Enhancing human activity recognition with machine learning: insights from smartphone accelerometer and magnetometer data

Technological advancements and machine learning for HAR

Traditional machine learning (ML) methods, such as decision trees, random forests, and support vector machines (SVMs), have demonstrated effectiveness with hand-engineered features (Ronao & Cho, 2016). Recently, deep models such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) (e.g., long short-term memory (LSTM)) have gained popularity for their ability to extract features from raw data (Yang et al., 2015; Zhou et al., 2020). Hybrid CNN-LSTM models and Transformer-based networks (Gu et al., 2021) enhance sequential modeling, but they require substantial computational resources. Lightweight deep models optimized for smartphones (e.g., MobileNet, pruning) aim to reduce energy use (Shehab et al., 2022; Warden & Situnayake, 2019).

Sensor fusion and edge considerations

Our work builds on this literature by introducing a lightweight ML-based HAR system that integrates features from accelerometer and magnetometer sensors, a less commonly explored fusion. By extracting statistical and signal-derived features from each modality, we achieve rich data representation with minimal computational overhead. Unlike most recent works on deep learning, our model demonstrates that traditional ML algorithms, particularly random forests, can still outperform complex DL architectures in specific scenarios with appropriate feature engineering.

Edge computing is another growing trend. Modern smartphones now include neural processing units (NPUs) that support on-device inference for privacy, latency, and energy efficiency reasons (Huang et al., 2023). Our model design is aligned with this paradigm by requiring minimal processing resources and offering real-time responsiveness, making it suitable for deployment on consumer-grade devices.

Challenges and considerations

Smartphone-based HAR faces several key challenges. Ensuring user privacy and ethical compliance is paramount, requiring adherence to data protection laws such as the GDPR (European Union, 2016), along with secure data storage, transmission, and consent mechanisms. Techniques like differential privacy and federated learning are being explored to enhance data security (Jung, 2020). Variability in smartphone hardware leads to inconsistent sensor quality and availability, making it challenging to build universal HAR systems. This necessitates robust and adaptable algorithms, as well as the development of diverse benchmark datasets (Das Antar, Ahmed & Ahad, 2019). Another major concern is balancing computational load with battery life, as continuous sensing can rapidly deplete device resources. Lightweight and energy-efficient models, including those inspired by Tiny Machine Learning (TinyML), are gaining traction for deployment on low-power devices (Warden & Situnayake, 2019). Adaptability to real-world environments and individual user behaviors is essential, driving research into continuous learning and adaptive systems (Taylor et al., 2018). Finally, HAR is inherently interdisciplinary, requiring integration of knowledge from fields such as psychology, healthcare, and urban planning to effectively interpret human behavior and implement practical solutions (Jiang & Yin, 2015).

Applications and impact

• Healthcare and Rehabilitation: facilitates remote monitoring, fall detection, and physical therapy tracking (Mirjalali et al., 2022; Zhou et al., 2020).

• Fitness and Lifestyle: provides personalized activity feedback, promotes healthy living, and engages users through gamification (Phukan et al., 2022).

• Industrial Applications: enhances workplace safety and monitors worker ergonomics (Svertoka et al., 2021).

• Smart Homes and Urban Planning: supports energy optimization, safety systems, and public infrastructure planning (Bouchabou et al., 2021).

• Educational and Behavioral Research: enables new insights into student behavior, attention patterns, and mental health (Bosch et al., 2015).

Recent studies and future directions

Recent studies have achieved breakthroughs in creating real-time, efficient HAR systems that adapt to user feedback (Gu et al., 2021; Shehab et al., 2022). Integrating multimodal data and context-aware models expands the range and robustness of HAR applications. As HAR systems move toward seamless daily integration, future research will focus on user-personalized models, privacy-enhancing computation, and deeper fusion with IoT and wearable technologies. Our proposed system aligns with this direction by enabling on-device, efficient HAR through a novel fusion of magnetometer and accelerometer data, providing accurate yet resource-efficient performance.

Algorithms and code

The experiments in this study were conducted using various machine learning algorithms, including neural networks, random forest, support vector machine (SVM), CN2 rule inducer, Naive Bayes, and AdaBoost. These algorithms were implemented using the Orange data mining library (Team, 2024), a Python-based tool that provides an intuitive interface for data analysis and machine learning. The following algorithms were employed for activity classification:

• Neural Networks: configured with hidden layers and optimized using stochastic gradient descent (SGD).

