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

Due to the inconvenience of drawing blood and the possibility of infection associated with invasive methods, research on non-invasive glycated hemoglobin (HbA1c) measurement methods is increasing. Utilizing wrist photoplethysmography (PPG) with machine learning to estimate HbA1c can be a promising method for non-invasive HbA1c monitoring in diabetic patients. This study aims to develop a HbA1c estimation system based on machine learning algorithms using PPG signals obtained from the wrist. We used a PPG based dataset of 22 subjects and algorithms such as extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), Categorical Boost (CatBoost) and random forest (RF) to estimate the HbA1c values. Note that the AC-to-DC ratios for three wavelengths were newly adopted as features in addition to the previously acquired 15 features from the PPG signal and a comparative analysis was performed between the performances of several algorithms. We showed that feature-importance-based selection can improve performance while reducing computational complexity. We also showed that AC-to-DC ratio (AC/DC) features play a dominant role in improving HbA1c estimation performance and, furthermore, a good performance can be obtained without the need for external features such as BMI and SpO. These findings may help shape the future of wrist-based HbA1c estimation (e.g., via a wristwatch or wristband), which could increase the scope of noninvasive and effective monitoring techniques for diabetic patients.