On the Limitations of Hand-Engineered PPG Features for Hypertension Classification: Evidence from the VitalDB Subset of PulseDB

Photoplethysmography (PPG) has been widely used for cuffless blood pressure estimation and hypertension detection, yet no consensus exists regarding a standard set of discriminant PPG features. This study evaluates whether (1) an extensive feature space or (2) a clinically validated subset of commonly used PPG features can reliably distinguish hypertensive and healthy segments using only PPG signals. Experiments were conducted on the VitalDB-derived subset of the PulseDB database, which comprises 465,480 waveform segments. After signal preprocessing, 183 hand-engineered time-domain, frequency-domain, and derivative-based features were extracted in the first trial and reduced via correlation and variance filtering. The second trial employed 28 clinically supported features associated with blood pressure. After an evaluation of a logistic regression and a random forest classifier, both approaches yielded limited balanced accuracy (≈0.64–0.71) and demonstrated substantial overlap between classes despite extensive feature engineering and model tuning. Visual inspection of waveform morphology suggested structural differences between the classes which were not captured. These findings indicate that discriminative PPG characteristics for hypertension may be dataset-specific and that universally optimal hand-engineered feature sets are difficult to establish. The results highlight the limitations of fixed feature pipelines and suggest that dataset-adaptive strategies or end-to-end representation learning may be required for robust PPG-based hypertension classification.