• Random Forest: utilized for its robustness in classification tasks, employing a default number of trees.

• Support Vector Machine (SVM): configured with a radial basis function (RBF) kernel.

• CN2 Rule Inducer: applied for rule-based classification.

• Naive Bayes: utilized for its simplicity in probabilistic classification.

• AdaBoost: employed with decision trees as the base estimator to boost weak classifiers.

The source code for the data preprocessing, feature extraction, model training, and evaluation steps was created using Orange workflows. These workflows are reproducible and can be accessed by importing the workflow files into the Orange interface. The code allows researchers to preprocess the dataset, train multiple machine learning models, and evaluate their performance using various metrics.

Dataset

The dataset used in this study is publicly available at Pires & Garcia (2020). The dataset contains raw accelerometer, gyroscope, and magnetometer data for various human activities with motion. It includes multiple labeled activities such as walking, running, sitting, and more, and provides comprehensive data for implementing and testing machine learning models.

Implementation

This section introduces the code and workflows used in the study, providing a brief description of the dataset and its purpose within the context of human activity recognition.

The complete Orange workflow (.ows) used in this study is publicly available (Silva, 2025).

Open benchmark data

To support reproducible research and enable statistically valid future comparisons, we have made all raw cross-validation results publicly available. These include per-fold performance metrics (AUC, accuracy, F1-score, precision, recall) for each classifier and activity class. The dataset also contains confusion matrices and Orange workflow files used for training and evaluation. The data is accessible at Silva (2025).

Design and structure of the activity recognition study

In the present research, a mobile application was developed for the Android operating system. This application was installed on a BQ Aquaris 5.7 smartphone, strategically positioned in the anterior pocket of the participants’ trousers, to facilitate nonintrusive data acquisition during various activities of daily living (ADLs). The application was designed to capture sensor data pertinent to the performance of ADLs, with participants actively labeling the dataset to correspond with the activity being performed.

The data collection process involved 25 volunteers, comprising 15 males and 10 females, who used the application in various environments in Covilhã, Portugal. The dataset was gathered over approximately 180 h, with each ADL—specifically, moving downstairs, upstairs, running, standing, and walking—being recorded for 36 h. Each activity was documented in 5-s intervals at a sampling frequency of 40 Hz. The data acquired for this study is available at Pires et al. (2020a).

The participants were diverse, ranging in age from 16 to 60 years, and encompassed a spectrum of lifestyles: 40% of the individuals were engaged in regular physical exercise, while the remaining 60% had sedentary habits. The analysis of the data and the subsequent machine learning experiments were conducted utilizing the Anaconda software environment (Rolon-Mérette et al., 2016). The tests were conducted on a MacBook Pro in a setup specifically designed for machine learning research. The system features a Radeon Pro 560 graphics card with 4 GB of memory, a 2.6 GHz 6-core Intel Core i7 processor, and 16 GB of Random Access Memory (RAM) operating at 2,400 MHz DDR4.

All participants provided written informed consent, allowing us to disclose the test results anonymously. Additionally, the agreement provided the participants with informed consent based on the study’s purpose and potential hazards. Only the information about the people who signed the consent form to participate in the study was documented. The ethical considerations of this study were duly reviewed and approved by the Ethics Committee of the University of Beira Interior under the reference CE-UBI-Pj-2020-035:ID1965.

Data preprocessing strategies for sensor-based analysis

We executed two critical phases during the data preprocessing stage: data cleaning and feature extraction.

Configuration and training of machine learning models for activity recognition

The configuration and training of the machine learning models followed a methodical and rigorous approach. The preprocessed dataset was evaluated using a stratified 10-fold cross-validation strategy, in which the data was split into ten equal parts. Each fold was used once as a test set, while the remaining nine folds were used as the training set. This approach ensured that all observations were utilized for both training and validation, thereby reducing variance in performance estimation. Significantly, it enhanced the robustness and generalizability of the results by minimizing overfitting and providing a reliable evaluation of model performance on unseen data.

We implemented and configured six machine learning algorithms to recognize five daily human activities from smartphone sensor data: neural networks, random forest, support vector machines (SVM), CN2 rule inducer, Naive Bayes, and AdaBoost. Each classifier was implemented using the Orange data mining platform, which internally utilizes the scikit-learn library, allowing fine-grained control over model parameters.

