Early Hospital Mortality Prediction Using Routine Vital Signs in ICU Patients

Abstract

In a clinical setting, there are countless scenarios in which a statistical prognosis for patients can be extremely beneficial to medical professionals so that they may better allocate resources to provide the best patient care. The purpose of this paper is to identify when in a patient’s stay a meaningful prediction of hospital mortality can be made to provide that prognosis. In order to accomplish this, eight clinical variables were extracted from the MIMIC-III database for ICU patients and were supplied to a XGBoost model, an advanced Decision Tree Classifier that employs gradient boosting. Because of the imbalanced data, the positive values were weighted more heavily along with other optimized parameter values found from the use of GridSearchCV. A static model demonstrated an average accuracy of 80.50% with an AUC-ROC of 0.800 and an AUC-PR of 0.429. However, a time-series analysis using extracted statistics from twelve-hours of compounded, time-varying data generated a model with an 83.28% accuracy with an AUC-ROC of 0.846 and an AUC-PR of 0.562. Additionally, the model demonstrated the importance of GCS and airway management in the prediction of mortality indicating the need to focus more on these vitals in emergency situations. The time-series model was shown to be most effective in predicting mortality, exemplifying the importance of providing time-series data that can detail the progress/decline of the patient. This implementation especially could be very impactful in clinical settings to provide healthcare professionals with the means to make quick and effective decisions.