Abstract WP300: Predicting infarct location and stroke severity from symptoms using language models: towards the next generation of vision models in acute stroke

Introduction: symptoms observed during clinical presentation offer critical cues that guide image interpretation, aligning findings with physical exam results. Integrating clinical presentation data into image analysis models may improve performance and generalizability in tasks where current approaches fall short. Despite being widely available on admission records, analysis of clinical presentation is challenging due the unstructured nature of reports. This study evaluates the use of large language models (LLMs) to retrieve symptoms from admission notes and infer occlusion location and stroke severity based on their encoded representations. Methods: reports from patients who received endovascular treatment due to an anterior circulation vessel occlusion between 2018-2023 in a comprehensive stroke center were retrospectively collected. Decoder-only LLMs were used for preprocessing, deidentification and symptoms retrieval from unstructured reports. A set of encoder-only LLMs were used to generate text embeddings from retrieved symptoms. Models were assessed according to the embeddings' value for prediction of affected region (side, proximal ICA/M1 vs. distal M2/M3) and severity (patients dichotomized at NIHSS=6, 10, 16). Location and severity prediction was tested using embeddings for binary class prediction on logistic regression. Cluster separability was evaluated with silhouette scores based on semi-supervised universal manifold approximation (UMAP) coefficients reduced from embeddings. Results: reports from 811 (median age 80 [69-88] years, 45.0% male) patients were included. Embeddings from the best model led to effective prediction of side (AUROC [mean±SD] 0.92±0.02) and severity (NIHSS≥6: 0.84±0.08; NIHSS≥10: 0.84±0.09; NIHSS≥16: 0.81±0.10), and moderate performance on distal occlusion prediction (0.69±0.01). Silhouette scores revealed reasonable cluster structures for side (0.535±0.030) and weak to poor structures across severity thresholds (NIHSS≥6: 0.120±0.045; NIHSS≥10: 0.244±0.068; NIHSS≥16: 0.190±0.044) and proximal vs. distal occlusions (0.109±0.036). Conclusions: LLM-derived text embeddings enabled effective extraction and semantic encoding of stroke symptoms from unstructured admission reports. This proof-of-concept demonstrates their potential for predicting location and severity. Effective integration with imaging data may lead to more intelligent and capable AI models, potentially expanding their use cases in the acute stroke setting.