pipette
ESEspañol

Manifestation-Guided Adaptation of CT Foundation Model for Early Neurological Prognosis after ECPR

A. Kaviany, P. Mutirangura, S. Goyal, O. Bentho, A. Elliott, M. Bayat

PreprintReal-world use

In the authors' words

Purpose. To develop a data-efficient, clinically guided transfer-learning framework for hyperacute neurological outcome prediction from head CT in patients undergoing extracorporeal cardiopulmonary resuscitation (ECPR) after cardiac arrest by leveraging free-text radiology report. Methods. This retrospective study included 683 patients with available CT examination and radiology reports from December 2015 to November 2025. The primary outcome distinguished favorable neurological outcomes (Cerebral Performance Category [CPC] 1-2) from neurological death due to brain death or cerebral herniation (CPC 5). A cross-attention adaptor jointly learned to predict neurological outcome and five report-derived manifestations of brain injury from a frozen pretrained head CT encoder. Outcome prediction used CT images alone at inference. The manifestation-guided adaptor, trained with and without additional patients who contributed only manifestation labels, was compared with outcome-only adaptor and two pooled-feature baselines. Performance was summarized as mean {+/-} SD across five held-out folds. Results. Of 683 patients, 197 were eligible for outcome evaluation (63 with favorable outcomes and 134 with neurological death), and 486 formed the manifestation-only cohort. The area under the receiver operating characteristic curve (AUROC) was 0.918 {+/-} 0.044 for the outcome-only adaptor. The manifestation-guided adaptor achieved an AUROC of 0.927 {+/-} 0.042. With additional manifestation-only training data, the adaptor achieved the mean AUROC (0.929 {+/-} 0.037), specificity (93.7% {+/-} 6.5%), with a sensitivity of 79.9% {+/-} 9.2%. Conclusion. Manifestation-guided adaptation outperforms outcome-only adaptation for neurological outcome prediction from hyperacute head CT after ECPR. Report-derived manifestation labels enabled training with additional examinations without eligible outcome labels while preserving CT-only prediction.

Main resultThe abstract does not state a limitation.

Appeared: Thursday, November 19. medRxiv. Preprint, not yet peer-reviewed.

DOI: 10.64898/2026.11.15.26368165