Streamlining Motor Competence Assessments via a Machine Learning Approach

Abstract

Strong competencies in actual motor competence (AMC) and perceived motor competence (PMC) support lifelong physical activity. However, assessing MC is time-consuming, requiring multiple AMC and PMC evaluations. Streamlining these assessments would improve efficiency at a national level. This study used machine learning (ML) classification to (1) identify AMC assessments that can be accurately predicted in an Irish context using other AMC and PMC assessments, and (2) examine prediction accuracy differences between genders. AMC was measured using the Test of Gross Motor Development (3rd Edition) and the Victorian Fundamental Motor Skills Manual, while PMC was assessed with the Pictorial Scale of Perceived Movement Skill Competence. Five ML classification models were trained and tested on an Irish MC dataset (n = 2098, mean age 9.2 ± 2.04) to predict distinct AMC assessment outcomes. The highest prediction accuracies (>85%) were found for the Catch (female and gender-combined subsets) and Bounce (male subset) AMC assessments. These assessments could potentially be removed from the current Irish testing battery for their respective gender groups. Our findings highlight the effectiveness of ML classification in optimising Irish MC assessment procedures, reducing redundancy, and enhancing efficiency.

1. Introduction

Physical activity (PA) is essential for youth development, as it plays a critical role in reducing the risks of obesity, cardiovascular disease, and mental health issues (Eime et al., 2013; Katzmarzyk et al., 2019). PA has also been linked to improved cognitive performance, social skills, and overall well-being in children and adolescents (Erickson et al., 2019). Despite these well-documented benefits, global trends indicate a significant decline in PA levels among youth, with approximately 85% of young individuals failing to meet the World Health Organization’s recommended 60 min of moderate-to-vigorous PA per day (Guthold et al., 2020). This decline has been attributed to various factors, including increased screen time, urbanisation, and a lack of access to structured physical education and recreational facilities (Hesketh et al., 2017). Consequently, there is an urgent need for effective interventions that can enhance PA participation among youth and counteract these negative trends.

One key avenue for intervention lies in the assessment and development of motor competence (MC). MC is widely recognised as a fundamental determinant of PA engagement (Lima et al., 2017), as individuals with higher MC levels are more likely to participate in and sustain an active lifestyle (Estevan et al., 2023). The reciprocal relationship between MC and PA suggests that individuals with greater proficiency in movement skills are more confident and motivated to engage in various physical activities, thereby promoting a positive cycle of MC and participation (Stodden et al., 2008). Ultimately, the lack of an adequate foundation of MC may be linked to a hypothetical ‘proficiency barrier’ (Seefeldt, 1980) where low-MC individuals may not demonstrate health-enhancing levels of PA and health-related physical fitness later in life (Stodden et al., 2009). Such poorly proficient individuals may also be at greater risk for obesity across childhood and adolescence (Robinson et al., 2015). Therefore, it is key for practitioners to understand the components that comprise MC.

MC is composed of two distinct but interrelated components: actual motor competence (AMC) and perceived motor competence (PMC). AMC is associated with fundamental movement skills (FMS), which are the building blocks of more advanced, complex movements required to participate in games, sports or other context specific PA (Logan et al., 2018). FMS is made up of three components, namely locomotor skills (running and jumping, etc.), object control skills (throwing and kicking, etc.), and stability skills (i.e., balance) (Goodway et al., 2010). PMC, on the other hand, relates to an individual’s self-perception of their motor abilities and has been shown to influence PA participation and motivation (Estevan & Barnett, 2018). Children who perceive themselves as competent are more likely to develop the confidence and self-efficacy required to engage consistently in PA across a range of settings (Peers et al., 2020). PMC has also been shown to exert a significant influence on future PA behaviour, potentially surpassing the influence of AMC alone (Bardid et al., 2016). Research demonstrates a robust association between high levels of PMC and higher levels of PA, underscoring the critical roles that both AMC and PMC play in developing youth physical literacy (Barnett et al., 2008; Stodden et al., 2008). Given this established relationship, both AMC and PMC should be accorded equal consideration in the assessment of MC among children and adolescents. Recent international studies further support this approach, consistently identifying significant correlations between PMC and AMC (den Uil et al., 2025; Niemistö et al., 2023). Findings indicate that children displaying low PMC during early childhood face increased risks of persistently low PMC and limited AMC progression into middle childhood (Niemistö et al., 2023). Accordingly, prioritising the development of children’s MC proficiency from an early age is essential.

