Deep learning methods for apnoea detection based on pulse and oximetry data

This paper proposes a deep learning approach for contactless detection of sleep apnoea using pulse and blood oxygen saturation (SPO2) data. Three convolutional neural network architectures are adopted for apnoea classification purposes by fusing different features of the available time series signals. A conventional convolutional neural network (CNN), a CNN with a support vector machine (CNN-SVM), and a CNN combined with a recurrent neural network (CNN-RNN) are compared. The RNN includes Gated Recurrent Units (GRU) and Bidirectional GRU (BiGRU). The CNN is utilised to extract features, whilst the SVM and RNN are used for classification. In addition, we compare two different fusion methods, signal-level and feature-level fusion. The performance is validated and evaluated on a public dataset obtain from St. Vincent University Hospital. The results show that the concatenation of SPO2 and pulse signal at the signal level enhances the classification performance compared to using the individual signal. In addition, the classification sensitivity with signal-level fusion is higher than that with feature-level fusion. Overall, the proposed CNN-RNN with GRU (CNN-GRU) architecture gives the best performance with an accuracy of 85.4%, a sensitivity of 61.5%, a specificity of 91.9%, an F1 score of 0.64, and a κ score of 0.551 with a dropout rate of 0.5 and a 20- second overlap. The results demonstrate that the proposed deep learning approach offers a promising solution for non-invasive detection of sleep apnoea using affordable physiological signals.