Explainable machine learning versus logistic regression for diabetes risk prediction in Pima Indian women: a calibration- and decision-focused internal validation study
Journal of Prevention Diagnosis and Management of Human Diseases · 09.09.2026 · 73
DOI: 10.55529/jpdmhd.62.73.86 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Prevention Diagnosis and Management of Human Diseases, online erschienen am 09.09.2026. Verfasst von Mainazarova Elmira und 2 weiteren Autor:innen. Die Arbeit nennt 25 Literaturangaben.
Themen: Künstliche Intelligenz in der Pflege, Diabetes mellitus
Abstract (Originalsprache)
Background
Machine learning (ML) models for diabetes prediction frequently emphasise accuracy while overlooking calibration and clinical utility. This study compared logistic regression (LR) with ML approaches and laboratory-based with non-laboratory predictor sets.
Methods
The Pima Indians Diabetes Database comprising 768 women aged ≥21 years, including 268 (34.9%) with diabetes, was analysed. Five models were evaluated: glucose-only LR, non-laboratory LR, full LR, random forest and gradient boosting. Missing physiologically implausible values were imputed within training folds. Model performance was assessed using repeated 5-fold cross-validation (10 repetitions), discrimination (AUROC), calibration slope, calibration-in-the-large, Brier score and decision curve analysis. SHapley Additive exPlanations (SHAP) assessed predictor contributions. Reporting followed TRIPOD+AI guidelines.
Results
Full LR achieved an AUROC of 0.837 (95% CI 0.809–0.864), comparable to random forest (0.838; 95% CI 0.809–0.865) and gradient boosting (0.831; 95% CI 0.800–0.859). Glucose-only and non-laboratory LR achieved AUROCs of 0.790 and 0.749, respectively. LR and random forest demonstrated acceptable calibration (slopes 0.93 and 1.07), whereas gradient boosting showed overextreme predictions (0.69). All models provided greater net benefit than testing all women across thresholds of 0.10–0.60. Plasma glucose and body mass index were the strongest predictors.
Conclusions
ML approaches provided no meaningful discrimination advantage over LR. Gradient boosting required recalibration. Non-laboratory prediction may support screening decisions where laboratory access is limited. External validation in diverse populations is essential before clinical implementation.
Zitieren
APA
Elmira, M., Ainagul, A., & Nurbek, M. (2026). Explainable machine learning versus logistic regression for diabetes risk prediction in Pima Indian women: a calibration- and decision-focused internal validation study. Journal of Prevention Diagnosis and Management of Human Diseases, 73. https://doi.org/10.55529/jpdmhd.62.73.86
Vancouver
Elmira M, Ainagul A, Nurbek M. Explainable machine learning versus logistic regression for diabetes risk prediction in Pima Indian women: a calibration- and decision-focused internal validation study. Journal of Prevention Diagnosis and Management of Human Diseases. 2026:73. doi:10.55529/jpdmhd.62.73.86
BibTeX
@article{Elmira2026,
author = {Elmira, Mainazarova and Ainagul, Arzykulova and Nurbek, Mamashov},
title = {Explainable machine learning versus logistic regression for diabetes risk prediction in Pima Indian women: a calibration- and decision-focused internal validation study},
journal = {Journal of Prevention Diagnosis and Management of Human Diseases},
year = {2026},
pages = {73},
doi = {10.55529/jpdmhd.62.73.86}
}
EUVIMED