Inteligencia artificial en cirugía: creación y validación de una calculadora colombiana de riesgo de mortalidad perioperatoria

Introduction. An estimated 4.2 million deaths occur annually in the first 30 postoperative days. The Lancet Commission on Global Surgery highlights the importance of measuring and reducing this mortality. This study developed a perioperative mortality calculator specific to the Colombian population, aiming to identify and intervene early in patients at high risk.

Methods. We used data from the multicenter ColSOS study, which included 3807 patients from 54 centers in Colombia. Clinical, sociodemographic and perioperative variables were collected; missing data were handled with multiple imputation. Variables were selected by bivariate analysis, Lasso regression and Recursive Feature Elimination (RFE). Predictive models were compared using logistic regression and XGBoost, evaluating their performance with cross-validation.

Results. The XGBoost model was selected because it showed better sensitivity and fewer false negatives than logistic regression. The importance in predicting ASA classification, chronic obstructive pulmonary disease, hemodynamic instability and urgency of the procedure was highlighted. The model predicted mortality with an area under the curve (AUC) of 0.87.

Conclusion. The present study has developed the first perioperative mortality calculator designed for the Colombian population, including multiple surgical specialties. The selected machine learning model presents a sensitivity and specificity that make it comparable to the best international tools. The implementation of this tool allows early identification and management of patients at risk, which could improve surgical care in Colombia.