Uncertainty-Aware Validation of PPG as a Surrogate for ECG: A Clustering-Based Study

Photoplethysmography (PPG) offers a simpler and more cost-effective acquisition pipeline than electrocardiography (ECG), making it attractive for wearable and remote-monitoring applications. A key open question, however, is whether PPG can reliably act as a surrogate for ECG in early cardiovascular assessment, particularly when uncertainty affecting both data and models is taken into account. This work proposes an uncertainty-aware, unsupervised framework to evaluate the validity of PPG as a substitute for ECG. ECG and PPG signals are acquired simultaneously and represented at three levels of abstraction: raw waveforms, filtered signals, and compact cardiovascular parameters extracted from the filtered data. For each representation, eight clustering algorithms are applied independently to ECG, PPG, and their combination, and the resulting partitions are compared. Uncertainty is explicitly addressed along two complementary dimensions. Data uncertainty is modeled through controlled perturbations of the raw signals, emulating realistic sensor-level variability, while model uncertainty is assessed via bootstrap-based stability analysis. Clustering outcomes are evaluated using both global agreement metrics and subject-level stability indicators. The results show that PPG-based clustering remains highly consistent with ECG-based partitions and exhibits strong robustness under both data and model uncertainty, particularly when parameter-based features are employed. Overall, the findings support PPG as a practical and reliable surrogate for ECG in early assessment and wearable monitoring, demonstrating that its inferred subject groupings are both coherent and uncertainty-resilient