**3. Results**

As pictured in Figure 1, we identified 151 papers with three duplicates, that were removed. The other 148 articles were evaluated according to the title, keywords, and abstract, excluding 133 citations. After full-text evaluation, two papers were removed from the remaining 15 papers. The qualitative and quantitative synthesis included information related to the remaining 13 articles. In conclusion, we examined 13 documents.

**Figure 1.** Articles analysis.

To find relevant information about the implementations presented in the different studies analysed in this review, the reader should find the information in the original cited works. Table 1 shows the year of publication and the resume of the papers and final results. Table 2 shows the population, the purpose of the study, devices, settings of the papers, pros, and cons. When the datasets used in a study is publicly available, or the population information is provided, it is considered as a positive aspect. In many cases, the evaluation uses a cross-validation scheme (regular or stratified per class). However, the studies do not consider different subsets of the population for training and testing (i.e., train/test split based on subjects or patients). This is generally a more rigorous evaluation scheme and is expected to hurt the reported accuracy. Other more specific pros and cons are provided for each study.

The papers were published between 2012 and 2018, where two studies were published in 2018 (15%), four studies were published in 2017 (31%), two studies were published in 2016 (15%), two studies were published in 2015 (15%), two studies were published in 2014 (15%), and one study was published in 2012 (8%). Regarding the used devices, it was split among 43% for smartphones and the remaining 57% for mobile devices. The source code is not available for all studies analysed. Moreover, 69% of the studies have the raw data available. Finally, we verified that there are no studies that shared the source code.

#### *Methods for Identification of Activities in Daily Living*

In the study [57], the authors tried to use different classifiers for the recognition of activities with sensors to find the best method. Ten classifiers were utilized with the AdaBoost method. The dataset used was publicly available. The settings were investigated using nine inertial sensors from seventeen individuals taking into account 33 fitness activities. The used sampling rate was 50 Hz. After checking accuracies of the AdaBoost method, authors came to conclude that its implementation with random forest gives the best accuracy, with a value of 99.98%.

Authors of [49] have proposed harmonized authentication based on ThumbStroke dynamics (HATS) for mobile devices. The performance of HATS was tested, taking into account the different screen sizes of several mobile devices. Laboratory experiments were conducted to collect data for testing. Participants were required prior experience with touch screen devices and a qwerty keyboard. The study selected some features for learning ThumbStroke models, and these are timing features, spatial features, movement direction features, and operation features. The phrases, entered by the participants, were adopted from MacKenzie and Soukoreff and varied from 16 to 43 characters. Based method across all settings and classification models, the final results showed that HATS outperformed the keystroke dynamics. Among all the classification methods used, AdaBoost reported a maximum accuracy of 41.8%.

Li et al. [54] talks about an indoor/outdoor detection system (IOS). This method is split by the machine learning-based IOS-detector and the lightweight WiFi sub-detector. The first part infers indoor, outdoor, or semi-open environments based on the classification results. The second part focuses on the implementation of mobile devices. Finally, the other part consists of the IOS detection that shows high accuracy for the system. In conclusion, the proposed IOS detector achieves around 96% for the aggregated IOS detector and over 85% accuracy for the lightweight WiFi-based sub-detector.

In the study [50], the authors introduce a method for re-authenticating users taking into account a behavioral biometric-based on users' document scrolling traits. More specifically focused on identifying abnormal scrolling behavior on users while interacting with protected or read-only documents. Dataset was obtained from a previous project aimed to detect document access activities that indicate cyber attacks. Features for this paper were slit in vectors, being vector one derived from scrolling traits, vector two a representation of the polarity of scrolling, and vector 3 treats the dataset as a bipartite graph with two node sets. k-means clustering achieved the best performance with an 83.5% success rate in predicting the authenticated user.

The paper [48] presents a highly efficient method for the automatic detection of asthmatic wheezing in breathing sounds. The process is suitable for personal asthma monitoring via mobile devices since its not computationally complex. Most of the used data came from online databases of Human lung sounds. However, the authors also used several of their recordings of regular and wheezy breaths. The authors also confirmed the optimality of the audio spectral envelope (ASE) plus the value of the tonality index (TI) as a feature detector, using the mRMR (minimal redundancy–maximal relevance) method. Thousands of experiments were performed, and the best results were obtained from the fluctuation of the Audio Spectral Envelope descriptor adopted from the MPEG-7 standard, reporting an accuracy around 100%.

Authors of [53] developed a method to collect the sensor data, acceleration, gyroscope, geomagnetic, and atmospheric pressure were the four kinds of sensors used. The shallow feature extraction of the raw data happens before the CNN learning deep feature, which will reduce the complexity of the network and training time of the model. This process is critical for smartphones because of their limited resources. Three classes of features are extracted from each frame, including statistical, time, and frequency domains. Namely, the features used are: Mean, standard deviation, variance, median, minimum, maximum, range, interquartile range, kurtosis, skewness, root mean square, integral, double integral, autocorrelation, mean-crossing rate, fast Fourier transform, spectral energy, spectral entropy, spectrum peak position, wavelet entropy, and wavelet magnitude. Final results show that the proposed method can achieve 98% accuracy, meaning it outperforms the SVM (support vector machine) and AdaBoost classification in efficiency and computational cost, reporting accuracy of 93.6% with AdaBoost.