Unfortunately, research worldwide consistently reports poor levels of AMC among children and adolescents, with persistent deficits observed through adolescence (Chen et al., 2024; Hardy et al., 2013). Furthermore, research has identified notable gender differences in AMC proficiency, further emphasising the need for targeted intervention strategies. Studies have reported that boys tend to score significantly higher in object control skills, whereas girls typically outperform boys in locomotor and stability skills and balance-related movements (Behan et al., 2019; Zheng et al., 2022). Thus, there is a clear need for assessment and intervention to help all children enhance their AMC trajectory.

Given the well-documented benefits of MC development in promoting active lifestyles, there is a clear need for reliable and efficient assessment tools that can guide appropriate interventions. While there are a wide range of AMC measures, PMC is generally assessed using a reliable and valid self-report pictorial scale developed by Barnett et al. (2015), and enables alignment between the measurement of PMC and AMC competence in young children (Barnett et al., 2015). In the assessment of AMC, a wide range of FMS assessments are currently used to evaluate children’s overall AMC proficiency, with the Test of Gross Motor Development (TGMD, 2nd and 3rd editions) (Ulrich, 2000; Ulrich, 2013) being the most commonly used in related research (Behan et al., 2019; Duncan et al., 2020). As an assessment of FMS, the TGMD-3 can be used to identify developmental delay in relation to gross motor performance, to evaluate programmes aimed at enhancing AMC through intervention, and to assess changes as a function of increasing age, experience, instruction, or intervention (Ulrich, 2000; Ulrich, 2013).

However, while reliable and valid as a measure of AMC, employing the TGMD-3 requires a considerable amount of time to administer and interpret (Barnett et al., 2013). In addition to a considerable time burden on participants and trained staff, specific space and equipment are required for assessment, which can make it logistically difficult to administer in school settings (Tamplain et al., 2020). As a result, researchers have been exploring alternative strategies, such as confirmatory factor analysis and machine learning (ML) approaches, to streamline assessment procedures while maintaining accuracy and reliability (Bandeira et al., 2020; Duncan et al., 2022; Valentini et al., 2018). These efforts aim to identify key assessments that can predict overall MC, allowing practitioners to optimise evaluation processes and ensure that children receive the necessary support to enhance their motor development.

The integration of ML in MC research offers a promising avenue for refining assessment protocols. ML involves the development of algorithms and models that enable computers to learn and make predictions or decisions independently, without the need for specific instructions from the end-user (Yaqoob et al., 2023). By leveraging data-driven models, ML techniques can analyse complex patterns in movement proficiency and predict performance outcomes based on a combination of AMC and PMC assessments (Yaqoob et al., 2023). This approach has the potential to reduce the number of assessments required while maintaining predictive accuracy, ultimately improving the feasibility of large-scale MC evaluations (Barnett et al., 2013; Duncan et al., 2022). Furthermore, ML driven assessment tools can facilitate the identification of at-risk children who may benefit from targeted interventions, ensuring that resources are allocated efficiently and effectively in educational and community sports settings (Bardid et al., 2019). For instance, recent MC research (Britton et al., 2023; Duncan et al., 2024) examined the predictive value of AMC features in relation to children’s technical soccer skill performance and well-being, respectively. Both studies employed five distinct ML algorithms, initially introducing a baseline algorithm and subsequently introducing more complex algorithms, resulting in superior predictions, with Duncan et al. (2024) producing superb predictions of 99% whilst Britton et al. (2023) produced strong predictions of 87%. Thus, to address the ongoing challenges in MC assessment and intervention, it is imperative to explore innovative methodologies that balance efficiency, accuracy, and practicality. By harnessing the power of ML classification models, this study aims to contribute to the development of streamlined assessment tools that can support the promotion of MC and PA participation among children and adolescents.

