Comparative performance of machine learning and conventional scoring systems for neuroprognostication in out-of-hospital cardiac arrest survivors.

Background: Accurately predicting the neurological outcomes in out-of-hospital cardiac arrest (OHCA) survivors is crucial. Conventional prediction scores should be validated across different settings. Additionally, machine learning (ML) models may provide improved predictive performance. This study aimed to externally validate these conventional scores and compare their performance with the developed ML model.

Methods: This multicenter retrospective cohort study included non-traumatic, adult OHCA survivors from the National Taiwan University Hospital Hsinchu and Yunlin Branch Out-of-Hospital Cardiac Arrest Research Database between January 2016 and July 2021. The primary outcome was neurological outcomes at discharge. We externally validated four conventional prediction scores-OHCA, SWAP, PROLOGUE and CAHP-by performing a receiver-operating characteristic (ROC) analysis and using calibration plots. A new ML model was developed using the XGBoost algorithm. The area under the ROC (AUROC) of the conventional scores and ML model was compared using DeLong's test.

Results: This study included 1253 non-traumatic adult OHCA survivors, of whom 279 (22.3 %) achieved a favorable neurological outcome. Among the conventional scores, PROLOGUE had the highest AUROC of 0.89, followed by CAHP, OHCA and SWAP (0.84, 0.78, and 0.78, respectively) on the complete dataset. The ML model outperformed the conventional scores, with AUROC slightly higher than the best performing conventional score PROLOGUE (0.94 vs. 0.88, p = 0.032).

Conclusions: PROLOGUE was the best performing conventional score. CAHP, although slightly less accurate, used simpler variables. The ML model outperformed the conventional scores, but its advantage over PROLOGUE was not substantial. Clinicians should select predictive tools based on the clinical context.