Forecasting Dropout in Home-Based Movement Rehabilitation After Stroke With Sensors and Machine Learning

Adherence to home-based rehabilitation can support recovery after stroke, yet many patients disengage within the first few weeks. While prior studies have examined perseverance in small samples or under supervised settings, little is known about early behavioral signals of long-term perseverance in large, unsupervised cohorts. This study analyzes one of the largest datasets on home rehabilitation perseverance measured with a sensorized system (N = 2,747 FitMi users) and applies interpretable machine learning models to forecast dropout risk. We focused on generalizable features (features that are readily available in many other home exercise systems). As a starting point for analysis, we used the first three weeks of exercise behavior to predict persistence into the fourth week, the highest-performing model achieved an AUC of 0.810, indicating strong separability between persisters and non-persisters. The precision-recall trade-off was also strong, with prAUC = 0.717, and F1-score = 0.703. For this model, feature importance analysis identified average weekly active days (24.7%), exercise consistency (16.4%), average selected session length (15.2%), and repetition rate (10.1%) as the most predictive behavioral signals. These findings demonstrate that early engagement patterns, which are easily measurable with a wide range of sensor-based systems, contain robust information for identifying users at risk of stopping home movement practice. Such identification could provide a scalable pathway for adaptive, personalized interventions in autonomous rehabilitation systems.