Implementation of Spiking Neural Networks for Sleep Apnea Classification Using Pulse Oximetry

Introduction: Obstructive sleep apnea is a prevalent respiratory disorder, which is conventionally diagnosed in a clinical setting involving overnight polysomnography. Transition to enhanced autonomy is facilitated through applications involving pulse oximetry, which can be amalgamated with artificial intelligence for augmented diagnostics. Spiking neural networks constitute an advanced form of artificial intelligence representing brain-like neuromorphic processing that facilitates low power requirements and edge computing opportunities. The objective of this research endeavor is to apply spiking neural networks to classify simulated apnea events based on signal data from pulse oximeters.

Methods: The dataset was comprised of 20 subjects from the PhysioNet Respiratory Oximetry Apnoea database: “Respiratory and Pulse Oximetry Waveforms from Healthy Adults During Simulated Apnoea Events”. Raw time-series signals were transformed into sparse spatiotemporal spike trains using derivative-based delta modulation. The implemented spiking neural network integrated a 1D temporal convolutional layer paired with a leaky integrate-and-fire neuron model, optimized via surrogate gradient backpropagation.

Results: The spiking neural networks successfully distinguished between normal breathing periods and simulated apnea events based on pulse oximeter signal data. From a global perspective in consideration of the 20 subjects, a mean classification accuracy of 98.28% was achieved.

Conclusions: These findings indicate that spiking neural networks can accurately detect and distinguish respiratory apnea anomalies using sensor signal data amenable for a wearable context. Future studies should also consider other neuromorphic algorithms, such as the Izhikevich neuron model, and further optimization. This achievement indicates a pathway to edge computing with the inherent low power requirement for spiking neural networks with opportunities for real-time diagnostic telemedicine systems capable of advanced, continuous, non-invasive home sleep monitoring.