STRUCTURING RADIOLOGY AT SCALE: MACHINE-LEARNING CLASSIFICATION OF CTA/CTV REPORTS FOR PE/DVT AND INCIDENTAL FINDINGS

Abstract

The core of assessing the suspected cases of pulmonary embolism (PE) and deep vein thrombosis (DVT) is CT pulmonary angiography (CTA) and CT venography (CTV), but free-text reporting does not allow population-wide analytics and consistent monitoring of its incidental findings. The natural language processing (NLP) and machine-learning pipeline that we designed and tested was used to structure the CTA/CTV radiology reports, thromboembolism outcome classification, and unveil clinically interesting incidentalomas. The annotation of a de-identified corpus was done with the help of a schema that included entities (anatomy, thromboembolic patterns), relations (Location_of), and modalities (positive/negative/known/incidental/hypothetical). Plain text, concept/relationship annotations, and section typing feature sets were all trained on the Naive Bayes and Maximum Entropy models. Precision, recall and F-measure were used to measure performance. CTA+/CTV+/CTV− 24.8% vs CTA+/CTV−/CTV− 18.2% complementary diagnostic yield was emphasized in 5,000 reports CTA−/CTV+ 10.4%. Incidental findings that are of clinical significance were found in 32.0 percent of examinations. Entity annotation was very agreeable but relation extraction was relatively difficult. Baseline plain-text modeling was showing PE with F-measure 0.78 but worse in DVT and incidentalomas. Compounding concepts, modalities and relations led to significant accuracy gains: Naive Bayes made gains across categories and maximum entropy made gains to a maximum of 0.98 on thromboembolic classification and 0.80 on incidentalomas with section typing and critical-section features. Transforming narrative CTA/CTV reports into formalized representations allows to classify PE/DVT accurately and reliably, as well as identify important incidental findings. The framework advocates a semiautomated cohort introduction, choice assistance, and track-down, and elucidates the discerning added worth of CTV in high-risk groups (e.g., ICU, postoperative, malignancy, postpartum). Future work priorities are better relation/event modeling, expanding multilingual imaging lexicon and future assessing the impact of the intervention on time-to-action, following up recommendations adherence and patient outcomes. As demonstrated by this work, scalable narrative radiology based on integrating CTA/CTV data with NLP and machine learning can be converted into actionable clinical intelligence.