Predictors of brain death in computed tomography reports in a Portuguese cohort from 2011 to 2024

Introduction: Timely recognition of patients progressing to brain death (BD) remains a major bottleneck in donor identification, as manual surveillance  is vulnerable to delays and missed opportunities.
Objective: To develop and validate a calibrated natural language processing (NLP) model to estimate the probability of progression to BD from admission CT reports, prioritising sensitivity over specificity to minimise missed donor opportunities.
Methods: A single-centre retrospective cohort (2011–2024; n=163) was used to train and evaluate four machine learning classifiers: logistic regression, random forest, support vector machine, and XGBoost. Models were trained on curated TF–IDF representations of 1–3 word n-grams, with probability calibration using isotonic regression. A sensitivity-constrained threshold was optimised on training folds and applied to the test set.
Results: XGBoost achieved the best performance (AUC=0.786). At the pre-specified threshold, sensitivity reached 100%, ensuring no potential donor was missed, with specificity of 43%, positive predictive value of 70%, and negative predictive value of 100%.

Top predictive terms corresponded to catastrophic neuroimaging features, including herniation, subarachnoid haemorrhage, and basilar involvement.

Conclusions: A calibrated NLP model can safely flag patients at risk of BD progression from CT reports. Prospective validation and integration with structured data are warranted for clinical translation.