Nutrients, Vol. 17, Pages 3832: Prediction of Postprandial Blood Glucose Variability Using Machine Learning in Frequent Insulin Injection Therapy with a Simplified Carbohydrate Counting Model
Nutrients doi: 10.3390/nu17243832
Authors:
Hiroyuki Tominaga
Masahide Hamaguchi
Youji Hamaguchi
Ren Yashiki
Aki Yamaguchi
Tadaharu Arai
Masahiro Yamazaki
Noriyuki Kitagawa
Yoshitaka Hashimoto
Hiroshi Okada
Michiaki Fukui
Background/Objectives: Postprandial glucose variability is a key challenge in diabetes management for patients receiving multiple daily insulin injections (MDI). This study evaluated transformer-based machine-learning models for predicting post-prandial glucose peaks and nadirs using pre-meal glucose, insulin dose, and nutritional input. Methods: In this observational study, 58 adults with diabetes provided dietary records, insulin logs, and continuous glucose monitoring data. After preprocessing and participant-level splitting (64:16:20), model-ready datasets comprised 6155/1449/1805 (train/validation/test) meal events for the Full-Nutrition model and 6299/1484/1849 for the Carbohydrate and Available-Carbohydrate models. We evaluated three transformer-based models and assessed performance using MAE, R2, and the Clarke error grid. Results: The Full Nutrition Model achieved MAEs of 32.2 mg/dL (peak) and 21.8 mg/dL (nadir) with R2 values of 0.58 for both. Carbohydrate-based models showed similar accuracy. Most predictions fell within Clarke error grid Zones A and B. Conclusions: Transformer-based machine-learning models can accurately predict postprandial glucose variability in MDI-treated patients. Carbohydrate-only inputs performed comparably to full-nutrient data, supporting the feasibility of simplified dietary inputs in clinical applications.
Background/Objectives: Postprandial glucose variability is a key challenge in diabetes management for patients receiving multiple daily insulin injections (MDI). This study evaluated transformer-based machine-learning models for predicting post-prandial glucose peaks and nadirs using pre-meal glucose, insulin dose, and nutritional input. Methods: In this observational study, 58 adults with diabetes provided dietary records, insulin logs, and continuous glucose monitoring data. After preprocessing and participant-level splitting (64:16:20), model-ready datasets comprised 6155/1449/1805 (train/validation/test) meal events for the Full-Nutrition model and 6299/1484/1849 for the Carbohydrate and Available-Carbohydrate models. We evaluated three transformer-based models and assessed performance using MAE, R2, and the Clarke error grid. Results: The Full Nutrition Model achieved MAEs of 32.2 mg/dL (peak) and 21.8 mg/dL (nadir) with R2 values of 0.58 for both. Carbohydrate-based models showed similar accuracy. Most predictions fell within Clarke error grid Zones A and B. Conclusions: Transformer-based machine-learning models can accurately predict postprandial glucose variability in MDI-treated patients. Carbohydrate-only inputs performed comparably to full-nutrient data, supporting the feasibility of simplified dietary inputs in clinical applications. Read More