BioEdgeNet: A compact deep residual network for stress recognition on edge devices

• Highly accurate and compact deep network for human stress recognition. • Integration of multiple design optimizations for edge device deployment. • Post-training quantization evaluated on Raspberry Pi Zero and Pi 4 devices. • Over 97 % accuracy achieved with a 22 KB model size across two datasets. • Foundational model applicable to diverse biomedical signal types. The effectiveness of deep learning algorithms in discerning stress levels from biomedical signals has attracted considerable attention. However, the computational demands of these algorithms often make them unsuitable for deployment on resource-limited devices. This paper introduces a highly accurate and compact deep residual network specifically designed for human stress recognition, optimized to operate with significantly reduced computational power, thereby facilitating deployment on edge devices. Its general-purpose architecture is intended to serve as a robust baseline model adaptable to various biomedical signal types, including photoplethysmography and accelerometer data. The network design integrates several optimization techniques, such as reduced kernel sizes, the substitution of pooling layers with strided convolutions, and the incorporation of inverted residual bottlenecks. Post-training quantization further enhances the model’s efficiency, as validated on Raspberry Pi Zero and Pi 4 devices. Performance evaluation with two publicly available datasets demonstrated over 97 % accuracy while maintaining a compact size of 22 KB, effectively balancing accuracy, the number of parameters, and inference time. These results surpass many existing methods in accuracy while requiring substantially fewer computational resources, highlighting potential for integration into edge devices.