Automated identification of cancer-associated thrombosis events <i>via</i> natural language processing: a systematic review of the literature

Accurate identification of cancer-associated thrombosis (CAT) in electronic health records is essential for disease surveillance, trial design, and the development of risk stratification models. Manual chart review is impractical at scale, while ICD-code or similar coding system-based extraction often misses events, fails to distinguish incident from prevalent events, or equates bland and tumor thrombi. Natural language processing (NLP) is a promising solution, offering scalable extraction of structured CAT events directly from clinical text. We systematically reviewed the literature for NLP-based pipelines for CAT identification according to PRISMA guidelines. PubMed, Embase, and Web of Science were queried for English language studies from 2010-2025. NLP methodology, training strategy, and model performance were extracted from relevant studies. Seven studies, implementing NLP approaches ranging from lexicon or rules-based pipelines to transformer models, met inclusion criteria. Most studies developed and evaluated models using text from a single institution, and only one distinguished incident CAT events from prevalent events. Reported metrics varied between studies, though models tended to exhibit higher specificity and negative predictive value than sensitivity and positive predictive value. Overall, current NLP systems for CAT identification have achieved desired results to assist but not supplant manual chart review. Major limitations include small and institution-specific datasets, absence of external validation, and lack of distinction between incident vs prevalent events. Development of generalizable models will require large, multi-institutional datasets as the field moves towards transformer-based models, and standardized evaluation metrics with shared benchmark test sets are needed for an unbiased measure of progress.