Key result
Novel PPG feature framework classifies blood pressure categories with ~0.78 AUC using XGBoost.
Why the study?
Photoplethysmography provides waveform features for blood pressure category classification, but classification performance may vary across different recording conditions.
Does a photoplethysmography feature framework incorporating Step-Morphological Decomposition accurately classify blood pressure categories across multiple datasets?
Observational
Does a photoplethysmography feature framework incorporating Step-Morphological Decomposition accurately classify blood pressure categories across multiple datasets?
Interpretable photoplethysmography features combined with machine learning show moderate discriminative performance for blood pressure category classification within a dataset, though performance degrades across external cohorts and removing Step-Morphological Decomposition features minimally impacts accuracy.
Answers are grounded in this paper and cite their sources.
SMD removal minimally affects PPG BP classification AUC; leaves open whether decomposition features enhance wearable monitoring accuracy.
Photoplethysmography (PPG) provides waveform features for blood pressure (BP) category classification, but performance may vary across recording conditions. We evaluated a framework incorporating Step-Morphological Decomposition (SMD) on Aurora-BP, PPG-BP, PulseDB-Vital and the Synchronized Signals from Wearable Cardiovascular Monitoring Sensors dataset (SWCS). Signals were processed at 125 Hz. Conventional and neural classifiers were compared using subject-grouped evaluation, feature-family ablation, Shapley additive explanations (SHAP) and cross-cohort adaptation. On Aurora-BP, XGBoost achieved a mean three-class macro one-versus-rest area under the receiver operating characteristic curve (AUC) of 0.778 ± 0.011 across five folds. Subject-aggregated binary prediction achieved AUC 0.841. Source-only AUCs were 0.662 ± 0.096 on PPG-BP, 0.556 ± 0.049 on PulseDB-Vital and 0.696 ± 0.083 on SWCS. A restricted PPG-only pipeline achieved AUC 0.704. Removing the SMD feature family changed AUC from 0.778 to 0.776. Period morphology had the largest summed SHAP attribution, whereas demographic predictors had the largest average attribution per feature. Adaptation effects varied across cohorts and available target data. These retrospective results characterize the performance and limitations of the evaluated feature framework. Prospective clinical screening performance was not evaluated.
No takes yet. Share an insight, caveat, or question.
Msokar et al. (2026) conducted an observational in Blood pressure category classification. Photoplethysmography (PPG) feature framework incorporating Step-Morphological Decomposition (SMD) with XGBoost vs. Restricted PPG-only pipeline, SMD feature-family ablation, and neural/conventional classifiers across external cohorts was evaluated on Three-class macro one-versus-rest area under the receiver operating characteristic curve (AUC) for blood pressure category classification on Aurora-BP (AUC 0.778 ± 0.011). A PPG feature framework with XGBoost achieved a three-class macro AUC of 0.778 ± 0.011 on Aurora-BP for BP classification, vs source-only AUCs of 0.556–0.696 on external cohorts.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: