Development of a Reproducible Risk Score Based on Machine Learning for Cervical Cancer Triage in Latin American Primary Care

Resumen

Every year more than 75,000 women in Latin America are diagnosed with cervical cancer, and thousands die because the diagnosis arrives too late. The cause is not scientific but structural: healthcare systems in the region lack simple tools that allow identification of women at highest risk from the very first medical consultation. This work develops such an instrument through a data engineering pipeline that included winsorization without record loss, class weighting according to real prevalence, and stratified cross-validation, where each technical decision was justified by the clinical context. Three machine learning algorithms Logistic Regression, RF, and XGBoost competed under identical and reproducible conditions; XGBoost was selected for its greater robustness. That model was transformed into a scorecard consisting of 6 routine clinical questions and a maximum of 15 points, completable in less than two minutes without additional technology. The instrument operates under two adaptable scenarios: Scenario A, with 83% specificity to reduce unnecessary referrals where resources are scarce, and Scenario B, with 73% sensitivity and 53% fewer missed cases for mass screening campaigns. Applied at a regional scale, it could generate more than 31,000 additional diagnoses annually. The pipeline is open and replicable in any Latin American hospital using its own data.