Predictive model to assess survival and neurologic outcomes in patients with out-of-hospital cardiac arrest has potential to improve clinical management in the emergency department

Study of all out-of-hospital cardiac arrest (OHCA) cases in Sweden from 2010-2020 using the Swedish CPR Registry

We had 393 candidate predictors describing the circumstances at the time of cardiac arrest, critical time intervals, patient demographics, initial presentation, spatiotemporal data, socioeconomic status, medications, and pre-arrest comorbidities.

To develop, evaluate, and test a range of predictive models, we created stratified (outcome measure) random samples of the study population

Created training set (60% of the data), evaluation set (20% of the data), and test set (20% of the data)

Thirty-day survival and cerebral performance category (CPC) scores at discharge were assessed using multiple machine learning frameworks with hyperparameter tuning

Parsimonious models with top 1 to 20 strongest predictors were tested

Decision thresholds were calibrated to assess survival sensitivity at a 95% cutoff

The final model is deployed as a web application

The imbalance in SRCR is evident as approximately 10% of all patients survive

The imbalance is addressed by downsampling the number of deaths, e.g. for one survivor we include four deaths, thus reducing the class imbalance

Only the training data set is artificially balanced

55,615 OHCA cases included

Initial presentation, prehospital intervention, and critical time interval variables were most important

In addition, the calibration curve and DCA results are also relatively good

The test data for predicting 30-day survival had a sensitivity of 95%, a specificity of 89%, a positive predictive value of 52%, and a negative predictive value of 99%

AUC of 0.97 in test data using all 393 predictors or only the 10 most important predictors

The final model shows excellent calibration

This web app allows for near-instantaneous survival calculations

30-day survival and neurologic outcomes of OHCA during ongoing cardiopulmonary resuscitation in the emergency department can be quickly and reliably estimated using a machine learning model that incorporates a wide range of available variables

Baseline characteristics of 55,615 patients with out-of-hospital cardiac arrest

Outcomes for 55,615 patients with out-of-hospital cardiac arrest

(a): Relative variable importance plot: top 50 predictor variables ranked by relative importance for 30-day survival. N=55615

(b): Relative variable importance plot: 39 predictor variable categories ordered by relative importance for 30-day survival. N=55615

AUC-ROC with corresponding 95% confidence intervals for predicting 30-day survival for 20 models, including top 1 to top 20 survival predictors

ROC curve predicting 30-day survival based on (a) evaluation data and (b) test data

Calibration plot illustrating the agreement between predicted and observed survival probabilities with 95% confidence intervals

Description of the Swedish EMS system and hospital system

Grid search hyperparameter tuning

Tuning the XGB model using a grid of 144 different hyperparameter combinations

Missing top 20 predictors in final model (non-imputed data)

Baseline characteristics of 55,615 patients with out-of-hospital cardiac arrest stratified by 30-day survival

Adjust decision thresholds (cutoffs) to evaluate data sets

Model performance on test data

Use baseline data from the input dataset

Research flow chart

Model evaluation of training data and ROC curve of CPC-score test data

Figure S3 shows the model’s evaluation on the training data, and its performance on the CPC-score test data (roc curve)

SCARS-1 web application with 3 patients

Crude ROSC rate, hospitalization rate, and 30-day survival rate

Description of software used