Skip to Content

Machine Learning-Based Prediction of Cardiac Arrest Outcome Using a Large Multi-Center Database

Team Members:
  • Tatiana Gelaf Romer
  • Qingchu Jin
  • Hanbiehn Kim
  • Hieu Nguyen
  • Sharmila Tamby
  • Eric Sung
  • Robert Stevens, MD
  • Jose Suarez, MD
  • Christian Storm
  • Raimond Winslow, PhD
  • Joseph Greenstein, PhD
  • Ran Liu


Cardiac arrest (CA) is a leading cause of death and poses a significant risk of long-term neurological disability due to hypoxic-ischemic encephalopathy. There is a large unmet need for accurate and reliable methods to predict post-CA neurological outcomes and treatment responses. Our goal is to predict the probability of neurological recovery using a database of patients admitted to 208 hospitals in the US. Aims are twofold: (1) to compare the predictive performance of models using features extracted from electronic health records (EHR), from physiological time series (PTS), and models combining EHR and PTS features; and (2) to contrast predictive performance obtained with different statistical and machine learning models.

We analyzed data collected in the first 24 hours following ICU admission. Outcome was the motor subscore of the Glasgow Coma Scale (GCS) at the time of discharge from the ICU (good: six, poor: one through five). From 240,000 ICU admissions, 2,216 CA patients were selected based on the following inclusion criteria: alive for > 24 hours, mechanical ventilation, and motor GCS recorded within 24 hours of ICU discharge. Missing data was imputed using population mean and linear interpolation. 452 predictive features were identified based on prior knowledge of variables perceived as impactful on CA outcome. Feature selection was performed using random forest and LASSO regularization. We evaluated the predictive performance of a generalized linear model (GLM) and several machine learning models, including random forest, support vector machine, gradient boosting, and neural networks.

Our best performing model was GLM that achieves an average 0.846 AUC, 0.798 sensitivity, and 0.742 specificity. This model had higher discrimination than models trained with APACHE IV variables (AUROC ~0.70). The top ranked features selected using LASSO regularization include GCS score upon admission; worst motor and eye GCS score within 24 hours; maximum lactate, albumin, MCHC, and alkaline phosphate; and mean respiratory rate.

These findings indicate that a machine learning model applied to a large clinical dataset can predict post-CA outcome with a level of accuracy that surpasses APACHE, the current standard for ICU prediction. Results also indicate PTS features boost model performance. In addition, several previously overlooked predictive features were identified which merit further investigation.

Read the Johns Hopkins University privacy statement here.