity Norm - Foot VT Acceleration - Foot AP Acceleration - Foot ML Acceleration - Foot Acceleration Norm - Foot Acceleration Norm Jerk (1st Derivative of Acceleration Norm)	Instrumented treadmill, T-170-FMT Arsalis (Louvain la-Neuve, Belgium) Fs = 1000 Hz	Most precise feature: Minimum of foot pitch angular velocity within the first and second half of a mid-swing to mid-swing cycle Inter-trial [median ± IQR] IC: 2 ± 1 ms TO: 4 ± 2 ms CT, FT, step and swing: <4 ± 3 ms	Conclusions The two minimum values of the pitch angular velocity in the first half and second half of a mid-swing to mid-swing cycle provide the best estimation of IC and TC Considerations Running speed can significantly alter the estimations
Mo & Chow []	2018	Evaluate the accuracy of three common methods used for detecting gait events during jogging and running	Jogging (3.1 ± 0.1 m/s) Running (4.1 ± 1.2 m/s)	n = 11 (7 M; 4 F) Healthy volunteers Samples 220 steps	MyoMOTION MR3 (Noraxon, Scottsdale, AZ, USA) Fs = 200 Hz	Sacrum, shank (left, right) and foot (left, right) n = 5	(1) S-method IC and TO: Foot Resultant Acceleration (2) M-method IC and TO: Shank VT Acceleration (3) L- method IC and TO: Pelvis AP Acceleration (4) MS-method IC: Foot Resultant Acceleration TO: Shank VT Acceleration	Force-platform, (Bertec Inc., Colombus, OH, USA) Fs = 2000 Hz	(1) S-method IC, MAD = 4.7 ± 4.1 ms TO, MAD = 26.3 ± 7.5 ms (2) M-method IC, MAD = 18,45 ± 8.8 ms TO, MAD = 7.0 ± 3.5 ms (3) L-method IC, MAD = 7.6 ± 3.3 ms TO, MAD = 17.7 ± 7.1 ms (4) MS-method ST, MAD = 8.95 ± 3.85 ms	Conclusions S-method was the most accurate for IC detection M-method was the most accurate for TO detection MS-method was the most accurate for ST estimation Considerations The S- and L-method affected by running speed; the S-method revealed less variance at both speeds
Benson []	2019	Provide detailed and automated gait events detection algorithms for running using accelerometers on the foot and low back and examine their accuracy across various running conditions	Running Exp1 instrumented treadmill) (2.7, 3.3, 3.6 m/s) Exp2 indoor track (preferred speed: 25% faster, 25% slower) Exp3 outdoor (preferred speed)	Exp1 n = 12 (8 M; 4 F) Healthy runners Exp2 n = 20 (10 M; 10 F) Healthy runners Exp3 n = 22 (11 M; 11 F) Healthy runners Samples 1942 steps	Shimmer3 (Shimmer Inc., Dublin, Ireland) Fs = 50, 100, 200 Hz	Dorsum of the right foot and lower back n = 2	(1) Foot Accelerometer IC and TO: Foot Resultant Acceleration (2) Back Accelerometer IC and TO: Trunk AP acceleration	Instrumented treadmill (Bertec Inc., Colombus, OH, USA) Fs = 1000 Hz	Back-Foot Difference Exp 1 IC: [0.065; 0.069] ± [0.014; 0.018] s TO: [−0.015; −0.008] ± [0.023; 0.041] s Exp 2 IC: [0.047; 0.066] ± [0.029; 0.049] s TO: [−0.079; −0.029] ± [0.061; 0.088] s Exp 3 IC: [0.074; 0.080] ± [0.032; 0.033] s TO: [−0.048; −0.015] ± [0.064; 0.067] s	Conclusions The consistency of the back-foot difference suggests the algorithms are robust across conditions Considerations Speed didn’t significantly affect the average gait detection difference, but it did impact its variability in some running conditions
Van Werkhoven []	2019	Validate a foot-mounted IMU sensor to determine foot strike angles and pattern	Treadmill run	n = 12 (Gender distribution not reported)	BioStampRC sensor (MC10, Lexington, MA, USA)	Right foot n = 1	IC: Peak of the Resultant Foot Acceleration TO: Maximal Foot Angle magnitude	High-speed camera (Sentech Technologies America, Inc, Carrollton, TX, USA) Fs = 120 Hz	Accuracy: 92.5%	Conclusions The algorithm provided accurate identification of foot strikes and foot strikes patterns (rear vs. non-rear foot) Considerations This approach underestimates the angle measured via video analysis, so improvement is needed
Fadillioglu []	2020	Investigate whether a novel rule-based gait event detection algorithm, applicable to various locomotion tasks, would provide comparable estimation accuracies as existing task-specific algorithms	Straight walking, moderate running, fast running, 90° walking turns, 45° and 90° running cuts (self-selected speeds, except fast running, which was set at 150% of moderate running speed)	n = 13 (13 M; 0 F) Healthy injury-free volunteers Samples 234 trials	Analog Devices Inc. (Norwood, MA, USA) Fs = 1500 Hz	Right shank n = 1	IC and TO: Shank AP Angular Velocity	AMTI force plates (Advanced Mechanical Technology Inc., Watertown, MA, USA) Fs = 1000 Hz	IC (average of the tasks): MAE = 11 (3) ms RAME = 3.07 (1.33)% TO (average of the tasks): MAE = 29 (11) ms RAME = 7.27 (2.92)%	Conclusions The algorithm is capable of detecting gait events for a variety of locomotion tasks Considerations RAME for IC and TO was highest for fast running or running cuts and lowest for straight walking
Aubol & Milner []	2020	Develop and validate a new technique for identifying the time of foot contact during rearfoot strike running using a single accelerometer	Running (3.0 ± 5% m/s)	n = 19 (9 M; 10 F) Healthy, injury-free runners Samples 190 foot contacts	PCB Piezotronics (Model 356A45, Depew, NY, USA) Fs = 1000 Hz	Right tibia n = 1	IC: Tibial Resultant Acceleration TO: Not reported	AMTI force plate (Advanced Mechanical Technology Inc., Watertown, MA, USA) Fs = 1000 Hz	IC: MAE = 2.3 ± 4.7 ms earlier with LoA of [−6.8; 11.5] ms	Conclusions The technique shows excellent concurrent validity for identifying foot contact during running. Considerations The algorithm’s performance for non-rearfoot strike runners needs further testing. The technique is recommended for field-based studies
Tomita []	2021	Estimate the accuracy of IMUs for phase identification of long-track speed skating, specifically focusing on the identification of foot contact and foot-off	Full-speed 1000 m speed skating trial on a 400 m ice rink	n = 12 (7 M; 5 F)) Healthy competitive athletes Samples 1036 strokes (86.3 ± 10.4 strokes per participant)	myoMOTION, (Noraxon, Scottsdale, AZ, USA) Fs = 100 Hz	Lower thorax and pelvis, and bilaterally on the thighs, shanks, and feet n = 8	(1) Acceleration Detection Method IC and TO: Foot AP Acceleration (2) Integrated Detection Method IC and TO: Foot AP Acceleration and Knee Flexion Angle	Portable foot pressure measurement system F-Scan System (TeckScan, South Boston, MA, USA) Fs = 100 Hz	Stance time detection (1) Acceleration Detection Method ICC(2,1) [95% CI] Straight: Right: 0.927 [0.906; 0.943] Left: 0.882 [0.852; 0.907] Curve: Right: 0.904 [0.875; 0.926] Left: 0.657 [0.582; 0.721] (2) Integrated Detection Method ICC(2,1) [95% CI] Straight: Right: 0.948 [0.925; 0.963] Left: 0.868 [0.834; 0.895] Curve: Right: 0.891 [0.529; 0.956] Left: 0.700 [0.633; 0.757]	Conclusions High degree of agreement between the IMU-based methods and the kinetic gold standard Considerations A limitation noted was that the IMU methods may be biased when stance time is greater and skating speed is slower
Donahue & Hahn []	2021	Compare a heuristic feature identification algorithm with output from the Beta Process Auto Regressive Hidden Markov Model (BP-AR-HMM) for the identification of gait events prior to their occurrence	Level ground walking, running ramps and stairs across speeds (0.8 m/s to 3.0 m/s)	n = 16 (8 M; 8 F) Able-bodied subjects Samples Not reported	Vicon (Centennial, CO, USA) Fs = 100 Hz Fs = 25 Hz	Dominant foot n = 1	(1) Heuristic Algorithm IC and TO: Foot Angular Velocity (2) Machine Learning Algorithm Model: unsupervised BP-AR-HMM with reversible jump MCMC Input: Foot Angular Velocity Output: IC and TO	Insole force sensors (Novel Electronics, St. Paul, MN, USA) Fs = 100 Hz	(1) Heuristic Algorithm Accuracy IC = 94.87% TO = 94.87% Temporal difference IC = 186.32 ± 86.70 ms TO = 63.96 ± 46.30 ms (2) Machine Learning Accuracy IC >97% TO >97%	Conclusions The BP-AR-HMM was more accurate but potentially more computationally intensive Considerations These models show promise for locomotion mode classification, prediction, and assistive device control
Blauberger []	2021	Describe a method for extracting the stride parameter ground contact time (GCT) from inertial sensor signals in sprinting	Maximum 50 m Sprinting Maximum 100 m Sprinting (No speed reported)	n = 5 (3 M; 2 F) Healthy elite runners Samples 889 steps	Physilog5 (Gait Up SA, Lausanne, Switzerland) Fs = 512 Hz	Foot (left, right) n = 2	IC: Foot Resultant Acceleration TO: Foot Angular Velocity	Photoelectric bars (Optogait, Microgate, Bolzano, Italy) Fs = Not reported	FCD: Correct rate = 97.1% Ground contact time: Mean Offset = 3.55 ms Total RMSE = 7.97 ms MAD = 5.46 ± 4.55 ms Bland-Altmann LoA = [−9.28; 16.14] ms	Conclusions FCD with IMU data was highly reliable; GCT deviation varied by run section Considerations Early and late sprint stages had lower GCT deviations, likely due to algorithm issues
Reenalda []	2021	Validate the method using the peak downward velocity of the pelvis, to detect initial contact at different speeds and foot strike patterns	Running (3.1, 3.6, and 4.2 m/s)	n = 20 (15 M; 5 F) Healthy runners Samples ~ 400 steps	Xsens Technologies (Movella Inc., Henderson, NV, USA) Fs = 240 Hz	Sacrum n = 1	IC: Pelvis VT Downward Velocity TO: Not applicable	Instrumented treadmill (C-Mill (ForceLink, Culemborg, the Netherlands) Fs = 1000 Hz Optical motion capture system (Vicon, Oxford, UK) Fs = 100 Hz	IC: Offset, at 4.2 m/s = 21.7 ± 0.2 ms Offset, at 3.1 m/s = 25.1 ± 0.1 ms	Conclusions The algorithm was accurate across foot strike patterns, but speed affected accuracy, with greater delays at lower speeds Considerations Offset decreased with speed increase
Young []	2022	Investigate the validity of a zero-crossing (ZC) based methodology for running gait assessment using a foot-mounted IMU	Overground and treadmill running (8, 10, 12, 14 km/h and a self-selected pace with an average of 15.1 ± 0.8 km/h)	n = 31 (20 M; 11 F) Well-trained healthy runners Samples 148 trials	AX6 (Axivity, Newcastle upon Tyne, UK) Fs = 60 Hz	Foot (left, right) n = 2	Zero-Crossing (ZC) gradient maxima algorithm IC and TO: Foot VT acceleration	Optical motion capture system (Vicon, Oxford, UK) Fs = 200 Hz	Ground Contact Time Lower range of speeds: ICC(2,1) > 0.9 Mean Error (ms) = [9; 15] Mean Error (Hz) = [2.15; 3.58] Mean Error (%) = [0.32; 2.65] Higher range of speeds: ICC(2,1) > 0.75 Mean Error (ms) = [21; 27] Mean Error (Hz) = [4.21; 5.48] Mean Error (%) = [3.94; 10.24]	Conclusions The evaluated method effectively quantifies foot strike location, pronation severity, and ground contact time Considerations The method’s deterioration at higher speeds highlights the need for new approaches in faster running analysis
Khandan []	2022	Estimate the temporal and spatial parameters of skating using wearable IMUs and validate the system against in-lab reference systems	Forward skating (0.88 to 2.63 m/s)	n = 10 (6 M; 4 F) Able-bodied subjects Samples 50 trials	Xsens Technologies (Movella Inc., Henderson, NV, USA) Fs = 100 Hz	Pelvis, shank (left, right), skate (left, right) n = 5	11 algorithms to detect IC (SS, Skate Strike) and TO (BO, Blade-Off): T1—VT Skate Acceleration T2—Horizontal Skate Acceleration T3—Skate VT Velocity T4—Norm of Skate Acceleration T5—Shank Angular Velocity T6—Norm of Shank Angular Velocity T7—Norm of Skate Acceleration T8—Norm of Skate Acceleration T9—Shank Angular Velocity T10—Norm of Skate Acceleration T11—Norm of Skate Acceleration	Insole pressure measurement system (Pedar-X, Novel, Munich, Germany) Fs = 100 Hz Optical motion capture system (Vicon, Oxford, UK) Fs = 100 Hz	Mean Error ± SD per event detection (SS, BO) SS_T1 = 0.00 ± 0.03 s BO_T1 = −0.03 ± 0.08 s SS_T2 = −0.01 ± 0.03 s BO_T2 = −0.05 ± 0.04 s SS_T3 = −0.17 ± 0.09 s BO_T3 = 0.00 ± 0.05 s SS_T4 = −0.01 ± 0.04 s BO_T4 = −0.10 ± 0.04 s SS_T5 = 0.03 ± 0.20 s BO_T5 = 0.09 ± 0.03 s SS_T6 = −0.12 ± 0.06 s BO_T6 = −0.05 ± 0.07 s SS_T7 = −0.19 ± 0.11 s BO_T7 = −0.06 ± 0.05 s SS_T8 = −0.03 ± 0.04 s BO_T8 = −0.08 ± 0.04 s SS_T9 = 0.04 ± 0.18 s BO_T9 = 0.05 ± 0.22 s SS_T10 = 0.04 ± 0.29 s BO_T10 = −0.09 ± 0.13 s SS_T11 = −0.25 ± 0.23 s BO_T11 = −0.05 ± 0.04 s	Conclusions Wearable IMUs, specifically a configuration of two on the skates and one on the pelvis, can estimate skating temporal and spatial parameters with high accuracy and precision comparable to gait analysis Considerations The most effective methods in finding SS events were T1, T2 and T4. Also T3, T2 and T1 were more effective in detecting BO events in skating
