Desarrollo de un modelo predictivo para la estimación del riesgo metabólico asociado a la diabetes

Abstract

Abstract: A person can accumulate metabolic risk for years without noticing type 2 diabetes rarely warns before it settles in. Detecting that early stage through accessible tools is exactly what drove this work. Using 1,582 participants from the NHANES 2017-2018 program, evenly split between both classes, a model was built to estimate that risk using only non-invasive variables: age, waist circumference, lifestyle habits, family history, nothing that required a blood test. Five algorithms entered the comparison Logistic Regression, SVM, Random Forest, XGBoost, and CatBoost, all under identical training conditions, and the selection criterion went beyond the single best score: the F2-Score was prioritized, designed not to trade sensitivity away for specificity. XGBoost came out ahead, with a decision threshold of 0.39 that yielded a Recall of 0.7911, specificity of 0.5157, and an AUC-ROC of 0.7274. SHAP, the technique used to interpret its decisions, pointed to age and waist circumference as the most influential variables. The model works, then, as a first screening filter not a diagnosis and its next step, before any real-world use, is external validation on a population different from the one used to build it.