AI has become central to diabetes complication prediction, yet adoption of newer methods remains slow. A recent review in the Journal of Diabetes Science and Technology examined current trends in AI-based predictive modeling and found major gaps that limit progress in this growing field.

The review searched PubMed, Scopus, Ovid MEDLINE, CINAHL and IEEE databases. It included studies that developed or evaluated AI algorithms to predict diabetes-related microvascular and macrovascular complications. Predicted conditions, population features, modeling strategies, model performance and feature importance were assessed.

A total of 49 studies met inclusion criteria. Eye-related outcomes were most frequently predicted, appearing in 59% of studies. Among 48 studies applying AI specifically for prediction tasks, 54% used only Machine Learning models, 8% used only Deep Learning and 38% used both. No study applied foundation models or generative AI. Only 10% incorporated unstructured data such as imaging or physiologic signals. Age and glycated hemoglobin emerged consistently among the most important predictors.

The findings emphasize that current predictive research relies heavily on conventional AI approaches. Expanding the use of diverse data sources and modern architectures is necessary to advance prediction science and improve clinical relevance.