Detecting sleep anomalies from SpO2 data using autoencoder-based neural networks

Abstract

SpO2 is a vital indicator for diagnosing OSA. This study presents an automated detection framework integrating an adaptive VMD denoising algorithm with deep learning-based autoencoder models to identify abnormal signals in SpO2 data. We evaluated four autoencoder architectures (VAE, CAE, LAE, and CLAE) on PPG and OPEN datasets. The VMD algorithm achieved an RMSE of 0.862% compared to the SOMNOtouch device, while the LAE model exhibited superior detection performance, achieving the highest precision, F1-score, and accuracy, attributed to its LSTM-based temporal modelling capabilities, albeit at greater computational cost. Study finds there is a non-linear relationship between architectural complexity and performance gains, where increased model sophistication may not necessarily yield proportional improvements in effectiveness. Dataset characteristics significantly influenced model performance, with limited differences observed in the small, homogeneous PPG dataset, while the diverse OPEN dataset highlighted the advantages of temporal modelling. While effective, the framework's current limitations include dataset size constraints and class imbalance issues, suggesting directions for future optimization. These findings advance our understanding of automated sleep apnoea detection and provide insights for developing more robust monitoring systems.