A deep hybrid learning framework with attention-enhanced feature extraction for BMI prediction based on physical fitness.

Background: Body Mass Index (BMI) assessment remains a critical challenge in university health monitoring, where traditional methods fail to capture dynamic relationships between physical fitness indicators and body composition. This study develops a novel predictive framework to address this gap through advanced machine learning techniques applied to longitudinal fitness data from Chinese university students. Methods: We analyzed 6,698 male students' fitness records (2018-2022) using a hybrid CNN1D-Attention-LightGBM architecture. The model integrates temporal pattern recognition via sliding windows, multi-kernel convolutional operations for physiological coupling analysis, and dynamic attention weighting. Performance was validated through 10-fold cross-validation against SVM and XGBoost benchmarks. Results: The model achieved 94.5% accuracy (F1 = 0.93), significantly outperforming conventional methods (XGBoost: 90.1%). Cardiorespiratory endurance (3000 m run, r = 0.2009) and upper-body strength (pull-ups, r = -0.2786) emerged as primary BMI determinants. The framework successfully classified four BMI categories (normal weight: 4,991; obese: 82) . Conclusion: This study establishes the first unified solution for fitness-informed BMI prediction, though limited by male-only sampling. Implementation should prioritize integration with campus health systems and expansion to diverse populations. Future work should incorporate psychosocial factors and multi-regional validation.