TL;DR
An AI-driven calorie burn prediction system that leverages machine learning to provide an accessible, low-cost alternative to conventional methods, with potential applications in personalized fitness analytics and digital health platforms.
Precise step length estimation (SLE) is a key necessity for not only navigation systems design but also gait health monitoring in neurological conditions. Among existing solutions, non-invasive inertial sensor-based approaches operating without dedicated infrastructure are more cost-effective. Many such methods, however, rely on bodily affixed inertial sensors rather than freely held smartphone sensors. Traditional signal processing approaches, on the other hand, offer varying accuracy across diverse gait patterns due to user parameter calibration. This study, thus, proposes a regression-based SLE framework employing eight regression algorithms: linear regression (LR), k-nearest neighbors, support vector machine, decision tree, elastic network, random forest, histogram-based gradient boosting (HGB) regressor, and artificial neural network (ANN). Their extensive and rigorous evaluations across varied window sizes, using a dataset collected in normal and fast walking modes with two device positions (hand-held and trouser-pocket) during three evaluation scenarios, demonstrate the HGB regressor’s outstanding performance, achieving the lowest mean absolute error (MAE) below 1 cm across four different contexts under leave-one-out cross-validation-based evaluation and three in the seen test evaluations. Moreover, the findings report the ANN’s exceptional generalization capacity over other models and the previous method IRT-SD-SLE in unseen test evaluations, with an MAE not exceeding 6.3 cm. The extensive evaluations of training and testing times reveal the highest computational efficiency for LR, moderate efficiency for the HGB regressor, and the highest training cost for the ANN, indicating a clear trade-off between MAE and computational expense. Additionally, this study includes an insightful discussion on the performance results, including the trade-offs between accuracy and efficiency.
This paper presents a real-time food recognition and calorie estimation system that leverages the YOLOv8 nano object-detection architecture coupled with a curated nutritional knowledge base and outperforms prior approaches based on Faster R-CNN and SSD MobileNet on both speed and accuracy metrics.
Highlights What are the main findings? Explainable machine learning model, especially XGBoost, outperformed the conventional linear regression in predicting six-minute walk distance (6MWD), as it can better model complex and non-linear relationships between anthropometric parameters in children with obesity. Age is the main predictor of 6MWD in all models, with sex-specific different patterns of body mass index, height and waist circumference. What are the implications of the main finding? The use of SHAP-based explainable AI with machine learning offers can bypass the limitations of conventional linear models and provide a transparent and accurate approach for the assessment of functional exercise capacity in pediatric obesity. These newly derived ML-informed prediction equations are practical tools for field-based clinical assessment tailored to each subgroup, but external validation in large cohorts is required before routine implementation. Abstract Background: The six-minute walk test (6MWT) assesses functional exercise capacity, but existing reference equations for children with obesity rely on traditional linear regression, potentially overlooking complex, non-linear relationships between anthropometric characteristics and functional exercise capacity. Objective: This study aimed to develop and internally validate machine-learning (ML) prediction models and preliminary prediction equations derived from explainable ML models for six-minute walk distance (6MWD) in Tunisian school-aged children with obesity and to compare their predictive performance with a conventional regression-based approach. Methods: We analyzed data from 236 school-aged children with obesity (104 females, 132 males; 6–12 years). Anthropometric measurements included body mass (BM), height, body mass index (BMI), waist circumference (WC), and hip circumference (HC). Five models were evaluated: linear regression, Ridge, Lasso, Elastic Net, and XGBoost. Performance was assessed using five-fold cross-validation and evaluated by the mean absolute error (MAE) and root mean square error (RMSE). Predictor importance was assessed using SHapley Additive exPlanation (SHAP) and Gini importance. Results: XGBoost achieved the best predictive performance, with the lowest MAE (17.4 ± 2.5 m in females and 19.5 ± 1.9 m in males) and RMSE (25.3 ± 3.1 m in females and 26.4 ± 2.2 m in males). Age was the strongest predictor across all models (SHAP: 54.2–62.0%; Gini importance: 0.52–0.69), followed by height and BMI. Sex-specific analyses indicated that, in females, age and BMI contributed ~80% to the cumulative SHAP analysis; whereas, in males, age, height, and WC were the primary factors. Conclusions: ML, particularly XGBoost, significantly improves 6MWD prediction in school-aged children with obesity compared with traditional linear regression. Explainable ML increases model interpretability by evaluating the relative importance of anthropometric predictors. These obesity-specific prediction models may better capture complex non-linear associations between anthropometrics and functional exercise capacity. These initial population-specific models should be validated in larger independent samples before routine use clinical or field settings.
This study proposes an ensemble-learning-based prediction framework for physical fitness assessment using physiological and performance data acquired from conventional testing systems and wearable sensing platforms. Student attributes, including gender, height, weight, and vital capacity, are utilized as input features, while physical fitness indicators are treated as prediction targets. To improve prediction robustness and anomaly sensitivity, a two-stage ensemble architecture combining decision trees and multilayer perceptrons is developed. The decision tree provides preliminary predictions and suppresses the influence of outliers, whereas the multilayer perceptron performs nonlinear feature refinement to enhance predictive accuracy. Experimental evaluation based on 3,000 university students demonstrates that the proposed framework consistently achieves lower prediction errors than individual learning models in both 50-meter sprint and standing long jump assessments. Furthermore, anomaly analysis enables the identification of potential measurement deviations and performance-improvement opportunities. Longitudinal validation indicates that students selected through the proposed framework and subjected to personalized interventions achieve significantly greater improvements than control groups. By interpreting physical fitness assessment as a wearable sensing and data-analysis problem, the proposed approach establishes a closed-loop process linking data acquisition, anomaly detection, performance prediction, and intervention optimization. The framework provides an engineering-oriented methodology for intelligent health monitoring, signal-based performance evaluation, and personalized training support in large-scale monitoring environments.
Objective. The first objective of this study was to refine previously designed machine learning models that predict energy expenditure (EE) of preschool children by modifying the method used to calculate metabolic equivalents (METs). The secondary objective was to compare estimates of time spent in different physical activity intensities across the newly developed models, previously published METs models, existing METs-based models from the literature, and models calibrated using direct observation. Approach. The model training dataset included 35 Canadian children (aged 3.0–5.99 years) equipped with GT9X accelerometers on their right hip. A portable metabolic unit was used to measure EE during a semi-structured protocol consisting of activities ranging from low- to high-intensity. The resulting models were applied to a sample of Canadian preschool children (n = 118; aged 3.0–5.99 years) to estimate time spent in sedentary (SED), light (LPA), moderate-to-vigorous (MVPA), and total physical activity (TPA). A repeated measures ANOVA was used to compare time estimates across models and according to three different configurations of METs activity thresholds. Main results. Results indicated that the newly developed models from Objective 1 produced significantly different estimates of time spent in SED, LPA, MVPA, and TPA compared to both previously published models and other existing METs-based models, highlighting the impact of different approaches to calculating METs. Significance. Model selection and METs calculation methods markedly influenced activity intensity estimates, underscoring the need for consistent methodology. Classification models yielded the most plausible free-living estimates.