How photoplethysmography can be used to detect major depressive disorder among patients with obstructive sleep apnea during sleep

Abstract

Major Depressive Disorder (MDD) frequently coexists with Obstructive Sleep Apnea (OSA), yet it remains underdiagnosed in OSA populations due to overlapping symptoms and limited access to psychiatric sleep evaluations. Earlier studies have explored photoplethysmography (PPG) for screening either OSA or MDD individually but have not investigated the use of PPG to analyze comorbid MDD in patients with OSA. Additionally, beat-level entropy and complexity features, which quantify subtle nonlinear variations in pulse morphology and may reflect autonomic nervous system dysregulation associated with this comorbidity, have been largely overlooked. This study investigates whether beat-to-beat PPG-derived features can distinguish among healthy controls (CO), individuals with OSA only, referred to as OSA-, and those with OSA and comorbid MDD, referred to as OSA+. PPG recordings from 60 participants (CO: 25, OSA-: 20, OSA+: 15) were preprocessed to extract artifact-free segments. For each segment, skewness (Sk) and kurtosis (Ku) were computed on a beat-to-beat basis, followed by the extraction of Approximate Entropy (ApEn), Hjorth Activity (HA), Hjorth Mobility (HM), and Hjorth Complexity (HC) parameters to quantify signal variability. Feature selection was conducted within a 5-fold nested cross-validation framework using Spearman correlation, and only features that satisfied both a correlation threshold and statistical significance (p < 0.05) were retained for SVM classification. For the CO vs. OSA task, ApEn(Sk) and ApEn(Ku) emerged as the most discriminative features, achieving an accuracy of 84 % and an AUC of 0.91. For the OSA− vs. OSA + task, three features—ApEn(Sk), ApEn(Ku), and HM(Sk)—were selected, yielding an accuracy of 76 % and an AUC of 0.88. This study demonstrates that beat-to-beat variability in PPG morphology can effectively identify MDD within the OSA population. Unlike prior work, which did not investigate comorbid classification or entropy-based features, our approach addresses this gap and supports the feasibility of sleep-based PPG for mental health screening.