Systematic review finds diabetes risk models after gestational diabetes fall short

Women who develop gestational diabetes face a dramatically elevated chance of developing type 2 diabetes later in life, with evidence suggesting a nearly tenfold higher risk compared with women who never had the pregnancy complication. Because much of that risk depends on individual clinical and demographic characteristics, many of them modifiable, researchers have spent years building prediction models designed to flag which women need the closest postpartum follow-up. A new systematic review now takes stock of those models, and its verdict is sobering: while many tools look statistically impressive on paper, their methodological foundations and real-world applicability leave substantial room for doubt.

The review, published in Diagnostic and Prognostic Research, was led by Nigus Bililign Yimer and colleagues and registered prospectively on PROSPERO. The team searched five electronic databases, including MEDLINE, EMBASE, Emcare, CINAHL and SCOPUS, from their inception through January 24, 2026, following an initial search to May 12, 2025. After removing 1,583 duplicates from 5,504 retrieved records and screening nearly 4,000 articles at the title and abstract stage, the reviewers whittled the field down to 108 full-text candidates and ultimately included 19 studies describing 20 prediction models. Two reviewers independently screened and quality-assessed every study, with disagreements resolved through consensus or by a third reviewer.

The models themselves span a wide methodological spectrum. Logistic regression remains the workhorse, appearing in nine of the twenty models, followed by machine learning approaches in six and Cox regression in five. Among the machine learning entries, random forest and decision tree classifiers were the most frequently used algorithms. Geographically, the United States and Europe each contributed five studies, while Asia and Australia each accounted for four, and one study was multinational. More than half of the studies drew on prospectively collected data, and fifteen of the models were developed in hospital settings. Follow-up durations ranged from as short as two years to as long as thirty, with most studies tracking participants for at least five years.

Performance figures looked encouraging at first glance. Discrimination, the ability of a model to separate women who later developed type 2 diabetes from those who did not, was reported in all but one study, with area under the curve values ranging from 0.67 to 0.92. Yet the review authors caution that these numbers may be optimistic. Several models had very small numbers of events per predictor, spanning from as low as 1.4 to 121.6, a classic setup for overfitting. Split-sample validation, used in some studies, can also inflate apparent discrimination. And in five models, the reported c-statistics came without 95 percent confidence intervals, making it hard to judge precision.

The deeper problem lies in what was not measured. Calibration, which assesses whether predicted probabilities match observed outcomes, was reported using a slope or plot in only four models, with a handful of others offering alternatives such as the Hosmer-Lemeshow test or mean absolute error. Overall performance metrics like the Brier score or R-squared appeared in just two models. Clinical utility, the measure of whether a model actually improves decisions in practice, was rarely assessed at all. Internal validation was performed in thirteen models, but only one study validated its model externally, a striking gap given that external validation is the key test of whether a model generalises beyond its original dataset.

Sample size practices raised particular concern. Only one of the nineteen studies reported an appropriate sample size determination, despite enormous variability across the field: sample sizes ranged from 75 to 3,203 participants, event counts from 10 to 608, and the number of candidate predictors from 5 to more than 1,100. The reviewers note that inadequate sample size leads to overfitting, unstable estimates and poor clinical utility, and they point to established sample size determination methods and the TRIPOD + AI reporting guideline as routes to improvement. Missing data handling was similarly underreported, with some studies using complete case analysis without justifying the pattern of missingness, which the authors warn can bias predictor effects and distort discrimination estimates when data are not missing at random.

Quality appraisal used PROBAST + AI, a recently updated tool that this review was the first to apply to prognostic models in this field. The results revealed a telling split. During model development, quality concerns were generally low for the predictors, participants and outcome domains, but the analysis domain and overall quality carried higher proportions of unclear or high concern. The most striking finding concerned applicability: 60 percent of studies were judged to have high concern for overall applicability in both the development and evaluation phases. In other words, many models were methodologically sound in narrow respects but poorly positioned for transfer to diverse real-world populations, partly because of limited external validation, small samples, and specialised predictors not routinely available in clinical practice.

The predictors that dominated the models were refreshingly ordinary. Maternal age was the most commonly used variable, appearing in eleven models, followed by fasting glucose during pregnancy in ten and early pregnancy body mass index in eight. Family history of diabetes and oral glucose tolerance test results also featured prominently. A minority of studies incorporated metabolic and lipidomic signatures or genetic and other biomarkers, and while some of these omics-enhanced models reported improved discrimination, the review authors stress that this potential remains preliminary given the methodological limitations across the field. Notably, no included study employed hybrid approaches integrating machine learning with causal inference, despite these being proposed as a way to improve interpretability.

The review also positions itself against two recent systematic reviews that, in the authors’ assessment, failed to adhere to established best-practice standards such as the TRIPOD-SRMA and CHARMS guidelines, and inappropriately pooled heterogeneous model performances in meta-analysis. The new review identified six additional eligible studies missed by earlier work and excluded two studies that did not belong, one examining a general pregnant population and another evaluating postpartum glucose screening attendance rather than diabetes onset. Because of substantial heterogeneity in follow-up durations and predictor mix, and the scarcity of externally validated models, the authors deliberately refrained from meta-analysis, consistent with expert guidance.

What does this mean for women with a history of gestational diabetes? The review’s authors argue that appropriately developed and evaluated risk prediction tools could genuinely help, by sharpening risk perception, supporting adherence to postpartum screening, and identifying high-risk women who would benefit from lifestyle interventions such as healthy diet and physical exercise, which can prevent or delay the onset of type 2 diabetes. But they are equally clear that substantial methodological improvements are required before clinical deployment can be recommended. Their prescriptions include adherence to best-practice reporting guides such as TRIPOD + AI, far more extensive external validation across diverse populations, and exploration of novel prediction modelling techniques. Most existing models were developed in high-income settings, and with essentially no externally validated models available for developing countries, the current tools cannot be generalised to those populations, underscoring the need for context-specific development. Until then, the promise of personalised post-GDM risk prediction remains exactly that: promising, but not yet ready for the clinic.

Subject of Research: Systematic review of statistical and artificial intelligence-based prognostic models for predicting type 2 diabetes after gestational diabetes

Article Title: Traditional statistics and artificial intelligence-based prognostic models for predicting type 2 diabetes mellitus after gestational diabetes: a systematic review

Article References: Yimer, N. B., Belsti, Y., Seifu, B. L., Goldstein, R., Teede, H., & Enticott, J. (2026). Traditional statistics and artificial intelligence-based prognostic models for predicting type 2 diabetes mellitus after gestational diabetes: a systematic review. Diagnostic and Prognostic Research, 10(1), Article 12. https://doi.org/10.1186/s41512-026-00229-8

Image Credits: AI Generated

DOI: 10.1186/s41512-026-00229-8

Keywords: gestational diabetes, type 2 diabetes, prediction models, systematic review, machine learning, logistic regression, PROBAST + AI, external validation, calibration, risk prediction, TRIPOD + AI, postpartum screening