Neural networks were configured as multi-layer perceptrons (MLPs) with a single hidden layer of 100 neurons. The logistic (sigmoid) activation function was used, and the model was optimized using stochastic gradient descent (SGD) with regularization set by $\alpha =0.0001$. The training was capped at 200 iterations to avoid overfitting.

Random forest classifiers used 10 decision trees with default maximum depth (unlimited). Splitting was stopped when a node contained fewer than five instances. Gini impurity was used as the splitting criterion.

Support vector machines (SVM) used a radial basis function (RBF) kernel, defined as $K(x,y)=\exp (-\gamma \Vert x-y{\Vert }^2)$, with $\gamma$ automatically selected. The regularization parameter was $C=1.0$, and $\epsilon =0.1$ was used for convergence. Training was limited to 100 iterations.

CN2 rule inducer was applied for interpretable, rule-based classification. The rule learning process used a beam width of 5, entropy as the evaluation metric, and a maximum rule length of 5. Rules were ordered and required a minimum coverage of 1 instance. The parameter $\gamma =0.7$ was used to balance accuracy and simplicity.

Naive Bayes was applied with default uniform priors. Variance smoothing was set to ${10}^{-9}$ to prevent numerical instabilities in probability estimation.

AdaBoost used 50 decision stumps (depth-1 trees) as weak learners. The boosting method used was SAMME.R, which incorporates class probability estimates to improve performance over discrete output.

Each model was meticulously configured with specific parameters to optimize the performance, as detailed in Table 2.

These models were trained on a MacBook Pro, which leveraged its advanced hardware capabilities. This hardware setup included a 2.6 GHz 6-core Intel Core i7 processor, 16 GB of RAM operating at 2,400 MHz DDR4, and a Radeon Pro 560 with 4 GB of graphics memory, providing the computational power necessary to handle the intensive demands of training multiple machine-learning models.

Upon completion of the training phase, the models were assessed for their efficacy in recognizing the specified activities of daily living. This assessment used a range of performance metrics, including area under the curve (AUC), classification accuracy (CA), F1-score, precision, and recall. These metrics were derived from the fundamental cardinalities of the confusion matrix, namely true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN), providing a comprehensive evaluation of each model’s performance.

Through this structured approach to model configuration and training, the study aimed to establish a robust framework for accurately recognizing and classifying human activities based on accelerometer and magnetometer data.

Assessment and validation of model performance in activity detection

Upon completing the training phase, the study conducted a comprehensive assessment and validation of the machine learning models. This evaluation was pivotal in determining the efficacy of the models in accurately recognizing human daily living activities. The assessment used five key performance parameters: area under the curve (AUC), classification accuracy (CA), F1-score, precision, and recall. These metrics were derived from the fundamental cardinalities of the confusion matrix, namely true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). The performance of each model was meticulously analyzed using these metrics, enabling a thorough validation of its capability to recognize the specified activities of daily living. This rigorous assessment ensured that the models were accurate and reliable in their predictions, thereby contributing significantly to the advancement of activity recognition using machine learning techniques.

• Classification Accuracy (CA): this metric was defined as the proportion of accurate predictions (both TP and TN) about all predictions made by the model. It provided a holistic view of the model’s overall accuracy in classifying activities. Mathematically, it is calculated as in Eq. (1). (1) $$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}=\frac{TP+TN}{TP+TN+FP+FN}.$$

• Precision: precision was calculated to determine the percentage of actual true positives among all optimistic predictions (TP and FP). This metric was particularly useful in understanding the models’ Type-I errors (FP). Mathematically, it is calculated as shown in Eq. (2). (2) $$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}=\frac{TP}{TP+FP}.$$

• Recall: also known as sensitivity, recall measures the proportion of actual true positives correctly identified by the model, offering insights into the model’s ability to detect positive instances. Mathematically, it is given as in the Eq. (3). (3) $$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}=\frac{TP}{TP+FN}.$$

• F1-score: the F1-score was employed as a harmonic mean of precision and recall, providing a balanced measure of the model’s performance, especially when the class distribution is imbalanced. Mathematically, it is expressed as in Eq. (4): (4) $$\mathrm{F}1\text{-}\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}=2\times \frac{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\times \mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}.$$

• Area Under the Curve (AUC): auc was used to evaluate the model’s ability to discriminate between the classes. A higher AUC indicated a better model performance in distinguishing between different activities. Mathematically, it is expressed as in Eq. (5). (5) $$\mathrm{A}\mathrm{U}\mathrm{C}={\int }_0^{1}TruePositiveRate(FalsePositiveRat{e}^{-1}(u))du$$

Data were collected as previously described in Hussain et al. (2023). Specifically, the binary classification issue leverages the parameters provided above. Following the one-vs-all (OVA) approach, which considers one class at a time, the multi-class classification performance parameters for each class were computed. Finally, all the results were computed by averaging every parameter for every class.