Whilst previous studies have presented original contributions to the TGMD literature base, there are numerous areas in which ML classification analyses could build upon such contributions. Firstly, the ML approach in network analysis has been utilised in prior research (Bandeira et al., 2020; Duncan et al., 2022), and whilst it is highly effective at assessing AMC relationship patterns, it is less effective than classification techniques when the aim of a study is predictive modelling (Yaqoob et al., 2023). The scoring criteria associated with the TGMD-3 (Ulrich, 2013) results in participants being awarded a score of either ‘1’ (‘Mastery/Near Mastery’) or ‘0’ (‘Not Mastered’). Such scoring criteria is ideal for ML classification analysis given that the variable’s data is binary, thus requiring each variable to be classified into its predefined group. Secondly, whilst Bandeira et al. (2020) and Duncan et al. (2022) were capable of reducing their original 13-12 AMC skill models to a final model of 6–7 AMC skills, both studies did not include any PMC variables within their analysis. Indeed, the omission of such PMC variables from the aforementioned analyses is key given the impact a positive AMC–PMC relationship can have on a child’s motivation to continuously improve their AMC capabilities (Barnett et al., 2008; Stodden et al., 2008).

Building on these efforts, this study explores the potential of ML classification models as a novel approach to optimising MC evaluation. ML techniques have been increasingly utilised in sports science and health research due to their ability to analyse complex patterns in large datasets and generate predictive models with high accuracy. By applying ML classification methods, this study aims to determine which MC assessments can be accurately predicted based on the results of other AMC and PMC evaluations. If successful, this approach could enable educators, coaches, and health practitioners to implement more efficient assessment protocols, ultimately facilitating broader participation in MC testing and fostering more inclusive physical literacy initiatives for youth populations.

2. Materials and Methods

2.1. Study Design

Cross-sectional data were collected as part of an Irish AMC study ‘Moving Well-Being Well’ (Behan et al., 2019). Participants (n = 2098, 47% girls, ranging from 5–12 years of age, mean age 9.2 ± 2.04) were recruited from 44 schools across twelve counties, encompassing all four provinces in Ireland and Northern Ireland. Data was collected from March to June 2017 across the full range of the Irish primary school cycle of typically developing children aged five to twelve. The subjects were classified per age and sex for data analysis. Ethical approval was obtained from the institutional Research Ethics Committee (DCU/REC/2017/029). The principals of each of the participating schools initially consented to participation. Consent procedures ensured that all individuals provided informed written consent, as well as parental consent, prior to data collection, thus supporting participant autonomy. An age-appropriate child assent form was supplied to participants in the younger classes (5–8 years of age). Additionally, robust anonymisation procedures were employed, including assigning participants a unique numerical code, which not only protected participant confidentiality but also enhanced the integrity and reliability of the dataset by minimising potential biases or risks associated with participant identification.

2.2. Procedure

AMC was measured using the TGMD-3 (Ulrich, 2013) alongside an additional locomotor skill test, the vertical jump, from the Victorian Fundamental Motor Skills manual (Walkley et al., 1996). Both have been used in previous research and have a high degree of validity and reliability (Walkley et al., 1996; Cools et al., 2009). In the dataset such AMC skills were labelled as ‘Mastery/Near Mastery’ (MNM) features in which participants were awarded a score of one if the participant fulfilled the necessary skill criteria (considered as ‘Mastery/Near Mastery’) whilst a score of zero (‘Not Mastered’) indicated that they failed to meet such skill criteria (Ulrich, 2013). Further details of these protocols can be found in Behan et al. (2019). PMC was measured using the Pictorial Scale of Perceived Movement Skill Competence (PMSC) for Young Children (Barnett et al., 2015). The PMSC administered fifteen skills, of which seven related to perceptions of various object-control skills and eight related to perceptions of locomotor skills. Each skill item was evaluated from 1 to 4 points leading to a maximum sum score for locomotor skills of 32 points (8 × 4) and 28 points (7 × 4) for perceptions of object control skills, using a binary two-step choice process to arrive at one of four responses (1 = not that good, 2 = sort of good, 3 = pretty good, 4 = really good) per each skill. The maximum total score a child could receive was 60 points. The higher the child scored, the higher the child’s PMC. Previous studies have shown good reliability and validity for assessing young children’s PMC with PMSC in different cultures for locomotor (Diao et al., 2018) and object-control skills (Diao et al., 2018; Johnson et al., 2016).

