Efficient and Secure µ-Training and µ-Fine-Tuning for TinyML Optimization and Personalization at the Edge

This study introduces a novel approach for training and fine-tuning machine learning models for bio-signal data analysis on edge medical devices. The technique can be used in all physiological signals, and in this paper, we used electrocardiogram (ECG) signals as a case example to demonstrate its capability. The proposed methodology combines full training and a novel technique termed µ-Training, in which the encoder and decoder layers of the tiny model are frozen. In contrast, the middle layer weights remain trainable. We investigate the effectiveness of this approach across different stages, including full training, µ-Training, and µ-Fine-Tuning. The model’s performance is evaluated using both in-sample data from the Telehealth Network of Minas Gerais (TNMG) dataset and an out-of-sample test on the China Physiological Signal Challenge 2018 (CPSC) dataset, with the results demonstrating that the combined training approach performs similarly to or better than traditional full training and fine-tuning while providing significant advantages in computational efficiency. Furthermore, the model is deployed on an edge device for µ-Fine-Tuning, showcasing its effectiveness even under resource-constrained conditions. For demonstration and deployment purposes, we used Radxa Zero hardware, while a range of other edge devices can be used. The results demonstrate that the proposed method outperforms traditional approaches, improving computational efficiency and resource utilization, making it a promising solution for real-time bio-signal processing on edge devices.