Machine learning detects diabetes with 94 percent accuracy from demographic data: a study on 56,000 people in 14 low-income countries
Listen to this article
Read by Anchor
A study published in Scientific Reports developed machine-learning models to classify diabetes using WHO STEPS survey data from 14 low- and middle-income countries, comprising 56,560 adults (16,357 with diabetes, 45,423 without). The aim: early detection of diabetes in settings lacking conventional diagnostic tools.
The best model, XGBoost, achieved an accuracy of 0.90, balanced accuracy of 0.87, a F1 score of 0.93, and an area under the curve (AUC) of 0.94, outperforming random forests, gradient boosting, and traditional classifiers. The confusion matrix showed identification of 2,586 diabetic cases and 8,549 non-diabetic cases with high precision. SHAP analysis revealed the five most important predictive variables: cholesterol, age, residence (urban/rural), tobacco use, and waist circumference.
The study addressed class imbalance with SMOTEENN and evaluated ten algorithms including logistic regression, decision trees, SVM, MLP, XGBoost, and AdaBoost. Feature selection of the ten most important variables was performed via recursive feature elimination (RFE).
What this means for the Gulf and the Middle EastSaudi Arabia (18.3 % diabetes prevalence among adults), Kuwait, Qatar, the United Arab Emirates and Bahrain rank among the world’s highest for type-2 diabetes prevalence. The region possesses massive national health databases (Saudi “Seha”, Emirati “MyFile”, Qatari “Nara”) but lacks deployable, automated population-screening tools for primary care. This model, trained on diverse population data, interpretable and open-access, is an ideal candidate for local adaptation: train it on your data, obtain regulatory approval, and deploy it in primary-care centres as a triage tool before costly laboratory tests.
External validation on independent data, interpretability via SHAP values, and the use of standardized WHO survey data are all factors that make this framework transferable and verifiable across different health contexts. The region that holds the world’s highest diabetes rates should be a leader in AI-driven early-detection tools, not merely a consumer of them.
The paper is open access (CC BY-NC-ND 4.0); the code and data are available. It is an explicit call to researchers in the region: take the framework, apply it to your demographic and clinical data, and evaluate it in your contexts.