A Systematic Review of TinyML Applications in Rehabilitation and Human Activity Recognition (HAR): From Gait Monitoring to Intelligent Prosthesis Control

Intelligent rehabilitation systems and Human Activity Recognition (HAR), as fundamental pillars of telemedicine and personalized healthcare, are heavily dependent on data from wearable sensors. However, transmitting massive volumes of raw data (such as accelerometer and gyroscope readings) to the cloud introduces serious challenges, including latency, high energy consumption, and privacy concerns. With the advent of TinyML (machine learning on resource-constrained devices), these challenges have been partially addressed or mitigated. This systematic review examines research conducted between ۲۰۱۷ and ۲۰۲۶ on the application of TinyML in rehabilitation and HAR. In this research, the reviewed articles are categorized by application domain into gait disorders, upper limb rehabilitation, daily activity recognition, and prosthetic control. Regarding model architectures and optimization, three techniques-quantization, pruning, and knowledge distillation are employed. Furthermore, concerning target hardware platforms, three categories are utilized: STM۳۲, ESP۳۲, and Arduino. The findings of this research indicate that TinyML models, with accuracy comparable to cloud-based networks (often above ۹۵%) and power consumption of less than ۵۰ mW, hold great potential for home-based rehabilitation and real-time patient monitoring. Finally, key challenges such as model personalization, domain shift, and the lack of standard datasets are examined, and future research directions are outlined.

کلیدواژه ها:

Tiny ML ، Rehabilitation ، Human Activity Recognition