Bach []	2022	Prove that a single reservoir computer can accurately predict vertical GRF waveforms	Comfortable walking (1.24 ± 0.12 m/s) and running (2.20 ± 0.14 m/s)	n = 21 (13 M; 8 F) Healthy young adults Samples 3020 walking strides 1771 running strides	Mini wave plus, Zerowire (Cometa, Bareggio, Italy) Fs = 142.86 Hz	Shank (left, right) n = 2	Machine learning algorithm with reservoir computing Input: normalized Tibial Acceleration, Velocity and Position Output: GRFs Data splitting: (1) 50/25/25% training, validation, test sets with 100 repetitions with random sampling (2) leave-M-out cross validation IC and TO: from predicted GRFs	Instrumented dual-belt treadmill (Motek Medical BV, Culemborg, Netherlands) Fs = 1000 Hz	Prediction of vertical GRF R² = 0.96 ± 0.00 RMSE = 6.8 ± 0.3% with M = 1 (LOOCV) R² = 0.91 ± 0.12 RMSE = 9.1 ± 3.6% Gait event detection MAE_IC = 21.9 ± 6.5 ms MAE_TO =29.1 ± 16.0 ms With M = 1 (LOOCV) MAE_IC = 63.7 ± 167.1 ms MAE_TO =140.9 ± 224.3 ms	Conclusions The approach works well with a reduced set of input data, reducing computational time and model complexity Considerations Limitations include testing only on a flat horizontal track (except for stairs) and the time-consuming video analysis
Donahue & Hahn []	2022	Validate the identification of running gait events using data from Inertial Measurement Units in a semi-uncontrolled environment	Running (2.4 to 5.4 m/s)	n = 15 (9 M; 6 F) Healthy volunteers Samples Not reported	(Casio Computer Co., Ltd., Tokyo, Japan) Fs = 200 Hz	Foot (left, right) and sacrum n = 3	(1) Foot Acceleration IC: Foot Resultant Acceleration TO: Foot VT Acceleration (2) Pelvis Acceleration IC and TO: Pelvis AP acceleration	Insole force sensors(Novel Electronics, St. Paul, MN, USA) Fs = 100 Hz	(1) IC, offset: −63 to −5 ms TO, offset: −78 to 2 ms GCT: 4 ms 95% LoA [−0.005; 0.013] (2) GCT, offset: 1 ms 95% LoA of [−0.018; 0.021]	Conclusions Both algorithms were validated across speeds and skill levels; foot-mounted IMUs were more accurate than sacral-mounted IMUs Considerations The sacral IMU underestimated CT below 2.56 m/s and overestimated it above 3.16 m/s
Nazarahari []	2022	Present a novel approach for initial contact and terminal contact detection in real-time based on foot orientation	Walking on ground (0.9 ± 0.1 m/s) Walking on treadmill (1.0 ± 0.2 m/s) Running on treadmill (1.9 ± 0.4 m/s) Incline Walking on treadmill (10% slope) (1.0 ± 0.2 m/s)	n = 7 (7 M: 0 F) Healthy volunteers Samples 5555 ICs/TCs	Xsens Technologies (Movella Inc., Henderson, NV, USA) Fs = 100 Hz	Foot (left, right) n = 2	IC and TO: Foot Pitch Inclination Angle	Wearable pressure insoles (Novel, St. Paul, MN, USA) Fs = 100 Hz	IC, offset: Walking (ground), −40 ± 20 ms Walking (treadmill), −10 ± 40 ms Running (treadmill), 0 ± 30 ms Incline walk (treadmill), 0 ± 40 ms TO, offset: Walking (ground), 40 ± 30 ms Walking (treadmill), 40 ± 30 ms Running (treadmill), 60 ± 50 ms Incline walk (treadmill), 40 ± 30 ms	Conclusions The algorithm achieved 100% sensitivity and precision, correctly detecting all reference events (no false positives) Considerations Magnitude constraints on acceleration or angular velocity are motion-intensity dependent, while the foot pitch angle pattern not affected by intensity
Patoz []	2022	Propose a method that uses data recorded by a single sacral-mounted IMU to estimate ground reaction force, contact time (CT) and flight time (FT)	Running (2.5, 3.1, and 3.6 m/s)	n = 100 (73 M; 27 F) Healthy recreational runners Samples 3000 steps	Movesense (Suunto, Vantaa, Finland) Fs = 208 Hz	Sacrum n = 1	IC and TO: Sacrum VT Acceleration	Instrumented treadmill (Arsalis, Louvain la-Neuve, Belgium) Fs = 200 Hz	RMSE: Fz (GRF): 0.15 BW (6%) CT (contact time): 20 ms (8%) FT (flight time): 20 ms (18%)	Conclusions RMSEs were below the smallest real differences, indicating no clinically significant difference between the gold-standard and IMU method Considerations TO accuracy varied with speed, estimation was most accurate at 3.1 m/s
Yang []	2022	Propose intelligent analysis system from micro-Inertial Measurement Unit data, placed on ankle, to estimate contact time (CT) and flight time (FT) during running	Running (75%, 85%, and 95% of maximum speed)	n = 36 (36 M) Healthy football players Samples: training set 53,280 steps validation set 180 steps	Blue Thunder (I Measure U, Auckland, New Zealand) Fs = 500 Hz	Ankle (left, right) n = 2	(1) Accelerometer-Based Algorithm IC: Foot Resultant Acceleration TO: Foot VT Acceleration (2) Gyroscope-Based Algorithm IC and TO: Ankle Angular Rate	iPad camera system (Apple Inc., Sydney, Australia) Fs = 30 FPS	(1) IC, MAE: 4.5 ms to 6.7 ms TO, MAE: 23.8 ms to 31.4 ms (2) IC, MAE: 11.3 ms TO, MAE: 16.7 ms	Conclusions The algorithms were consistent overall, though TO detection varied at higher speeds; combining accelerometer and gyroscope ensured accuracy Considerations Accuracy and consistency decreased with speed
Apte []	2023	Develop and validate a method for detecting change of direction (COD) based on the wavelet-denoised antero-posterior acceleration signal	90° Change of direction 180° Change of direction	n = 23 (23 M) Healthy professional soccer players Samples Not reported	AdMos (Archinisis, Düdingen, Switzerland) Fs = 200 Hz	Upper back n = 1	IC and TO: Upper back AP Acceleration	GoPro Camera, Hero 5 (GoPro, San Mateo, CA, USA) Fs = 60 Hz Photocell (Microgate, Bolzano Italy) Fs = Not applicable	COD: Error (mean ± standard deviation) = − 0 ± 66 ms CI 95% = [−130; 130] ms Completion time: Error (mean ± standard deviation) = 160 ± 220 ms	Conclusions The algorithm detects COD events and phase durations in a T-test, using a single trunk-worn GNSS-IMU Considerations Foot IMUs would distinguish leg contact time and analyse ground impact transfer to the trunk
Donahue & Hahn []	2023	Estimate gait events from inertial measurement units and utilize machine learning for the estimation of ground reaction force (GRF) waveforms	Running (2.93 to 3.72 m/s)	n = 16 (8 M; 8F) Healthy runners Samples = 90,537 steps	(Casio Computer Co., Ltd., Tokyo, Japan) Fs = 200 Hz	Foot and sacrum n = 3, one on each foot and one on the sacrum	(1) Heuristic Algorithm IC: Foot Angular Velocity about the x-axis & Foot Resultant Acceleration TO: Foot VT Acceleration (2) Machine Learning Algorithm Model: BD-LSTM neural network Input: Foot and Pelvis Accelerations and Angular Velocities Output: vertical GRFs Data splitting: LOOCV IC and TO: from predicted GRFs	Insole force sensors, Novel Electronics, St. Paul, MN, USA) Fs = 100 Hz	(1) Heuristic Algorithm IC, RMSE: 11 to 51 ms TO, RMSE: 20 to 53 ms CT, RMSE: 20 to 66 ms (2) Machine Learning Algorithm IC, RMSE: 16 to 39 ms TO, RMSE: 14 to 59 ms CT, RMSE: 21 to 40 ms	Conclusions Both IMU data and machine learning algorithms accurately estimated contact time Considerations The machine learning algorithm appears to have limited transferability to a novel participant, as evidenced by the inferior performance in the estimation of kinetic variables, especially at faster running velocities
Lucot []	2024	Develop and test an original algorithm capable of discriminating steps in multi-activity scenarios using IMUs embedded into the midsole and deep learning algorithms	3 min free-walking sequence (multi-activity: walking, running, stomping, high knees, butt kicks and descending stairs)	n = 21 (13 M; 8 F) Healthy volunteers Samples 4017 steps	MetaMotionR (Mbient Lab, SanJose, CA, USA) Fs = 200 Hz	Right shoe (in the back of the sole) n = 1	Deep learning algorithm Model: LSTM neural network Input: Foot VT Acceleration, Acceleration Norm, Gyroscope Norm Output: binary classification per frame (foot-flat, not foot-flat) Data splitting: 7-fold cross validation IC and TO: from binary classification	Video recording using Xiami Mi 9T (Xiaomi Corporation, Beijing, China) Fs = 60 fps Video analysis using Kinovea software (v0.9.5, 2019)	All data MAPE_DL = 3.7% MAPE_{n° of steps} = 1.7% AccNorm + GyroNorm + AccZ MAPE_DL = 4.8% MAPE_{n° of step} = 2.1% Rest Delay = 375 ms MAPE_G = 0.75 (0.47)% Rest Delay = 1000 ms MAPE_{n° of step} = 53.02 (1.83)%	Conclusions The approach works well with a reduced set of input data, reducing computational time and model complexity Considerations Limitations include testing only on a flat horizontal track (except for stairs) and the time-consuming video analysis required for data enrichment
Ramli []	2024	Present methods for estimating foot velocity and trajectory during stair running using foot-mounted IMUs	Walking and running gait activities, including speed-calibration tests at slow walk to running speeds (SC-L1, SC-L2, SC-L3, SC-L4, and SC-L5)	n = 30 (Gender not reported) Children (15 muscular dystrophy; 15 healthy) Samples 35,531 steps	iPhone 11 (Apple Inc., Sydney, Australia) Fs = 100 Hz	Near the body’s center of mass n = 1	IC: AP Acceleration TO: estimated from IC	Video recording using GoPro camera (San Mateo, CA, USA) Fs = 30 FPS	MAPE Correlation (p-Value): 0.9987 (p < 0.0001) Mean (SD): 1.29% (4.59%)	Conclusions The technique performs an accurate FCD and estimates step length based on individual gait style Considerations The method performs well across structured and free-roaming activities, enabling the extension of gait analysis to community settings using consumer devices
Santicchi []	2024	Develop and evaluate a gyroscope-based algorithm for FCD and distance estimation using IMU equipped shin guards, applied to soccer	Walking Jogging Running Change of direction (Angle/Speed not reported)	n = 15 (15 M) Healthy skilled athletes Samples 13,202 steps	A pair of smart shin guards XSEED (Soccerment, Milan, Italy) Fs = 200 Hz	Shank (left, right) n = 2	IC and TO: Shank Pitch Angular Rate	High-quality video camera-based system (N/D) Fs = Not reported	FCD accuracy: Walking = 96.4% Jogging = 95.4% Straight line sprinting = 93.6% CCC = 0.94 CI 95% = [0.92; 0.95]	Conclusions The algorithm captured various soccer-specific movements but had some limitations Considerations Distance was overestimated at low intensity and slightly underestimated at high intensity
Miqueleiz []	2025	Assess the agreement between running stride variables measured with an inertial sensor using a specific algorithm and a floor-based photo-electric system	Overground and treadmill running at speeds (9, 15, 18, and 21 km/h)	n = 28 (14 M; 14 F) Well-trained endurance runners Samples Not reported	Xsens Technologies (Movella Inc., Henderson, NV, USA) Fs = 120 Hz	Lumbar spine level (L4-L5) n = 1	IC: Pelvis AP Acceleration TO: Pelvis AP Acceleration and VT acceleration	Optojump (OJ) photoelectric system (Microgate, Bolzano, Italy) Fs = 1000 Hz	Contact Time (CT) and Flight Time (FT) r = 0.81–0.93; TEE% = 3.2–7.5% Stride Frequency (SF), Stride Length (SL), and Stride Time (ST) r = 0.91–0.99; TEE% = 0.2–1.7%	Conclusions The MTw IMU system with its specific algorithm showed high agreement with the OJ system when the OJ used the 4_4 filter setting. Considerations Lower or higher OJ filter settings should be used cautiously, especially for CT and FT, as they may not align with values reported in literature.
Chebbi []	2025	Demonstrate the method’s accuracy in identifying the right and left-side impacts during level ground, incline, and decline runs across a range of speeds, applied in outdoor running scenarios	Indoor Running (3.16 to 4.88 m/s) Outdoor Running (3 to 5.5 m/s)	(1) Indoor n = 10 (3 M; 7 F) Healthy recreational runners Samples 13 running bouts (2) Outdoor n = 7 (4 M; 3 F) Healthy recreational runners Samples Not reported	(Casio Computer Co., Ltd., Tokyo, Japan) Indoor Fs = 200 Hz Outdoor Fs = 100 Hz	Sacrum n = 1	Right-Left identification algorithm IC: Sacrum Resultant Acceleration and Angular Velocity TO: Not reported	Force-instrumented treadmill (Bertec, Columbus, OH, USA) Fs = 1000 Hz	Accuracy in identifying the side of foot contact: Indoor Level ground run = 99.2% Incline run = 95.8% Decline run = 99.8% Outdoor Level ground run = 97.2% Incline run = 96.5% Decline run = 96.0%	Conclusions In indoor running scenarios, this method demonstrated excellent accuracy in identifying the side of foot contact Considerations A limitation of the current method is the difficulty in identifying occasional peaks in the resultant acceleration across all speed ranges

