Emotional artificial intelligence at the edge for tele-rehabilitation

Tele-Rehabilitation (TR) has gained increasing attention in recent years due to its potential to reduce costs and overcome physical barriers compared with traditional in-person treatments. Within the framework of the Tele-Rehabilitation as a Service (TRaaS) research project, funded by the Italian Ministry of University and Research, previous work focused on developing a Cloud/Edge infrastructure to deliver remote rehabilitation to patients with Severe Acquired Brain Injuries (SABI), including cerebral stroke, Alzheimer's, and Parkinson's diseases. Building upon that work, this study aims to evaluate the applicability of Emotional Artificial Intelligence (AI) in TR environments through an AI-based Facial Expression Recognition (FER) system designed to monitor patients' emotional states during therapy exercises.

The proposed FER pipeline was deployed and tested on two computational environments: a Raspberry Pi 4B representing an edge device, and a Virtual Machine (VM) emulating a cloud environment. Several experiments were conducted to compare their performance in terms of processing time, resource usage, and feasibility for real-time deployment in TR scenarios.

The experiments characterize the computational cost of the pipeline rather than its recognition quality or its clinical effect. On the Raspberry Pi 4B, the complete pipeline sustained between 1.9 and 10.8 frames per second (FPS) depending on the Face Feature Map and the capture resolution, within a power envelope of 5 to 7 W. The lightweight feature maps reached approximately 9 to 11 FPS at 480p and 600p, above the 5 FPS threshold we adopt here as the minimum for expression-level monitoring, whereas the full-mesh configurations fell below it. The cloud counterpart was between 2.8 and 9 times faster, at roughly twenty times the power draw.

These results indicate that Emotional AI is computationally feasible on low-cost edge hardware within a power budget compatible with always-on operation in a patient's home, and they therefore support its potential future integration into adaptive, patient-centered remote therapies. The evaluation used public datasets of acted and synthetic expressions from healthy subjects and included neither rehabilitation patients nor any form of clinical validation, so it establishes technical feasibility rather than clinical effectiveness.