Yuan et al. [58] propose an indoor localization system using sensors for smartphones and smartwatches. Over 36,000 samples of data were collected in a 185.12 m<sup>2</sup> real indoor environment by a user using two different devices. Looking with the experimental results, the authors concluded that Twi-AdaBoost outperforms the state-of-the-art indoor localization algorithms. The localization error of position x and y achieved was 0.387 m and 0.398 m, respectively. The used datasets include the features: Place ID, Timestamp, Accelerometer\_X, Accelerometer\_Y, Accelerometer\_Z, MagneticField\_X, MagneticField\_Y, MagneticField\_Z, X\_Axis Angle (Pitch), Y\_Axis Angle (Roll), Z\_Axis, Angle (Azimuth), Gyroscope\_X, Gyroscope\_Y, and Gyroscope\_Z, reporting an accuracy around 99%.

In the paper [55], a novel technique based on the Bayesian voting algorithm that can be used with low-power sensors for transportation mode detection is presented. The authors used a set of data that consists of 400 h from eight individuals. Five sensors were used, being those: Acceleration, gyroscope, geomagnetic, barometer, and base station obtain by using AdaBoost classification to improve the results. Besides, the Bias algorithm was used to extract the features to reduce the adaptive boosting feature dimensions and determine the critical factors for identifying different transportation modes. The features used are: Mean, standard deviation, variance, median, minimum, maximum, range, interquartile, kurtosis, skewness, root mean square, time integral, double integral, auto-correlation, mean-crossing rate, fast Fourier transform, spectral energy, spectral entropy, spectrum peak position, wavelet entropy, wavelet magnitude, peak volume, intensity, length, variance of peak features, peak frequency, stationary duration, stationary frequency. Taking into account the final results, authors concluded that their algorithm could supply and replace some traffic pattern recognition algorithms and fix the problem that different mobile phones have various sensors, reporting accuracy between 64.54% and 96.83%.

In [51], the authors presented a contextual multi-armed bandits (MAB) approach that enables activity classification. This method makes context adaptation, continuous online learning, and active learning. Since the cost of extracting specific features is very high, the authors decided to use side information as the context. Since features can be used as contexts, this is not a limitation for the project. The proposed algorithm with active learning outperformed the benchmark algorithms by an average of 35%, reporting, and accuracy between 70% and 85%.

Xu et al. [47] focuses on three challenges, including the ability to accurately detect context using sensors and machine learning. The selection of activities for classification is performed by using context, reducing the complexity and improving the accuracy, speed, and energy usage, and the ability for experts in prescribing sets of physical activities under different environments. The features used for the project were: kNN (k-Nearest Neighbor) with time, kNN with wireless media access control (MAC) address and signal strength, and AdaBoost with audio peak frequency, peak energy, average power, and total energy. These were extracted from raw sensor data using a java program implementing the IContextFeatureExtractor interface. The data used was acquired by 14 participants that carried an Android mobile phone, and four 9-DOF devices were placed on dominant wrists, knee, ankle, and mid-waist. Each subject performed every required activity under every context for 2–5 min. The data were split into training (30%) and testing (70%) sets. Authors concluded that despite the methodology demonstrating effectiveness, efficiency, and potential, a more extensive study needs to be performed to improve privacy, security, and user-friendliness, reporting accuracy between 59% and 100%.

In [56], the problem of occupancy detection in a domestic environment was studied using machine learning techniques and their boosting versions on a dataset collected from electricity and water consumption smart meters. These features were selected using the Mutual Information technique. The dataset contains energy and water consumption (during summer) time data of 1-minute resolution for 16 consecutive days. The features included in the used dataset were: Central power, refrigerator, television, washing machine, dryer, cold water-kitchen, hot water-kitchen, dishwasher-water and washing machine-water, reporting accuracy higher than 70%.

Authors of [52] evaluated ten representative classifiers in the identification of two available datasets. The first dataset consists of accelerometer readings of walking patterns from 22 participants. The second one contains activity and postural transition data collected from the accelerometer and magnetometer data acquired from 30 participants. For the Walking dataset, the authors split the data into fixed-width sliding windows with a 50% overlap and extract nine features from every window and scale the features to [ −1, 1]. The authors obtained the mean, standard deviation, and median absolute deviation from the different axis of the sensors. The authors of the study already pre-processed the sensor signals by noise filter and partitioned the data into fixed-width sliding windows with a 50% overlap as well and constructed a 561-feature vector for every window. From those features, authors extracted 24 features, including mean, standard deviation from the different axis of body acceleration, gravity acceleration, jerk signals of body acceleration, angular velocity, and jerk signals of angular velocity. In conclusion, the authors reported an accuracy between 95.6% and 97.8%.

The study [46] focuses on using mobile devices for the detection of cardiovascular autonomic neuropathy. The authors concentrated on the task of the detection and monitoring of cardiovascular autonomic neuropathy. After all the studies, they concluded that best outcomes were obtained by the novel combined ensemble of AdaBoost and Bagging based on the J48 decision tree, reporting the highest accuracy of 94.53%.
