Uncertainty Reliability Under Domain Shift: An Investigation for Data-Driven Blood Pressure Estimation in Photoplethysmography
Researchers investigated the reliability of uncertainty quantification in deep learning models for blood pressure estimation from photoplethysmography (PPG) signals. The study found that deep ensembles (DE) offer greater robustness under domain shift compared to Monte Carlo dropout (MCD). Recalibrated Gaussian negative log-likelihood (GNLL) methods, particularly with DE and conformal prediction or temperature scaling, provided the best uncertainty calibration for both systolic and diastolic blood pressure. AI
IMPACT Enhances trustworthiness of AI for critical healthcare applications by improving uncertainty quantification under real-world conditions.