Foundation models promise a diabetes AI revolution—without the evidence to match

Artificial intelligence in diabetes care is undergoing its most dramatic architectural shift in a decade, moving from bespoke networks trained for a single task to enormous foundation models pretrained on millions of images, glucose readings, patient records, and text tokens. A critical narrative review published in Frontiers in Artificial Intelligence maps this transition across every diabetes-relevant data modality—retinal imaging, continuous glucose monitoring, electronic health records, and clinical text—and reaches an uncomfortable conclusion: the models have grown far faster than the evidence behind them. In the literature the authors reviewed, spanning January 2016 through December 2025, they found no randomized evaluation of any foundation model in a diabetes workflow, and prospective validation was, in their words, effectively absent.

The review formalizes the field’s history as three eras. Era I belonged to fixed risk scores such as the UKPDS risk engine, whose coefficients were estimated once on small cohorts from hand-engineered features like age, HbA1c, and blood pressure. Era II, roughly 2016 to 2020, brought task-specific deep learning: convolutional networks for diabetic retinopathy grading, exemplified by Gulshan and colleagues’ 2016 work, and the IDx-DR system, which in 2018 became the first autonomous AI diagnostic authorized by the US Food and Drug Administration in any field of medicine. Era III, dated from 2021, rests on a different premise—self-supervised pretraining on broad corpora, followed by adaptation to many downstream tasks with few or no new labels.

The emblematic ophthalmic example is RETFound, a Vision Transformer with roughly 300 million parameters pretrained by masked autoencoding on 1.6 million retinal photographs from the UK Biobank and Moorfields Eye Hospital. Its developers reported that it outperformed task-specific networks on diabetic retinopathy grading, retinal disease prognosis, and even prediction of systemic events such as myocardial infarction, heart failure, and stroke from retinal images alone. But the review’s authors dissect what those headline claims rest on. The systemic-prediction results were validated only on held-out UK Biobank data drawn from the same biobank used for pretraining, and the reported AUROCs for cardiac endpoints, roughly 0.58 to 0.64, sit far below clinical deployment thresholds and below established non-imaging risk scores.

Independent comparisons complicate the story further. Head-to-head evaluations published as preprints found that RETFound’s advantage over conventional deep learning is clearest when labeled data are scarce, and that it narrows or disappears when labels are plentiful—ImageNet-pretrained supervised networks were largely comparable for ocular disease detection when labeled data were ample. The label-efficiency dividend, the review notes, is consistently reported but chiefly by the model developers themselves, and the authors identified no comparable independent comparisons against tuned classical baselines for glucose-monitoring or health-record models.

On the continuous glucose monitoring side, pretrained temporal transformers in the lineage of PatchTST and masked time-series autoencoders are being adapted for glucose forecasting and hypoglycemia anticipation. Cross-dataset zero-shot forecasting is a genuinely new capability relative to era-II per-patient models, but the evidence is predominantly benchmark-level, on public datasets, and has not been shown to translate into clinical-alert performance such as sensitivity for clinically significant hypoglycemia in ambulatory patients. The review also flags a structural blind spot: nearly all CGM foundation-model work is type 1 diabetes-centric, and the authors found no evaluation stratified by diabetes type, age band, pregnancy status, or sensor generation—axes that change both the signal and the clinical question.

Clinical large language models present the most visible frontier. Med-PaLM reached 67.6 percent on the USMLE-style MedQA benchmark and Med-PaLM 2 reached 86.5 percent, with GPT-4-family models reporting comparable or higher scores on suites containing substantial diabetes pharmacotherapy and screening content. Yet the review argues these scores are an upper bound on, not an estimator of, operational competence. Licensing-exam content circulates on the same public web scraped for pretraining, so benchmark contamination is plausible and largely unmeasurable for closed-weight models whose corpora are undisclosed. Closed models are also updated without versioned clinical re-evaluation, meaning a benchmark score from one generation does not bind its successor.

The safety analysis concentrates on what the authors call the glycemic-critical zone. They classify deployments into four ascending risk classes, from clinician-facing documentation up to patient-facing autonomous advice on insulin dosing and hypoglycemia response. LLM capability is roughly uniform across these classes, but safety evidence is concentrated almost entirely in the lowest one; for autonomous patient-facing dosing advice, they identified no prospective evidence at all. Hallucination—fluent, confident generation unsupported by input—concentrates precisely where training data are sparse or rapidly changing: insulin titration in patients with declining kidney function, perioperative management, and steroid-induced hyperglycemia. Retrieval grounding and constrained decoding over structured dose spaces offer partial mitigation, but neither has been evaluated in any prospective diabetes workflow.

The review’s evidence grading, an interpretive L1-to-L4 ladder from prospective trials down to benchmark-only evaluation, reveals a consistent pattern: most diabetes foundation-model studies sit at level 3, internal held-out data, or level 4, benchmark-only. External validation, when it exists, usually means a different institution or device rather than a different population or a later time period—distinctions the authors argue must be reported separately, because passing one kind of external test does not demonstrate transport across the others. Calibration reporting and fairness subgroup analysis were rare, and a perfectly discriminative but miscalibrated model is dangerous at a referral threshold. A ClinicalTrials.gov search in August 2026 found one relevant registration, a cluster-randomized trial of an LLM-enabled management tool for hypertension, diabetes, and dyslipidemia, not yet recruiting at the last check.

Privacy exposures also change shape at foundation-model scale. Fine-tuning a general model on institutional records can leak memorized patient content through prompted extraction; patient data submitted to closed, API-hosted models leaves the institutional boundary; and synthetic CGM and health-record data are increasingly marketed as privacy-preserving even though de-identification, differential privacy, and memorization audits answer different questions and are frequently conflated or unreported. The authors also describe a feedback ratchet of algorithmic inequity: models deployed first in well-resourced, well-documented health systems will absorb pretraining-scale demographic skews and reinforce the very data asymmetries that made representations weakest for the populations carrying the highest diabetes burden.

The proposed remedy is not new architectures but new evaluation culture. The review lays out a nine-point standard—pre-registered temporally separated test sets, contamination disclosure, mandatory calibration curves and decision-curve analysis, subgroup analysis as default, version-pinned results for closed models, shadow-mode silent trials, externally validated multimodal representations, explanation faithfulness testing, and tuned era-II baselines—none of which, the authors note, is technically difficult. Their research agenda culminates in coupling foundation-model representations with treatment-effect estimation, since the decisive clinical question in diabetes, which drug sequence for which patient, is a causal one that no current foundation-model architecture answers. The field’s center of gravity, they conclude, must shift from building representations to building accountable pipelines; whether that happens before or after wider clinical deployment is the central open question for the coming years.

Subject of Research: A critical review of foundation models in diabetes care and their evaluation and safety gaps

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

Article References: Nanda Kishore, M., & Navaneethan, C. (2026). From task-specific networks to foundation models in diabetes care: a critical review of opportunities, evaluation gaps, and safety frontiers. Frontiers in Artificial Intelligence, 9, Article 1983619. https://doi.org/10.3389/frai.2026.1983619

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1983619

Keywords: foundation models, diabetes care, artificial intelligence, diabetic retinopathy, continuous glucose monitoring, large language models, RETFound, model validation, benchmark contamination, clinical safety, algorithmic bias, electronic health records