A Multi-modal Teacher-student Framework for Improved Blood Pressure Estimation.

Area of Science:

• Biomedical Engineering
• Medical Informatics
• Signal Processing

Background:

• Accurate blood pressure (BP) monitoring is crucial for managing hypertension.
• Existing BP estimation methods often require multiple physiological signals or invasive procedures.

Purpose of the Study:

• To develop a novel framework for blood pressure estimation using a uni-modal approach (photoplethysmogram - PPG).
• To transfer knowledge from a complex multi-modal BP estimation model to a simpler uni-modal model using teacher-student training.

Main Methods:

• A multi-modal model incorporating gated recurrent units, attention mechanisms, and a regression layer was developed using PPG, electrocardiogram (ECG), age, height, and weight.
• A uni-modal model using only PPG signals was trained to mimic the embeddings generated by the multi-modal model.
• Teacher-student knowledge distillation was employed to train the uni-modal model.

Main Results:

• The proposed uni-modal model achieved low absolute prediction errors for systolic blood pressure (5.24±6.41 mmHg) and diastolic blood pressure (3.49±3.85 mmHg) on the MIMIC-III dataset.
• The model performance satisfied the Association for the Advancement of Medical Instrumentation (AAMI) standard for BP monitoring devices.

Conclusions:

• Knowledge distillation effectively enables accurate BP estimation from a single PPG signal.
• The developed uni-modal model offers a promising non-invasive approach for regular blood pressure monitoring in hypertensive patients.