Analysis of lifestyle factors associated with physical activity participation among university female students using...

Female university students often report lower levels of physical activity compared to their male counterparts. While prior research has examined psychological, sociocultural, and physiological factors associated with these differences, many quantitative approaches rely on linear models that may not capture nonlinear relationships. This study proposes a Deep Belief Network (DBN) framework to model associations between lifestyle factors and physical activity status among female university students. A sample of 1,032 female students was analyzed using four predictor variables: Stress Level, Study Hours, Sleep Hours, and Social Hours. These variables were preprocessed through encoding and Min-Max normalization, and physical activity was represented as a binary classification outcome. The DBN model was trained using unsupervised pre-training followed by supervised fine-tuning. It outperformed baseline models, achieving a mean accuracy of 92.4% (±1.2), F1-score of 91.3% (±1.5), and Area Under the Curve (AUC) of 0.96 (±0.01) across five-fold cross-validation. Bootstrap analysis on out-of-fold predictions yielded 95% confidence intervals of [90.1, 94.6%] for accuracy, [89.0, 93.5%] for F1-score, and [0.94, 0.98] for AUC. Latent feature analysis indicated that stress level showed the strongest association with model predictions, followed by social and physiological variables, while study hours showed a comparatively weaker association. Sensitivity analysis revealed nonlinear patterns, including a threshold effect for stress and a U-shaped association between social hours and physical activity. These findings demonstrate the ability of deep learning models to capture nonlinear associations within a limited set of observable lifestyle variables. However, the study does not directly measure psychological motivation constructs, and the results reflect predictive associations rather than causal relationships; therefore, causal inferences cannot be drawn from this analysis. Future research should validate these findings using longitudinal data to examine temporal dynamics and potential causal relationships. Incorporating validated psychological constructs and objective measurements, as well as evaluating the framework across diverse populations, may further improve interpretability and generalizability.