Sean McManus
Anesthesiology Resident Physician
Cleveland Clinic Foundation
Disclosure information not submitted.
Reem Almuqati, MD
Critical Care Fellow
The Cleveland Clinic, United States
Disclosure information not submitted.
Reem Khatib, MD
Anesthesiologist & Intensivist
Cleveland Clinic Foundation, United States
Disclosure information not submitted.
Ashish Khanna, MD, FCCP, FASA,FCCM
Anesthesiologist & Intensivist, Associate Professor of Anesthesiology
Wake Forest Baptist Medical Center
Winston Salem, North Carolina
Disclosure information not submitted.
Jacek Cywinski, MD
Anesthesiologist
Cleveland Clinic Foundation, United States
Disclosure information not submitted.
Francis Papay, MD
Plastic Surgeon
Cleveland Clinic Foundation, United States
Disclosure information not submitted.
Piyush Mathur, MD, FCCM
Anesthesiologist & Intensivist
The Cleveland Clinic
Cleveland, Ohio, United States
Disclosure information not submitted.
Title: Machine Learning Based Early Mortality Prediction at the Time of ICU Admission
Introduction: Triage and appropriateness of care decisions are important for critical care clinicians at the time of ICU admission. Prediction scores such as APACHE (Acute Physiologic Assessment and Chronic Health Evaluation) have good mortality prediction performance, but lack practical application. Thus, there is a need for a dynamic model which can be used upon patient admission to the ICU for prediction of early mortality using readily available data.
Methods: We used the multicenter e-ICU collaborative dataset, containing 139,367 unique patient records from ICU admissions 2014-2015 across 208 hospitals in the United States. We excluded patients with mortality within 4 hours of ICU admission, ICU readmissions and those with missing data. We used APACHE data variables including mean arterial pressure (MAP), heart rate (HR), Glasgow Coma Scale (GCS) variables (verbal, eyes, motor, intubated), presence of mechanical ventilation, first documented SpO2 upon admission, ICU mortality and discharge status. After splitting the dataset into training (80%) and testing (20%) sets, we created a Support Vector Machine (SVM) classifier model to predict mortality with balanced class weight. SVM performs the classification tasks with a high degree of accuracy by finding the hyper-plane that differentiates the two classes. Areas under receiver-operator curve (AUROC), F1 score, Precision and Recall were measured for accuracy of predictions on the test set. Feature importance was used to understand the relative weights of physiologic variables in prediction of mortality.
Results: In our cohort of 97,382 patients, 93,975 survived the ICU admission, and 3,407 patients died within the first 48 hours of admission to the ICU. Our model performance achieved an AUROC of 0.81 and F1 score of 0.89 on the test dataset. Precision for the predicted class was 0.14 with recall of 0.80. Feature importance demonstrated higher prediction weights of HR, MAP, respiratory rate, and SpO2 followed by GCS components (motor response with highest predictive weight).
Conclusions: Patients with prediction of early mortality using this model, represent the failure to rescue population despite intensive treatment early in the ICU admission. Predictions from this model may help guide appropriate resource utilization and earlier palliative care interventions.