Evaluate the value of MRI DWI and FLAIR sequences in predicting outcomes among comatose post-cardiac arrest patients.  

Neuroimaging can aid early prognostication in comatose post-arrest patients. We investigated how abnormalities on MRI DWI and FLAIR sequences at different time points correlate with outcomes and contribute to predictive modeling.  

We conducted a retrospective analysis of adult patients (2011-2021) initially comatose after arrest who underwent MRI between 2-7 days following arrest. Neuroimaging was independently reviewed by two board-certified neurointensivists, blinded to clinical data, for abnormalities in 12 individual regions on DWI and FLAIR sequences (frontal cortex, parietal cortex, temporal cortex, occipital cortex, insula, hippocampus, caudate, putamen, globus pallidus, thalamus, brainstem, cerebellum). Composite scores were calculated for cortex and deep gray nuclei. Univariate logistic regression assessed associations between imaging features and mortality. We trained three machine-learning models using demographics, clinical exam findings, EEG features, and MRI scores to predict mortality.  

Among 222 patients, 69% had out-of-hospital arrests and 31% survived to discharge. Median age was 60 years (IQR 49-69) and 59.5% were male. Brainstem hyperintensity on DWI and FLAIR (p = 0.01, 0.001 respectively) and DWI abnormalities in the thalamus (p < 0.001) were most strongly associated with mortality. Higher overall DWI scores, cortical and deep gray nuclei DWI scores, and the combined total FLAIR and DWI score all showed strong associations with mortality (p < 0.001). The Gradient Boosting Machine model demonstrated the highest performance (sensitivity 0.91, specificity 0.86, AUC 0.82). Shapley analysis identified thalamic diffusion abnormalities, Charlson Comorbidity Index, and shockable rhythm as the most impactful features in outcome prediction.  

MRI DWI and FLAIR sequences are important prognostic features for comatose post-arrest patients, both individually and within multivariate machine learning models. We seek to further assess how these imaging findings can be used effectively and reliably in end-of-life decision making.