An extended Cox regression model showed moderate ability to distinguish which adults with normal baseline fasting plasma glucose would have recorded prediabetes at three and four years. In a separate hospital-based external cohort, its area under the curve (AUC), a measure of how well predictions separate higher-risk from lower-risk participants, was 0.750 at three years and 0.764 at four years. The model was chosen after a random-forest approach performed best in training but declined in internal validation and the external cohort.
The study set out to develop, compare and externally evaluate machine-learning and Cox-regression models for predicting FPG-defined incident prediabetes. It focused on adults with baseline FPG below 5.6 mmol/L. Incident prediabetes meant a recorded follow-up FPG of 5.6 to less than 7.0 mmol/L, with follow-up ending at detection or the last event-free visit.
A model tested across two cohorts
In the development cohort, 10,592 participants were followed for a mean of 3.05 years. During that period, 1,287, or 12.2%, developed the study's FPG-defined prediabetes. The separate hospital cohort contained 671 participants; 75, or 11.2%, developed the outcome over a mean follow-up of 3.02 years.
To build the candidate models, the researchers used LASSO, a variable-selection method, with 10-fold cross-validation of 17 candidate variables. The procedure retained 11 predictors with nonzero coefficients. Six machine-learning models and an initial proportional-hazards Cox model were compared using discrimination and threshold-based measures. The development cohort was randomly split into a 65% training set and a 35% internal validation set, with the separate hospital cohort used for external evaluation.
Why the Cox model was extended
Random forest had the highest apparent performance in training, but declined in the internal validation set and external cohort. The initial Cox model showed more consistent discrimination across the evaluation datasets. The researchers then tested whether the Cox framework needed to account for a changing FPG association over time.
That testing identified nonproportionality for FPG, meaning its association with the recorded outcome did not stay constant through follow-up. The final model, M2, included an interaction between baseline FPG and log time and had the lowest Akaike information criterion, or AIC, among the alternatives. It was selected over more complex models.
At the three-year reference time, each 1-mmol/L higher baseline FPG was associated with a hazard ratio of 6.33, a measure comparing event rates over time; the 95% confidence interval was 5.23 to 7.67. The adjusted hazard ratio declined from 11.60 at two years to 2.95 at five years, indicating that the association weakened with time. Because these were observational data, the result is an association, not evidence that higher FPG caused prediabetes.
Useful prediction, but not a clinical verdict
At three years, M2's AUC was 0.800 in training, 0.778 in internal validation and 0.750 externally. At four years, the figures were 0.796, 0.755 and 0.764. Calibration, the agreement between predicted and observed risk, was generally favorable, and decision-curve analysis showed greater net benefit than treat-all or treat-none strategies at the primary horizons.
Those results are useful as measures of prediction, but they do not establish clinical effectiveness. At five years, AUCs were 0.761 in training, 0.739 in internal validation and 0.667 in the external cohort, where performance was considered less stable and exploratory.
The external hospital cohort also contributed to model comparison and selection, so it was not reserved as a fully independent validation set. That matters when interpreting the external figures: they test performance in a different cohort, but not in a setting untouched by the model-building decisions.
The outcome itself was narrower than the everyday word prediabetes suggests. The study counted recorded follow-up FPG between 5.6 and less than 7.0 mmol/L after normal baseline glucose, so the findings should not be read as covering every possible prediabetes definition or the biological moment when the condition began.
Before the model could be used in practice, the authors call for independent multicenter validation, recalibration where needed and prospective evaluation of clinical impact. The paper presents M2 as a potentially practical, interpretable framework for three- and four-year risk estimation, not proof that a calculator prevents diabetes or improves patient care.
Paper data and sources
Original title: Development and comparative evaluation of machine learning algorithms and Cox regression for predicting fasting plasma glucose-defined incident prediabetes: a longitudinal cohort study
Authors: Hu S, Yan C, Tian X et al.
Journal/Repository: Frontiers in endocrinology
Status: Peer-reviewed
First online: 2026-10-09
DOI: Not available
Original paper · Full text