From task-specific networks to foundation models in diabetes care: a critical review of opportunities, evaluation gaps, and safety frontiers

Artificial intelligence in diabetes care has historically relied on task-specific supervised learning, such as bespoke networks for retinopathy grading, kidney disease prediction, and glucose forecasting. The emergence of foundation models (FMs: large models pretrained on broad data, typically with self-supervision, and adapted to many downstream tasks) is displacing this paradigm across diabetes-relevant modalities: ophthalmic models such as RETFound, pretrained on 1.6 million retinal photographs, have been reported to outperform task-specific networks on retinopathy grading and systemic disease prediction, especially when labeled data are scarce; transformer-based time-series models are replacing long short-term memory (LSTM) pipelines for continuous glucose monitoring (CGM) forecasting; clinical large language models (LLMs) such as Med-PaLM and GPT-4 attain expert-level scores on medical question-answering benchmarks; and generative models supply synthetic CGM and electronic health record (EHR) data. This review provides a critical, modality-organized synthesis of this transition. We (i) formalize a three-era taxonomy: rule-based risk scores (era I), task-specific deep learning (era II, 2016–2020), and foundation models (era III, from 2021), and map the diabetes literature onto it; (ii) survey foundation-model architectures, pretraining objectives, and adaptation strategies across retinal imaging, CGM, longitudinal EHR, and multimodal data; (iii) expose an evaluation crisis: benchmark contamination of LLM medical evaluations, a near-universal absence in the literature we reviewed of external and prospective validation for retinal and CGM foundation models, miscalibration under dataset shift, and, to our knowledge and within the reviewed literature (January 2016–December 2025; negative findings are time-sensitive), the absence of any randomized evaluation of a foundation model in a diabetes workflow; and (iv) analyze safety risks: hallucinated insulin-dosing advice, automation bias, inherited algorithmic inequity, and privacy exposure during adaptation. We conclude that the decisive open problem is no longer representation learning but grounded, calibrated, causally informed deployment, requiring contamination-resistant evaluation, prospective trials, and hybrid architectures coupling foundation-model representations with treatment-effect estimation.