Machine Learning-Based Mortality Prediction for Mechanically Ventilated Liver Failure Patients: A Retrospective Study

To develop a mortality prediction model for patients with liver failure who require mechanical ventilation (MV) in ICU.

This retrospective study utilized data from the MIMIC-III and MIMIC-IV (v3.0) databases. Variables included demographics, liver function markers, ventilator parameters, severity scores, comorbidities, use of vasoactive drugs and continuous renal replacement therapy (CRRT) as well as length of ICU and hospital stay. The primary outcome was in-hospital mortality. Five machine learning algorithms were applied: logistic regression, decision tree, Bagging, neural network, and extreme gradient boosting (XGBoost). The dataset was randomly split into a training set (70%) and a test set (30%). Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) and calibration plots. Predictive factors were ranked for the best-performing model by importance.

A total of 2336 patients were identified from the MIMIC-III and MIMIC-IV databases. On the test set, the AUC values for each model were as follows: logistic regression (0.836), decision tree (0.768), neural network (0.774), Bagging (0.854), and XGBoost (0.875). Calibration curves showed good agreement for all models except the neural network. Among the 5 algorithms, XGBoost demonstrated the highest predictive performance. The top 5 predictors identified were length of hospital stay, SAPS II score, use of vasoactive drugs, total bilirubin and length of ICU stay.

Machine learning-based models can effectively predict in-hospital mortality in liver failure patients receiving mechanical ventilation. Among the 5 algorithms tested, the XGBoost model achieved the best overall performance.