BACKGROUND/OBJECTIVES: Early identification of patients progressing toward brain death may reduce missed organ donation opportunities. We aimed to develop and internally validate an interpretable natural language processing (NLP) model using admission brain computed tomog-raphy (CT) reports. METHODS: This single-center retrospective cohort included 163 adults ad-mitted between 2011 and 2024 with acute neurological injury; 96 subsequently fulfilled brain death criteria and 67 did not. Predictors were derived exclusively from admission CT report text using rule-based preprocessing and term frequency–inverse document frequency word 1–3-gram features. Logistic regression, random forest, XGBoost, and support vector machine models were evaluated using repeated stratified 5-fold cross-validation with 20 repetitions. Vectorization, cal-ibration, model fitting, and threshold selection were restricted to training folds. A prespecified sensitivity target of at least 95% guided threshold selection. Interpretability was assessed using SHapley Additive exPlanations (SHAP). RESULTS: XGBoost achieved the highest discrimination, with a mean area under the receiver operating characteristic curve of 0.807 (95% confidence in-terval, 0.793–0.820) and a pooled out-of-fold value of 0.791. At the high-sensitivity operating point, mean sensitivity was 0.947 and specificity was 0.275. SHAP identified clinically plausible descriptors related to cerebral edema, mass effect, ventricular distortion, ischemic injury, and midline abnormalities. CONCLUSIONS: Admission CT reports contain prognostic information that can be extracted using NLP to support earlier identification of patients at risk of progression to brain death and potential organ donation. External multicenter validation is required before clinical implementation.