Systematic Review of Artificial Intelligence Techniques for Blood Glucose Predicting in Diabetes

Abstract

Background/Objectives: The accurate prediction of Blood Glucose Levels (BGLs) is important for the effective management of diabetes. Artificial Intelligence (AI) has become an effective tool for forecasting BGLs and offering decision support and therapeutic interventions. In this study, existing AI techniques for predicting BGLs were reviewed systematically to identify their limitations and strengths. Methods: Following the Preferred Reporting Items for Systematic Reviews and Meta Analyses (PRISMA) guidelines, a systematic review of relevant studies was performed. Research was identified through a detailed search of several databases including PubMed, Institute of Electrical and Electronics Engineers (IEEE) Explore, and Scholar with a focus on publications ranging from 2010 to 2025. Search terms included “AI for predicting glucose levels in diabetes” and related keywords. Non-English and irrelevant studies were excluded. Results: Fifty-seven studies met the inclusion criteria. They used AI models such as Bidirectional Long Short-Term Memory (Bi-LSTM) and Convolutional Neural Network (CNN), decision trees, and hybrid models like CNN-Gated Recurrent Unit (GRU). Many models demonstrated high accuracy, especially Bi-LSTM and CNN-GRU hybrids, but were often limited by small datasets, lack of generalizability, and low interpretability. Non invasive sensing technologies and real-time systems using Internet of Things (IoT) and wearable devices showed significant promise but require further validation. Conclusions: AI-based glucose prediction systems have advanced significantly, with some models achieving Root Mean Squared Error (RMSE) values below 10 mg/dL in short-term forecasting. However, no single approach consistently outperforms others across all prediction horizons and patient scenarios. Future research should focus on integrating multimodal inputs, enhancing model interpretability, and validating systems using diverse, real-world datasets to ensure clinical reliability.