A head CT within six hours of ECPR can flag neurological outcome: report-derived injury signs as auxiliary supervision lift AUROC to 0.929
Synopsis
For patients who had a head CT within six hours of extracorporeal cardiopulmonary resuscitation (ECPR), a transfer-learning scheme that uses five brain-injury signs extracted from radiology reports as auxiliary supervision raised mean AUROC for neurological outcome from 0.918±0.044 with outcome labels alone to 0.927±0.042, and to 0.929±0.037 after adding 486 examinations that carried only manifestation labels and no eligible outcome label, while inference uses CT images alone.
Interpretation
Hypoxic-ischemic injury signs on early head CT are often subtle and incompletely developed, conventional CT markers have limited sensitivity and discrimination within the first six hours after cardiac arrest, and outcome-labeled cases are scarce. Earlier approaches learned the outcome directly on the same outcome-labeled patients, or pooled features from a frozen encoder and classified them with logistic regression and a multilayer perceptron. In five-fold patient-level cross-validation on 197 outcome-eligible patients, pooled-feature logistic regression and the multilayer perceptron reached mean AUROCs of 0.894±0.045 and 0.891±0.043, and the outcome-only cross-attention adaptor reached 0.918±0.044.
The scheme makes one frozen head CT encoder serve six tasks at once: one neurological outcome and five brain-injury signs labeled from radiology reports. Six learned queries each aggregate a separate representation from shared image features through cross-attention, and each feeds its own prediction head, trained jointly with weighted binary cross-entropy losses. Manifestation labels were extracted zero-shot from free-text reports by Qwen3.8-27B under a structured prompt; only explicitly present and absent findings entered the losses, uncertain and unmentioned findings were masked, and outcome loss was computed only for patients with eligible outcome labels.
Report-derived sign supervision lets examinations without eligible outcome labels contribute to training while inference still needs only CT images. Within the same outcome-eligible patients, adding sign supervision moved mean AUROC from 0.918±0.044 to 0.927±0.042, and adding the 486 manifestation-only patients brought it to 0.929±0.037. With the additional examinations, mean specificity rose from 82.9% to 93.7%, mean sensitivity fell from 83.6% to 79.9%, and mean balanced accuracy rose from 83.3% to 86.8%; against the outcome-only adaptor, sensitivity was 79.9% versus 75.4% and specificity 93.7% versus 89.1%.
Perspective
The findings rest on 683 consecutive patients who underwent ECPR at a single center between December 2015 and November 2025 and had non-contrast head CT within six hours of cannulation with a corresponding radiology report, of whom 197 entered outcome evaluation; performance was estimated by internal five-fold cross-validation, and the authors plan to expand the development cohort and pursue external validation at other ECPR centers.
Report labeling trades annotation burden against label completeness: masking uncertain and unmentioned findings avoids treating unreported signs as absent but reduces usable supervision and cannot correct inaccuracies in the reports or errors in label extraction, so manifestation performance was estimated only in the subset with explicit labels. The outcome analysis included 197 patients from a single center because it required both the CT timing and a discharge outcome meeting prespecified criteria, leaving generalizability to other centers uncertain; inference uses CT alone, whereas guidelines recommend multimodal neuroprognostication.
