Interpretable Two-Stage PPG Framework for Cerebrovascular Risk Assesment via Intermediate Blood Pressure Modeling

Le résumé fourni par la source

Cerebrovascular disorders are a major cause of death and long-term disability worldwide. Photoplethysmography (PPG), an optical modality that is already embedded in many consumer wearables, offers a low-cost, noninvasive window into vascular physiology. Rather than attempting end-to-end classification from large sets of handcrafted or learned PPG features, we hypothesize that chronic elevation of arterial blood pressure is the principal physiological mediator that links cerebrovascular pathology to observable changes in peripheral PPG waveforms. To test this hypothesis, we propose a transparent two-stage pipeline. In Stage I, short PPG segments are transformed into spectral-cepstral descriptors and used to estimate systolic and diastolic blood pressure via an ensemble of bagged regression trees. In Stage II, the mean estimated systolic and diastolic pressures for each subject serve as the sole inputs to a conventional classifier that predicts cerebrovascular risk (three classes: healthy, cerebral hypoperfusion, confirmed disease). On the public PPG-BP dataset, this minimal two-feature representation yielded 94.3% classification accuracy and AUC = 0.96 substantially outperforming typical single-stage models. Our findings emphasize the dominant role of blood pressure as a mediator of PPG-detectable cerebrovascular alterations and raise concerns about potential label leakage in common benchmarks. We discuss implications for model interpretability and for the design of robust wearable-based screening tools.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.