Natural language processing for triage of cerebral large-vessel occlusion
Bibliographic record
Abstract
Timely identification of large-vessel occlusion (LVO) in ischemic stroke is essential for optimizing prehospital triage and enabling rapid mobilization of thrombectomy-capable teams. Traditional LVO screening tools are often lengthy and reliant on neurological examination skills that may be inaccessible to nonspecialists.To assess the ability of large language models (LLMs) to detect LVO using only free-text summaries, with or without National Institutes of Health Stroke Scale (NIHSS) data, in a national teleneurology service.We conducted a retrospective analysis of 2,887 suspected stroke cases across 21 spoke hospitals within a national TeleStroke network. Neurologist-authored case summaries were processed using natural language processing techniques, including text embedding and supervised machine learning classification. Contextual LLMs (BERTimbau, BioBERTpt, GPorTuguese-2) were evaluated with five algorithms. The Bootstrap method was employed to mitigate class imbalance, with performance averaging over 100 iterations.Of 1,060 cases included in the final dataset, 143 had confirmed proximal occlusions. Median Alberta Stroke Program Early CT Score (ASPECTS) was 9 and mean National Institutes of Health Stroke Scale (NIHSS) was 5.4 ± 2. AdaBoost paired with BioBERT yielded the highest accuracy (89.82%), precision (98.37%), and AUC (89.86%). Incorporating NIHSS as a numerical feature improved recall (87.60% with multilayer perceptron) and F1-score (89.05% with Dense Neural Network). BioBERT consistently outperformed other models, regardless of NIHSS inclusion.The LLM-based models demonstrated strong performance in identifying LVO using routine clinical narratives. These findings support the integration of NLP and ML in TeleStroke systems and underscore the need for further validation across larger, multilingual datasets to ensure generalizability and clinical applicability.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.003 | 0.010 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.001 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.001 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".