Data-knowledge co-driven feature based prediction model via photoplethysmography for evaluating blood pressure.

Area of Science:

• Biomedical Engineering
• Physiological Signal Processing
• Machine Learning for Healthcare

Background:

• Photoplethysmography (PPG) knowledge features (KF) are vital for blood pressure prediction.
• Existing KF often overlook PPG's global characteristics, limiting prediction accuracy.

Purpose of the Study:

• To develop a novel feature extraction method for enhanced blood pressure prediction using PPG.
• To evaluate the efficacy of data-knowledge co-driven features (DKCF) against traditional methods.

Main Methods:

• Functional data analysis (FDA) was employed to extract data features (DF).
• DKCF was proposed by integrating FDA with KF constraints.
• Random forest, AdaBoost, gradient boosting, SVM, and DNN models were used for comparison.

Main Results:

• DKCF achieved the lowest mean absolute errors (MAE) for systolic blood pressure (SBP) and diastolic blood pressure (DBP) in dataset 1.
• DKCF demonstrated the smallest MAE for SBP prediction and the second smallest for DBP in dataset 2.

Conclusions:

• Low-dimensional DKCF derived from PPG are significantly correlated with blood pressure.
• DKCF presents a potential, non-invasive indicator for comprehensive health assessment.