Optimized Neural Architecture for Wearable Health Devices using Edge-Level TinyML

Wearable health devices enable real-time monitoring of physiological signals such as heart rate, SpO₂, and ECG. However, their limited computational resources and battery life necessitate efficient machine learning models suitable for edge deployment. This paper presents an optimized neural network architecture for edge-level TinyML inference on wearable platforms. The model utilizes quantization-aware training, pruning, and lightweight convolutional layers to reduce memory and energy consumption without compromising accuracy. Evaluated on real-world biomedical datasets and deployed on commercial MCUs like ARM Cortex-M, the proposed model achieves 91.8% classification accuracy across various health signal categories, uses under 45 KB of memory, and maintains inference latency below 80 milliseconds. These results highlight its suitability for real-time, low-power applications, enabling privacy-preserving, always-on health monitoring. The proposed architecture addresses critical hardware constraints and contributes to the development of scalable, intelligent wearable health systems in remote or resource-limited environments.