Automated readings of imaging reports by natural language processing (NLP) supplemented by clinical data predict disease course in children and adults with Crohn's disease: report from the epi-IIRN.
TL;DR
An NLP tool to automatically extract imaging-based variables from radiological reports and integrate them with clinical data to predict disease course in children and adults with Crohn's disease (CD) enables scalable and standardized phenotyping of CD in large datasets and populations.
Abstract
INTRODUCTION The implementation of imaging features in administrative databases and electronic health records is limited by non-standardized free-text radiology reports. We developed a natural language processing (NLP) tool to automatically extract imaging-based variables from radiological reports and integrate them with clinical data to predict disease course in children and adults with Crohn's disease (CD).
Methods
Free-text reports from Magnetic-Resonance Enterography and Computed-Tomography Enterography of patients with newly diagnosed CD were linked to clinical data from the nationwide epi-IIRN cohort and processed using Hierarchical Structured Matching Prediction BERT (HSMP-BERT), an NLP model. The primary outcome was difficult-to-treat course, defined by steroid-dependency, the need for ≥2 classes of biologics or surgery. Predictors were identified using Cox proportional hazards and Gradient Boosting Survival Analysis (GBSA) machine learning models.
Results
Among 780 newly diagnosed patients, 160 (20%) developed difficult-to-treat course. Imaging-based predictors, including stricturing/penetrating disease, disease location, disease extent and the MaRIAs score, demonstrated modest discrimination for difficult-to-treat course (0.59 [95%CI 0.53-0.65]). Clinical predictors (laboratory values, induction treatment, age, sex and perianal involvement), showed better discrimination (AUC of 0.68 [95%CI 0.64-0.73]), while combining imaging and clinical variables resulted in only marginal improvement (AUC of 0.69 [95%CI 0.64-0.74]). In GBSA, the AUC was 0.57 (0.52-0.65) for radiologic model, 0.65 (0.62-0.72) to clinical model and 0.67 (0.61-0.72) to the integration model. For surgery, the improvement was more pronounced, in both, Cox regression and GBSA. Across models, the most influential predictors were induction treatment with systemic steroids, and stricturing or penetrating disease.
Conclusions
Automated NLP-based extraction of imaging reports linked to clinical and laboratory data enables scalable and standardized phenotyping of CD in large datasets and populations.