Deep learning model utilization for mortality prediction in mechanically ventilated ICU patients

Abstract

Background:: The requirement for mechanical ventilation has increased in recent years. Patients in the intensive care unit (ICU) who undergo mechanical ventilation often experience serious illness, contributing to a high risk of mortality. Predicting mortality for mechanically ventilated ICU patients helps physicians implement targeted treatments to mitigate risk. Methods:: We extracted medical information of patients with invasive mechanical ventilation during ICU admission from the Medical Information Mart for Intensive Care III (MIMIC-III) dataset. This information includes demographics, disease severity, diagnosis, and laboratory test results. Patients who met the inclusion criteria were randomly divided into the training set (n = 11,549, 70%), the test set (n = 2,475, 15%), and the validation set (n = 2,475, 15%). The Synthetic Minority Over-sampling Technique (SMOTE) was utilized to resolve the imbalanced dataset. After literature research, clinical expertise and an ablation study, we selected 12 variables which is fewer than the 66 features in the best existing literature. We proposed a deep learning model to predict the ICU mortality of mechanically ventilated patients, and established 7 baseline machine learning (ML) models for comparison, including K-nearest Neighbors (KNN), Logistic Regression, Decision Tree, Random Forest, Bagging, XGBoost, and Support Vector Machine (SVM). Area under the Receiver Operating Characteristic Curve (AUROC) was used as an evaluation metric for model performance. Results:: Using 16,499 mechanically ventilated patients from the MIMIC-III database, the Neural Network model outperformed existing literature by 7.06%. It achieved an AUROC score of 0.879 (95% Confidence Interval (CI) [0.861-0.896]), an accuracy of 0.859 on the test set, and was well-calibrated with a Brier score of 0.0974, significantly exceeding previous best results. Conclusions:: The proposed model demonstrated an exceptional ability to predict ICU mortality among mechanically ventilated patients. The SHAP analysis showed respiratory failure is a significant indicator of mortality prediction compared to other related respiratory dysfunction diseases. We also incorporated mechanical ventilation duration variable for the first time in our prediction model. We observed that patients with higher mortality rates tended to have longer mechanical ventilation times. This highlights the model’s potential in guiding clinical decisions by indicating that longer mechanical ventilation may not necessarily enhance patient survival chances.

Keywords

• Mechanical ventilation
• ICU mortality
• Deep learning
• Machine learning