Comparison of parametric and nonparametric methods for outcome prediction using longitudinal data after cardiac arrest

Predicting outcome after cardiac arrest is challenging. We previously tested group-based trajectory modeling (GBTM) for prognostication based on baseline characteristics and quantitative electroencephalographic (EEG) trajectories. Here, we describe implementation of this method in a freely available software package and test its performance against alternative options.

We included comatose patients admitted to a single center after resuscitation from cardiac arrest from April 2010 to April 2019 who underwent ≥6 h of EEG monitoring. We abstracted clinical information from our prospective registry and summarized suppression ratio in 48 hourly epochs. We tested three classes of longitudinal models: frequentist, statistically based GBTMs; non-parametric (i.e. machine learning) k-means models; and Bayesian regression. Our primary outcome of interest was discharge CPC 1-3 (vs unconsciousness or death). We compared sensitivity for detecting poor outcome at a false positive rate (FPR) <1%.

Of 1,010 included subjects, 250 (25%) were awake and alive at hospital discharge. GBTM and k-means derived trajectories, group sizes and group-specific outcomes were comparable. Conditional on an FPR < 1%, GBTMs yielded optimal sensitivity (38%) over 48 h. More sensitive methods had 2-3 % FPRs.

We explored fundamentally different tools for patient-level predictions based on longitudinal and time-invariant patient data. Of the evaluated methods, GBTM resulted in optimal sensitivity while maintaining a false positive rate <1%. The provided code and software of this method provides an easy-to-use implementation for outcome prediction based on GBTMs.