Intensive care photoplethysmogram datasets and machine-learning for blood pressure estimation: Generalization not...

Area of Science:

• Biomedical Engineering
• Physiological Signal Processing
• Medical Informatics

Background:

• The MIMIC dataset, from intensive care units, is widely used for developing Photoplethysmography (PPG) based blood pressure (BP) estimation algorithms.
• Concerns exist regarding anomalous BP-PPG relationships in MIMIC data due to patient conditions and drug effects.

Purpose of the Study:

• To investigate the impact of dataset characteristics on PPG-based BP estimation.
• To compare the relationship between PPG signal features and BP in critical care versus controlled settings.

Main Methods:

• Compared 12,000 MIMIC records with 219 PPG-BP records, analyzing BP distributions and 31 PPG pulse morphological features.
• Assessed correlations between PPG features and BP, and feature inter-correlations.
• Trained and cross-validated regression models on both datasets.

Main Results:

• Significant differences found in diastolic BP and 20/31 features between datasets after heart rate adjustment.
• Features highly correlated with systolic BP in PPG-BP showed weaker correlations in MIMIC.
• PPG-BP yielded twice the baseline predictive power compared to MIMIC; cross-dataset validation showed near-complete loss of predictive power.

Conclusions:

• MIMIC and PPG-BP datasets exhibit distinct BP-PPG relationships.
• BP estimation models trained on MIMIC data may have limited predictive power for the general population.
• Dataset selection is crucial for developing robust and generalizable PPG-based BP estimation algorithms.