Artificial Intelligence based Glucose Forecasting and Data Augmentation for Personalized Advanced Diabetes Management

Abstract

Background: Diabetes, a chronic condition affecting millions globally, requires continuous glucose monitoring and precise insulin dosing. The development of the artificial pancreas (AP) can automate glucose regulation, but widespread adoption remains limited due to several challenges, lack of automated physical activity integration, manual meal recording and declaration, continuous glucose monitoring sensor errors such as missing data points, high costs of CGM sensors and insulin pumps. These limitations underscore the need for a more intelligent, accessible, and affordable diabetes management system. CGM sensors track the glucose levels in interstitial fluid (ISF) using electrochemical methods. These sensors provide real-time glucose data and alerts but are limited by the physiological delay between blood glucose and ISF glucose levels, leading to delayed insulin administration. Inaccurate or untimely insulin dosing can result in hyperglycemia or hypoglycemia, posing severe health risks. Furthermore, the absence of multimodal data, including physical activity and precise meal intake information, reduces forecasting accuracy, limiting the effectiveness of decision support systems and closed-loop insulin delivery. Objectives: The major objective of the work is to develop a personalized, data-driven, interpretable, deep learning-based glucose forecasting and insulin dose tuning solution for closed-loop diabetes management devices or artificial pancreas that takes into account abiotic factors including diet, stress, and physical activity. Methods: This research develops a series of AI-driven models for glucose forecasting, starting with WDNet (Wide Deep LSTM-GRU), which integrates content-based attention mechanisms to enhance prediction accuracy even with limited training data. WDNet outperforms baseline models across both clinical and in-silico datasets by effectively capturing temporal dependencies. To further improve interpretability and forecasting reliability, multihead attention learning was embedded within neural basis expansion and hierarchical interpolation networks, enabling more transparent and user-centric predictions. Additionally, GAN-based generative AI was employed to address data scarcity in CGM sensor systems by synthesizing realistic glucose, physical activity, and meal data. This augmentation technique, validated across multiple datasets, significantly enhances prediction accuracy by leveraging both real and synthetic data. Results: All the proposed AI algorithms were tested on both the OhioT1D dataset and an indigenous dataset from SCB Medical College, Cuttack India. Among them, the Neural Basis Expansion model with Multihead Attention outperformed other models across both datasets in terms of MARD, RMSE, MAE, and Clarke Error Grid Analysis (CEGA), ensuring superior accuracy in glucose forecasting. For the OhioT1D dataset, at 30-minute PH, proposed model achieved MARD of 6.81 ± 1.39% and RMSE of 16.57 ± 2.56 mg/dL, while at 60-minute PH, these values were 12.15 ± 3.15% and 29.25 ± 6.02 mg/dL, respectively. CEGA confirmed that >91% of predictions fell in Region A, indicating high clinical accuracy with minimal safe deviations in Region B. For the indigenous dataset, at 30- minute PH, MARD was 7.25 ± 0.16% and RMSE was 16.82 ± 0.54 mg/dL, while at 60-minute PH, MARD reached 13.14 ± 0.26% and RMSE was 28.96 ± 0.43 mg/dL. CEGA showed that most predictions remained in Region A, ensuring reliable clinical decision-making. Furthermore, incorporating Generative AI for data augmentation significantly improved prediction accuracy, reducing data gaps in CGM readings and other lifestyle input parameters. This enhancement is crucial for glycemic control, ensuring more precise insulin dosing in real-time closed-loop diabetes management. Conclusions: The proposed models effectively integrate multimodal inputs, including physical activity, meal intake, and CGM data, to improve prediction accuracy and clinical applicability. The results confirm the potential of AI-enhanced closed-loop diabetes control systems, providing a foundation for personalized and automated insulin dosing. The findings contribute toward the clinical translation of AI-based decision support systems, addressing key challenges in diabetes management, including data scarcity, interpretability, and real-world adaptability.