Fourteen PPG-derived features were significantly associated with left ventricular dysfunction (EF ≤40%), achieving a mean AUC of 0.824±0.073 for detection.
Can photoplethysmography-derived waveform features accurately detect left ventricular systolic dysfunction in hospitalised patients?
Specific photoplethysmography-derived waveform features can identify left ventricular systolic dysfunction with good discriminative performance, highlighting their potential as scalable digital screening biomarkers.
Absolute Event Rate: 0% vs 0%
Abstract Background Left ventricular (LV) systolic dysfunction is a critical precursor to heart failure and a major contributor to premature mortality, often identified late due to a lack of scalable screening tools. Photoplethysmography (PPG), a low-cost and widely accessible signal, holds significant promise for remote cardiovascular assessment; however, its application in detecting structural heart disease is limited by the absence of robust, confounder-resistant, and explainable biomarkers. Purpose This study aimed to identify PPG-derived waveform features independently associated with reduced ejection fraction (EF) and to establish their potential as digital biomarkers for LV systolic dysfunction, for early, non-invasive detetion using digital devices. Methods In this observational study, 92 of 116 screened, hospitalised patients undergoing routine echocardiography were enrolled. Standardised 10-minute red and infrared finger PPG recordings were acquired. After automated quality control, 232 morphological and HRV features were derived from PPG using pyPPG toolbox. To control for major clinical confounders, propensity score matching (PSM) was applied to balance age, sex, and atrial fibrillation (AF) status across EF thresholds (≤50%, ≤45%, ≤40%). A paired t-test or Wilcoxon signed-rank test was performed on the matched pairs. Rigorous statistical analysis included correction for multiple comparisons using the Benjamini-Hochberg procedure to control the False Discovery Rate (FDR). Discriminative performance was assessed using 5-fold cross-validated logistic regression on PSM-selected features and, for independent validation, 5x5 nested cross-validated LASSO logistic regression. Results Post-PSM, 14 PPG-derived features demonstrated significant associations with EF ≤40% (incl. Tp2, Tpw25/Tpiₛtd, Ab/Aa). The logistic regression model incorporating these features (and clinical confounders) achieved a mean AUC of 0. 824±0. 073 (accuracy 0. 744±0. 075; F1-score 0. 624±0. 145). A parallel analysis, where age, sex, and AF were excluded as direct predictors, yielded comparable performance (AUC 0. 828±0. 087). Across decreasing EF thresholds (≤50%, ≤45%, ≤40%), the number of significant PPG features increased (6, 9, and 14, respectively). Specifically, features such as Tp2 (increasing from 0. 7 to 1. 0) and Tb/Tpi (increasing from 0. 6 to 0. 9) showed a progressive increase in standardised effect sizes. An independent LASSO model achieved an AUC of 0. 79±0. 061 (acc. 0. 70±0. 063). LASSO consistently selected three features with non-zero coefficients overlapping with PSM: Tp2, Tpw25/Tpiₛtd, and Ab/Aa. Conclusions Several consistent PPG-derived digital biomarkers demonstrated association with LV systolic dysfunction, supporting their potential for non-invasive cardiac screening independently from the other confounders. Despite the limited sample size, these findings provide a strong rationale for external validation in larger cohorts to enable clinical translation.
Basza et al. (Thu,) reported a other. Fourteen PPG-derived features were significantly associated with left ventricular dysfunction (EF ≤40%), achieving a mean AUC of 0.824±0.073 for detection.