Implicit HbA1c Achieving 87% Accuracy within 90 Days in Non-Invasive Fasting Blood Glucose Measurements Using Photoplethysmography

Figures

• 

  Figure 1. Schematic illustration of the concept of overfitting or underfitting using circle size, along with the transition from explicit HbA1c to implicit HbA1c for an improved machine learning prediction of blood glucose levels. To reduce the circle size of the predicted blood glucose level with the PPG feature vector, we introduce pretest HbA1c. It significantly reduces instances of overfitting or underfitting that lead to scattered blood glucose level predictions derived solely from PPG data (from ① to ②). Implicit HbA1c is related to pretest HbA1c but contains systematic correction items to account for personal deviation, thereby further enhancing the accuracy of blood glucose level prediction (from ② to ③). The input features for each stage are as follows: ① PPG features, ② PPG features and pretest HbA1c, ③ PPG features and implicit HbA1c.

• 

  Figure 2. Illustration of the sequential workflow, starting from the pretest phase to derive the implicit HbA1c, followed by its application during testing to generate the final blood glucose level prediction. During pretest, the required inputs are PPG measurement and a finger-prick blood glucose level. The finger-prick value is only used at the final step of the pretest to determine implicit HbA1c value. During testing, the required inputs are the pretest determined implicit HbA1c and a PPG measurement.

• 

  Figure 3. Illustrative example of the process for determining implicit HbA1c. A range of HbA1c values, ranging from 4 to 12, is employed to estimate blood glucose levels (BGL). Among these values, the HbA1c value corresponding to the predicted BGL closest to the pretest reference BGL is identified and designated as the implicit HbA1c value.

• 

  Figure 4. (a) Difference in implicit HbA1c and corresponding measured reference HbA1c, and (b) distribution of the BG–HbA1c (measured) relationship for the training and testing sets.

• 

  Figure 5. CEG analysis of using implicit HbA1c and explicit HbA1c for model prediction on subjects not affected by drugs.

• 

  Figure 6. Percentage error distribution of the prediction while using implicit and explicit HbA1c for conducting the nonparametric Wilcoxon signed-rank test. The implicit HbA1c method demonstrates a distribution pattern that resembles a normal distribution. The explicit HbA1c method exhibits a left-skewed distribution with systemically lowered prediction.