Lightweight Multi-task Hyperdimensional Computing Framework Driven by Binary Neural Network for Sleep Apnea Detection

Sleep apnea syndrome (SAS) is a prevalent sleep disorder that impacts neurocognitive and cardiovascular health. The need for effective long-term management of SAS in everyday conditions has led to the adoption of wearable devices based on photoplethysmography (PPG) technology as a viable monitoring solution. However, the computational demands of current detection algorithms often surpass the capabilities of low-power, miniaturized wearable devices. To overcome this challenge, this study introduces a novel, lightweight detection algorithm for SAS that utilizes hyperdimensional computing. This algorithm is a multi-task hyperdimensional computing framework that is enhanced by binary neural network and incorporates metric learning. The proposed model demonstrates inference performance with an accuracy of $\mathbf{7 5. 7 2 \%}$, sensitivity of 73.74%, and specificity of $\mathbf{7 6. 4 3 \%}$. Compared to conventional machine learning approaches, this model also offers better performance on embedded platforms, achieving up to $25 \times$ reduction in memory usage, $39 \times$ decrease in latency, and 64× lower energy consumption on an ARM CortexM4 processor.