Table 3. Details of IMU-based foot contact detection algorithms utilised in the articles included in the systematic review (when description was available). IC—initial contact; TO—toe off; COD—change of direction; GRF—ground reaction force.

Reference	Detailed Algorithm
Purcell []	IC: Minima in the ML acceleration trace which corresponded to the peak of the resultant 3D acceleration TO: Mean of the indexes at which local minima in the ML acceleration and local maxima in the AP acceleration
Lee J. []	IC: Acute positive peaks in Sacrum AP Acceleration
Bergamini []	IC: Negative peaks of Second Derivative of the Angular Velocity TO: Positive peaks of Second Derivative of the Angular Velocity
Sinclair []	IC: Determined as the onset of the peak tibial shock using a threshold of zero which was employed and had to be crossed by a minimum of 20 frames to be implemented to prevent false detection TO: Determined using target pattern recognition with a 2% tolerance as the first plateau in the descent phase of the second peak of the axial tibial acceleration time curve
Fortune []	IC: (a) Filtering with a low-pass Butterworth filter at 6 Hz the foot anteroposterior acceleration signal (b) Calculate initial adaptive thresholds (c) Detect local minima and maxima peaks (d) Check if: (min. peak) > \|th₁\| & (preceding max. peak) > ((min. peak) + \|th₂\|) (e) Check if: (consecutive min. peaks) > t1 (f) If all checks passed: valid peak detection, otherwise invalid TO: Not detailed
Harrison & Whelan []	IC: Identified by the peak impact acceleration in the anterior–posterior (Z) axis TO: Indicated by a second peak acceleration forward in the anterior-posterior axis
Gindre []	Contact time is defined as the time during which the vertical force is greater than body weight, and aerial time as the time during which the vertical force is below body weight IC: Not detailed TO: Not detailed
Lee H. []	A step is defined as a peak and its adjacent valley. It uses adaptive magnitude thresholds (based on step average and step deviation) and adaptive temporal thresholds (based on the statistics of time intervals between adjacent peaks/valleys) IC: Not detailed TO: Not detailed
Ammann []	IC and TO: Not detailed
Stetter []	- A wavelet based approach was used to filter the raw signal (150 Hz low pass). - If the squared difference, between the raw signal and the low-pass filtered one, exceeded a predetermined threshold (0.03 voltage) in 1 of the 3 directions, the value in a new binary signal was set to high level. - The point in time of the previously identified peak of the stride detection, which was consistently located within the first third of the swing phase, was used as the initial start point for searching initial contact and blade-off. - To detect the beginning (ICacc) of a high-frequency section in the binary signal, a threshold of 10 high-level samples within a forward moving window (of length 30 samples) had to be reached IC: The first high-level signal in foot VT acceleration within this window was selected as ICacc TO (BO): (a) A potential blade-off frame was identified, and assigned to the first high-level signal in foot VT accelearation within a 10 samples backward moving window when more than 5 high-level samples were detected (b) If 100 of the 200 subsequent samples in positive direction were low-level samples, potential blade-off was considered as the real blade-off (BOacc)
Schmidt []	IC: Detected when ankle VT acceleration (Ax) exceeds a critical threshold (> 5 g) and angular velocity (Gz) shows a continuous slope TO: After a individually scalable dead time (usually 90 ms) starting at initial ground contact, a global minimum in ankle VT acceleration within a definable time window (usually 150 ms) is set as take-off
Khandelwal & Wickström []	IC and TO: Not detailed
Brahms []	Stance: simultaneously detect thresholds of <3.5 rad/s resultant angular velocity and 0.3–1.7 g resultant acceleration IC: Not detailed TO: Not detailed
Falbriard []	IC detection zone: period between the first zero-crossing of the pitch angular velocity and mid-stance TO detection zone: period between mid-stance and the last zero-crossing of the pitch angular velocity Mid-stance: time instant when the angular velocity norm is minimum within the 30–45% time range of each mid-swing to mid-swing cycle IC, individual rules: - Minimum of the pitch angular velocity - First zero-crossing of the pitch angular velocity - First local minimum smaller than 100°/s on the pitch angular velocity - Maximum of the pitch angular acceleration - Minimum of the pitch angular acceleration before Maximum of the pitch angular acceleration - Last zero-crossing of the pitch angular jerk before Maximum of the pitch angular acceleration - Maximum on the angular velocity norm - Maximum vertical acceleration - Maximum of the acceleration norm - Minimum of the acceleration norm before Maximum of the acceleration norm - First local minimum of the acceleration norm	TO, individual rules: - Minimum of the pitch angular velocity - First zero-crossing of the pitch angular acceleration after Minimum of the pitch angular velocity - First zero-crossing of the roll angular velocity after Minimum of the pitch angular velocity - Maximum of the angular velocity norm - First local maximum of the vertical acceleration after Minimum of the pitch angular velocity - Minimum of the longitudinal acceleration - First local maximum of the coronal acceleration after Minimum of the pitch angular velocity - Maximum of the acceleration norm - First local maximum of the acceleration norm after Minimum of the pitch angular velocity
Mo & Chow []	(1) S-method IC: Instant of peak foot-resultant acceleration TO: When foot acceleration exceeds a threshold of 2 g in the region of interest	(3) L-method IC: Instant of peak pelvis anteroposterior acceleration TO: Maximum in pelvis anteroposterior acceleration in the region of interest
Mo & Chow []	(2) M-method IC: Minimum before the positive peak shank vertical acceleration TO: Minimum of shank vertical acceleration in the region of interest	(4) MS-method IC: Instant of peak foot-resultant acceleration TO: Minimum of shank vertical acceleration in the region of interest
Benson []	(1) Foot Accelerometer Method IC: Major positive peak in foot’s resultant acceleration (min 0.5 s between peaks) TO: (a) Find lower bound: 0.1 s after IC (b) Find upper bound: midpoint of current and next IC, or last frame (c) Greatest positive peak in vertical acceleration between the bounds	(2) Back Accelerometer Method IC: (a) Find major peak in vertical acceleration (min 0.25 s between peaks) (b) Find negative peak in anteroposterior acceleration between previous and current vertical acceleration peak TO: (a) Find max peak in AP between previous and current VT peak (b) Find largest peak slope between max AP and IC
Van Werkhoven []	IC: Peak of the Resultant Foot Acceleration TO: Maximal Foot Angle magnitude
Fadillioglu []	IC: Identified by the first zero crossing point after mid-swing peak TO: (a) A 15 Hz 4th order Butterworth low pass filter was applied to the GRF and angular velocity signals (b) A complementary signal was calculated by calculating the difference signal using the unfiltered signal and the previously described 15 Hz low-pass filtered signal (c) A 10 Hz low-pass filter (2nd order Butterworth) was used to smooth the complementary signal (d) For the slow tasks: TO is identified as the minimum of complementary signal in searching window (it started at 50% of the gait cycle duration and ended at the negative peak plus 10% of the gait cycle duration) For the fast tasks: TO is identified as the maximum of complementary signal in searching window (it started at 50% of the gait cycle duration and ended at the negative peak)
Aubol & Milner []	IC: (a) The algorithm first identified peaks in the resultant jerk waveform that were within 150 ms of a peak in the resultant acceleration waveform (b) Local minima in the resultant tibial acceleration waveform that occurred in the 75 ms preceding each resultant jerk peak were then identified (c) Local minima were removed if they occurred after a peak in the resultant tibial acceleration wave form that had a frequency greater than 40 Hz or were below a prominence threshold equal to 20% of the third highest resultant acceleration peak (d) The earliest of the remaining local minima that occurs in the resultant acceleration signal prior to the peak resultant tibial acceleration was identified as IC TO: Not detailed
Tomita []	IC: Local minimum in acceleration TO: Local minimum in combined angular velocity, between the two peaks characteristic of the end of the contact phase (TO determined with the help of the ground truth data)
Donahue & Hahn []	(1) Heuristic Algorithm IC: Minimum peak in foot’s angular velocity (<−100 rad/s) TO: Maximum peak in foot’s angular velocity (>150 rad/s)	(2) Machine Learning Algorithm - Input: single stream of angular velocity data derived from a single gyroscopic sensor mounted on the dorsum of the participant’s dominant foot IC: (a) For most locomotion modes, impending IC was identified as the first occurrence of a state that was not state 5 or state 6 (b) For running, the rule differed slightly, identifying IC as the first instance when the state was not state 13 TO: (a) The first occurrence of either state 2 or state 1 If neither state 2 nor state 16 was observed, the first occurrence of state 6 was used instead to identify impending TO
Blauberger []	IC: Local minimum in acceleration TO: Local minimum in combined angular velocity, between the two peaks characteristic of the end of the contact phase (TO determined with the help of the ground truth data)
Reenalda []	IC: minimum of the integral of linear acceleration in the global vertical axis of the IMU on the sacrum
Young []	IC: Peaks are found with the zero-crossing gradient maxima peak detection algorithm TO: It is identified using the same ZC gradient maxima algorithm to find an inverse peak in the acceleration signal following IC
Khandan []	(1) T1-method IC (SS): The minimum of the time series occurred between two consecutive BOs TO (BO): Positive peaks of the time series	(2) T2-method IC (SS): The maximum of the time series occurred between two consecutive BOs TO (BO): The minimum of the time series occurred before the positive peaks
Khandan []	(3) T3-method IC (SS): Positive peaks of the time series TO (BO): Negative peaks of the time series	(4) T4-method IC (SS): The maximum of the time series occurred between two consecutive BOs TO (BO): The last minimum of the time series occurred before the dominant positive peaks
Bach []	IC: The foot contact was defined as the last point below a threshold (12.5% of the maximum of the data) nearest the ascend of the GRF of the stance phase TO: The foot off was defined as the first point crossing the same threshold nearest the descending GRF
Donahue & Hahn []	(1) Foot Accelerometer Method IC: (a) minimum resultant acceleration of 50 m s⁻² (b) minimum duration of 500 ms between estimated consecutive IC TO: (a) temporal window beginning 100 ms after the estimated IC, ending at the half-width of the estimated stride time (b) local maxima of vertical acceleration or the first instance when the vertical acceleration is greater than three times gravity	(2) Sacral Accelerometer Method IC: (a) local minima with a maximum value of 5 m s⁻² in the posterior direction (b) minimum temporal difference of 200 ms between the identified IC TO: (a) maximum acceleration in the anterior direction or the maximum positive slope of the acceleration in the anterior direction
Nazarahari []	TO: (a) the current sample is at least 150 ms after the previous TO (b) −sin (x) < −0.2 (corresponding to foot angle of ~11° with respect to. the initial angle, i.e., 0°) (c) the previous sample was a local minimum IC: (a) at least one TO is detected before the last IC (b) the current sample is at least 150 ms after the previous IC (c) −sin (x) > −0.2 (d) the previous sample was a local maximum Local minima and maxima: (a) −sin (x)k < −sin (x)k − 1tok-15 and −sin (x)k < −sin (x)k + 1 (b) at least 80% of −sin (x)k − i (i = 1⋯15) show a decreasing trend with respect to. −sin (x)k − i − 1 x = foot pitch angle + C18:D20
Patoz []	Approximate the vertical ground reaction force multiplying the filtered vertical acceleration signal by body mass IC and TO: detected using a 20 N threshold
Yang []	(1) Accelerometer-Based Algorithm IC: the peak of foot resultant acceleration TO: First point to exceed the threshold of 2 g in resultant acceleration, within the region of interest Region of interest: defined within the 25% to 75% range of a full stride (IC to IC), starting when the acceleration magnitude shows an upward trend that exceeds 2 g after the peak point (IC), and terminated when the signal finished with a downward trend and dropped below 2 g	(2) Gyroscope-Based Algorithm TO, conditions: - it is the local maximum - a local minimum MS is after TO - a local maximum IC is after MS - it is the local maximum between MSn and MSn − 1 MS, conditions: - it is the local minimum - MS(n,1) − MS(n − 1,1) > 300 ms IC: - a MS is identified before the location of IC - IC(n,1)–TC(n,1) > 100 ms
Apte []	Utilize macro analysis to identify the start, finish, and transient phases of the COD test; then, utilize microanalysis to accurately detect patterns within each COD phase IC: (a) Detect the first two local maxima above 4 m/s2 in the anterior–posterior acceleration signal (b) IC is defined as the local minimum preceding these peaks TO: (a) Analyze the acceleration signal following the identified IC (b) TO is defined as the first local minimum in the anteroposterior acceleration signal within the temporal window after IC
Donahue & Hahn []	(1) Heuristic Algorithm IC: - identification of minimum angular velocity about the x-axis of the IMU with a minimum of 0.500 s between identified minima - resultant acceleration > 50 m s⁻² found within the temporal window relative to each minimum, ranging from 0.005 s to 0.045 s TO: - Search for a specific temporal window beginning 0.010 s after IC and ending at the half-width of the estimated stride time - in this window, TO is either identified as the local maxima of vertical acceleration or the first instance that vertical acceleration was > 3 g	(2) Machine Learning Algorithm Input: 1 s windows of inertial data, 3-D accelerations, angular velocities, and their respective resultants, from three anatomical locations (dorsum of both feet and the sacrum) Output: 1 s intervals of estimated GRF data IC: the first instance of force > 5% BW TO: the last instance of force greater than > 5% BW
Lucot []	IC: Not detailed TO: Not detailed
Ramli []	- Applied a low-pass filter to individual participants’ data to smooth the signal and remove short-term fluctuations while preserving the longer-term trend indicating each step - Identify the peak values of the filtered signal as the peaks occur only once per step in the filtered signal. The number of peaks corresponds to the number of steps taken by the participant. IC: Maximum peak value of the anteroposterior acceleration original signal in each step window TO: Find the midpoint between two peaks in the filtered signal
Santicchi []	IC and TO: Determined by observing the angular rate of the human shank along the pitch axis Search for the maximum peak corresponding to TO, followed by detecting the Mid Swing (MSW) and completing the FCD upon identifying the IC peak Thresholds: (a) magnitude threshold of 1 rad/s to identify the IC and TO events (b) temporal threshold of 250 ms
Miqueleiz []	IC: Last positive peak in the filtered anteroposterior acceleration signal before the pronounced braking (negative AP acceleration peak) that occurred when the foot touched the ground TO: The first point where anteroposterior acceleration was negative again and vertical acceleration was below g/2.
Chebbi []	IC: (a) Sort by speed and by grade the polynomial function (b) For each speed and grade, differentiate the Sacral resultant acceleration three times to obtain a Crackle (c) Find peaks in Crackle (d) Find Sacral resultant acceleration on impact peaks in a window of 0.1 s around the Crackle (e) Compare the location of each Sacral resultant acceleration impact peak found with the closest positive or negative peaks of ω_y (f) If the closest ω_y is positive/negative, the Sacral resultant acceleration impact peak is happening during the right/left foot impact TO: Not detailed

