Publikation
Privacy-Preserving Edge AI for Rehabilitation Support: Lessons from a Field Study
Sabine Janzen; Prajvi Saxena; Muhammad Ebad Ullah Khan; Wolfgang Maaß
In: Proceedings of "Digital Health Perspectives - Bridging Healthcare, Technology and Society" at Informatik 2026. Jahrestagung der Gesellschaft für Informatik (INFORMATIK-2026), Dresden, 9/2026.
Zusammenfassung
AI-supported rehabilitation promises more individualized therapy by accounting for subjective factors such as perceived difficulty, expectations, motivation, fear of pain, and situational constraints. However, such systems must be evaluated not only for predictive validity, but also for privacy-preserving deployment feasibility. We report a one-day public field study of an Artificial Mental Model (AMM)-based rehabilitation support system with 220 valid study sessions. Participants completed a short profile, received an AI-based prediction of perceived exercise difficulty, performed a standardized sit-to-stand exercise, and reported their actual difficulty. The system was deployed locally on a constrained edge setup without cloud processing which enabled the study to be conducted without relying on cloud services for sensitive health data. Results show operationally robust no-cloud deployment, while predictive validity remained below pre-specified targets. Agreement responses provided a plausibility check, as participants were more likely to agree when predictions were close to their post-exercise ratings. The study highlights the need for context-sensitive calibration, human-centered evaluation, and low-latency privacy-preserving deployment.