Blood pressure stratification using photoplethysmography and light gradient boosting machine.

Area of Science:

• Biomedical Engineering
• Cardiovascular Health
• Machine Learning in Healthcare

Background:

• Hypertension (HT) is a major global risk factor for cardiovascular disease and mortality.
• Early identification and treatment of HT are critical for patient outcomes.
• Current blood pressure monitoring often requires invasive methods or cuff-based devices.

Purpose of the Study:

• To evaluate the efficacy of the LightGBM machine learning model for blood pressure stratification using photoplethysmography (PPG) signals.
• To develop a noninvasive, rapid method for early hypertension detection.
• To assess the performance of PPG-derived features in classifying different blood pressure categories.

Main Methods:

• Utilized 121 records of PPG and arterial blood pressure (ABP) signals from the MIMIC-III database.
• Extracted features from PPG, velocity plethysmography, and acceleration plethysmography.
• Trained an Optuna-tuned LightGBM model on seven feature sets for three classification trials (NT vs. PHT, NT vs. HT, NT+PHT vs. HT).

Main Results:

• Achieved high F1 scores of 90.18% (NT vs. PHT), 97.51% (NT vs. HT), and 92.77% (NT+PHT vs. HT).
• Demonstrated that combining multiple PPG-derived features improved classification accuracy compared to using PPG features alone.
• The model effectively stratified blood pressure categories with high precision.

Conclusions:

• The proposed LightGBM model accurately stratifies hypertension risk using PPG signals.
• This noninvasive, rapid, and robust method shows promise for early hypertension detection.
• Potential for integration into wearable devices for cuffless blood pressure monitoring.