Table 4. Summary of articles reporting Mean Absolute Error (MAE) values in milliseconds (ms). Values are reported for initial contact (IC), toe-off (TO), or ground contact time (GCT), depending on the algorithm, across different tasks and studies. Values are presented as mean (±standard deviation), unless otherwise specified. For articles reporting multiple algorithms (*), only the best-performing results were included.

Reference	Task(s)	Event	MAE [ms]
Bergamini []	running	IC & TO	5.0
Sinclair []	running	IC	5.5
Sinclair []	running	TO	5.0
Falbriard []	running	IC	[median ± IQR] 2.0 ± 1.0
Falbriard []	running	TO	[median ± IQR] 4.0 ± 2.0
Fadillioglu []	walking/running/sprinting/ 45° and 90° cuts	IC	11 ± 3.0
Fadillioglu []	walking/running/sprinting/ 45° and 90° cuts	TO	29.0 ± 11.0
Aubol & Milner []	running	IC	2.3 ± 4.7
Bach []	walking/running	IC	21.9 ± 6.5
Bach []	walking/running	TO	29.1 ± 16.0
Yang [] *	running	IC	4.5
Yang [] *	running	TO	16.7

Table 5. Summary of articles reporting Mean Absolute Deviation (MAD) in milliseconds (ms). Values are reported for initial contact (IC), toe-off (TO) or ground contact time (GCT). Values are presented as ranges or as mean (±standard deviation). For articles reporting multiple algorithms, (*) only the best-performing results were included.

Reference	Task(s)	Event	MAD [ms]
Harrison & Whelan []	running	IC & TO	[−17.0; 15.0]
Schmidt []	sprinting	IC & TO	4.3
Mo & Chow [] *	walking/jogging/running	IC	4.7 ± 4.1
Mo & Chow [] *	walking/jogging/running	TO	7.0 ± 3.5
Blauberger []	sprinting	GCT	5.5 ± 4.6

Table 6. Summary of articles reporting Root Mean Square Error (RMSE) in milliseconds (ms). Values are reported for initial contact (IC), toe-off (TO), ground contact time (GCT), or contact time (CT) an flight time (FT). For articles reporting multiple algorithms (*), only the best-performing results were included.

Reference	Task(s)	Event	RMSE [ms]
Blauberger []	sprinting	GCT	8.0
Patoz []	running	CT & FT	20.0
Donahue & Hahn [] *	running	IC	11.0
		TO	14.0
		CT	20.0

Table 7. Summary of articles reporting systematic error measures. Average offset or bias and Limits of Agreement (LoA) are reported in milliseconds (ms), for initial contact (IC), toe-off (TO), ground contact time (GCT), or foot contact window (FC). Values are presented as ranges or as mean (±standard deviation), unless otherwise specified. Where available, confidence intervals (CI) are also reported. Positive and negative values indicate delay or anticipation of the predicted events compared to the reference. For articles reporting multiple algorithms (*), only the best-performing results were included.

Reference	Task(s)	Event	Offset/Bias/LoA [ms]
Purcell []	jogging	GCT	0.0 ± 12.0
	running		−2.0 ± 3.0
	sprinting		−1.0 ± 1.0
	at 1st step of maximal sprinting		−8.0 ± 9.0
Sinclair []	running	IC	1.7 CI: [−2.9; 6.3]
Sinclair []	running	TO	−3.6 CI: [−5.4; 1.8]
Gindre []	running	GCT	Δ% = 31–38
Stetter []	running	CT	−0.1 LoA: [−18.7; 18.5]
Stetter []	running	ST	−0.8 LoA: [−16.8; 15.2]
Schmidt []	sprinting	IC & TO	−2.5 ± 4.8 LoA: [−11.8; 6.8]
Benson [] *	running	IC	[47.0; 66.0] ± [29.0; 49.0]
Benson [] *	running	TO	[−15.0; −8.0] ± [23.0; 41.0]
Donahue & Hahn [] heuristic approach	running	IC	186.3 ± 86.0
Donahue & Hahn [] heuristic approach	running	TO	64.0 ± 46.3
Blauberger []	sprinting	GCT	3.6 LoA: [−9.3; 16.1]
Reenalda []	running	IC (at 3.1 m/s)	25.1
Reenalda []	running	IC (at 4.2 m/s)	21.7
Young []	running	GCT (at higher speeds)	[9.0; 15.0]
Young []	running	GCT (at lower speeds)	[21.0; 27.0]
Khandan [] *	skating	IC	0.0 ± 30.0
Khandan [] *	skating	TO	−30.0 ± 80.0
Donahue & Hahn [] *	running	IC	[−63.0; −5.0]
		TO	[−78.0; 2.0]
		GCT	1.0 LoA: [−0.0, 0.0]
Nazarahari []	running	IC	0.0 ± 30.0
Nazarahari []	running	TO	60.0 ± 50.0
Apte []	90° and 180° change of direction	IC & TO	−0.0 ± 66.0 [−130.0; 130.0]

Table 8. Summary of articles reporting reliability and agreement measures, including intraclass correlation coefficients (ICC), Pearson’s correlation coefficients (r), accuracy (%), F1 scores, and Mean Absolute Percentage Error (MAPE). Metrics are reported for initial contact (IC), toe-off (TO), ground contact time (GCT), contact time (CT) and flight time (FT), or foot contact window (FC). For articles reporting multiple algorithms (*), only the best-performing results were included.

