PPGNet: Deep Multiple Instance Learning for Automated Hypertension Detection in OSA Patients Using PPG Signals

Abstract

Systemic hypertension (HTN) is a prevalent cardiovascular comorbidity in patients with obstructive sleep apnea (OSA), yet no scalable framework exists for HTN detection using routinely collected polysomnography (PSG) signals. We present PPGNet, a deep learning framework for diagnostic detection of established HTN in OSA patients using photoplethysmography (PPG) signals extracted from standard overnight PSG recordings. The framework employs a Multiple Instance Learning (MIL) approach to model long-term hemodynamic and autonomic patterns across full-night recordings, aggregating 30-second instances into 30-minute bags enriched with sleep stage and apnea context. Evaluated under Leave- One-Patient-Out (LOPO) cross-validation on a clinically annotated cohort of 33 OSA patients, PPGNet achieved AUROC of 86.72%, AUPRC of 88.01%, Accuracy of 79.03%, Sensitivity of 83.33%, and MCC of 0.5842. Attention-based interpretability analysis confirms that the model captures clinically meaningful vascular signatures, including arterial stiffness markers and differential vasomotor responses to apnea events. This work establishes a proof-of-concept for cardiovascular risk stratification as a secondary outcome of routine sleep studies, enabling earlier HTN identification in this highrisk population without additional instrumentation.