Explainable Artificial Intelligence (XAI) for Survival Prediction after Cardiac Arrest and Intervention Analysis

Keywords: Explainable Artificial Intelligence, Cardiac Arrest, Survival
Abstract Type: Paper Abstract

Abstract

Background: More than 350,000 cardiac arrests (CAs) occur in the United States each year of which only 6% survive. This study aims to identify and quantify key factors affecting the survival of CAs to alleviate the disease burden.

Methods: A total of 57223 patients treated by Georgia Emergency Medical System (EMS) included in Georgia Department of Public Health are utilized in this study from 2019 to 2021. To predict survival at the end of EMS event using 15 incident variables, Machine Learning classification models, including Logistic Regression, Gradient Boosting, eXtreme Gradient Boosting, Support Vector Machine and Neural Network are compared with global and local agnostic explainable tools. Using the model with highest predictability, the intervention analyses are conducted regarding 4 variables, including EMS response time, use Automated External Defibrillator (AED) or not, use CPR (Cardiopulmonary resuscitation) or not, AED used by layperson or EMS. Response time is intervened by using the time spent from nearest ambulance station to the incident location.

Results (Expected): Neural network shows the highest performance. Arrest witnessed or not, whom the arrest is witnessed by, EMS response time, use AED or not, use CPR or not, who use AED have certain effects on survival. Intervention analysis of patient outcome shows that additional patients would have survived if shorter response time, timely AED use, more professional AED manipulation are applied to them.

Conclusion (Expected): Planning EMS ambulance stations and AED sites more reasonably and promote the AED training program can help to improve CA survival.