A Deep Learning Model That Combines ResNet and Transformer Architectures for Real-Time Blood Glucose Measurement...

Recent advances in wearable devices and physiological signal monitoring technologies have motivated research into non-invasive glucose estimation for diabetes management. However, the existing studies are often limited by sample constraints, in terms of relatively small numbers of subjects, and few address personalized differences. Physiological signals vary considerably for different individuals, affecting the reliability of accuracy measurements, and training data and test data are both used from the same subjects, which makes the test result more affirmative than the truth. This study aims to compare the two scenarios mentioned above, regardless of whether the testing/training involves the same individual, in order to determine whether the proposed training method has better generalization ability. The publicly available MIMIC-III dataset, which contains 700,000 data points for 10,000 subjects, is used to create a more generalized model. The model architecture uses a ResNet CNN + Transformer block, and data quality is graded during preprocessing to select signals with less interference for training to increase data quality. This preprocessing method allows the model to extract useful features without being adversely affected by noise and anomalous data that decreases performance; therefore, the model's training results and generalization capability are increased. This study creates a model to predict blood glucose values from 70 to 250 for 180 classes, using mean absolute relative difference (MARD) as the evaluation metric and a Clarke error grid (CEG) to determine a reasonable error tolerance. For personalized cases (specific individual data during model training), the MARD is 11.69%, and the optimal Zone A (representing no clinical risk) in the Clarke error grid is 82.7%. Non-personalized cases (test subjects not included in the model training samples) using 60,000 unseen data yields MARD = 15.16%, and the optimal Zone A in the Clarke error grid is 75.4%. Across multiple testing runs, the proportion of predictions falling within Clarke error grid zones A and B consistently approached 100%. The small performance difference suggests that the proposed method has the potential to improve subject-independent estimation; however, further validation in broader populations is required. Therefore, the primary objective of this study is to improve subject-independent, non-personalized PPG-based glucose estimation and reduce the performance gap between personalized and non-personalized measurements.