Evaluating cross-dataset transfer learning for photoplethysmography-based blood pressure estimation.

Non-invasive, continuous blood pressure (BP) monitoring using photoplethysmography (PPG) remains challenging due to limited subject-specific data and domain shifts between recording environments. This study evaluates the performance of cross-dataset transfer learning for BP estimation. We developed a compact version of the BP-CRNN model with reduced trainable parameters to allow effective training on small datasets. Experiments were conducted using the MIMIC-III and VitalDB databases, which include PPG signals and continuous invasive BP measurements. Pre-trained models were first established on one dataset and then fine-tuned using subject-specific data from the other dataset. The cross-dataset transfer learning approach, evaluated on 200 patients, achieved mean absolute errors of 3.36 mmHg for systolic and 1.81 mmHg for diastolic BP, reflecting an approximately 13% improvement over models trained without transfer learning. These results satisfy the AAMI and BHS standards. The performance gap between cross-dataset and intra-dataset transfer learning was within 1% and not statistically significant. These findings demonstrate that cross-dataset transfer learning can effectively leverage knowledge from large-scale medical datasets to improve BP estimation accuracy, particularly in domains with limited data availability and varying measurement conditions.