We conducted a structured rapid review following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines. Consistent with a rapid review approach, the study adhered to PRISMA principles but did not aim for exhaustive coverage, duplicate data extraction at all stages, or meta-analytic synthesis characteristic of a full systematic review. Methodological streamlining, including a restricted time frame and narrative synthesis, was implemented to provide a timely overview of emerging AutoML applications in a fast-evolving field. Accordingly, the findings should be interpreted as a structured mapping of methodological approaches rather than a comprehensive appraisal of the evidence. Although some features overlap with exploratory or scoping reviews, the use of predefined eligibility criteria, a focused research question, and a formal risk-of-bias assessment support the classification of this study as a rapid review. The research question was structured using the population, intervention, comparator, and outcome (PICO) framework to guide eligibility criteria, search strategy, and data extraction.

Eligibility criteria were defined based on the PICO framework (Textbox 1).

Eligible studies were empirical investigations using AutoML to predict T2D onset or risk from observational or experimental datasets (eg, cohorts, case-control studies, clinical datasets, or open-access data).

A structured search strategy was developed, combining four thematic blocks: (1) diabetes (“diabet*,” “type 1,” “type 2,” “T1D,” “T2D,” “pre-diabet*,” “insulin resistance,” “impaired glucose tolerance,” “HbA1c,” “hyperglycemia,” and “glycemic control”), (2) AutoML and ML (“automated machine learning,” “machine learning,” “artificial intelligence,” “deep learning,” “neural network,” “random forest,” “support vector machine,” “decision tree,” “XGBoost,” “TPOT,” “H2O.ai,” “DataRobot,” “feature selection,” and “pipeline automation”), (3) data types of interest (“genom*,” “DNA,” “SNP,” “gene expression,” “epigenetic,” and “GWAS”; “diet,” “physical activity,” “sleep,” and “lifestyle”; “pollution,” “air quality,” “urban,” “deprivation,” and “exposure”; and “combined model,” “multi-input,” and “fusion model”), and (4) prediction (“predict*,” “forecast*,” “risk model*,” “classification model*,” and “risk prediction”). This strategy was adapted and applied across 6 databases: PubMed, Scopus, Web of Science, IEEE Xplore, Google Scholar, and Embase. Filters were used to limit the search to original scientific articles published between January 2015 and May 2025 in English or French.

All 851 retrieved studies were imported into Covidence (Veritas Health Innovation), and duplicates were removed. Title and abstract screening were performed independently by 2 reviewers, with ChatGPT (OpenAI) acting as an AI-based third reviewer for adjudication. Articles deemed potentially relevant (59/743, 7.9%) underwent full-text assessment against the predefined eligibility criteria using the same dual-reviewer process and AI-assisted adjudication.

Data extraction for the final 22.0% (13/59) of the studies was conducted in Covidence using a standardized form pretested on 3 to 5 studies. Extracted fields covered bibliographic details, study context, design, population, data types, predictors, AutoML implementation, performance metrics, interpretability, reproducibility, limitations, and clinical relevance. All data were cross-checked for completeness and consistency. ChatGPT served as an AI-assisted third reviewer solely to support adjudication in cases of disagreement by summarizing relevant methodological elements; final decisions were reached by consensus between human reviewers and were not determined by AI-generated outputs. All final inclusion decisions were verified and confirmed by human assessors, strengthening methodological rigor. The extracted study-level data are provided in Multimedia Appendix 1.

Extracted data were narratively synthesized and summarized in tables, detailing study methods, data types, AutoML approaches, and key performance metrics, including calibration. The analysis emphasized validation strategies, interpretability, and clinical applicability. All synthesis steps were cross-checked by ChatGPT as an AI-based third reviewer to ensure coherence. Risk of bias was assessed using the PROBAST (Prediction Model Risk of Bias Assessment Tool)+AI tool [7], evaluating participants, predictors, outcomes, and analysis domains.

The database search yielded 851 records, of which 743 (87.3%) were screened. Following title and abstract screening and full-text review, 13 studies published between 2017 and 2025 met the eligibility criteria (Figure 1; Checklist 1).

