Hybrid Machine Learning Predicts Exercise Calorie Burn and Guides Personalized Training

Researchers at Lovely Professional University have demonstrated that pairing supervised machine learning with rule-based reasoning can predict workout energy expenditure and classify physical activity levels with high precision, bypassing the constant requirement for wearable sensors. Evaluating retrospective records from 973 fitness facility members, the team found that gradient-boosted decision trees achieved a coefficient of determination of 0.91 when estimating caloric expenditure, while random forest classification algorithms sorted participants into activity intensity tiers with 92 percent accuracy. The hybrid architecture integrates tabular physiological measurements, anthropometric markers, cardiovascular indicators, and reported exercise routines to address key limitations in consumer fitness technology. While conventional platforms frequently depend on continuous data streams from proprietary hardware or provide uniform routines that disregard personal physiological baselines, the evaluated system translates discrete physiological inputs into actionable fitness guidance. Joining predictive modeling with deterministic physiological rules provides an interpretable mechanism designed to support individualized health recommendations and preventive lifestyle interventions.

The global expansion of sedentary living patterns and chronic metabolic conditions has intensified the demand for accessible, individualized physical activity interventions. Contemporary healthcare models increasingly emphasize preventive strategies over reactive care, prompting the development of artificial intelligence applications intended to promote sustainable behavioral modifications. Existing digital health literature shows that machine learning algorithms can dynamically adjust exercise objectives based on user behavioral patterns and physiological states, improving adherence over extended intervention periods. Simultaneously, consumer adoption of wearable activity monitors and smartphone applications has established the feasibility of tracking physiological indicators like heart rate and caloric expenditure. However, many commercial solutions treat exercise tracking as an isolated metric-collection task, focusing almost exclusively on step counts or energy estimates without evaluating broader contextual parameters such as training experience, baseline cardiovascular response, or workout consistency. By developing a hybrid modeling structure, the authors sought to unify multi-factor physiological profiling with structured rule-based reasoning to generate coherent, safe, and contextually grounded exercise prescriptions.

To construct and evaluate the hybrid recommendation system, the research team analyzed the publicly available Gym Members Exercise Dataset sourced from Kaggle, comprising de-identified records from 973 fitness facility attendees. The cross-sectional dataset captured a diverse spectrum of demographic variables, anthropometric measurements, cardiovascular observations, and documented workout habits without containing personally identifiable information. Anthropometric parameters included measured height and weight, from which body mass index values were recalculated using standard formulas. Cardiovascular features incorporated resting heart rate, average exercise heart rate, and maximum recorded heart rate during sessions. Detailed exercise attributes documented specific workout modes, individual session durations, weekly workout frequencies, and measured caloric expenditure, with calories burned functioning as the primary target variable for continuous regression tasks. Furthermore, the dataset incorporated body composition and lifestyle metrics, specifically body fat percentage, hydration levels, and subjective training experience levels. The dataset contained zero missing values across all tracked features, eliminating any necessity for synthetic data imputation during modeling.

Before training predictive models, the investigators executed a structured data preprocessing and feature engineering pipeline within an environment running Python version 3.10, scikit-learn version 1.3, and XGBoost version 1.7. Categorical features, including participant gender, workout mode, and reported experience level, underwent one-hot encoding to prevent artificial ordinal bias in mathematical estimators. To eliminate scale variations across disparate physiological measurements, continuous variables were standardized using z-score normalization. Outlier detection was conducted exclusively on the training split using the interquartile range rule, removing values outside 1.5 times the interquartile range to prevent data contamination from leaking into test partitions. The cleaned dataset was partitioned into training and testing subsets using an 80:20 stratified split based on activity level categories. In addition to primary features, the researchers engineered an activity intensity score derived from cardiovascular exertion and computed total weekly training loads, capturing physical effort and routine consistency across five-fold cross-validation folds.

Exploratory analysis of the 973 participant profiles demonstrated substantial physiological heterogeneity across recognized body mass index categories, providing a balanced foundation for downstream goal inference modeling. Normal weight individuals constituted the largest group at 41.3 percent of the sample with 402 members, followed by overweight individuals representing 29.7 percent with 289 members, obese individuals accounting for 16.9 percent with 164 members, and underweight participants comprising 12.1 percent with 118 members. Pearson correlation analyses revealed statistically significant linear associations between physiological parameters and caloric expenditure at p < 0.05. Average exercise heart rate exhibited the strongest positive correlation with caloric expenditure at r = 0.72, followed closely by workout session duration at r = 0.69. Weekly workout frequency demonstrated a moderate positive correlation of r = 0.53, while body mass index registered r = 0.48, and experience level showed r = 0.37. Conversely, body fat percentage demonstrated a negative correlation of r = -0.41, reflecting metabolic efficiency differences during physical exertion across diverse physiological profiles.

