Answers are grounded in this paper and cite their sources.
Background: Machine learning offers opportunities to improve risk stratification in type 2 diabetes by capturing complex clinical and sociodemographic interactions. Aim: To develop and evaluate machine learning models for predicting 1-year glycaemic control in patients with type 2 diabetes in Burkina Faso. Setting: A tertiary hospital in Burkina Faso, including patients followed up between 2020 and 2025. Methods: A retrospective cohort study included 424 patients. Glycaemic control was defined as glycated haemoglobin (HbA1c) < 7% at 1 year. Nine classifiers were trained using nested cross-validation. Performance was assessed using AUC-ROC, F1-score, recall, accuracy, precision and AUC-PR. Interpretability was assessed using SHAP (SHapley Additive exPlanations) with consensus feature importance ranking. Results: Overall, 42.9% of patients achieved glycaemic control at 1 year. Performance was moderate, with AUC-ROC ranging from 0.633 to 0.676 and no significant pairwise differences. AdaBoost performed best (AUC-ROC: 0.676). SHAP consistently identified baseline HbA1c, diabetes duration, total cholesterol, hypertension history and oral antidiabetic therapy as the most influential predictors. Conclusion: Machine learning models showed moderate, comparable performance, suggesting a ceiling imposed by routinely available data. Baseline HbA1c and diabetes duration were dominant predictors. Future research should prioritise external validation and longitudinal data integration. Contribution: This study provides evidence supporting machine learning for identifying key predictors of glycaemic control in West Africa, offering clinically actionable insights and highlighting the need for context-adapted decision-support tools in resource-limited settings.
No takes yet. Share an insight, caveat, or question.
Ouattara et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: