Sleep Apnea Events Recognition Based on Polysomnographic Recordings: A Large-Scale Multi-Channel Machine Learning approach

Abstract

Goal: The gold standard for detecting the presence of apneic events is a time and effort-consuming manual evaluation of type I polysomnographic recordings by experts, often not error-free. Such acquisition protocol requires dedicated facilities resulting in high costs and long waiting lists. The usage of artificial intelligence models assists the clinician's evaluation overcoming the aforementioned limitations and increasing healthcare quality. Methods: The present work proposes a machine learning-based approach for automatically recognizing apneic events in subjects affected by sleep apnea-hypopnea syndrome. It embraces a vast and diverse pool of subjects, the Wisconsin Sleep Cohort (WSC) database. Results: An overall accuracy of 87.2 $\pm$ 1.8% is reached for the event detection task, significantly higher than other works in literature performed over the same dataset. The distinction between different types of apnea was also studied, obtaining an overall accuracy of 62.9 $\pm$ 4.1%. Conclusions: The proposed approach for sleep apnea events recognition, validated over a wide pool of subjects, enlarges the landscape of possibilities for sleep apnea events recognition, identifying a subset of signals that improves State-of-the-art performance and guarantees simple interpretation.