A Multimodal Learning Framework for Detecting Systemic Hypertension in Sleep Apnea Patients Using ECG and PPG Signals.

Abstract

Systemic hypertension (HTN) is a major cardiovascular comorbidity in patients with obstructive sleep apnea (OSA), yet these conditions are often diagnosed and managed independently, limiting integrated cardiovascular risk assessment. There is a significant gap in scalable automated frameworks for the screening of hypertension using routinely collected polysomnography (PSG) signals. To address this, we introduce FusionNet, a novel deep learning architecture to detect established HTN in patients with OSA. The proposed framework leverages a Multiple Instance Learning (MIL) approach to capture long-term temporal dynamics from overnight PSG recordings. It integrates electrocardiogram (ECG) and photoplethysmogram (PPG) signals through a gated fusion mechanism for robust multimodal analysis, further enriched by frequency-domain heart and pulse rate variability (HRV/PRV) and clinical context features, including sleep stages and apnea events. The model was evaluated using Leave-One-Patient-Out (LOPO) cross-validation on a clinically annotated dataset of 33 patients to ensure subject-independent generalizability. FusionNet achieved an AUROC of 87.34%, accuracy of 79.03%, specificity of 81.25%, and an MCC of 0.5801, indicating a balanced and clinically conservative detection profile through adaptive integration of complementary cardiac and vascular information. This work establishes a proof-of-concept for a dual diagnostic paradigm, demonstrating the feasibility of deriving hypertension-related cardiovascular signatures from standard sleep studies. These findings present an initial step towards automated cardiovascular risk assessment and require validation in larger independent cohorts prior to clinical deployment.