Machine Learning Models for Predicting the Length of ICU Stay Using Perioperative Patterns

Abstract

Perioperative medicine encompasses the care of surgical patients from the preoperative to the postoperative phase, dealing with unpredictable complications that can be fatal and require monitoring in the Intensive Care Unit (ICU). Artificial intelligence (AI) presents itself as a crucial tool for predicting postoperative complications, improving outcomes, and reducing healthcare costs. This project employs the INSPIRE database to develop AI models for predicting ICU stay length for abdominal surgeries. The CRISP-DM methodology was applied for the selection, preprocessing, and analysis of patient data using machine learning. The models were evaluated using confusion matrices, precision, sensitivity, specificity, F1-score, the area under the curve (AUC), and ROC curve graphs. It was observed that the Random Forest and KNN models showed the best overall performance, standing out for their high precision and balance in classifying the duration of ICU stay. The analysis of feature importance using SHAP revealed that the most influential intraoperative parameters on the length of stay were estimated blood loss (ebl), heart rate (hr), inspired oxygen fraction (fio2), and positive end-expiratory pressure (peep). Machine learning models, especially Random Forest, proved to be an effective tool for predicting the length of ICU stay. This approach can significantly improve hospital resource management and patient care planning, enabling the early identification of those at risk for prolonged ICU stays and facilitating the implementation of preventive measures to improve clinical outcomes and reduce associated health care costs. © 2024 IEEE.