Artificial intelligence-based prediction of diseases among homeless populations in Bogotá: Implications for targeted interventions

Abstract

Background The homeless population in Bogotá exhibits complex social and health vulnerabilities, with a high prevalence of chronic and communicable diseases such as hypertension, diabetes, tuberculosis, HIV/AIDS, and cancer. These conditions hinder epidemiological surveillance and timely public health decision-making, particularly in contexts of social exclusion. Objective To develop and evaluate an artificial intelligence-based predictive model aimed at identifying disease occurrence risk profiles among the homeless population in Bogotá, in order to support public health decision-making. Methods Data from the 2024 Bogotá Homeless Census were analyzed, comprising 10 478 records and 46 demographic, clinical, and socioeconomic variables. Data processing and model development followed the CRISP-DM methodology. The extreme gradient boosting (XGBoost) algorithm was implemented to predict disease occurrence. Model performance was evaluated in terms of accuracy, sensitivity, and the F1-score, while interpretability was assessed through SHAP (SHapley Additive exPlanations) values. Additionally, metrics related to social trust, system response time, and potential health impacts were examined. Results The model achieved an accuracy of 0.91, a sensitivity of 0.57, and an F1-score of 0.70, striking an adequate balance between identifying high-risk individuals and reducing false positives. SHAP analysis identified hypertension, diabetes, and HIV/AIDS as the main predictors of classification. However, complementary metrics revealed limitations in social trust (TAS = 0.49), system response time (SRT = 24.86 hours), and potential health impact (PHIS = 0.24). Conclusions and implications These findings suggest that explainable artificial intelligence (XAI) models may support public health surveillance and intervention prioritization in homeless populations by identifying epidemiological risk profiles. However, the models’ applicability remains context-specific and requires external validation before implementation across other vulnerable populations or healthcare settings.