The studies spanned Brazil, Greece, India, Australia, South Korea, Pakistan, China, Spain, the United States, Japan, and Iraq and used a variety of designs. Most were observational (prospective or retrospective), although 1 study conducted in 2022 by Mora-Ortiz et al [13] used a randomized controlled trial dataset. Among the 13 included studies, 2 (15.4%) used fully automated pipelines, 1 (7.7%) used semiautomated workflows, and 10 (76.9%) relied predominantly on manual ML approaches. Sample sizes ranged from 183 participants in a targeted clinical trial to 25,406 participants in a large community-based cohort.

Not all studies predicted incident T2D. Several addressed related but distinct outcomes, including diagnosis, disease subtyping, or engagement in prevention programs. These outcomes differ fundamentally in clinical meaning, prediction horizon, and class balance. Consequently, performance metrics such as the area under the receiver operating characteristic curve (AUROC) or accuracy are not directly comparable across studies, and apparent differences in performance should be interpreted with caution. Therefore, AUROC values derived from these different prediction tasks should not be interpreted as directly comparable.

Clinical variables, such as anthropometry, biochemical measures, and demographics, were the most common predictors. Behavioral variables appeared in several studies; genomic and environmental features were rare or absent. Data sources included large public cohorts (eg, Brazilian Longitudinal Study of Adult Health, English Longitudinal Study of Ageing–UK, Korean Genome and Epidemiology Study, and University of California, Irvine repositories); commercial platforms (eg, Noom); and local hospital records. The number of predictors ranged from fewer than 10 to more than 200.

A few studies used general-purpose AutoML frameworks (eg, Tree-Based Pipeline Optimization Tool [TPOT], H2O AutoML, and Azure Machine Learning), while many relied on partially automated or traditional workflows. Common algorithms included random forests, gradient-boosted methods, artificial neural networks, and logistic regression. Most studies combined limited automation (eg, feature selection or tuning) with manual preprocessing, corresponding to semiautomated pipelines under the operational definition used in this review.

Internal validation was common, particularly 10-fold cross-validation and holdout splits (eg, 70/30 or 80/20), often with stratified or repeated sampling. External or prospective validation was reported in a small subset, and performance sometimes declined compared with internal results. A few studies incorporated external or prospective validation [14,15]; however, these were the exception.

Reported performance varied by dataset, outcome definition, and modeling approach. AUROC values ranged from approximately 0.74 to 0.99 across studies, and accuracy ranged from approximately 75% to 97.7%. In some cases, AutoML-derived models outperformed established clinical risk scores; for instance, Fazakis et al [4] achieved an AUROC of 0.884 compared with 0.821 for the Finnish Diabetes Risk Score (FINDRISC). AUROC values near 0.99 were observed primarily in training or test splits or for nonincidence tasks (eg, engagement prediction in digital prevention programs), and these often lacked external validation [4]. However, calibration measures and decision-analytic metrics were rarely reported, limiting the clinical interpretability of high AUROC values.

Most studies included some form of interpretability, such as feature importance rankings, Shapley additive explanations (SHAP) values, or Boruta-based feature selection. Reproducibility was rarely addressed: few studies provided open-source code, complete model parameters, or complete datasets.

Although several studies discussed potential clinical applications, only a minority incorporated elements aligned with real-world implementation, such as prospective validation, workflow integration, or usability testing for clinicians. Overall, methodological performance was often high, but concrete steps toward clinical adoption remained limited.

A risk-of-bias assessment using the PROBAST framework examined 4 domains: participants, predictors, outcomes, and analysis.

In the participants domain, most studies showed a high risk of bias, although populations were generally well defined and clinically relevant [16,17].

The predictors domain was more heterogeneous: some studies used automated feature selection methods such as least absolute shrinkage and selection operator (LASSO) or SHAP, while others relied on manual or poorly documented approaches, leading to moderate to high risk of bias. In the outcome domain, risk was low, as predictive targets were consistently well defined [4,14,16,18].

Regarding analysis, most studies applied appropriate validation techniques (cross-validation, training or test splits, or external validation), although a few lacked sufficient methodological detail [12,18-20].

Overall, approximately one-third of studies were at high risk of bias, one-third at moderate risk, and only a few at low risk across all PROBAST domains.

Across studies, the PROBAST domains most frequently associated with high risk of bias were predictor selection and analysis, often due to manual feature selection, limited reporting of preprocessing steps, or the lack of external validation. Studies using more automated and standardized pipelines tended to show lower risk in the predictors domain but not necessarily in the analysis domain. Notably, very high AUROC values were more frequently observed in studies with a higher overall risk of bias.

