Predicting and Interpreting Clinical Deterioration in the Intensive Care Unit Using Machine Learning and Explainable AI

• Vanderleicio Carvalho Leite Junior UEFS
• Matheus Giovanni Pires UEFS
• Fabiana Cristina Bertoni UEFS

Resumo

Intensive Care Units in hospitals have been the focus of significant research efforts, as patients require continuous monitoring of their physiological parameters due to their elevated risk of rapid clinical deterioration. In this study, we evaluated eighteen predictive models for clinical deterioration in ICU patients. Fifteen of these models were composed by combining three machine learning algorithms – LightGBM, Random Forest and XGBoost – with five data imputation techniques: linear interpolation, forward filling, carry forward, indicator imputation, and zero imputation. The use of data imputation techniques aims to address the challenge of missing clinical information. The remaining three were formed using each machine learning algorithm, with the inclusion of statistical metrics for each feature in the dataset. For all models, hyperparameter tuning was performed using the grid search method, and model efficiency was assessed using the G-mean metric. The results demonstrated that models based on the Random Forest algorithm achieved the best performance. Finally, we applied the SHAP technique to identify the most influential variables for predicting clinical deterioration in the best-performing models.