AI-DERIVED radiology phenotypes improved stroke mortality prediction, when added to structured EHR data in acute stroke admissions.
What the Researchers Set Out to Do
Clinical prediction in acute stroke often relies on structured variables, yet radiology reports contain high-value descriptors that are difficult to quantify at scale. This study evaluated whether automated stroke phenotyping from both imaging outputs and radiology language is feasible, and whether text-derived radiologic features can improve mortality prediction beyond structured electronic health record data.
To establish a quantitative reference, the investigators first analyzed MRI lesion masks from a public dataset of 60 cases. Using automated processing, they calculated lesion volumes and imaging features as a reproducible benchmark for lesion quantification. In parallel, they worked with 15,492 head CT and MRI reports from a large critical care database, applying a rule-based natural language processing pipeline to identify eight stroke phenotypes, including hemorrhage, infarct, midline shift, edema, chronic change, and vascular territory involvement across ACA, MCA, and PCA regions.
Stroke Mortality Prediction Improves with NLP Phenotypes
For report classification, the team trained logistic regression models using TF-IDF features, evaluating performance with AUC and F1-score. The strongest text classifiers achieved high discrimination for key findings, including hemorrhage and edema, supporting the practicality of extracting interpretable stroke phenotypes from routine radiology language at scale.
They then merged probabilistic NLP phenotype outputs with structured admission variables such as age, sex, ICD-9 codes, and length of stay to model in-hospital mortality among 3,999 stroke admissions. Adding text-derived phenotypes led to a modest improvement in model discrimination compared with structured data alone. In permutation analysis, ICD-9 codes contributed the most to mortality prediction, while edema and infarct phenotypes were among the most informative radiology-derived signals.
What This Could Mean in Practice
The authors emphasized that MRI lesion volume was not incorporated into mortality models because of dataset limitations, leaving open the potential benefit of combining direct imaging metrics with text-derived phenotypes in future work. Broader validation, additional modalities, and workflow-aware implementation will be important next steps before these models can support actionable clinical decision support.
Reference: Alotaibi A et al. AI-driven integration of imaging and radiology language improves stroke mortality prediction. Front Neurol. 2026;16:1722965. doi:10.3389/fneur.2025.1722965.
