Noninvasive detection of diabetes in obstructive sleep apnea based on overnight SpO₂ signal and deep learning

The prevalence of obstructive sleep apnea comorbid with diabetes is high while the awareness of diabetes is low. There is a strong need for new diagnostic biomarkers to detect diabetes at an early stage. Therefore, we aimed to establish an automatic, deep-learning based model that could be applied to assess diabetes risks using overnight SpO₂ signals. The samples were derived from the Sleep Heart Health Study including 5,021 middle-aged and older adults (6.9% diabetes). The deep-learning models were established to identify diabetes solely from SpO₂ or in combination with clinical factors (gender, age, and BMI). Class Activation Map (CAM) was utilized to determine the models' effects. By adding SpO₂ to clinical factors, the prediction performance was significantly improved from 0.646 ± 0.011 to 0.751 ± 0.006 in terms of the area under the receiver operator characteristic (AUC) after 10-fold cross-validation. CAM results showed significant subsequences for the classification decision were the hypoxic events. The findings suggest that the SpO₂ signals could provide substantial information. The deep learning model could be used to evaluate diabetes risks, which is beneficial for long-term health management monitoring.