Abstract / Summary
Background/Objectives: To evaluate a CatBoost-based machine learning (ML) approach for detecting previously unrecognized dysglycemia at the index primary-care visit and to assess the contribution of the Finnish Diabetes Risk Score (FINDRISC) and routinely available demographic, clinical, and laboratory variables using SHAP. Methods: This retrospective cross-sectional study included 2033 adults without previously diagnosed prediabetes or diabetes. The dataset was randomly split once into a training set of 1626 participants and an internal held-out test set of 407 participants. Three CatBoost feature-set configurations were evaluated: FINDRISC Only, SHAP-selected variables Without FINDRISC, and the same variables With FINDRISC. Five-fold stratified cross-validation and hyperparameter optimization were performed within the training set; the held-out test set was not used for model fitting or tuning. Raw FINDRISC was additionally evaluated as a non-ML benchmark. Test-set evaluation included discrimination, classification metrics, calibration, and paired-bootstrap comparison of AUCs. Results: Overall, 935 (46.0%) participants had screen-detected dysglycemia. On the held-out test set, AUCs were 0.758 for Without FINDRISC, 0.835 for FINDRISC Only, 0.840 for With FINDRISC, and 0.833 for raw FINDRISC. The incremental AUC of With FINDRISC versus FINDRISC Only was 0.005 (95% paired bootstrap CI −0.012 to 0.022; p = 0.556), indicating no statistically significant improvement. Brier scores were 0.203, 0.169, and 0.166 for Without FINDRISC, FINDRISC Only, and With FINDRISC, respectively. Conclusions: FINDRISC accounted for most of the observed discrimination for prevalent, previously unrecognized dysglycemia. Adding routine clinical and laboratory variables within CatBoost produced only a small, statistically non-significant incremental improvement. These findings should be regarded as preliminary internal evidence and require prospective external validation.