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

Area of Science:

• Biomedical Engineering
• Physiological Signal Processing
• Machine Learning

Background:

• Accurate cuffless blood pressure (BP) estimation is challenged by individual physiological differences.
• Training personalized models typically requires extensive data, which is often impractical.
• Developing methods for accurate, individualized BP estimation with limited data is crucial.

Purpose of the Study:

• To develop a personalized cuffless blood pressure estimation model using only a few labeled samples (few-shot learning).
• To enhance model personalization by integrating self-supervised and few-shot learning techniques.

Main Methods:

• A two-stage training strategy was employed.
• Stage 1: Self-supervised learning to extract shared physiological signal features across subjects.
• Stage 2: Few-shot learning to adapt the pre-trained model for individual subject personalization.

Main Results:

• The proposed method achieved low Mean Absolute Errors (MAE) for systolic BP (SBP) and diastolic BP (DBP) estimation: 6.57±6.22 mmHg and 3.66±3.99 mmHg (using PPG and ECG) with 5-shot learning.
• Performance was also strong using only photoplethysmogram (PPG) signals (6.77±6.43 mmHg SBP, 3.80±3.92 mmHg DBP).
• The approach outperformed existing non-personalized and transfer learning methods and demonstrated strong generalization across multiple datasets.

Conclusions:

• The developed two-stage method offers a novel approach for few-shot personalization of cuffless BP estimation.
• This technique facilitates accurate and individualized blood pressure monitoring, overcoming data limitations.
• The findings support the potential for more accessible and personalized BP management tools.