Predictive evaluations for energy expenditure regression revealed notable differences across model architectures tested across five-fold cross-validation on unseen test data. Baseline linear regression yielded a mean absolute error of 48.6, a root mean square error of 63.2, and a cross-validated coefficient of determination of 0.79 with a standard deviation of 0.03, failing to model non-linear interactions among cardiovascular metrics. In contrast, ensemble tree-based architectures demonstrated superior precision. A random forest regressor, configured with 200 trees and a maximum depth of 12, achieved a mean absolute error of 28.4, a root mean square error of 39.7, and a test R-squared of 0.89 with a standard deviation of 0.02, alongside a training R-squared of 0.93. The XGBoost regressor, using 300 estimators, a learning rate of 0.05, and a maximum depth of 6, achieved the highest accuracy, recording a mean absolute error of 24.9, a root mean square error of 34.8, a training R-squared of 0.95, and a test R-squared of 0.91 with a standard deviation of 0.02.

For the discrete classification task, investigators categorized participants into low, moderate, and high activity intensity groups using empirical percentile thresholds at activity scores below 0.35, between 0.35 and 0.65, and exceeding 0.65. Benchmark comparisons across classification algorithms highlighted varying capabilities in resolving overlapping physiological boundaries. Multinomial logistic regression provided baseline utility, achieving an accuracy of 0.84, precision of 0.83, recall of 0.82, and an F1-score of 0.82. A support vector machine utilizing a radial basis function kernel with C = 1.0 improved class separation, yielding an accuracy of 0.88, precision of 0.87, recall of 0.86, and an F1-score of 0.86. The random forest classifier, configured with 150 trees and a maximum depth of 10, achieved top performance across all metrics, delivering an accuracy of 0.92 with a standard deviation of 0.02, alongside precision, recall, and F1-scores of 0.91, 0.90, and 0.90, respectively. Confusion matrix analyses confirmed that misclassifications occurred primarily between adjacent moderate and high activity bands due to continuous physiological overlap.

Feature importance analyses conducted via Gini impurity calculations across tree-based models revealed that cardiovascular metrics exerted primary influence over energy expenditure predictions and activity classifications. Average exercise heart rate ranked as the most influential predictive variable with an importance value of 0.28, followed by workout session duration at 0.24 and maximum exercise heart rate at 0.14. Habitual and behavioral metrics demonstrated intermediate predictive weight, with weekly workout frequency contributing 0.11 to model decisions. In contrast, anthropometric and lifestyle indicators exhibited smaller direct effects on acute energy expenditure estimation, with body mass index registering an importance value of 0.09, while body fat percentage and participant experience level each contributed 0.07. The experimental findings indicate that while acute cardiovascular workload governs immediate metabolic demands during exercise, anthropometric characteristics like body mass index operate as contextual modulators. These anatomical features prove far more critical for downstream fitness goal determination and structural workout program design than for immediate mathematical calculations of session caloric expenditure.

The recommendation module converted predictive model outputs and user profiles into tailored fitness guidance using a transparent rule-based goal inference engine aligned with World Health Organization guidelines and exercise physiology principles. Individuals presenting with a body mass index of at least 25 alongside a body fat percentage of 25 percent or greater were assigned weight-loss goals, receiving moderate-intensity cardiovascular exercise prescriptions paired with caloric deficit guidance. Participants with a body mass index below 18.5 were assigned weight-gain objectives supported by structured resistance training and surplus nutritional guidance. Experienced participants demonstrating high training consistency with weekly training loads of at least 300 were categorized as muscle-gain candidates and prescribed progressive high-intensity resistance training. Individuals exhibiting resting heart rates below 60 beats per minute were identified as candidates for aerobic endurance interval conditioning, while all remaining balanced profiles were assigned general fitness maintenance routines. This explicit deterministic mapping ensures that recommendations remain biologically plausible, interpretable by fitness practitioners, and directly traceable to measurable physiological thresholds rather than unverified generative assumptions.

Despite the high mathematical accuracy achieved across both regression and activity classification benchmarks, the investigators emphasized several methodological boundaries inherent to their study design. Because the evaluation relied exclusively on retrospective secondary analysis of 973 Kaggle records without prospective field deployment, user satisfaction surveys, or longitudinal adherence tracking, the practical efficacy of the generated recommendations remains unverified in real-world environments. The empirical activity thresholds and physiological decision boundaries, though derived from percentile distributions and established clinical literature, were not tested through systematic sensitivity analyses across diverse external populations. Furthermore, the dataset lacked behavioral variables such as verified nutritional logs, sleep metrics, and psychological determinants known to influence long-term exercise compliance. The authors noted that future iterations could incorporate dynamic sensor streams and reinforcement learning to update recommendations as user fitness evolves over time. Until prospective clinical validation and end-user usability assessments are conducted in community centers and mobile health environments, the reported hybrid framework represents an interpretable proof-of-concept for scalable digital fitness coaching.

Subject of Research: Hybrid machine learning and rule-based modeling for personalized exercise and health recommendations

Article Title: AI-based personalized exercise and health recommendation system using hybrid modeling

Article References: Singh, S., Singh, A., Tiwari, A., Gurjar, H., Patel, S., & Kaur, A. (2026). AI-based personalized exercise and health recommendation system using hybrid modeling. Frontiers in Artificial Intelligence, 9, Article 1815026. https://doi.org/10.3389/frai.2026.1815026

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1815026

Keywords: machine learning, exercise recommendation, caloric expenditure, activity classification, rule-based reasoning, hybrid modeling, random forest, XGBoost, cardiovascular metrics, heart rate analysis, body mass index, digital health, preventive healthcare, fitness tracking, training load, feature importance, supervised learning