How much information is needed to predict outcomes after cardiac arrest?

Abstract

Introduction

Prognostication after cardiac arrest is challenging but may be improved with machine learning (ML). ML accommodates large quantities of data, but in practice these arise from heterogeneous sources that may be challenging to assemble. We compared ML performance with combinations of registry, electronic health record (EHR) and electroencephalography (EEG) data to test if only a subset of sources was sufficient.

Methods

We performed a cohort study including consecutive adults treated between January 2010 and February 2022 at a single hospital who were unresponsive after cardiac arrest. We developed ML models to predict poor outcome (discharge Cerebral Performance Category (CPC) of 4 or 5) from various combinations of registry, EHR and EEG data. We developed sequential models at presentation and 12-, 24-, 48- and 72-hours post-arrest, including only patients remaining hospitalized and information known at that timepoint. Our primary performance metric was sensitivity predicting poor outcome at perfect specificity (zero false positives).

Results

We included 1,106 patients of whom 773 (70 %) had poor outcome. Best performing models were random forests. At each timepoint, the best performing model included both registry and EEG data; after 12 h the best models used a combination of registry, EHR and EEG data. Peak median sensitivity at perfect specificity was 70 % (65–73 %) and occurred at 24 h. Discrimination of this model was excellent (median AUC 0.949 [0.947–0.951]).

Conclusion

Multiple data sources were needed to achieve optimal sensitivity. There is a need to develop large, comprehensive, multicenter datasets to improve post-arrest prognostication.