Reference	Task(s)	Event	ICC, r, Accuracy, MAPE
Purcell []	running/sprinting	GCT	r = [0.892–0.997]
Purcell []	running/sprinting	GCT, at 1st step of maximal sprinting	r = [0.951–0.991]
Lee J. []	running	FC (at 2.8 to 3.3 m/s)	r ≥ 0.900
Lee J. []	running	FC (at 4.4 to 5.6 m/s)	r = 0.760
Fortune []	walking/jogging	FC	ICC(A,1) > 0.97
Harrison & Whelan []	running	IC & TO	ICC > 0.99 r > 0.990
Lee H. []	running	IC & TO	Accuracy = 99.4%
Ammann []	running	GCT	ICC = 0.98
Ammann []	running	GCT, at maximal speed (9 m/s)	ICC = 0.81
Khandelwal & Wickström []	walking/running indoor	IC	F1 score = 0.98
	walking/running indoor	TO	F1 score = 0.94
	walking/running outdoor	IC	F1 score = 0.82
	walking/running outdoor	TO	F1 score = 0.53
Brahms []	running	FC	ICC = 0.96 r = 0.961
Van Werkhoven []	running	FC	Accuracy = 92.5%
Tomita [] *	skating	FC time	ICC(2,1) = 0.95
Donahue & Hahn [] *	running	IC & TO	Accuracy > 97.0%
Blauberger []	sprinting	FC	Accuracy = 97.1%
Young []	running	GCT, at lower speeds	ICC(2,1) > 0.90
Young []	running	GCT, at higher speeds	ICC(2,1) > 0.75
Lucot []	walking/running	FC	MAPE = 4.8%
Ramli []	walking/running	FC	MAPE = 1.29% r = 0.999
Santicchi []	sprinting	IC & TO	Accuracy = 93.6%
Miqueleiz []	running	CT & FT	r > 0.81
Chebbi []	running indoor	FC	Accuracy > 95.8%
Chebbi []	running outdoor	FC	Accuracy > 96.0%

4. Discussion

The present systematic review aimed to explore the current literature on IMU-based FCD algorithms for sports movements. FCD algorithms designed for sports applications were comprehensively inspected through critical aspects such as the tasks under investigation, the target population, the IMU systems adopted, the structure of each algorithm, the ground truth adopted for validation, and the main outcomes emerged. The most important finding of the review was the limited number of studies inspecting and validating FCD algorithms in sports movements (Figure 2). Although 37 studies were identified, approximately 85% (31/37, Table 2) focused on running or sprinting tasks. Only three studies investigated FCD algorithms in cutting manoeuvres or multidirectional actions. The latter movements have been associated with the occurrence of severe musculoskeletal injuries in athletic populations [,]. Of the three multidirectional tasks, two were sport-specific (skating). These findings suggest a clear lack in the current body of literature, which remains largely committed to gait, running, or (generically) linear movement analysis. The limited focus on high-dynamics sports movements reflects the complexity of accurate FCD in this context and the still limited adoption of IMUs for prevention, rehabilitation, and performance testing in the wild, despite a clear trend of interest in the last decade [,,]. To cope with sports biomechanics applications, future research is advisable to validate FCD algorithms in more complex, non-linear, sport-specific movements. The accurate detection of foot contact is pivotal to provide relevant IMU-based biomechanics in injury-inciting movements such as changes of direction and sprints with deceleration and inform injury prevention and rehabilitation strategies [,,,,].

Notably, only half of the studies reported the speed of the investigated movement tasks. Many of them evaluated only a narrow range of speeds. These studies consistently showed that FCD algorithm performance were affected by the speed of movement [,,,,,,,,]. Such an effect can be attributed to the fact that most algorithms are based on heuristic rules and rely on predefined thresholds for detecting key events such as IC and TO points. At different speeds, changes in signal amplitude and noise levels due to wobbling mass effect and drift can lead to these thresholds becoming either excessively permissive or restrictive. This critical aspect implies questioning the robustness of the FCD algorithms across varying speeds. Thresholds fluctuations likely explain the observed variations in the FCD outcomes: in the study by Reenalda et al., which evaluated an algorithm based on pelvis downward velocity, a systematic decrease in offset from ground truth was observed as speed increased []. The authors suggested that higher speeds may improve FCD []. The authors further proposed that at faster running speeds, the impact shock caused by body deceleration propagates more rapidly through the body, leading to a sharper sinusoidal vertical velocity pattern of the pelvis, which causes the IC to be detected earlier []. On the other hand, Patoz et al. found a significantly inverse effect of running speed on foot contact time was observed, with the most accurate estimation occurring at 3.1 m/s. The same authors suggest that this finding could not be readily explained and that more complex relationships exist between speed and FCD algorithm performance [].

No single FCD algorithm consistently outperforms the others across all multiple speeds. Some methods appear robust at higher speeds or in specific movement tasks, while others degrade in accuracy when applied outside their original validation parameters. For instance, approaches based on foot acceleration [,,,] showed high performance in detecting FC events at validated conditions, but their accuracy tended to deteriorate as speed increased. This reinforces the idea that algorithm performance is both speed- and task-dependent, and future research should aim to systematically identify strengths and limitations across different sport-specific contexts. Recent studies showed that the speed at which a movement is performed significantly influences the quality of the movement towards the risk of musculoskeletal injuries. As an example, the occurrence of ACL injury risk profile in 90° cut manoeuvre and deceleration tasks has been shown to increase with higher speed, which more closely resembles game situations []. Furthermore, the only three studies investigating cut manoeuvres failed to report change of direction angles. Higher cut angles have been associated with increased risk of knee overloading and ACL injury, with angles between 90° and 135° being the most dangerous ones [,]. According to the current review, it is therefore essential to report movement speed associated with a FCD algorithm’s performance and ideally test multiple speeds to ensure its robustness and capabilities to handle the complexities of high-dynamic movements.

Almost all reviewed studies exclusively examined healthy individuals. Only one study compared healthy children with those affected by Duchenne muscular dystrophy []. This finding implies that the influence of movement disorders on FCD performance has not been considered in the current literature. Biomechanical alterations have been reported in the presence of musculoskeletal disorders []. In athletes, severe injuries and surgical procedures often impair physiological foot contact [,]. Restricting the validation of FCD algorithms to healthy cohorts may limit generalizability to non-healthy individuals []. To ensure real-world applicability, future research should include both healthy and injured athletes and encompass larger sample sizes to maximize statistical and clinical relevance. Furthermore, the sample size of the studies included in the review varied consistently across studies (range 5–100 subjects). Small sample size might compromise statistical robustness and algorithms’ reliability. The only conspicuous athlete category was runners: the influence of sport specialization, e.g., football vs. basketball players, on FCD performance is therefore lacking. It has been shown that the situational patterns leading to an ACL injury are sport-specific [,,]. Dedicated, sport-specific FCD algorithms might be necessary to maximize foot window identification.

IMU system, frequency, and placement emerged distinctly from this systematic review. Various commercially available IMU systems were identified, with Xsens Technologies resulting in the most widely adopted (5/37 studies). The adoption of commercial vs. non-commercial IMU systems was overall unbalanced towards the former. This systematic review confirmed that commercial systems are relied on the most when assessing foot contact in sports context. Sample frequency also varied substantially among the studies (range 50–2400 Hz). Minimum sample frequency adopted was 50 Hz, thus not meeting the ISB recommendations from Cereatti et al., indicating that minimum 200 Hz are required for sports (running) biomechanics applications []. The most frequent IMU placement was on trunk (n = 17) and foot (n = 16). These locations allow stable positioning and reliable metrics identification, in line with what was found by Pacini Panebianco et al. for gait timing estimation and Horsley et al. for running [,]. Interestingly, despite about half of the articles adopting a multi-sensor setup (≥2 IMUs), none of them adopted the combination of metrics from multiple sensors to develop a FCD algorithm. A qualitative synthesis across subgroups, based on the GRADE scoring system, revealed that IMU placement influenced algorithm performance and methodological quality (Table A1). Algorithms using shank- or foot-mounted IMUs were more often associated with higher certainty of evidence and lower risk of bias. In contrast, trunk-mounted IMUs exhibited more variability in outcomes, particularly during more dynamic tasks or under non-standard conditions (e.g., outdoor environments).

The combination of signals from multiple sensors could potentially enhance the robustness of FCD algorithms in high-dynamic movements by capturing complementary biomechanical signals [,]. The complementary strengths of, e.g., signal from sacrum- and foot-mounted sensor units could help minimizing the motion artifacts at high-speeds and during multidirectional tasks. It should be noted that this approach may be challenging to implement in ecological, on-field settings due to increased equipment costs, complexity in data management and processing, and athlete discomfort. Several studies demonstrated that single-sensor setups can achieve satisfactory FCD results, while keeping the system more practical for day-to-day sport applications [,,]. However, when feasible, leveraging data from multiple sensors remains a promising strategy to improve detection capabilities and algorithm robustness, particularly in high-dynamic movements where signal fluctuations and environmental noise are more pronounced.

The level of detail in the description of the algorithms’ structures varied significantly across the retrieved articles. Particularly, 19/71 algorithms only provided a general description of the implemented approach. Among the wide variety of algorithm metrics retrieved, linear acceleration emerged as the most used feature. Transparency in sharing algorithmic details allows other researchers to replicate and test the algorithms, fostering continuous improvement. It might be argued that, considering the critical challenge to achieve accurate FCD, intellectual property issues may have limited the full disclosure of algorithmic structures. Within these constraints, greater transparency is advisable to promote independent validations and broader adoption of dedicated FCD algorithms.

The finding on ground truth systems adopted to assess the performance of the IMU-based algorithms was peculiar. Force-sensing systems were the most utilized reference, despite being adopted only in 22/37 of the retrieved studies. Force-sensing systems have been shown to reliably detect foot contact [], providing trustworthy reference data that enhances the credibility of the study’s findings. The remaining articles used other solutions, e.g., infrared or high-speed cameras, consumer-grade cameras, photoelectric bars. When applied to FCD in high-dynamics movements, the latter systems may not yield as reliable ground-truth data as force plates and suggest a more cautious interpretation of the validation findings []. Future studies should therefore prioritize the use of force-sensing systems as ground-truth data for high-dynamics movement tasks.

This review also suggested a low consistency in performance metrics adopted to inspect FCD algorithms’ validity. A wide variety of dispersion metrics (MAE, ICC, RMSE, MAD, r) and graphical representations (e.g., Bland-Altman plots) were inconsistently adopted across the studies. There were notable omissions and inconsistencies in the reporting of performance metrics. For instance, when reporting ICC results, essential contextual information such as the timing offset in milliseconds was often missing. These findings are consistent with previous literature regarding FCD in running, showing the importance of assessing a complete summary of outcomes and validity statistics []. On top of that, the interpretation of algorithms’ performance lacked standardization: most of the studies addressed their findings as “adequate”, despite the band scores being different even for the same metric. TO detection was often poorly defined or excluded from performance evaluations, despite its importance for accurate FCD. These omissions limit the ability to compare algorithm performance. The inclusion of clear offset values (in milliseconds) for IC and TO, reported separately and at different speeds (when available), is advisable as the most reproducible and transparent way to present FCD algorithm performance.

The analysis of study certainty (GRADE) and risk of bias (PROBAST) revealed improved overall quality over time. The oldest retrieved studies (before 2020) have been more frequently rated as low certainty and high risk of bias, often due to limited sample sizes, unclear outcome definitions, or lack of a robust validation system and method. In contrast, recent studies (post-2020) showed a higher prevalence of moderate to high certainty ratings and low risk levels, reflecting methodological improvements and standardized validation protocols. This finding underscores the positive trend in maximizing the quality of studies focusing IMU-based FCD algorithms.

In the current review, only four articles included a machine learning-based FCD algorithm, and two of them compared their machine learning models to heuristic approaches [,]. Recent advancements in machine learning tools highlight the potential for their application in IMU-based algorithms for FCD. Machine learning models can detect complex patterns in large datasets, potentially improving FCD accuracy beyond traditional heuristic rule-based models. The results of the current review suggest that machine learning solutions have not yet overtaken heuristic algorithms. Arguably, some heuristic algorithms already demonstrate very high performance while offering greater interpretability and lower computational demands. Machine learning models, although promising, are highly dependent on the availability of extensive training datasets, which remains a concern in the field of wearable devices. Most machine learning models adopt a “black box” approach, are more complex to implement, and might require significant computational processing power, making them less practical compared to heuristic models for sports applications []. Sensitivity to movement variability across different populations can further reduce generalizability, as suggested by Trabassi et al. []. Larger sample sizes and greater consistency in output metrics are needed to support the widespread application of machine learning models in high-dynamics movements [,].

