Artificial intelligence in clinical endocrinology: A narrative review of predictiveanalytics, therapeutic optimization, and diagnostic applications

Background: Endocrine disease is governed by non-linear, multi-organ feedback loops and longitudinalphysiological data, which makes it a natural fit for machine learning (ML) and deep learning (DL) methods. Thisnarrative review synthesizes current evidence on the application of artificial intelligence (AI) across fivedomains of clinical endocrinology: diabetes risk prediction, metformin response and the gut microbiome,continuous glucose monitoring (CGM) with closed-loop insulin delivery, thyroid nodule ultrasound, and adrenallesion imaging.

Methods: PubMed, Scopus, and Web of Science were searched for studies and systematic reviews publishedmainly between 2020 and 2026, using combinations of terms including ML, DL, "diabetes prediction,""metformin," "gut microbiome," "continuous glucose monitoring," "closed-loop insulin," "thyroid nodule," and"adrenal incidentaloma." Priority was given to meta-analyses, randomized controlled trials, and large validatedcohorts; small single-center case series were included only where no larger evidence existed for that domain,and are flagged as such. As a narrative rather than systematic review, no PRISMA flow diagram or formal recordcount is reported; this is stated as a limitation.

Results: Ensemble models (random forest, XGBoost) predicted incident type 2 diabetes with 98-99% accuracyin one large longitudinal cohort in which most participants did not develop diabetes; because accuracy can beinflated in such imbalanced outcome data, this figure should be interpreted alongside positive predictive value,which was not reported in the source study, and alongside the fact that external validation remains uncommonin this literature. DL models applied to thyroid ultrasound reached pooled sensitivities of 87-91% andspecificities of 83-95% across meta-analyses covering more than 146,000 patients. Randomized trial data showthat closed-loop insulin delivery significantly improves time-in-range and reduces hypoglycemia, whilereinforcement-learning-based dosing remains largely at the simulation stage. ML models linking gutmicrobiome composition to drug pharmacokinetics achieved moderate accuracy (AUROC approximately 0.75),a distinct question from separate work linking gut microbiome signatures to type 2 diabetes risk itself. AdrenalCT radiomics models reported AUCs frequently above 0.85, though largely from small, single-center cohorts (asfew as 19-40 patients).

Conclusion: AI shows strong, evidence-backed value in endocrine risk prediction and image-based diagnosis,though the diabetes-prediction accuracy figures above warrant cautious interpretation given class imbalance.Evidence for therapeutic personalization, particularly metformin response, is earlier-stage. Larger prospectivetrials, external validation, standardized outcome reporting (including PPV alongside accuracy), and improvedinterpretability are needed to support broader clinical adoption.