MAF-PPGNet: A Morphology-Aware Attention Fusion Network for Cardiovascular Risk Stratification from Fingertip Photoplethysmography

Abstract

Hypertension drives much of global cardiovascular mortality, yet most cases are found late because cuff measurement is episodic and unsuited to continuous screening. Fingertip photoplethysmography (PPG) is a low-cost optical alternative that fits naturally into wearable and smartphone workflows. This paper presents MAF-PPGNet, a morphology-aware attention fusion network that stratifies cardiovascular risk from a single 2.1 s fingertip PPG segment. The model joins two views of the pulse: a convolutional branch with squeeze-and-excitation recalibration and temporal attention pooling that learns waveform shape, and a parallel branch encoding 24 interpretable morphological, derivative, and spectral descriptors. A gated fusion layer merges the two embeddings before classification, and the whole network needs only 37,278 trainable parameters. Evaluation uses the public PPG-BP database of 657 segments from 219 subjects under strict subject-independent five-fold cross-validation, so no segment of a test subject is seen during training. MAF-PPGNet delivers the best accuracy and area under the curve on the hardest contrast, normotension versus prehypertension, and matches the strongest classical baselines on normotension versus hypertension at 72.84% accuracy. Two further findings are reported plainly: a class-weighted logistic regression over the same descriptors remains the strongest hypertension screener, and relaxing the split to segment level inflates the compact model by at most seven accuracy points, so loose protocols alone cannot explain far higher published figures.