2.3. Data Collection

The data collection protocol utilised by the research team is outlined fully in Behan et al. (2019). As such, all research team members completed formal training to ensure a comprehensive understanding of the skill assessment procedure as well as consistency during assessment of the test subjects. To ensure the data was consistent and of high quality, the research team was required to meet a 95% inter-observer agreement on a pre-coded data set, which was pre-coded by the lead researcher. The research team members were blind to the conditions of coding.

2.4. Experimental Setup

The original dataset consisted of 2378 rows and 492 columns. The initial cleaning process involved removing null values and removing variables that were deemed surplus for the purpose of the study. Features were only included if they were pure measures of AMC or PMC, for instance ‘Actual Kick MC’ and ‘Perceived Kick MC’. This procedure for removing redundant features was applied uniformly across all gender subsets. Below in

Table 1. Brief Synopsis of Features Analysed.

Feature Category	Feature Description	Count (Perceived & Actual Features Combined)
Locomotor-Skill	Actual MC—How capable children are to run, jump, skip etc. Full marking criteria are available in TGMD-3 (Ulrich, 2013). Perceived MC—The child’s perception of their ability to run, jump, skip etc. Full marking criteria available in (PMSC) for Young Children (Barnett et al., 2015).	16
Object-Control Skill	Actual MC—How capable children are at manipulating objects such as balls (i.e., kicking, catching, throwing). Full marking criteria available in TGMD-3 (Ulrich, 2013). Perceived MC—The child’s perception of their ability to kick, catch, throw etc. Full marking criteria available in (PMSC) for Young Children (Barnett et al., 2015).	14

, a brief synopsis is provided on several of the analysed variables, including their data type and a description of the variable. After cleaning the dataset, there were a total of 2378 rows and 50 columns remaining, which comprised the final dataset.

As such, 16 locomotor skill and 14 object-control skill features were analysed, which resulted in a total of 30 variables being included in the analysis. Following this the dataset was split into different subsets of data in which the dataset was split by a specific gender subset.

Table 2. Example of Splitting Data by Gender Into several Subsets for Testing.

Dependent Variable	Gender Subset	Count
Run_MNM	Female Only	969
Run_MNM	Male Only	1107
Run_MNM	Gender Combined	2076

below gives a brief example of what this splitting process looked like for all dependent variables.

Before proceeding with data sampling, several crucial feature engineering steps were undertaken to refine the dataset for analysis. The process began by encoding categorical variables through one-hot encoding, converting non-numeric attributes into a machine-readable format. This approach generated binary indicator variables for each categorical feature while discarding the original columns, ensuring that classification models could interpret categorical data without imposing any unintended ordinal relationships.

Following this, numerical features such as height and weight were standardised using StandardScaler to maintain consistency across varying scales. Since ML models—particularly those that rely on distance-based computations—are sensitive to differences in feature magnitudes, standardisation ensured that all continuous variables were centered at zero with a standard deviation of one. This prevented certain features from disproportionately influencing the model due to differences in scale.

Additionally, a feature selection process was implemented to remove irrelevant, redundant, or null-indicator columns that did not add value to the predictive model. A filtering mechanism identified and excluded these unnecessary features, enhancing both computational efficiency and model interpretability. After these transformations were finalised, the cleaned dataset was stored in Parquet format, preserving its structured and optimised form for the subsequent data sampling phase. Numerous ML methods have proposed that the target classes (i.e., dependent variables) hold similar distributions (Mohammed et al., 2020). However, this assumption does not hold true in several domains, for instance diagnosis of illness (Krawczyk et al., 2016), as most of the instances are labelled with one class, while few instances are labelled as the other. As a result, the models gravitate towards the majority class, thus eliminating the minority class. This reflects on the model’s performance, as these models will perform poorly when the datasets are imbalanced. This is known as a class imbalance problem. Indeed, one of the common approaches in counteracting this issue is to incorporate resampling methods to balance the dataset.