The current systematic review highlighted notable interest and advancements in IMU-based FCD algorithms in recent years. A gap in the literature regarding their application to high-dynamic movements remains evident when compared to linear daily motion (e.g., gait). These aspects mark the need for further development of FCD algorithms that can handle more complex, sport-specific tasks, such as changes of direction or decelerations. Commercial IMU systems at sampling rates (at least) 100 Hz, ideally in multi-sensor configuration, validated against force platform could be suggested as the most appropriate and robust methodological features. These algorithms should be tested across a broader range of speeds, angles, sport-specific contexts, and populations. Future research should focus on addressing these gaps to enhance the accuracy and clinical utility of IMU-based FCD algorithm in biomechanics testing for sport performances, injury prevention, and return to sport after surgery. Dedicated guidelines for IMU-based FCD could increase consistency across scientific literature and promote further Level I studies.

The present review has some limitations to acknowledge. Such limitations could be summarized in three main thematic categories: population and sample constraints, methodological heterogeneity, and bias and evidence gaps. Regarding population and sample constraints, most of the included studies involved small, homogeneous cohorts and lacked non-healthy athletes, which restricted generalizability. Sport-specific and cutting maneuvers were only minimally analyzed, limiting ecological validity for the FCD algorithms and partially affecting study purpose. A further limitation to acknowledge is a potential gender bias in the retrieved studies: no study differentiated FCD algorithms outcomes for male and female, less than 40% of the studies had a nearly 1:1 gender ratio, and 3 studies did not even report the participants’ gender distribution. These aspects should be considered in future research to maximize generalizability, robustness, and inclusiveness.

With respect to methodological heterogeneity, several issues restrict the comparability and reproducibility of the findings. The limited detailing of the algorithms’ steps hindered reproducibility and transparency. The absence of a uniform gold standard (force platform) for validation may have influenced reported accuracy and biased comparability across studies. Despite an attempt to standardize reported metrics by grouping algorithms into five categories based on the validation measure used, the heterogeneity of protocols, analyzed signals, reported events, and movement tasks ultimately prevented a formal meta-analysis. These aspects constraints the current review to descriptive rather than quantitative FCD algorithms comparisons. Moreover, most of the identified articles did not strictly focus on FCD algorithms but instead evaluated IMU-based methods for related biomechanical features such as stance time, ground contact time, and flight time. Nonetheless, in all these cases, accurate foot contact detection represents a crucial part of these analyses, and biomechanical results heavily depend on it. This key aspect not only justifies the inclusion of such studies in this review but highlights the central role of FCD as a foundational element for assessing multiple other biomechanical features.

Under bias and evidence gaps, potential publication bias must be considered. The predominance of studies reporting positive findings suggests that less successful validations may have been underreported. This limitation may lead to an overestimation of algorithm performance, and ultimately of their generalizability and effectiveness. Lastly, the lack of machine learning based algorithms might derive from the exclusion of articles describing movement pattern recognition (Figure 1). However, it should be noted that the latter did not directly fit the scope of the current review in identifying the performances of key foot contact window features such as IC and TO frames.

5. Conclusions

This systematic review highlighted the strengths and limitations of the current IMU-based FCD algorithm in high-dynamics sports movements. Current FCD algorithms might lack robustness and clinical utility towards sports injury prevention. This was evidenced by the limited assessment of sport-specific tasks (e.g., cut manoeuvres), the impact of speed on algorithm performance, the inconsistency in outcome metrics adopted, and the target population (healthy athletes only). Validated FCD algorithms for high-dynamic movements are required to support the widespread use of IMUs in sports-specific biomechanics testing for injury prevention.

Author Contributions

Conceptualization, S.D.P.; methodology, J.M.P.d.A.d.S., P.M., M.M. and S.D.P.; software, J.M.P.d.A.d.S. and M.M.; writing—original draft preparation, J.M.P.d.A.d.S. and M.M.; writing—review and editing, M.M. and S.D.P.; supervision, S.Z. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Data Availability Statement

No new data were created or analysed in this study. Further inquiries can be directed to the corresponding author.

Conflicts of Interest

The authors declare no conflicts of interest.

Abbreviations

The following abbreviations are used in this manuscript:

IMUs	Inertial Measurement Units
FCD	Foot Contact Detection
ACL	Anterior Cruciate Ligament
IC	Initial Contact
TO	Toe-off
PRISMA	Preferred Reporting Items for Systematic Reviews and Meta-Analyses
PROBAST	Prediction model Risk Of Bias ASsessment Tool
GRADE	Grading of Recommendations, Assessment, Development and Evaluation
RMSE	Root Mean Square Error
MAD	Mean Absolute Deviation
MAE	Mean Absolute Error
MAPE	Mean Absolute Percentage Error
RAME	Relative Mean Absolute Error
TEE	Typical Error of the Estimate
SD	Standard Deviation
IQR	Interquartile Range
LoA	Limits of Agreement
CI	Confidence Interval
M	Male
F	Female
Fs	Sample Frequency
CT	Contact Time
ST	Stride Time
GCT	Ground Contact Time
COD	Change of Direction
FT	Flight Time
VT	Vertical
AP	Anteroposterior
ML	Mediolateral

Appendix A

Table A1. Quality assessment through PROBAST (Prediction model Risk Of Bias ASsessment Tool) and GRADE (Grading of Recommendations, Assessment, Development and Evaluation).

Study	PROBAST Risk	GRADE Certainty
Purcell []	High	Low
Lee J. []	Low	Moderate
Bergamini []	High	Low
Sinclair []	Low	Moderate
Fortune []	Low	Moderate
Harrison & Whelan []	Moderate	Low
Gindre []	Moderate	Low
Lee H. []	Low	Moderate
Ammann []	Low	High
Stetter []	Low	Moderate
Schmidt []	Low	Moderate
Khandelwal & Wickström []	Moderate	Low
Brahms []	Low	Moderate
Falbriard []	Low	High
Mo & Chow []	Low	Moderate
Benson []	Low	Moderate
Van Werkhoven []	Moderate	Low
Fadillioglu []	Low	Moderate
Aubol & Milner []	Low	Moderate
Tomita []	Low	Moderate
Donahue & Hahn []	Low	Moderate
Blauberger []	Low	Moderate
Reenalda []	Low	High
Young []	Low	Moderate
Khandan []	Low	Moderate
Bach []	Low	High
Donahue & Hahn []	Low	High
Nazarahari []	Low	High
Patoz []	Low	High
Yang []	Moderate	Low
Apte []	Low	Moderate
Donahue & Hahn []	Low	Moderate
Lucot []	Low	High
Ramli []	Low	High
Santicchi []	Low	Moderate
Miqueleiz []	Low	High
Chebbi []	Low	Moderate

