Challenges in calculating the AHI to diagnose sleep apnoea using deep learning and portable monitors

Abstract Background Automatic detection of the apnoea–hypopnoea index (AHI) using a portable monitor (PM) with artificial intelligence (AI) represents a significant challenge. The objective of this study was to examine factors that affect the performance of an AI algorithm that had been previously trained in calculating the AHI with polysomnography (PSG) data using signals collected by a PM. Methods A deep learning model, specifically a version of U‑Net, had been previously trained on a PSG dataset comprising 480 patients for training, 96 for evaluation and 65 for testing. The dataset comprised three variables: oxygen saturation (SpO 2 ), heart rate (HR) and abdominal respiratory effort (ARE). The deep learning model was employed to analyse data collected from 7 patients by the PM, with the objective of classifying obstructive sleep apnoea events and subsequently calculating the AHI. Subsequently, the results obtained by the PM and PSG were compared. Results The present study sought to verify the automatic recognition system for calculation of the AHI. The data for this study were collected from 7 patients using the PM, and the mean absolute error was found to be 19.95% when compared to the results obtained in PSG. The absolute error was found to be greater for patients with a more severe form of sleep apnoea, which resulted in an increase in absolute inaccuracy. Conclusion The findings of this study indicate that training AI models for use with new data obtained from different automatic detection systems is a challenging process that requires careful consideration of various factors, including the quality and quantity of data as well as their preprocessing prior to feeding into the AI models.