A benchmark for machine-learning based non-invasive blood pressure estimation using photoplethysmogram

Area of Science:

• Cardiovascular Health
• Biomedical Engineering
• Machine Learning Applications

Background:

• Blood Pressure (BP) is a critical cardiovascular health indicator.
• Cuff-based BP monitoring is inconvenient; portable, continuous monitoring using photoplethysmography (PPG) is desirable.
• Machine Learning (ML) models show promise for estimating BP from PPG, but lack standardized evaluation.

Purpose of the Study:

• To establish a standardized benchmark for comparing ML-based BP estimation models.
• To facilitate model comparison by providing open datasets, shared preprocessing, and robust validation strategies.
• To improve the interpretability of ML model performance across diverse BP datasets.

Main Methods:

• Developed a benchmark incorporating four open datasets with consistent preprocessing.
• Implemented a validation strategy to prevent information shift and leakage.
• Adapted the Mean Absolute Scaled Error (MASE) for enhanced evaluation interpretability.
• Compared 11 ML-based approaches across three categories using the benchmark.

Main Results:

• The benchmark provides a reproducible framework for ML-based BP estimation.
• Standardized evaluation revealed performance differences among 11 ML models.
• The adapted MASE metric improved cross-dataset performance interpretability.

Conclusions:

• The proposed benchmark enables objective comparison of ML models for PPG-based BP estimation.
• This work facilitates advancements in non-invasive, continuous blood pressure monitoring.
• Open datasets and code promote further research and development in the field.