The abstract is read at the publisher; the summary is JClub's.
Manifestation-guided adaptation of a CT foundation model predicts favorable versus neurological death outcomes after ECPR from hyperacute head CT with AUROC 0.927 ± 0.042.
- Why it matters: Early neurological prognosis after ECPR is difficult and data-limited, creating a need for transfer learning that can use routinely available radiology text to improve CT-based prediction.
- What they did: In a retrospective cohort (December 2015–November 2025) of 683 patients, a cross-attention adaptor learned outcome and five report-derived brain-injury manifestations from a frozen CT encoder, with CT-only inference.
- The result: Compared with an outcome-only adaptor (AUROC 0.918 ± 0.044), the manifestation-guided adaptor improved AUROC to 0.927 ± 0.042 and, with extra manifestation-only training, reached 0.929 ± 0.037 AUROC and 93.7% ± 6.5% specificity.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.