170 An Integrated Voice Recognition and Natural Language Processing Platform to Automatically Extract Thoracolumbar Injury Classification Score (TLICS) Features from Radiology Reports

Abstract

INTRODUCTION: Many predictive models have been reported in the literature for clinical outcomes after spine surgery. However, the real-time implementation of these predictive scores in clinical practice is limited by the time intenstive nature of manually abstracting relevant predictors from the electronic health record. Natural language processing (NLP) provides a framework to automate data abstraction for commonly used predictive scores in spine surgery. METHODS: We retrieved and annotated the radiology reports of all Mayo Clinic patients with an International Classification of Diseases(ICD)-9/10 code corresponding to a fracture of the thoracolumbar spine between January 2005 and October 2020. Annotated data was used to train an n-gram NLP model using machine learning (ML) methods including random forest(RF), stepwise linear discriminant analysis (sLDA), k-nearest neighbors (kNN), and penalized logistic regression (pLR) models. RESULTS: 1,085 spine radiology reports were included in our analysis consististing of 868 to train our ML-NLP model and 217 for validation. Our dataset included 483 compression, 401 burst, 103 translational/rotational, and 98 distraction fractures. 103 reports had documented an injury of the PLC. The overall accuracy of the RF model for fracture morphology feature detection was 76.96% versus 65.90% in the stepwise LDA, 50.69% in the kNN, and 62.67% in the pLR. The overall accuracy to detect PLC integrity was highest in the RF model at 83.41% versus 70.97% in sLDA, 70.05% in kNN, and 67.74% in pLR model. Our RF model was implemented in the backend of a web application where users can dictate reports and have TLICS score features be automatically extracted. CONCLUSION: We have developed a robust ML-NLP model for extracting TLICS score features from radiology reports, which we deployed in a convenient web application that can be integrated prospectively into clinical practice and scaled to other commonly used predictive scores.