Background: Cardiovascular diseases (CVDs) remain a leading cause of global morbidity and mortality. Early identification of the development of conditions such as hypertension (HTN), type 2 diabetes (T2DM), or the detection of new major adverse cardiovascular events (MACE) is critical for timely intervention and can lead to vastly different health outcomes for patients. Photoplethysmography (PPG) is a non-invasive, scalable method for with promising utility for cardiovascular risk assessment, but most studies rely on short-term or single-timepoint recordings, limiting their ability to capture dynamic physiological changes relevant to disease risk.
Aim: To determine whether longitudinal changes in PPG-derived features improve prediction of HTN, T2DM, and MACE compared to single-timepoint models.
Methods: We analysed PPG waveforms from 27,403 UK Biobank participants with multiple visits, separated by an average of 2400 days. Over 100 morphological features were extracted per subject. XGBoost models were trained using three feature sets: baseline visit features (V1), follow-up visit features (V2), and baseline features plus temporal changes (V1+Δ), including time between visits and feature deltas. Models were evaluated using 80:20 train-test splits with oversampling. Performance was assessed using AUROC, AUCPR, and Brier scores.
Results: Models incorporating temporal changes (V1+Δ) consistently outperformed single-timepoint models. For HTN (prevalence 29.2%), AUROCs were 0.746 (V1), 0.780 (V2), and 0.879 (V1+Δ); AUCPRs were 0.577, 0.639, and 0.813; Brier scores were 0.200, 0.182, and 0.121. For T2DM (2.7%), AUROCs were 0.851, 0.833, and 0.888; AUCPRs 0.250, 0.201, and 0.468; Brier scores 0.141, 0.150, and 0.021. For MACE (2.5%), AUROCs were 0.702, 0.731, and 0.814; AUCPRs 0.055, 0.119, and 0.333; Brier scores 0.205, 0.200, and 0.021.
Conclusions: Longitudinal changes in PPG features significantly enhance prediction of cardiovascular conditions. These findings highlight the potential of wearable PPG devices for continuous risk monitoring and more accurate early detection through temporal feature tracking.