This study trained and tested a low-cost machine learning model to flag patients at risk of chronic kidney disease, defined as an eGFR below 60 mL/min/1.73m² (CKD-EPI 2021), using a multicenter dataset of 203,067 anonymized Latin American records. The model uses eight routine, non-invasive variables: age, sex, systolic and diastolic blood pressure, body mass index, hypertension, type 2 diabetes and diabetes duration. After oversampling to balance cases, the data were split 70/12/18 for training, validation and testing, and the Arkangel AutoML platform generated 424 candidate models, ranked first by sensitivity above 90%.
The selected model was a decision tree trained with SMOTE augmentation, reaching 90.2% sensitivity, 92.7% specificity, 89% precision and an AUC of 0.914 on the test set. SHAP analysis identified age and diabetes duration as the most influential features. The authors present it as a screening tool that needs no specialized lab tests and can sit inside electronic health records to prioritize confirmatory testing. This is a preprint and has not yet been peer reviewed.