As such, a class imbalance issue existed in the current study’s dataset, which resulted in several resampling methods being incorporated, with those methods being: undersampling (the process of decreasing the amount of majority target instances or samples), oversampling (the process of increasing the amount of minority class instances or samples by producing new instances or repeating some instances (Mohammed et al., 2020)), no sampling, and SMOTE (the process of analysing the minority samples and integrating new synthetic minority samples according to the minority samples, in order to add them to the data set (Shi et al., 2021)). These resampling methods resulted in the creation of four instances of the dataset for each individual dependent variable. Figure 1 below shows an example of the distribution difference for one of these imbalanced dependent variables.

Five classification methods were chosen as candidates for predicting the outcome of individual AMC assessments: decision tree (DT), gradient boosting decision tree (GBDT), k-nearest neighbour (KNN), logistic regression (LR), and XGBoost (XGB). Detailed descriptions of such classifiers are provided within the following references: Alzubi et al. (2018), Carmona et al. (2022), and Yoon (2021). The metrics used to validate the classifiers were: precision (percentage of observations labelled as positive that are actually positive), recall (percentage of positive observations classified as positive), and F₁ (weighted average of precision and recall). As such, these metrics were calculated from four distinct counts: true positives (TP), a count of the correctly predicted positive values; true negatives (TN), a count of the correctly predicted negative values; false positives (FP), which count the occurrences where the actual class is false but is predicted as true; and false negatives (FN), which count the occurrences where the actual class is true but is predicted as false. The sum of these four counts is the total number of observations (Britton et al., 2023). Additionally, the F₁ metric was utilised as it is a recognised metric enabling comparison between classifiers trained using differing methodologies, which was present in our case with an imbalanced distribution existing between non-mastery and mastery–near-mastery values. We performed hyperparameter optimisation using GridSearch to determine the optimal model configuration for each method, data subset, and target AMC assessment. The data was split using an 80:20 ratio for training and testing. In total, 780 predictive classification models were trained.

3. Results

Results for the top performing target features in terms of highest predictive accuracies (F₁ score) amongst each subset alongside their corresponding resampling method, precision, and recall scores are shown in

Table 3. Top Performing Target Features—Gender Combined Subset.

Feature	Classifier	Sampling	Precision	Recall	F1 Score
Catch MNM	XGB	SMOTE	0.83	0.88	0.85
Run MNM	KNN	SMOTE	0.89	0.79	0.83
Slide MNM	XGB	Unsampled	0.79	0.87	0.83
Bounce MNM	XGB	Undersampled	0.80	0.80	0.80
Hop MNM	XGB	SMOTE	0.75	0.85	0.80
Kick MNM	DT	Oversampled	0.74	0.66	0.70
HJ MNM	LR	SMOTE	0.67	0.68	0.68
Skip MNM	DT	Unsampled	0.66	0.64	0.65
Roll MNM	DT	Unsampled	0.68	0.58	0.63
VJ MNM	DT	Unsampled	0.68	0.53	0.60
Gallop MNM	XGB	Unsampled	0.57	0.44	0.50
Throw MNM	XGB	SMOTE	0.49	0.43	0.46

,

Table 4. Top Performing Target Features—Female-Only Subset.

Feature	Classifier	Sampling	Precision	Recall	F1 Score
Catch MNM	KNN	Unsampled	0.85	0.88	0.87
Hop MNM	KNN	SMOTE	0.93	0.78	0.85
Run MNM	XGB	Unsampled	0.80	0.91	0.85
Slide MNM	XGB	Unsampled	0.83	0.85	0.84
Kick MNM	DT	Unsampled	1.00	0.62	0.77
Bounce MNM	XGB	Unsampled	0.72	0.73	0.72
Gallop MNM	XGB	Unsampled	1.00	0.57	0.72
HJ MNM	XGB	Undersampled	0.71	0.69	0.70
Throw MNM	LR	Unsampled	1.00	0.54	0.70
Skip MNM	XGB	SMOTE	0.66	0.68	0.67
VJ MNM	XGB	SMOTE	0.61	0.62	0.61
Roll MNM	GB	Oversampled	0.56	0.55	0.56

and

Table 5. Top Performing Target Features—Male-Only Subset.