References

Della Villa, F.; Buckthorpe, M.; Grassi, A.; Nabiuzzi, A.; Tosarelli, F.; Zaffagnini, S.; Della Villa, S. Systematic Video Analysis of ACL Injuries in Professional Male Football (Soccer): Injury Mechanisms, Situational Patterns and Biomechanics Study on 134 Consecutive Cases. Br. J. Sports Med. 2020, 54, 1423–1432. [Google Scholar] [CrossRef]
Tosarelli, F.; Buckthorpe, M.; Di Paolo, S.; Grassi, A.; Rodas, G.; Zaffagnini, S.; Nanni, G.; Della Villa, F. Video Analysis of Anterior Cruciate Ligament Injuries in Male Professional Basketball Players: Injury Mechanisms, Situational Patterns, and Biomechanics. Orthop. J. Sports Med. 2024, 12, 23259671241234880. [Google Scholar] [CrossRef]
Della Villa, F.; Buckthorpe, M.; Pellegrini, A.; Ranzini, A.; Esposito, F.; Crescenzo, C.; Nanni, G.; Zago, M. A Comparative Video Analysis of Hamstring Injuries Mechanism and Situational Pattern in Men’s and Women’s Football (Soccer). Knee Surg. Sports Traumatol. Arthrosc. Off. J. ESSKA 2024, 32, 2610–2621. [Google Scholar] [CrossRef]
Della Villa, F.; Buckthorpe, M.; Tosarelli, F.; Zago, M.; Zaffagnini, S.; Grassi, A. Video Analysis of Achilles Tendon Rupture in Male Professional Football (Soccer) Players: Injury Mechanisms, Patterns and Biomechanics. BMJ Open Sport Exerc. Med. 2022, 8, e001419. [Google Scholar] [CrossRef]
Sundberg, A.; Högberg, J.; Tosarelli, F.; Buckthorpe, M.; Della Villa, F.; Hägglund, M.; Samuelsson, K.; Hamrin Senorski, E. Sport-Specific Injury Mechanisms and Situational Patterns of ACL Injuries: A Comprehensive Systematic Review. Sports Med. 2025. [Google Scholar] [CrossRef]
Lloyd, D. The Future of In-Field Sports Biomechanics: Wearables plus Modelling Compute Real-Time in Vivo Tissue Loading to Prevent and Repair Musculoskeletal Injuries. Sports Biomech. 2024, 23, 1284–1312. [Google Scholar] [CrossRef]
Gokeler, A.; Grassi, A.; Hoogeslag, R.; van Houten, A.; Lehman, T.; Bolling, C.; Buckthorpe, M.; Norte, G.; Benjaminse, A.; Heuvelmans, P.; et al. Return to Sports after ACL Injury 5 Years from Now: 10 Things We Must Do. J. Exp. Orthop. 2022, 9, 73. [Google Scholar] [CrossRef]
Rodriguez-Merchan, E.C.; Encinas-Ullan, C.A. Knee Osteoarthritis Following Anterior Cruciate Ligament Reconstruction: Frequency, Contributory Elements, and Recent Interventions to Modify the Route of Degeneration. Arch. Bone Jt. Surg. 2022, 10, 951–958. [Google Scholar] [CrossRef] [PubMed]
Di Paolo, S.; Bragonzoni, L.; Della Villa, F.; Grassi, A.; Zaffagnini, S. Do Healthy Athletes Exhibit At-Risk Biomechanics for Anterior Cruciate Ligament Injury during Pivoting Movements? Sports Biomech. 2024, 23, 2995–3008. [Google Scholar] [CrossRef] [PubMed]
Silvers-Granelli, H.J.; Bizzini, M.; Arundale, A.; Mandelbaum, B.R.; Snyder-Mackler, L. Does the FIFA 11+ Injury Prevention Program Reduce the Incidence of ACL Injury in Male Soccer Players? Clin. Orthop. 2017, 475, 2447–2455. [Google Scholar] [CrossRef] [PubMed]
Xu, D.; Zhou, H.; Quan, W.; Ma, X.; Chon, T.-E.; Fernandez, J.; Gusztav, F.; Kovács, A.; Baker, J.S.; Gu, Y. New Insights Optimize Landing Strategies to Reduce Lower Limb Injury Risk. Cyborg Bionic Syst. 2024, 5, 0126. [Google Scholar] [CrossRef]
Dos’Santos, T.; Thomas, C.; McBurnie, A.; Donelon, T.; Herrington, L.; Jones, P.A. The Cutting Movement Assessment Score (CMAS) Qualitative Screening Tool: Application to Mitigate Anterior Cruciate Ligament Injury Risk during Cutting. Biomechanics 2021, 1, 83–101. [Google Scholar] [CrossRef]
Dos’Santos, T.; Thomas, C.; Comfort, P.; Jones, P.A. The Effect of Angle and Velocity on Change of Direction Biomechanics: An Angle-Velocity Trade-Off. Sports Med. 2018, 48, 2235–2253. [Google Scholar] [CrossRef]
Niswander, W.; Kontson, K. Evaluating the Impact of IMU Sensor Location and Walking Task on Accuracy of Gait Event Detection Algorithms. Sensors 2021, 21, 3989. [Google Scholar] [CrossRef]
Della Villa, F.; Di Paolo, S.; Crepaldi, M.; Santin, P.; Menditto, I.; Pirli Capitani, L.; Boldrini, L.; Ciampone, L.; Vassura, G.; Bortolami, A.; et al. Kinematics of 90° Change of Direction in Young Football Players: Insights for ACL Injury Prevention from the CUTtheACL Study on 6008 Trials. Knee Surg. Sports Traumatol. Arthrosc. Off. J. ESSKA 2024, 32, 2666–2678. [Google Scholar] [CrossRef]
Reenalda, J.; Zandbergen, M.A.; Harbers, J.H.D.; Paquette, M.R.; Milner, C.E. Detection of Foot Contact in Treadmill Running with Inertial and Optical Measurement Systems. J. Biomech. 2021, 121, 110419. [Google Scholar] [CrossRef]
Wade, L.; Needham, L.; McGuigan, P.; Bilzon, J. Applications and Limitations of Current Markerless Motion Capture Methods for Clinical Gait Biomechanics. PeerJ 2022, 10, e12995. [Google Scholar] [CrossRef]
Di Paolo, S.; Lopomo, N.F.; Della Villa, F.; Paolini, G.; Figari, G.; Bragonzoni, L.; Grassi, A.; Zaffagnini, S. Rehabilitation and Return to Sport Assessment after Anterior Cruciate Ligament Injury: Quantifying Joint Kinematics during Complex High-Speed Tasks through Wearable Sensors. Sensors 2021, 21, 2331. [Google Scholar] [CrossRef] [PubMed]
Nijmeijer, E.M.; Heuvelmans, P.; Bolt, R.; Gokeler, A.; Otten, E.; Benjaminse, A. Concurrent Validation of the Xsens IMU System of Lower-Body Kinematics in Jump-Landing and Change-of-Direction Tasks. J. Biomech. 2023, 154, 111637. [Google Scholar] [CrossRef]
Heuvelmans, P.; Benjaminse, A.; Bolt, R.; Baumeister, J.; Otten, E.; Gokeler, A. Concurrent Validation of the Noraxon MyoMotion Wearable Inertial Sensors in Change-of-Direction and Jump-Landing Tasks. Sports Biomech. 2022, 24, 184–199. [Google Scholar] [CrossRef]
Prasanth, H.; Caban, M.; Keller, U.; Courtine, G.; Ijspeert, A.; Vallery, H.; von Zitzewitz, J. Wearable Sensor-Based Real-Time Gait Detection: A Systematic Review. Sensors 2021, 21, 2727. [Google Scholar] [CrossRef]
Horsley, B.J.; Tofari, P.J.; Halson, S.L.; Kemp, J.G.; Dickson, J.; Maniar, N.; Cormack, S.J. Does Site Matter? Impact of Inertial Measurement Unit Placement on the Validity and Reliability of Stride Variables During Running: A Systematic Review and Meta-Analysis. Sports Med. Auckl. NZ 2021, 51, 1449–1489. [Google Scholar] [CrossRef]
Cowin, J.; Nimphius, S.; Fell, J.; Culhane, P.; Schmidt, M. A Proposed Framework to Describe Movement Variability within Sporting Tasks: A Scoping Review. Sports Med.-Open 2022, 8, 85. [Google Scholar] [CrossRef]
van der Kruk, E.; Reijne, M.M. Accuracy of Human Motion Capture Systems for Sport Applications; State-of-the-Art Review. Eur. J. Sport Sci. 2018, 18, 806–819. [Google Scholar] [CrossRef]
Poitras, I.; Dupuis, F.; Bielmann, M.; Campeau-Lecours, A.; Mercier, C.; Bouyer, L.J.; Roy, J.-S. Validity and Reliability of Wearable Sensors for Joint Angle Estimation: A Systematic Review. Sensors 2019, 19, 1555. [Google Scholar] [CrossRef]
Preatoni, E.; Bergamini, E.; Fantozzi, S.; Giraud, L.I.; Orejel Bustos, A.S.; Vannozzi, G.; Camomilla, V. The Use of Wearable Sensors for Preventing, Assessing, and Informing Recovery from Sport-Related Musculoskeletal Injuries: A Systematic Scoping Review. Sensors 2022, 22, 3225. [Google Scholar] [CrossRef]
Adesida, Y.; Papi, E.; McGregor, A.H. Exploring the Role of Wearable Technology in Sport Kinematics and Kinetics: A Systematic Review. Sensors 2019, 19, 1597. [Google Scholar] [CrossRef]
Prill, R.; Walter, M.; Królikowska, A.; Becker, R. A Systematic Review of Diagnostic Accuracy and Clinical Applications of Wearable Movement Sensors for Knee Joint Rehabilitation. Sensors 2021, 21, 8221. [Google Scholar] [CrossRef]
Moons, K.G.M.; Wolff, R.F.; Riley, R.D.; Whiting, P.F.; Westwood, M.; Collins, G.S.; Reitsma, J.B.; Kleijnen, J.; Mallett, S. PROBAST: A Tool to Assess Risk of Bias and Applicability of Prediction Model Studies: Explanation and Elaboration. Ann. Intern. Med. 2019, 170, W1–W33. [Google Scholar] [CrossRef]
Guyatt, G.H.; Oxman, A.D.; Vist, G.E.; Kunz, R.; Falck-Ytter, Y.; Alonso-Coello, P.; Schünemann, H.J.; GRADE Working Group. GRADE: An Emerging Consensus on Rating Quality of Evidence and Strength of Recommendations. BMJ 2008, 336, 924–926. [Google Scholar] [CrossRef]
Cochrane Handbook for Systematic Reviews of Interventions; Higgins, J.P.T., Thomas, J., Chandler, J., Cumpston, M., Li, T., Page, M.J., Welch, V.A., Eds.; John Wiley & Sons: Hoboken, NJ, USA, 2024. [Google Scholar]
Lee, J.B.; Mellifont, R.B.; Burkett, B.J. The Use of a Single Inertial Sensor to Identify Stride, Step, and Stance Durations of Running Gait. J. Sci. Med. Sport 2010, 13, 270–273. [Google Scholar] [CrossRef]
Sinclair, J.; Hobbs, S.J.; Protheroe, L.; Edmundson, C.J.; Greenhalgh, A. Determination of Gait Events Using an Externally Mounted Shank Accelerometer. J. Appl. Biomech. 2013, 29, 118–122. [Google Scholar] [CrossRef]
Khandelwal, S.; Wickström, N. Evaluation of the Performance of Accelerometer-Based Gait Event Detection Algorithms in Different Real-World Scenarios Using the MAREA Gait Database. Gait Posture 2017, 51, 84–90. [Google Scholar] [CrossRef]
Fortune, E.; Lugade, V.; Morrow, M.; Kaufman, K. Validity of Using Tri-Axial Accelerometers to Measure Human Movement—Part II: Step Counts at a Wide Range of Gait Velocities. Med. Eng. Phys. 2014, 36, 659–669. [Google Scholar] [CrossRef]
Lee, H.; Choi, S.; Lee, M. Step Detection Robust against the Dynamics of Smartphones. Sensors 2015, 15, 27230–27250. [Google Scholar] [CrossRef]
Gindre, C.; Lussiana, T.; Hebert-Losier, K.; Morin, J.-B. Reliability and Validity of the Myotest^® for Measuring Running Stride Kinematics. J. Sports Sci. 2016, 34, 664–670. [Google Scholar] [CrossRef]
Harrison, D.; Whelan, N. A Comparison of Foot Strike Events Using the Force Plates and Peak Impact Acceleration Measures. In Proceedings of the 33th International Society of Biomechanics in Sport (ISBS 2015), Poitiers, France, 29 June–3 July 2015; Volume 33. [Google Scholar]
Ammann, R.; Taube, W.; Wyss, T. Accuracy of PARTwear Inertial Sensor and Optojump Optical Measurement System for Measuring Ground Contact Time During Running. J. Strength Cond. Res. 2016, 30, 2057–2063. [Google Scholar] [CrossRef]
Mo, S.; Chow, D.H.K. Accuracy of Three Methods in Gait Event Detection during Overground Running. Gait Posture 2018, 59, 93–98. [Google Scholar] [CrossRef]
Brahms, C.M.; Zhao, Y.; Gerhard, D.; Barden, J.M. Stride Length Determination during Overground Running Using a Single Foot-Mounted Inertial Measurement Unit. J. Biomech. 2018, 71, 302–305. [Google Scholar] [CrossRef]
Falbriard, M.; Meyer, F.; Mariani, B.; Millet, G.P.; Aminian, K. Accurate Estimation of Running Temporal Parameters Using Foot-Worn Inertial Sensors. Front. Physiol. 2018, 9, 610. [Google Scholar] [CrossRef]
Benson, L.C.; Clermont, C.A.; Watari, R.; Exley, T.; Ferber, R. Automated Accelerometer-Based Gait Event Detection During Multiple Running Conditions. Sensors 2019, 19, 1483. [Google Scholar] [CrossRef]
Aubol, K.G.; Milner, C.E. Foot Contact Identification Using a Single Triaxial Accelerometer during Running. J. Biomech. 2020, 105, 109768. [Google Scholar] [CrossRef]
Nazarahari, M.; Khandan, A.; Khan, A.; Rouhani, H. Foot Angular Kinematics Measured with Inertial Measurement Units: A Reliable Criterion for Real-Time Gait Event Detection. J. Biomech. 2022, 130, 110880. [Google Scholar] [CrossRef]
Patoz, A.; Lussiana, T.; Breine, B.; Gindre, C.; Malatesta, D. A Single Sacral-Mounted Inertial Measurement Unit to Estimate Peak Vertical Ground Reaction Force, Contact Time, and Flight Time in Running. Sensors 2022, 22, 784. [Google Scholar] [CrossRef]
Donahue, S.R.; Hahn, M.E. Validation of Running Gait Event Detection Algorithms in a Semi-Uncontrolled Environment. Sensors 2022, 22, 3452. [Google Scholar] [CrossRef]
Yang, Y.; Wang, L.; Su, S.; Watsford, M.; Wood, L.M.; Duffield, R. Inertial Sensor Estimation of Initial and Terminal Contact during In-Field Running. Sensors 2022, 22, 4812. [Google Scholar] [CrossRef]
Bach, M.M.; Dominici, N.; Daffertshofer, A. Predicting Vertical Ground Reaction Forces from 3D Accelerometry Using Reservoir Computers Leads to Accurate Gait Event Detection. Front. Sports Act. Living 2022, 4, 1037438. [Google Scholar] [CrossRef]
Young, F.; Mason, R.; Wall, C.; Morris, R.; Stuart, S.; Godfrey, A. Examination of a Foot Mounted IMU-Based Methodology for a Running Gait Assessment. Front. Sports Act. Living 2022, 4, 956889. [Google Scholar] [CrossRef]
Donahue, S.R.; Hahn, M.E. Estimation of Gait Events and Kinetic Waveforms with Wearable Sensors and Machine Learning When Running in an Unconstrained Environment. Sci. Rep. 2023, 13, 2339. [Google Scholar] [CrossRef]
Ramli, A.A.; Liu, X.; Berndt, K.; Chuah, C.-N.; Goude, E.; Kaethler, L.B.; Lopez, A.; Nicorici, A.; Owens, C.; Rodriguez, D.; et al. Gait Event Detection and Travel Distance Using Waist-Worn Accelerometers across a Range of Speeds: Automated Approach. Sensors 2024, 24, 1155. [Google Scholar] [CrossRef]
Miqueleiz, U.; Aguado-Jimenez, R.; Lecumberri, P.; Gorostiaga, E.M. Agreement between an Inertial Measurement Unit (IMU) Algorithm and a Photoelectric System for Analysing Spatiotemporal Variables during Overground and Treadmill Running. Sports Biomech. 2025, 1–19. [Google Scholar] [CrossRef]
Chebbi, A.; Robinson, R.M.; Donahue, S.R.; Hahn, M.E. Identifying Right and Left Impact Using the Derivative of Linear Resultant Acceleration from a Single Sacrum-Mounted IMU. Wearable Technol. 2025, 6, e14. [Google Scholar] [CrossRef]
Purcell, B.; Channells, J.; James, D.; Barrett, R. Use of Accelerometers for Detecting Foot-Ground Contact Time during Running. In Proceedings of the BioMEMS and Nanotechnology II, Brisbane, Australia, 19 January 2006; SPIE: Bellingham, WA, USA, 2006; Volume 6036, pp. 292–299. [Google Scholar]
Donahue, S.R.; Hahn, M.E. Feature Identification With a Heuristic Algorithm and an Unsupervised Machine Learning Algorithm for Prior Knowledge of Gait Events. IEEE Trans. Neural Syst. Rehabil. Eng. 2022, 30, 108–114. [Google Scholar] [CrossRef]
Lucot, Q.; Beurienne, E.; Behr, M. Multi-Activity Step Counting Algorithm Using Deep Learning Foot Flat Detection with an IMU Inside the Sole of a Shoe. Sensors 2024, 24, 6927. [Google Scholar] [CrossRef]
van Werkhoven, H.; Farina, K.A.; Langley, M.H. Using A Soft Conformable Foot Sensor to Measure Changes in Foot Strike Angle During Running. Sports 2019, 7, 184. [Google Scholar] [CrossRef]
Bergamini, E.; Picerno, P.; Pillet, H.; Natta, F.; Thoreux, P.; Camomilla, V. Estimation of Temporal Parameters during Sprint Running Using a Trunk-Mounted Inertial Measurement Unit. J. Biomech. 2012, 45, 1123–1126. [Google Scholar] [CrossRef]
Schmidt, M.; Rheinländer, C.; Nolte, K.F.; Wille, S.; Wehn, N.; Jaitner, T. IMU- Based Determination of Stance Duration During Sprinting. Procedia Eng. 2016, 147, 747–752. [Google Scholar] [CrossRef]
Blauberger, P.; Horsch, A.; Lames, M. Detection of Ground Contact Times with Inertial Sensors in Elite 100-m Sprints under Competitive Field Conditions. Sensors 2021, 21, 7331. [Google Scholar] [CrossRef]
Santicchi, G.; Stillavato, S.; Deriu, M.; Comi, A.; Cerveri, P.; Esposito, F.; Zago, M. Validation of Step Detection and Distance Calculation Algorithms for Soccer Performance Monitoring. Sensors 2024, 24, 3343. [Google Scholar] [CrossRef]
Fadillioglu, C.; Stetter, B.J.; Ringhof, S.; Krafft, F.C.; Sell, S.; Stein, T. Automated Gait Event Detection for a Variety of Locomotion Tasks Using a Novel Gyroscope-Based Algorithm. Gait Posture 2020, 81, 102–108. [Google Scholar] [CrossRef]
Stetter, B.J.; Buckeridge, E.; von Tscharner, V.; Nigg, S.R.; Nigg, B.M. A Novel Approach to Determine Strides, Ice Contact, and Swing Phases During Ice Hockey Skating Using a Single Accelerometer. J. Appl. Biomech. 2016, 32, 101–106. [Google Scholar] [CrossRef]
Khandan, A.; Fathian, R.; Carey, J.P.; Rouhani, H. Measurement of Temporal and Spatial Parameters of Ice Hockey Skating Using a Wearable System. Sci. Rep. 2022, 12, 22280. [Google Scholar] [CrossRef] [PubMed]
Tomita, Y.; Iizuka, T.; Irisawa, K.; Imura, S. Detection of Movement Events of Long-Track Speed Skating Using Wearable Inertial Sensors. Sensors 2021, 21, 3649. [Google Scholar] [CrossRef]
Apte, S.; Karami, H.; Vallat, C.; Gremeaux, V.; Aminian, K. In-Field Assessment of Change-of-Direction Ability with a Single Wearable Sensor. Sci. Rep. 2023, 13, 4518. [Google Scholar] [CrossRef]
Dos’Santos, T.; Thomas, C.; Comfort, P.; Jones, P.A. Role of the Penultimate Foot Contact During Change of Direction: Implications on Performance and Risk of Injury. Strength Cond. J. 2019, 41, 87–104. [Google Scholar] [CrossRef]
Jeon, H.; Kim, S.L.; Kim, S.; Lee, D. Fast Wearable Sensor-Based Foot-Ground Contact Phase Classification Using a Convolutional Neural Network with Sliding-Window Label Overlapping. Sensors 2020, 20, 4996. [Google Scholar] [CrossRef]
Rivadulla, A.; Chen, X.; Weir, G.; Cazzola, D.; Trewartha, G.; Hamill, J.; Preatoni, E. Development and Validation of FootNet; a New Kinematic Algorithm to Improve Foot-Strike and Toe-off Detection in Treadmill Running. PLoS ONE 2021, 16, e0248608. [Google Scholar] [CrossRef]
Di Paolo, S.; Nijmeijer, E.; Bragonzoni, L.; Dingshoff, E.; Gokeler, A.; Benjaminse, A. Comparing Lab and Field Agility Kinematics in Young Talented Female Football Players: Implications for ACL Injury Prevention. Eur. J. Sport Sci. 2023, 23, 859–868. [Google Scholar] [CrossRef]
Di Paolo, S.; Nijmeijer, E.M.; Bragonzoni, L.; Gokeler, A.; Benjaminse, A. Definition of High-Risk Motion Patterns for Female ACL Injury Based on Football-Specific Field Data: A Wearable Sensors Plus Data Mining Approach. Sensors 2023, 23, 2176. [Google Scholar] [CrossRef]
Havens, K.L.; Sigward, S.M. Joint and Segmental Mechanics Differ between Cutting Maneuvers in Skilled Athletes. Gait Posture 2015, 41, 33–38. [Google Scholar] [CrossRef]
Trojaniello, D.; Cereatti, A.; Pelosin, E.; Avanzino, L.; Mirelman, A.; Hausdorff, J.M.; Della Croce, U. Estimation of Step-by-Step Spatio-Temporal Parameters of Normal and Impaired Gait Using Shank-Mounted Magneto-Inertial Sensors: Application to Elderly, Hemiparetic, Parkinsonian and Choreic Gait. J. Neuroeng. Rehabil. 2014, 11, 152. [Google Scholar] [CrossRef] [PubMed]
Webster, K.E.; Wittwer, J.E.; O’Brien, J.; Feller, J.A. Gait Patterns after Anterior Cruciate Ligament Reconstruction Are Related to Graft Type. Am. J. Sports Med. 2005, 33, 247–254. [Google Scholar] [CrossRef] [PubMed]
King, E.; Richter, C.; Daniels, K.A.J.; Franklyn-Miller, A.; Falvey, E.; Myer, G.D.; Jackson, M.; Moran, R.; Strike, S. Biomechanical but Not Strength or Performance Measures Differentiate Male Athletes Who Experience ACL Reinjury on Return to Level 1 Sports. Am. J. Sports Med. 2021, 49, 918–927. [Google Scholar] [CrossRef] [PubMed]
Della Villa, F.; Tosarelli, F.; Ferrari, R.; Grassi, A.; Ciampone, L.; Nanni, G.; Zaffagnini, S.; Buckthorpe, M. Systematic Video Analysis of Anterior Cruciate Ligament Injuries in Professional Male Rugby Players: Pattern, Injury Mechanism, and Biomechanics in 57 Consecutive Cases. Orthop. J. Sports Med. 2021, 9, 23259671211048182. [Google Scholar] [CrossRef]
Cereatti, A.; Gurchiek, R.; Mündermann, A.; Fantozzi, S.; Horak, F.; Delp, S.; Aminian, K. ISB Recommendations on the Definition, Estimation, and Reporting of Joint Kinematics in Human Motion Analysis Applications Using Wearable Inertial Measurement Technology. J. Biomech. 2024, 173, 112225. [Google Scholar] [CrossRef]
Pacini Panebianco, G.; Bisi, M.C.; Stagni, R.; Fantozzi, S. Analysis of the Performance of 17 Algorithms from a Systematic Review: Influence of Sensor Position, Analysed Variable and Computational Approach in Gait Timing Estimation from IMU Measurements. Gait Posture 2018, 66, 76–82. [Google Scholar] [CrossRef]
Encarnación-Martínez, A.; Pérez-Soriano, P.; Sanchis-Sanchis, R.; García-Gallart, A.; Berenguer-Vidal, R. Validity and Reliability of an Instrumented Treadmill with an Accelerometry System for Assessment of Spatio-Temporal Parameters and Impact Transmission. Sensors 2021, 21, 1758. [Google Scholar] [CrossRef]
Drobnič, M.; Verdel, N.; Holmberg, H.-C.; Supej, M. The Validity of a Three-Dimensional Motion Capture System and the Garmin Running Dynamics Pod in Connection with an Assessment of Ground Contact Time While Running in Place. Sensors 2023, 23, 7155. [Google Scholar] [CrossRef]
Trabassi, D.; Serrao, M.; Varrecchia, T.; Ranavolo, A.; Coppola, G.; De Icco, R.; Tassorelli, C.; Castiglia, S.F. Machine Learning Approach to Support the Detection of Parkinson’s Disease in IMU-Based Gait Analysis. Sensors 2022, 22, 3700. [Google Scholar] [CrossRef]
Xu, D.; Zhou, H.; Jie, T.; Zhou, Z.; Yuan, Y.; Jemni, M.; Quan, W.; Gao, Z.; Xiang, L.; Gusztav, F.; et al. Data-Driven Deep Learning for Predicting Ligament Fatigue Failure Risk Mechanisms. Int. J. Mech. Sci. 2025, 301, 110519. [Google Scholar] [CrossRef]

