Cardiac Signal Analysis for the Detection of Major Depressive Disorder in Obstructive Sleep Apnea Patients

<p dir="ltr">Obstructive Sleep Apnea (OSA) and Major Depressive Disorder (MDD) are commonly co-occurring conditions with overlapping symptoms, leading to underdiagnosis, especially in resource-limited settings. Traditional diagnostic approaches such as polysomnography and psychiatric interviews are resource-intensive and unsuitable for routine or large-scale screening. Despite growing awareness of their comorbidity, the dependence on specialized, high-cost diagnostic tools continues to limit timely identification, particularly in rural and underserved populations. Shared symptoms, such as fatigue, poor concentration, and mood disturbances, further complicate the accurate differentiation. </p><p dir="ltr">To address this clinical gap, this thesis presents a non-invasive, wearable-compatible framework for detecting MDD within the OSA population using cardiac signals, ECG, and PPG, recorded during sleep. The research progresses through three stages: single-modality ECG-based screening, beat-to-beat PPG-based morphological analysis, and sleep-stage-specific multimodal classification. </p><p dir="ltr">In the first phase, ECG recordings from individuals diagnosed with OSA, with and without comorbid MDD, were analyzed to extract features related to heart rate variability and signal morphology. A support vector machine classifier, trained on features selected using the ReliefF algorithm, achieved an accuracy of 78.18%, demonstrating the potential of ECG-based detection. However, the reliance on manual noise handling and the absence of healthy control subjects limited the model’s scalability and generalizability. </p><p dir="ltr">To address these limitations and investigate an additional cardiac signal, the second phase introduced a PPG-based classification framework. This stage incorporated healthy controls, employed an automated artifact rejection pipeline, and extracted beat-to-beat entropy and complexity features derived from waveform skewness and kurtosis. The resulting model achieved 84% accuracy and an area under the curve (AUC) of 0.91 for classifying controls versus OSA, and 76% accuracy with an AUC of 0.88 for distinguishing OSA-only from those OSA with MDD. This was the first study to evaluate the utility of PPG-derived features for detecting depressive symptoms in the context of OSA. </p><p dir="ltr">The third and final phase integrated both ECG and PPG modalities within a sleep-stage specific classification framework. Deep sleep was identified as the most informative stage for distinguishing clinical groups. Features extracted from uninterrupted five-minute deep sleep segments enabled the model to achieve 100% accuracy for control versus OSA classification, and 89.29% accuracy with an AUC of 0.91 for differentiating OSA-only from OSA with MDD. These results demonstrate the effectiveness of combining modalities within a stage-aware analytical approach. </p><p dir="ltr">Collectively, this thesis introduces a robust, scalable, and practical framework for sleep based mental health screening using cardiac signals. The findings support the practical feasibility of deploying such models through wearable technologies, enabling early and accessible detection of comorbid psychiatric conditions in sleep-disordered populations</p>