Enhancing few-shot personalized cuffless blood pressure estimation with self-supervised learning

Objective.Individual differences across subjects reduce the accuracy of physiological signal-based cuffless blood pressure (BP) estimation. However, training a personalized model with a large amount of data is impractical. This study aims to learn a personalized model using only a few labeled samples (e.g. 5 datapoints).Approach.This study introduced a two-stage training method to enhance the few-shot personalized model with self-supervised learning. In the first training stage, self-supervised learning is used to learn shared features from physiological signals across subjects. In the second stage, few-shot learning is used to adapt the model to each subject based on the pre-trained encoder from the first stage.Main result.Experiments were conducted on the PulseDB dataset. Under the 5-shot setting, the proposed method achieved mean absolute errors of 6.57 ± 6.22 mmHg and 3.66 ± 3.99 mmHg for systolic BP (SBP) and diastolic BP (DBP) estimation, respectively, when using photoplethysmogram (PPG) and electrocardiogram. Using only PPG signals, the method achieved 6.77 ± 6.43 mmHg and 3.80 ± 3.92 mmHg for SBP and DBP estimation, respectively. The proposed approach exceeded previous non-personalized and transfer learning methods. Its generalization capability was validated on two additional smaller datasets, demonstrating the generalization ability of the proposed method.Significant.Overall, the proposed method provides a new approach for few-shot personalization of cuffless BP estimation models, which is helpful for accurate and individualized BP estimation.