Factors relevant for the outcome and for the prediction of cardiac arrest

Abstract

Even when treated, 30-day survival after cardiac arrest is low for in-hospital cardiac arrest (IHCA) (15-35%) and out-of hospital cardiac arrest (OHCA) (8%). In this thesis we aimed to investigate factors that affect (I) and predict the outcome of cardiac arrest (II), as these are important for treatment and prognostication. Another aim was to investigate the efficacy of using machine learning (ML) models to predict cardiac arrest in high-risk patients (III & IV). Paper I included patients registered for IHCA in the Swedish Cardiopulmonary Resuscitation Registry (SRCR) between 2017 and 2020. Using multivariate logistic regression, we found that the pre-arrest sign arrhythmia was associated with a higher likelihood of 30-day survival after cardiac arrest. However, hypoxia and hypotension were associated with lower likelihood of 30- day survival. Paper II included comatose survivors of cardiac arrest examined with Somatosensory Evoked Potentials (SSEP) and Neurological pupil index (NPi) >48 hours after cardiac arrest. We found that NPi <3.4 was predictive of pathological result from SSEP, which is a robust modality used to predict unfavourable outcome after cardiac arrest. Paper III included patients with newly diagnosed heart failure from the Swedish National Heart Failure Registry from 2005-2021. A Random Survival Forest ML model was developed to predict cardiac arrest registered in SRCR but performed poorly. Paper IV included survivors of IHCA from SRCR from 2010-2021. A XGBoost model had a moderate capacity to predict recurrent cardiac arrest or death within one-year of IHCA. Specific resuscitation treatment is necessary for cardiac arrests preceded by hypoxia or hypoten sion to improve survival. If validated, the less resource-demanding NPi could be used instead of SSEP in smaller hospitals where SSEP is unavailable. Two different designs to machine learning prediction model resulted in different performances. Carefully designed models should be tested and calibrated in validation studies before clinical use.