Manifold Learning Approaches for Characterizing Photoplethysmographic Signals.

Area of Science:

• Biomedical Signal Processing
• Machine Learning in Healthcare
• Cardiovascular Monitoring Technologies

Background:

• Photoplethysmography (PPG) is crucial for cardiovascular monitoring but faces challenges due to signal variability and limited interpretability.
• Existing analysis methods struggle with diverse acquisition settings and complex signal characteristics.

Purpose of the Study:

• To investigate the efficacy of low-dimensional embeddings for enhancing downstream tasks in PPG signal analysis.
• To support anomaly detection, activity classification, and signal authenticity verification across different PPG modalities.

Main Methods:

• Developed a pipeline utilizing dimensionality reduction techniques: Autoencoder (AE), Fully Connected Neural Network (FCNN), and Uniform Manifold Approximation and Projection (UMAP).
• Evaluated methods on four diverse datasets: clinical (BIDMC, MIMIC-PERFORM), wearable (Wrist PPG), and remote PPG (UBFC).
• Assessed performance using clustering indices, classification metrics, and anomaly detection rates under varying noise levels.

Main Results:

• AE embeddings accurately distinguished neonatal and adult signals (MIMIC-PERFORM: F1=0.92, AUC=0.90).
• UMAP excelled in clustering physical activities (Wrist PPG: Davies Bouldin Index=5.40).
• The framework detected synthetic anomalies (BIDMC: AUC=0.77) and manipulated remote PPG signals (UBFC: F1=0.75, AUC=0.73).

Conclusions:

• Low-dimensional representations offer compact, task-relevant encodings for PPG signals, boosting classification and detection performance.
• Embedding-based approaches demonstrate utility and robustness across diverse PPG modalities and noise conditions.
• Interpretability benefits are task-dependent, highlighting the adaptive nature of these methods in biomedical signal analysis.