In this review, clinical readiness was defined as the presence of key elements required for real-world deployment, including external validation, calibration reporting, assessment of clinical utility, workflow integration, and interpretability for end users. The included studies spanned multiple regions and addressed heterogeneous objectives, such as early screening, cardiovascular risk stratification, T2D subtyping, and support for digital prevention programs [13,16-18]. This diversity reflects the breadth of current AutoML applications in diabetes research but also underscores the need for caution when comparing performance across fundamentally different prediction tasks.

From a methodological standpoint, supervised algorithms such as random forest, extreme gradient boosting (XGBoost), and light gradient boosting machine were predominant, balancing performance and interpretability. Several studies used SHAP or decision trees to identify key predictors and improve clinical usability [4,14]. However, most relied on retrospective, single-center data, with limited external validation and potential selection bias [12,14,16]. This gap between methodological rigor and generalizability remains a major barrier to clinical translation.

Bias also arose from manual predictor selection [16,18]. In contrast, studies embedding techniques such as LASSO or Boruta within automated or semiautomated pipelines demonstrated improved standardization [4,14,17]. In this review, such techniques were considered part of AutoML only when integrated into an automated workflow rather than applied as standalone methods. Fully automated AutoML pipelines—encompassing preprocessing, feature selection, and hyperparameter tuning—remained rare; only 1 study implemented Azure Machine Learning without complete pipeline automation [19]. Nevertheless, a gradual shift toward semiautomated and externally validated approaches was observed over time [12,14].

Reported performance was high (area under the curve=0.74-0.99), but comparability suffered from heterogeneous outcomes and metrics [13,18]. Few studies reported calibration or predictive value measures, despite their importance for clinical risk prediction. Reproducibility also remained a major weakness, with no study sharing full code or modeling pipelines, contrary to open science recommendations [4,14]. Overall, AutoML approaches did not consistently outperform non-AutoML models under rigorous validation conditions, underscoring the primacy of external validation and calibration over nominal performance gains, particularly across heterogeneous prediction tasks.

Finally, interpretability remains a key challenge, as AutoML systems are often perceived as “black boxes” [11]. Integrating explainable AI tools such as SHAP, local interpretable model-agnostic explanations, and integrated gradients directly into AutoML frameworks (eg, AutoGluon) could enhance transparency and usability, thereby supporting clinical adoption [11]. Future research should prioritize data diversity; prospective multicenter validation; and the development of standardized, shareable AutoML pipelines to facilitate reproducibility and real-world implementation [4,12,14].

Notably, none of the reviewed studies included genomic data, despite its potential to improve prediction accuracy. Beyond technical considerations, practical barriers such as cost, limited availability of genomic data in routine care, population-specific transferability of polygenic risk scores, and data governance constraints likely contribute to the absence of genomic predictors in current AutoML-based T2D models. Nevertheless, existing evidence suggests that integrating polygenic risk information could enhance future AutoML-based prediction models [17].

The included studies reported several recurring limitations, notably the reliance on regional or hospital-based datasets, limited external or prospective validation, and incomplete integration of behavioral or genomic predictors. Authors consistently emphasized the need for broader and more diverse predictor sets, multicenter study designs, and prospective validation. Taken together, these limitations mirror the methodological and clinical gaps identified in this review and underscore the need for standardized, explainable, and reproducible AutoML pipelines oriented toward real-world clinical use.

This rapid review underscores the promising yet underrealized potential of AutoML for predicting T2D risk. Across 13 studies (2017‐2025), models showed strong discrimination and a growing use of interpretable methods (eg, SHAP and decision trees). However, clinical translation remains constrained by local or retrospective data; limited external validation; and the absence of fully automated, reproducible pipelines.

To overcome these gaps, future research should adopt prospective, multicenter designs and integrate diverse data sources—clinical, behavioral, environmental, and omics data. Developing standardized, transparent, and shareable AutoML frameworks is essential for building trust and enabling real-world applications.

Some actionable recommendations are as follows:

Collectively, these actions can transform AutoML from a high-performing research tool into a transparent, reliable decision support system with tangible public health benefits.