Physical Activity Prediction for Patients in a Diabetes Telehealth System: Comparison of Different Data Driven...

Diabetes poses a substantial and growing burden on healthcare systems worldwide. Physical activity and telehealth are effective strategies for the prevention and management of diabetes. DiabMemory, an Austrian diabetes telehealth platform, enables patients to upload and monitor, among others, physical activity data. In this study, retrospective physical activity data from DiabMemory were used to develop and evaluate predictive models for predicting weekly physical activity levels on a given day of the week. Four approaches were compared: a baseline method, an autoregressive integrated moving average model, the gradient-boosted decision tree algorithm XGBoost, and a long short-term memory (LSTM) neural network. We observed that the LSTM outperformed the other algorithms on all days, with an error between 14% on Saturdays and 58% on Mondays compared to the mean ground truth, depending on the day of the prediction. The predictive models developed in this work provide a foundation for the integration of an activity-prediction module within the DiabMemory platform, enabling personalized, data-driven feedback to support patient-specific physical activity management.