Masked autoencoders for early neurological outcome prediction in post-cardiac arrest patients using brain CT scan

Cardiac arrest can cause catastrophic brain injury, manifested as coma, and is a leading cause of death. Accurate neurologic prognostication is complex and requires multiple diagnostic modalities over days to weeks of hospitalization, leaving clinicians uncertain of which patients may benefit from early goal-directed therapies in the critical care unit to mitigate brain injury. We propose a masked autoencoder (MAE) method to accurately predict neurological outcomes from brain computed tomography (CT) scans taken within hours of cardiac arrest. Our method leverages a fine-tuned MAE model (CT-MAE) to process the input CT scans. The CT-MAE network employs a self-supervised framework that randomly masks a large portion of the input CT scan and then reconstructs the original scan using a lightweight decoder. This process enables the encoder to learn strong spatiotemporal representations of critical anatomical features, enabling subsequent classification layers to predict neurologic outcomes (classified as dichotomized Cerebral Performance Category (CPC) at discharge). Our CTMAE achieved an AUC-ROC of 0.79 with 80% sensitivity, demonstrating superior workflow efficiency compared to manual measurement approaches while maintaining competitive predictive performance.