Manifold Learning Approaches for Characterizing Photoplethysmographic Signals

Photoplethysmography (PPG) is widely used for cardiovascular monitoring, but its analysis is challenged by signal variability, inconsistent acquisition settings, and limited interpretability. This study investigates the use of low-dimensional embeddings to support downstream tasks, including anomaly detection, activity classification, and signal authenticity verification across diverse PPG modalities.

We developed a pipeline leveraging dimensionality reduction techniques, Autoencoder (AE), Fully Connected Neural Network (FCNN), and Uniform Manifold Approximation and Projection (UMAP) to extract compact signal representations. These methods were evaluated across four datasets representing clinical (BIDMC, MIMIC-PERFORM), wearable (Wrist PPG), and remote PPG (UBFC) recordings. Performance was assessed through clustering indices, classification metrics, and anomaly detection rates under varying noise levels.

Quantitative evaluation demonstrated that AE-based embeddings enabled accurate discrimination between neonatal and adult signals in the MIMIC-PERFORM dataset (F1 = 0.92, AUC = 0.90), while UMAP outperformed AE and FCNN in clustering physical activities from Wrist PPG data (Davies-Bouldin Index = 5.40). In the BIDMC dataset, the framework detected synthetic anomalies with an AUC of 0.77 at 2 dB SNR, with detection rates declining consistently with reduced noise. On the UBFC dataset, UMAP embeddings supported the detection of manipulated rPPG signals with an F1 score of 0.75 and an AUC of 0.73.

Low-dimensional representations provide a compact and task-relevant encoding of PPG signals that enhances classification and detection performance in multiple scenarios. While interpretability gains remain task-dependent, these findings support the utility of embedding-based approaches in biomedical signal analysis and their robustness across modalities and noise conditions.