Analysis of the results

To summarize, the evaluation of all six machine learning classifiers trained to recognize the five ADLs on the test data reveals that the random forest and neural networks classifiers achieved the most outstanding scores compared to the performance of all other classifiers. Moreover, the performance of random forest and neural networks classifiers exhibits considerable similarity. Nevertheless, after a thorough comparison, the random forest classifier performed better than all other classifiers, including Neural Network classifiers, in accurately identifying the five ADLs. Conversely, the support vector machine classifier exhibits the lowest overall performance compared to all other classifiers.

Figure 2A displays the mean AUC values of all trained machine learning classifiers on the test data for identifying the five ADLs. The random forest and neural networks classifiers demonstrated superior performance in recognizing the five ADLs compared to the other classifiers. Simultaneously, the support vector machine classifier exhibited the worst performance: an average AUC score of 85.31% for identifying the five ADLs. Figure 2B displays the average CA ratings of all machine learning classifiers trained on the test data. Once again, the random forest and neural networks classifiers demonstrated superior performance compared to the other classifiers in accurately identifying the five ADLs. Once again, the support vector machine classifier exhibited considerably lower performance, with an average CA score of 76.50% in recognizing the five ADLs. Similarly, Fig. 2C displays the mean F1-scores of all trained machine learning classifiers on the test data. Once again, the random forest and neural networks classifiers outperformed all other classifiers in accurately identifying the five ADLs. The performance of the support vector machine classifier is considerably lower, achieving an average F1-score of 52.29% for the recognition of the five ADLs. Figure 2D displays the mean precision scores of all machine learning classifiers trained on the test data. The random forest and neural networks classifiers performed better than all other classifiers in identifying the five activities of daily living (ADLs). Nevertheless, the support vector machine classifier exhibited inadequate performance, with an average precision score of 58.52% in identifying the five ADLs. Figure 2E displays the average recall scores of all trained machine learning classifiers on the test data. Once again, the random forest and neural networks classifiers performed better than all other classifiers in recognizing the five ADLs. Conversely, the support vector machine classifier performed inferior to the other classifiers.

Compared to a previously published study (Pires et al., 2018) based on the same dataset (Pires et al., 2020a), this study considered non-normalized data to recognize patterns, thereby reducing the computational power required for processing on mobile devices. Regarding the use of accelerometer and magnetometer data, the previous study used three different types of neural networks, implemented with different frameworks, including Neuroph (Ševarac, 2012), Encog (Heaton, 2015), and DeepLearning4j (Team, 2016), reporting an average accuracy of 35.15%, 42.75%, and 70.43%, respectively. The present study globally improved the performance of HAR using machine learning on the same dataset.

While we compared our results to prior works using similar datasets and models, we acknowledge that a statistical comparison with related studies is limited by the lack of publicly available raw results. Without access to per-fold or per-class performance metrics from those studies, it is not feasible to perform rigorous statistical significance testing. We therefore encourage future work to adopt open data practices, and we aim to contribute to this by releasing our raw evaluation results.

Feature importance analysis

To increase the interpretability of our results, we examined feature importance using the Random Forest classifier, which provides built-in mechanisms to rank features based on mean decrease in impurity. The top-ranked features were predominantly time-domain statistical data, such as the standard deviation, mean, and interquartile range of the accelerometer’s Z-axis, followed by magnetometer-derived features like peak distance and signal variance.

It demonstrates that vertical mobility (Z-axis acceleration) and directional orientation (as measured by the magnetometer) are crucial in distinguishing between walking and stair-climbing activities. The feature importance analysis validates our design decision to include both accelerometer and magnetometer data. It demonstrates that basic statistical features can carry strong discriminative power, even without deep feature engineering.

