Predicting Chronic Critical Illness in Bone Trauma Patients: An AI-Based Approach for ICU Healthcare Providers.

Abstract

Background: Chronic critical illness (CCI) is a serious condition characterized by a prolonged course of illness, resulting in elevated morbidity and mortality. CCI presents significant challenges for healthcare providers in intensive care units (ICUs), particularly among patients with bone trauma. Accurate prediction of CCI in this patient population is essential for effective management and intervention. This study aims to develop a web-based artificial intelligence (AI) application designed to predict CCI in ICU patients suffering from bone trauma.

Methods: A cohort of 1049 patients were included in the study, with 775 patients from the Medical Information Mart for Intensive Care III (MIMIC-III) database and 274 patients from two tertiary hospitals. Five machine learning techniques and logistic regression were employed to develop the models, using 80% of the MIMIC-III cohort. The models' internal effectiveness was evaluated using the remaining 20% of the cohort, and external validation was performed on the 274 prospective patients. Eleven evaluation metrics were used to develop a scoring system for comprehensive performance evaluation.

Results: Among all the models evaluated, the eXGBoosting Machine (eXGBM) model demonstrated the highest performance in internal validation, with an area under the curve (AUC) value of 0.979 (95%CI: 0.970-0.991). It outperformed the Random Forest (RF) model, which had an AUC of 0.957 (95%CI: 0.941-0.967), and the Support Vector Machine (SVM) model, which achieved an AUC of 0.911 (95%CI: 0.878-0.928). The Logistic Regression (LR) model had a relatively lower AUC of 0.753 (95%CI: 0.714-0.793). In terms of various evaluation metrics, including accuracy (0.925), precision (0.906), recall (0.947), specificity (0.902), F1 score (0.926), Brier score (0.056), and Log loss (0.197), the eXGBM model consistently outperformed the other models. Additionally, based on the scoring system, the eXGBM model achieved the highest prediction score of 60, followed by the RF model with a score of 52 and the K-Nearest Neighbor (KNN) model with a score of 39. External validation of the eXGBM model resulted in an AUC of 0.887 (95%CI: 0.863-0.917), confirming its robust performance and generalizability. A user-friendly web-based AI application based on the eXGBM model was successfully developed and was freely accessible at the Internet.

Conclusions: The development of a web-based AI application utilizing the eXGBM model demonstrates a promising advancement in the prediction of CCI among ICU patients. With favorable performance in both internal and external validation, the AI application not only achieved high accuracy and reliability but also provided a user-friendly tool for clinicians. This application has the potential to enhance patient management and care by facilitating timely interventions for at-risk patients. Future research should focus on further refining the model and exploring its integration into clinical practice to improve outcomes in this patient population.