A Bidirectional LSTM Framework for Long-term Exercise Pattern Recognition and Physical Health Promotion Effect Prediction

This study developed a bidirectional LSTM neural network framework for analyzing long-term exercise patterns and predicting multidimensional health outcomes from wearable sensor data. The model demonstrated 92.7% accuracy in recognizing diverse exercise modalities while characterizing temporal exercise dynamics over a 12-month period. Seven distinct exercise pattern clusters were identified, with differential effects across physical fitness, physiological function, body composition, and psychological wellbeing dimensions. Exercise consistency and temporal distribution emerged as stronger predictors of health improvements than traditional volume metrics. The predictive framework enabled personalized exercise prescription optimization, demonstrating potential improvements of 21–35% in health outcomes compared to standardized approaches. These findings suggest the need for paradigm shifts toward consistency-prioritized rather than volume-maximized exercise approaches, with individualized optimization based on personal response patterns rather than population averages.