Confusion matrix and misclassification analysis

To better understand how each model performs, confusion matrices were created to give a detailed picture of how well the models classified each activity. These matrices illustrate the frequency of correctly identified activities (true positives) and the locations of errors (false positives and false negatives). For example, both the random forest (Fig. 3A) and neural network (Fig. 3B) models performed very well, with a strong diagonal in their matrices, indicating high precision and accuracy. Some confusion did arise between “going upstairs” (Label 0) and “going downstairs” (Label 1), which is understandable given the similarity between those movements. On the other hand, the support vector machine (SVM) had a more scattered matrix, suggesting it struggled more with distinguishing between several activities, particularly “walking” (Label 4) and “standing” (Label 3). Below, we present all the confusion matrices in Fig. 3.

Statistical significance analysis of classifier performance

To evaluate whether the observed differences in classification performance among the six machine learning models were statistically significant, we performed a statistical analysis based on the 10-fold cross-validation results. Specifically, we compared the average F1-scores of each model across all folds using a paired t-test and Wilcoxon signed-rank test.

In the paired t-test comparing the random forest and support vector machine, the p-value was found to be $p<0.001$, indicating that the superior performance of the random forest was statistically significant. Similarly, comparisons between random forest and CN2 Rule Inducer ( $p=0.002$) and Naive Bayes ( $p=0.003$) also confirmed statistical significance.

Since the normality assumption could not be confirmed for all models, we performed a non-parametric Friedman test across all six classifiers using fold-wise F1-scores. The Friedman test yielded a statistically significant result ( ${\chi }^2(5)=17.25$, $p=0.004$). A post-hoc Nemenyi test revealed that random forest and neural networks significantly outperformed support vector machines and Naive Bayes.

These results validate that the performance improvements observed in random forest and neural networks are numerically higher and statistically significant.

Computational performance analysis

To assess the practicality of our models for real-world deployment—particularly on resource-constrained mobile devices—we evaluated their computational performance in terms of training time, prediction (inference) time per instance and model complexity.

All experiments were conducted on a standard consumer-grade laptop with an Intel Core i5-1135G7 processor (2.4 GHz), 8 GB of RAM, and running Windows 11. The Orange data mining platform was used for model training and evaluation, which internally leverages the scikit-learn machine learning library.

Training Time: the training time for the random forest classifier (10 estimators) was approximately 1.8 s, while the neural network (MLP with one hidden layer of 100 neurons) trained in 3.1 s. Simpler models, such as Naive Bayes and CN2 Rule Inducer, completed training in under 1 s.

Inference Time: for real-time feasibility, we measured the average inference time per instance across 1,000 predictions. Random forest and neural network required approximately 0.75 and 1.2 ms per prediction, respectively. Naive Bayes and CN2 Rule Inducer achieved inference times of less than 0.5 ms. These times are well within the acceptable range for real-time smartphone classification.

Model Size and Memory: the serialized model size for random forest was 190 KB, while the neural network model was approximately 210 KB. These compact footprints make them well-suited for deployment on embedded systems or mobile apps.

Energy and Resource Efficiency: because our models use raw (non-normalized) data and do not rely on deep learning layers or GPU acceleration, they consume minimal CPU and battery power, making them preferable for continuous background operation on smartphones.

Overall, our models’ computational profile confirms that they are suitable for real-time human activity recognition on mobile devices without the need for extensive cloud offloading or GPU resources.

Comparative advantages over state-of-the-art HAR models

Especially in performance, computational efficiency, and practical deployment, this study presents several notable advantages over existing HAR models:

• Enhanced Classification Accuracy Within the Same Dataset: this study achieves an average accuracy of 90.40% using random forest and 90.14% with neural network classifiers—substantially surpassing prior results on the same dataset (Pires et al., 2018), which yielded an accuracy range of 35.15–70.43% depending on the neural network framework.

• Efficiency Compared to Deep Learning Methods: recent HAR approaches such as CNN-LSTM hybrids (Abbaspour et al., 2020) and CNN-GRU designs (Gupta, 2021) offer strong accuracy but require high computational resources. In contrast, our models achieve similar or better performance while being lightweight and highly suitable for real-time mobile applications.

• Direct Use of Non-Normalized Data: most HAR pipelines require preprocessing steps, such as normalization, which adds to energy and memory costs. Our method omits this step, directly processing raw accelerometer and magnetometer data for more efficient computation on edge devices.

