Optimizing Non-Invasive Blood Glucose Prediction Through Data Preprocessing and ResNet50

In recent years, non-invasive blood glucose measurement methods have garnered significant attention, leading to a surge in related research and numerous remarkable achievements. However, these studies often involve a limited number of participants and rarely address individual variability. For PPG signals, which exhibit significant individual differences, the accuracy and reliability still require validation. This study aims to address the issue of personalization. We utilized the MIMIC-III Waveform Database Matched Subset [1]–[3], employing 170,000 instances of 5-second PPG signals with blood glucose values ranging from 70 to 250 for training, covering 4,157 subjects to enhance the model's generalizability. We selected ResNet50 as our model and incorporated a data stratification preprocessing concept to select higher quality signals for training, thus improving data quality. This preprocessing approach helps in better feature extraction, reducing the need for complex structures to handle noise and outliers, thereby increasing the training efficiency and generalizability of the model. Our study aims to predict a blood glucose value ranging from 70 to 250, divided into 180 classes, and adheres to the reasonable error range according to the international ISO:15197 2013 standards for blood glucose accuracy. The accuracy for personalized data (specific individual data already trained in the model) reached 91.02%, with the best A zone in the Clarke Error Grid (representing no clinical risk) achieving 97.44%, meeting the ISO:15197 2013 medical equipment standards. For non-personalized data, we used 50,000 instances from 1,846 individuals not included in the model's training set, achieving an accuracy of 75.38%, with the best A zone in the Clarke Error Grid reaching 70.2%, demonstrating excellent results in non-personalized studies.