Multimodal contrastive prognostication framework for early neurological outcome prediction in post-cardiac arrest patients

Sudden cardiac arrest (SCA) remains a critical public health challenge with mortality rates close to 90%. Current prognostication methods commonly analyze data of individual modalities separately and delay assessment until 72 hours post-arrest, creating a critical gap in early decision-making. Here, we introduce contrastive language and image reasoning with masked autoencoders (CLAIR), a novel multimodal framework integrating head computed tomography (CT) imaging with non-imaging clinical patient information through a cross-attention mechanism and contrastive learning approach to predict cerebral performance category (CPC) score in patients after cardiac arrest. In a retrospective study of 208 patients, we evaluated CLAIR against CT-based imaging-only assessment, as well as clinical evaluation by two experienced ICU neurologists. Our method achieved an AUC-ROC of 0.94 (CI: 0.90-0.97) when trained on a combination of multiplanar CT reconstructions and non-imaging clinical data, significantly outperforming CT scan-based imaging-only methods (AUC-ROC: 0.80, CI: 0.74-0.86) with statistical significance (p = 0.03). In a structured evaluation, the clinicians suggested that CLAIR assisted assessments resulted in fewer prognostic errors than non-assisted evaluations. Further, we demonstrate the applicability of our approach for early neurologic outcome prediction using CT scans obtained within the first 24 hours post-arrest (median acquisition time: 3.1 hours). Our results suggest that CLAIR can contribute value as a clinical assistive tool aiming at reliable early prognostication for post-cardiac arrest patients, potentially enabling more timely clinical decision-making, family counseling, and resource allocation.