Development and validation of a deep learning-based prediction system for movement-related health risks
Physical activity patterns are crucial indicators of overall health and disease risk. Despite the availability of wearable devices and activity trackers, interpreting complex movement data to predict health outcomes remains challenging. This study aimed to develop and validate an interpretable deep learning model to assess health risks based on daily movement patterns.
We conducted a comprehensive study involving 15,000 participants with continuous activity monitoring data collected over 2 years. The innovative Movement Health Risk Prediction System integrates LSTM and Self-Attention mechanisms to analyze temporal activity patterns and predict health risks.
The system processes 192 time-series features representing hourly activity distributions across weekdays and weekends, categorized into four activity types: sedentary behavior, light activity, moderate-to-vigorous activity, and sleep. This enables precise investigation of activity patterns and individualized health risk assessment. Predictive performance was evaluated using metrics including area under the ROC curve (AUC), sensitivity, and specificity.
External validation was performed on an independent cohort of 800 participants to assess real-world applicability. The model architecture employs advanced deep learning techniques, combining Long Short-Term Memory (LSTM) networks for temporal pattern recognition with Self-Attention mechanisms for identifying critical time periods and activity patterns.
Model interpretability was enhanced using SHapley Additive exPlanations (SHAP), which provides insights into how specific activity patterns contribute to health risk predictions. This transparency helps users understand which aspects of their daily routine most significantly impact their health风险评估.
Feature importance analysis revealed that patterns including prolonged sedentary bouts, irregular sleep schedules, and insufficient moderate-to-vigorous physical activity were key predictors of health risks. The system specifically identifies risk factors for metabolic disorders, cardiovascular issues, and musculoskeletal problems.
Our system accurately classifies individuals into 3 distinct risk categories (low, medium, high) based on their movement patterns. Compared to traditional assessment methods, our model demonstrated superior performance with mean AUCs of 0.892 for metabolic risk, 0.865 for cardiovascular risk, and 0.878 for musculoskeletal risk. For high-risk individuals, the model achieved particularly strong performance (AUC = 0.915; Sensitivity: 0.845; Specificity: 0.872).
The implementation of this prediction system could significantly improve early detection of activity-related health risks, enable personalized intervention strategies, and ultimately enhance population health through targeted activity recommendations. The model's ability to identify at-risk individuals based on their daily movement patterns provides valuable insights for preventive healthcare and lifestyle interventions.
To facilitate accessibility, the model has been deployed as a web-based health assessment tool with real-time analysis capabilities. Users can upload their activity data or use our example data to receive immediate health risk assessments and personalized recommendations.
Key features of our system include:
- Analysis of 96-hour activity patterns (48 weekday hours + 48 weekend hours)
- Four-category activity classification every hour
- Personalized risk scores and improvement recommendations
- Interactive visualization of activity patterns and risk factors
- Comparative analysis against healthy activity benchmarks