Dynamic Difficulty Adjustment in IoT-Enabled Rehabilitation Exergames Using Temporal Convolutional Networks

Abstract

Most dynamic difficulty adjustment (DDA) systems for rehabilitation exergames depend on heart-rate monitors or manual therapist scoring, hardware patients frequently cannot use reliably at home. This paper presents a sensor-agnostic DDA framework that reads exclusively from the smartphone already in the patient’s pocket. A multi-task temporal convolutional network (MT-TCN) processes raw 12-channel IMU data at 50 Hz to produce two simultaneous outputs: physical activity classification and a three-level movement-intensity state. The movement-intensity labels are generated by unsupervised clustering of acceleration-derived features; they approximate physical exertion and are not validated against clinical fatigue measures such as Borg RPE or EMG. An ϵ-greedy multi-armed bandit controller adjusts game speed, obstacle density, and restcheckpoint frequency once per 2.56 s window. Under leave-onesubject-out cross-validation on MotionSense (24 subjects, frontpocket iPhone) the MT-TCN achieves 97.1% activity accuracy and 0.89 macro-F1 for movement-intensity detection. Crossdevice transfer to UCIHAR (waist-mounted Samsung) with 10% target-domain labels reaches 89.1% accuracy on the five activities shared between the two corpora. The entire framework was developed and validated through simulation on publicly available datasets; no physical IoT devices or clinical populations were tested.