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

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.

Authors

Institutions

Publication Details

Journal
medRxiv
Published
2026-11-18
DOI
https://doi.org/10.64898/2026.11.15.26368165
Primary Topic
Cardiac Arrest and Resuscitation
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

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

Andrea M. Elliott, Mahdi Bayat, Samarth Goyal, Oladi Bentho et al.
3 links
medRxiv
Cardiac Arrest and Resuscitation
preprint

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

Andrea M. Elliott, Mahdi Bayat, Samarth Goyal, Oladi Bentho, Pornthira Mutirangura, Ardavan Kaviany
preprint en

Abstract

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.

medRxiv
University of Minnesota (US)
Good health and well-being
Cardiac Arrest and Resuscitation
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon