A Comparative Analysis of Various Machine Learning Algorithms to Improve the Accuracy of HbA1c Estimation Using Wrist...

Area of Science:

• Biomedical Engineering
• Medical Informatics
• Data Science

Background:

• Invasive blood draws for glycated hemoglobin (HbA1c) monitoring pose inconvenience and infection risks.
• Non-invasive methods for HbA1c estimation are increasingly researched to improve patient care.
• Wrist-based photoplethysmography (PPG) offers a promising avenue for continuous, non-invasive monitoring.

Purpose of the Study:

• To develop and evaluate a machine learning system for estimating HbA1c levels using wrist-based PPG signals.
• To investigate the impact of novel AC-to-DC ratio features on HbA1c estimation performance.
• To compare the efficacy of various machine learning algorithms for non-invasive HbA1c monitoring.

Main Methods:

• Utilized a PPG dataset from 22 subjects to train and test machine learning models.
• Employed algorithms including XGBoost, LightGBM, CatBoost, and Random Forest for HbA1c estimation.
• Incorporated AC-to-DC ratios across three wavelengths as new features alongside 15 existing PPG signal features.

Main Results:

• Feature-importance-based selection enhanced model performance and reduced computational load.
• AC-to-DC ratio features were identified as dominant contributors to improved HbA1c estimation accuracy.
• Accurate HbA1c estimation was achieved without requiring external data like BMI or SpO2.

Conclusions:

• Wrist-based PPG combined with machine learning, particularly with novel AC-to-DC ratio features, enables effective non-invasive HbA1c estimation.
• This approach can lead to more accessible and convenient monitoring for diabetic patients.
• Future development of wrist-worn devices could significantly expand non-invasive HbA1c monitoring capabilities.