Identifying important clinical features for post-cardiac arrest neurological prognostication using clinical BERT

Cardiac arrest is a leading cause of mortality, resulting in severe brain injury and coma. Predicting neurological outcomes in post-cardiac arrest patients is complex and typically requires extensive observation periods of 72 hours or more, which complicates decision-making for healthcare providers regarding early, targeted interventions in the critical care setting. While recent deep learning-based methods show promise for early neurological prognostication, their black-box nature limits clinical adoption and interpretability. In this work, we introduce a systematic feature importance analysis framework using leave-one-feature-out (LOFO) methodology to identify the most critical clinical predictors for neurological outcome prediction in post-cardiac arrest patients. By removing each clinical variable and measuring the resulting impact on model performance, our approach quantifies the relative contribution of individual features through a sigmoid-based feature importance scoring system. To validate the applicability of this interpretable AI framework, we performed a comprehensive analysis of 12 key clinical variables for post-cardiac arrest patient prognosis. Using ClinicalBERT-based classifier, we measured the importance of these features by measuring performance degradation when individual variables were excluded. Our analysis shows that neuron-specific enolase at 48 hours demonstrated the highest feature importance score for neurological outcome prediction, followed by initial cardiac rhythm and time to return of spontaneous circulation as the most critical predictors. These three features accounted for approximately 75% of the total predictive capacity. Conversely, demographic factors, including age and gender, showed minimal contributions to model performance, with importance scores significantly lower in comparison to other variables. With the help of this feature importance framework we aim to provide clinicians with interpretable insights into AI-driven prognostication, potentially enabling more informed clinical decision-making and earlier therapeutic interventions for cardiac arrest survivors.