• Integration of Magnetometer Signals: while many models, including those by Carlos et al. (2024), achieve high accuracy (97–98%) using accelerometer and gyroscope data, they often neglect the use of magnetometers. Our study demonstrates that incorporating magnetometer data enhances discrimination, particularly for directionally sensitive activities such as stair climbing.

• Lightweight Architecture Compared to SETransformer and WSense Models: Liu et al. (2025) introduced SETransformer, which integrates attention mechanisms for enhanced feature learning, achieving 84.7% accuracy with high interpretability but moderate resource needs. Similarly, WSense (Ige & Noor, 2023) employed a 1D-CNN-based architecture across three sensor types, achieving robust accuracy improvements but at the cost of added sensor and computational demands. By contrast, our model delivers balanced performance (90.41% F1-score) with only two sensors and a simpler architecture, making it better suited for real-time, energy-efficient deployment.

• Reproducibility and Accessibility: implemented entirely within the Orange visual programming interface, our method facilitates quick prototyping and broad accessibility, even for users with limited coding experience (features often absent in models that require code-heavy environments or specialized hardware).

• Balanced Performance Across All Activities: our system maintains high precision and recall across both dynamic and static activities. Unlike deep learning models that may excel in running or walking but fail in transitional activities like standing or stair climbing, our approach consistently sustains strong performance across all five ADLs.

These distinctions are further summarized in Table 9, comparing model types, sensor configurations, performance metrics, real-time feasibility, and implementation complexity across major recent studies.

Merits and demerits of the study

This section outlines the primary strengths and limitations of the proposed approach, providing a balanced perspective. While the methodology demonstrates notable advantages in performance, efficiency, and accessibility, certain aspects, such as dataset scope and deployment validation, highlight areas for future improvement.

The study highlights the advantages of a lightweight model for high classification performance, its feasibility for real-time deployment on mobile devices, its use of magnetometer data for improved recognition, and its reproducible and accessible implementation in Orange.

However, the evaluation is limited to one dataset, has limited hyperparameter control in the Orange GUI, does not compare with standard public datasets, and has not yet been implemented for real-time system testing.

Explainability and model interpretability (XAI)

Although our study does not implement advanced post-hoc explainable artificial intelligence (XAI) frameworks such as SHapley Additive exPlanations (SHAP) or LIME, we deliberately selected and configured machine learning models that are inherently interpretable. Specifically, the CN2 Rule Inducer generates human-readable decision rules, making them directly explainable and suitable for scenarios requiring transparent decision logic. Random forests, while more complex, also allow insight into feature importance by calculating the contribution of each feature to the overall prediction accuracy.

We analyzed feature importance scores derived from the random forest classifier to gain further insight into model behavior. These scores highlighted that time-domain statistical features, such as the accelerometer’s mean, standard deviation, peak distance, and magnetometer signals, were the most influential in classifying activities, including walking and stair movement. This aligns with domain knowledge that motion intensity and directional changes are strong indicators of activity.

In future work, we plan to incorporate model-agnostic XAI tools, such as SHAP, to further enhance interpretability, particularly for more complex ensemble models. It will help identify how individual features influence specific activity predictions, improving model transparency and user trust, particularly in health-monitoring applications.

Reproducibility and open data

To facilitate fair statistical comparisons and enhance reproducibility, we have published our raw 10-fold cross-validation results for each classifier, including per-class F1-scores, AUC, and confusion matrices. These can be accessed at Silva (2025). We encourage future researchers to use this benchmark for statistically grounded comparisons.

Limitations

While this study provides a comprehensive workflow for human activity recognition, some limitations should be noted. Using a single dataset may restrict the generalization of the findings to other datasets or populations. Additionally, the reliance on the Orange library limits the flexibility to fine-tune algorithms beyond what is offered within its graphical user interface. This study did not account for hardware heterogeneity across smartphones, potential variations in battery consumption, or noisy conditions in uncontrolled environments, which could impact real-world performance.

Future work could address these limitations by incorporating diverse datasets and exploring more advanced machine-learning frameworks.

Future work

Future work will focus on extending the current approach by incorporating gyroscope data to improve the recognition of more complex and transitional activities. We also plan to enhance model transparency through explainable AI techniques, such as SHAP, to better understand feature contributions. To assess generalizability, we intend to validate our models on public HAR datasets like WISDM, PAMAP2, and UCI-HAR. Finally, we will explore real-time deployment and energy optimization to support practical use in mobile and wearable devices.