Figure 1. Flow chart describing the identification, screening, eligibility assessment, and inclusion process of the present systematic review, following the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA). A total of 328 records were identified (322 from databases and 6 from grey literature). After title and abstract screening, 275 studies were excluded, and a further 16 studies were excluded after full-text assessment. Reasons for exclusion are detailed within the diagram. In total 37 studies met all predefined inclusion criteria and were included in the systematic review.

Figure 2. Graphical summary of the systematic review findings on IMU-based foot contact detection algorithms in high-dynamic movements. The key information collected for each study included in the systematic review (movement task analysed, speed, gender, ground truth system adopted, IMU system brand, number of IMUs, IMU placement, and metrics adopted in the foot contact detection algorithms) were summarized in percentages (%) and ranges. Left panel: Most studies investigated running tasks (78%), with reported speeds ranging from 1.9 to 9 m/s; nearly half (49%) reported an effect of velocity on algorithm performance. Male participants were more frequently included (67%), while 38% of studies reported a balanced gender distribution. For ground-truth validation, the majority relied on force-sensing systems, mainly insoles (20%). Right panel: Most studies employed a single IMU (51%), placed mainly on the trunk, foot, or shank, with sampling frequencies between 50 Hz and 2400 Hz. Xsens Technologies was the most frequently adopted IMU system (used in 14% of studies). Algorithmic approaches varied, with 65% focusing on linear acceleration metrics, and only 41% providing a detailed description of the algorithm.

Table 1. Inclusion and exclusion criteria used in the article screening process.

Inclusion Criteria	Exclusion Criteria
research articles describing a method for foot contact detection using wearable technology (e.g., IMUs)	review articles, editorials, letters, conference abstracts without full-text
research articles involving high-dynamic movements (e.g., running, jumping, cutting manoeuvres, sport-specific actions, etc.)	studies involving non-high-dynamic movements (e.g., only walking, standing, postural tasks)
research articles using inertial measurement units or comparable wearable motion sensors	studies evaluating wearable sensors for activity classification rather than foot contact detection
	studies focusing on physiological signal estimation (e.g., heart rate, respiration, etc.)
	studies focusing on fatigue analysis, balance assessment, or fall detection
	studies conducted on non-human subjects or robots

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

© 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).

Detection of Foot Contact Using Inertial Measurement Units in Sports Movements: A Systematic Review

Featured Application

Abstract

1. Introduction

2. Materials and Methods

2.1. Search Strategy

2.2. Inclusion and Exclusion Criteria

2.3. Quality Assessment

2.4. Data Extraction

3. Results

3.1. Movement Tasks and Speed

3.2. Target Population & Sample Size

3.3. IMU System & Placement

3.4. Algorithm Structure

3.5. Ground-Truth Data

3.6. Main Results & Effect of Speed

3.7. Machine Learning

3.8. Quality Assessment

4. Discussion

5. Conclusions

Author Contributions

Funding

Data Availability Statement

Conflicts of Interest

Abbreviations

Appendix A

References

Article Metrics

Citations

Article Access Statistics