Feature	Classifier	Sampling	Precision	Recall	F1 Score
Bounce MNM	XGB	Unsampled	0.84	0.89	0.86
Run MNM	KNN	Unsampled	0.78	0.88	0.83
Catch MNM	KNN	SMOTE	0.88	0.77	0.82
Slide MNM	KNN	Oversampled	0.83	0.80	0.82
Kick MNM	KNN	SMOTE	0.88	0.76	0.81
Hop MNM	KNN	SMOTE	0.82	0.73	0.77
HJ MNM	XGB	Undersampled	0.69	0.69	0.69
Skip MNM	XGB	SMOTE	0.69	0.68	0.68
VJ MNM	DT	Unsampled	0.75	0.58	0.66
Roll MNM	DT	SMOTE	0.58	0.65	0.62
Throw MNM	DT	Unsampled	0.71	0.50	0.59
Gallop MNM	KNN	Unsampled	0.33	0.30	0.31

.

Results from the gender combined subset showed that the Catch MNM feature produced the highest F₁ score at 85%, whilst the Throw MNM feature produced the weakest F₁ score at 46%, resulting in a 39% discrepancy between the best and worst performing features. As such, the three best performing features (Slide, Run, and Catch) ranged from 83–85% respectively. The XGB classifier proved to be the most prominent classifier, appearing on six occasions, whilst the SMOTE and Unsampled resampling methods were most prominent, appearing on five occasions each.

Results from the female-only subset showed that the Catch MNM feature produced the highest F₁ score at 87%, whilst the Roll MNM feature produced the weakest F₁ score at 56%, resulting in a 31% discrepancy between the best and worst performing features. As such, the three best performing features (Run, Hop, and Catch) ranged from 85–87%, respectively. The XGB classifier proved to be the most prominent classifier, appearing on seven occasions, whilst the Unsampled resampling method was most prominent, appearing on seven occasions.

Results from the male-only subset showed that the Bounce MNM feature produced the highest F₁ score at 86%, whilst the Gallop MNM feature produced the weakest F₁ score at 31%, resulting in a 55% discrepancy in accuracy between the best and worst performing features. As such, the three best-performing features (Catch, Run, and Bounce) ranged from 82–86%, respectively. The KNN classifier proved to be the most prominent classifier, appearing on six occasions, whilst the Unsampled and SMOTE resampling methods were most prominent, appearing on five occasions each.

When resampling methods were considered, Unsampled data and SMOTE were associated with the best-performing classifiers. These findings suggest that the classification models employed were able to learn effectively without any alterations to the dataset (i.e., using the original, unsampled data), while the use of synthetic samples generated through SMOTE enhanced the models’ predictive performance. Given the consistent success of both Unsampled and SMOTE approaches in our results, it could be inferred that low number of instances where individuals received a mark of not mastered resulted in overfitting and underfitting for the Oversampling and Undersampling approaches, respectively, in the context of the current dataset. The analysis of F₁ Score performance across gender subsets revealed relatively small differences in standard deviation values, indicating consistent model performance within each gender subset. The male-only subset exhibited the highest variability with a standard deviation of 0.15, compared to 0.10 for the female-only subset, while the gender-combined subset showed an intermediate value of 0.13. On average, the standard deviation across all subsets was 0.12, reflecting generally low variation in F₁ Score performance. Similarly, the 95% confidence interval was consistent at 0.02 across all groups, highlighting a high degree of precision in the F₁ Score estimates. These findings suggest that while the male-only group demonstrated slightly more variability, the model’s overall performance remained stable and reliable across gender-based subsets.

Figure 2 provides a visual overview of the variation in F₁ Scores amongst the three distinct subsets. Considering the results from an overall perspective, most of the features performed quite well, with several features producing F₁ scores > 70%. Interestingly, the Gallop MNM feature produced the largest subset-to-subset discrepancy of 41% in favour of females versus males, whilst the male-only subset produced the largest accuracy discrepancy between the most- and least-accurate features at 55%. When the gender combined subset is compared against both the male and female subsets, the latter subsets are shown to essentially mirror and at times exceed the gender combined predictive power (i.e., Run MNM; GC Subset F₁ Score—83%, Male Subset F₁ Score—83%, Female Subset F₁ Score—85%).

