Hypertension Classification with PPG Bio-signal from the VitalDB Subset of PulseDB: Ensemble Learning

Hypertension affects a large proportion of the global population and is a leading contributor to cardiovascular mortality, yet its detection often relies on cumbersome cuff-based measurements. Photoplethysmography (PPG) offers an affordable, non-invasive alternative, but transforming PPG signals into reliable hypertension classification constitutes major obstacles. This paper addresses a methodological oversight in a prior study on the VitalDB subset of the PulseDB dataset, which attributed visually distinct population average PPG morphologies between healthy and hypertensive segments purely to physiological indicators of hypertension. Through per-subject analysis and Cohen’s d effect-size estimation, this paper demonstrates that these visual differences mainly derive from inter-subject signal variability rather than hypertension-specific morphology, while also showing that population-motivated hand-engineered PPG, VPG, and APG features retain meaningful within-subject discriminative signal. The dataset was processed into three versions via high-correlation and low-variance filtering, random forest feature importance selection, and recursive feature elimination with cross-validation (RFECV). Twelve models — Random Forest, Extremely Randomized Trees, Histogram-based Gradient Boosting, and Logistic Regression across all three dataset versions — were combined into a soft-voting ensemble. The best ensemble achieved a ROC-AUC of 0.891 and a hypertensive-class F1 of 0.57 at an optimized threshold, representing a substantial improvement over the baseline study. These results confirm that population-level morphological features do carry indicators of hypertension, although the signal-to-noise ratio of such features remain low.