Transparência em modelos de comitê: aplicação de inteligência artificial explicável em dados do SUS

Abstract

Public health increasingly relies on the responsible use of data to support hospital management. In this context, ensemble machine learning models achieve strong predictive performance but often operate as "black boxes," hindering their adoption in sensitive domains. This study investigates the application of Explainable Artificial Intelligence (XAI) techniques — specifically SHAP and LIME — to tree-based ensemble models (Random Forest, Gradient Boosting, and XGBoost). These models were applied to microdata from the Brazilian Unified Health System (SUS) Hospital Information System regarding hospital admissions in the state of São Paulo between 2024 and 2025. Models were developed to classify hospital length-of-stay ranges and to predict the number of days of hospitalization (regression) — evaluated using metrics appropriate for each task — as well as a model to classify cost ranges. The results indicate that the primary diagnosis, procedure, hospital type, municipality of residence, and secondary diagnosis are the most influential factors in the predictions, showing strong agreement between global (SHAP) and local (LIME) analyses. The study demonstrates how open government datasets can be transparently leveraged through modern computational methods, valuing both model performance and an understanding of the factors influencing their predictions.