4. Discussion

The present study aimed to determine which MC assessments can be reliably predicted based on AMC and PMC evaluations, with a particular focus on gender-based differences in prediction accuracy. Our findings suggest that ML offers a promising avenue for streamlining MC assessments by identifying assessments that can be inferred from existing test results with high predictive accuracy. This not only optimises efficiency in testing but also reduces redundancy, potentially allowing practitioners to administer fewer assessments while maintaining robust evaluation standards. Optimising AMC assessment batteries is therefore crucial, as regularly evaluating the most relevant skills for specific subgroups of children can help identify more realistic MC proficiency profiles. This is important because difficulties in mastering MC may hinder success in many sports, potentially leading to subsequent physical inactivity beyond early childhood (Stodden et al., 2008).

The Catch assessment emerged as the most accurately predicted test in the female-only and gender-combined subsets, whereas the Bounce assessment was the most accurately predicted in the male-only subset. These findings align with prior research indicating that object-control skills, which are highly practiced and reinforced in structured and unstructured physical activities, tend to exhibit strong predictive patterns (Behan et al., 2019). The high predictability of these assessments suggests that they may be candidates for removal in future testing batteries, enabling a more targeted approach to MC assessment. However, before such changes are implemented, further validation in diverse populations and across different cultural contexts is warranted.

The contextualisation of these findings within the Irish sports landscape is essential for understanding their broader implications. Sports participation trends in Ireland indicate a strong emphasis on object-control skills, particularly through popular indigenous sports such as Gaelic football and hurling, as well as widely played international sports like soccer, rugby, and basketball (Woods et al., 2023). These sports provide ample opportunities for children to develop object-control skills, contributing to the observed high levels of mastery in skills like Catch and Bounce (Behan et al., 2019). As a result, it is not surprising that these skills are among the most predictable by ML models. Nevertheless, the transferability of these findings to other contexts requires careful consideration, as different regions and cultures emphasise varying movement skills within their youth sports structures (Newell, 2020). Taking the kick as an example, it could be assumed that in cultures where soccer is the most prevalent sport (i.e., Brazil), AMC proficiency for kicking would likely be superior compared to cultural contexts where soccer is played less (Duncan et al., 2022). Indeed, the cultural differences in sport and movement environments between countries may contribute to the reason why the skills may vary in reconceptualisations of the original TGMD model. It can therefore be concluded that children from different countries and cultures display variations in their AMC proficiency (Hulteen et al., 2018). Accordingly, the specificities of each country and culture must be considered by researchers in their validation models, which can be beneficial in discriminating variations between countries, and in determining the priority of intervention proposals, while respecting cultural aspects, as previously suggested (Hulteen et al., 2018; Barnett et al., 2016).

In practical terms, these results have several implications for MC assessment practices. First, the ability of ML models to predict certain MC outcomes with high accuracy suggests that assessment protocols could be streamlined and adapted for gender-specific applications. The removal of the Catch AMC assessment for females and the Bounce AMC assessment for males, based on their high predictive accuracy, could lead to a more efficient assessment process while preserving the integrity of overall MC evaluation. This streamlining could benefit practitioners, including teachers, sports coaches, and MC assessors, by allowing them to focus on evaluating skills that require direct observation due to lower mastery rates or higher variability in proficiency (Woods et al., 2023). Equally, we acknowledge that while the primary focus of the current study was to explore ways in which the burden of practical assessments could be reduced, we appreciate that removing some skills from the assessment battery may leave a practitioner with gaps of nuanced motor skill competence. We do however believe that the current standard testing protocols are impractical and time-intensive, and we feel that any efforts to lighten the burden should be strongly considered. That said, practitioners should be aware that omitting certain skills may result in a lack of data in specific skill competencies that are important in their domain or context (i.e., Gaelic football or soccer).

Our findings also align with previous studies that have sought to optimise MC testing batteries through confirmatory factor analysis and network analysis (Bandeira et al., 2020; Duncan et al., 2022; Valentini et al., 2018). Unlike these studies, which primarily focused on AMC skills alone, our analysis incorporated PMC features, providing a more holistic understanding of the relationships between MC skills. The inclusion of PMC is particularly relevant given its role in influencing PA engagement, as children who perceive themselves as competent are more likely to participate in and persist with PA (Stodden et al., 2008). Future research should explore the relative influence of AMC versus PMC in ML prediction models to better understand the interplay between PMC and AMC.

