Screening Sleep Apnea Using Random Convolution Kernels of Acoustic Speech Representations

Obstructive sleep apnea (OSA), characterized by complete or partial airway obstruction during sleep, affects approximately 38% of the adult population. The clinical diagnosis of OSA is typically conducted using polysomnography (PSG), which has been reported to be cumbersome, requires technical expertise, and involves long waitlists. Consequently, alternative methods, such as questionnaires, have gained attention in recent years. However, while these methods are highly sensitive, they suffer from low specificity. Speech-based screening has emerged as an accessible alternative and has been widely explored in the literature. Despite this, existing technologies primarily rely on acoustic features and classical machine learning models, with the potential of deep neural networks largely unexplored. In this study, we employed random convolution kernels to simulate the effects of convolutional neural networks on speech acoustic features to predict the severity of sleep apnea based on two apnea-hypopnea index (AHI) thresholds: 10 and 15 events/hour. Our results, based on a sample of 35 individuals, achieved an F1-score of 0.83 for the 10 events/hour threshold and 0.71 for the 15 events/hour threshold, surpassing the performance reported in the literature with similar sample sizes.