A Machine Learning Framework for Sleep Apnoea Detection Based on Imbalanced Pulse and Oximetry Data.

Abstract

Sleep apnoea, a disorder impacting both children and adults, typically requires costly and time-intensive diagnostics. This paper introduces a novel framework that uses the wavelet transform to extract features from sleep signals and the RUSBoost algorithm to address the challenge of imbalanced data in detecting sleep apnoea, which enables home self-monitoring. Patient data features short apnoea epochs and long periods of normal breathing, creating imbalances that challenge classification algorithms. The framework was tested on three public datasets with varying imbalance ratios. Significantly, the Childhood Adenotonsillectomy Trial (CHAT) dataset with an ‘apnoea’ to ‘normal’ period ratio of 1:15, effectively reflects actual sleep apnoea signals from children. The proposed framework with the CHAT dataset achieved a maximum accuracy of 91.54%, sensitivity of 72.06%, specificity of 92.39%, and an AUC of 0.923, surpassing state-of-the-art home screening models. For the classification task, this study compared several machine learning techniques, including support vector machine (SVM), K-nearest neighbour (KNN), and Dirichlet process Gaussian mixture model (DPGMM) algorithms. It is found that the RUSBoost algorithm provides the most accurate results when the ratio of the ‘apnoea’ to the ‘normal’ period reaches an imbalance of 1:3 or greater.