Another important consideration is the variation in prediction accuracy across different MC skills. While the Catch and Bounce assessments were highly predictable, other skills, such as Gallop and Throw, exhibited poorer predictive accuracy. These findings are consistent with prior research indicating that mastery rates vary considerably across different MC skills, with certain locomotor and object-control skills being more challenging to develop (Behan et al., 2019). In practical terms, this suggests that future assessment batteries should allocate more time to evaluating skills with lower mastery rates, as they are less likely to be reliably predicted through ML models.

Moreover, our findings provide insights into gender-based differences in MC skill prediction. The female subset produced superior predictions for Catch and Throw AMC assessments compared to males, whereas males showed higher predictive accuracy for Bounce AMC. These results align with existing literature suggesting that gender differences in MC development are influenced by participation patterns and movement skill practice opportunities (Bolger et al., 2018). While these findings suggest potential pathways for gender-specific assessment refinement, further research is needed to explore the underlying factors contributing to these predictive discrepancies, including sociocultural influences, access to PA opportunities, and inherent biomechanical differences.

Limitations and Future Work

A number of limitations must be acknowledged in this study. Firstly, only one dataset was analysed, which may limit the generalisability of the findings. The F₁ Scores for each AMC feature might have been improved with the integration of additional datasets that capture a broader range of socio-economic variables. Socio-economic factors, such as access to PA opportunities, parental support, and school resources, can significantly influence MC development. By incorporating such features in future analyses, it may be possible to refine prediction accuracy and provide a more comprehensive understanding of how socio-economic disparities impact youth motor skill acquisition. Our future work will look to build on this by incorporating additional datasets and exploring a range of ensemble learning techniques, such as Random Forest, Gradient Boosting, and Stacking, to enhance predictive performance. These methods have shown promise in previous studies by combining the strengths of multiple models to reduce overfitting and improve classification accuracy. By leveraging ensemble learning, we anticipate achieving more robust F₁ Scores, thereby refining the predictive capability of MC assessments and ensuring greater generalisability across diverse populations. Another limitation of the study was the lack of age-related data, which could have given more insight into the predictions made by the models, allowing for comparison across the age span for all subsets. Our future work will seek to incorporate this going forward. Finally, whilst several classification models were utilised in our analysis, there were several that were not included (i.e., Random Forest, Naive Bayes, AdaBoost etc.). These models were excluded primarily due to computational constraints and the necessity of balancing model complexity with interpretability. However, their inclusion in future studies could provide valuable insights into classification performance, particularly in identifying non-linear patterns and interactions within the dataset. Random Forest, for example, could enhance robustness by aggregating multiple decision trees, while Naive Bayes might offer advantages in handling categorical data efficiently. AdaBoost, with its ability to improve weak classifiers, could further refine prediction accuracy. Future work will explore these models to assess their potential contributions to optimising MC assessments. Our future work will seek to add to this suite of models, eventually including more complex models (i.e., artificial neural networks).

5. Conclusions

In summary, our study demonstrates the potential for ML to optimise MC assessment procedures by reducing redundancy and improving efficiency. However, caution is warranted when considering assessment streamlining, as the removal of certain tests must be accompanied by rigorous validation across different populations and settings. Future research should focus on longitudinal applications of ML-driven assessments to explore the long-term impact of reduced assessment batteries on MC development and PA engagement. Additionally, further investigation into the role of PMC in MC prediction models is necessary to refine our understanding of how self-perceptions of competence influence motor skill acquisition and performance. Strengthening MC assessment procedures is critical to equipping children with the foundational skills necessary for lifelong PA engagement. This is particularly important given evidence that early childhood marks a crucial period in which children may follow either a positive developmental trajectory, characterised by increasing PA participation and improved health outcomes, or a negative one, marked by declining engagement and adverse health effects (Stodden et al., 2008). By addressing existing limitations in assessment methodologies, we can foster a more robust, equitable, and evidence-informed approach to MC evaluation, ultimately benefiting both practitioners and youth populations.