Abstract 17845: Logic-Based Natural Language Processing Identifies VT Recurrence in the Electronic Health Record

Abstract

Introduction: With the advent of integrated electronic healthcare record (EHR) systems, it is increasingly important to identify patients who may suffer a clinical endpoint to allocate resources and develop predictive models. For the endpoint of VT recurrence after ablation, outcomes are currently assessed manually by experts, which is time consuming with few automated tools. We set out to develop a robust natural language processing model to parse cardiology and electrophysiology notes to detect VT. Hypothesis: We hypothesize that natural language processing (NLP) with logic-based algorithms can be used to accurately extract specific information on VT recurrence from complex clinical notes, compared to experts. Methods: We studied N=639 full-text clinical notes written by board certified cardiologists from N = 125 patients (32.0% female, LVEF 48.9±13.9%, 61±14.0 years). Notes were labeled as “No” or “Yes” for VT recurrence, and split into a development set of N=303 (median word count = 236) from N=75 patients and a test set of N=336 (median words = 297) from N=50 patients. We developed an NLP rule based expert system that uses hierarchical groups of keywords to identify sentences related to VT, then uses negation and quantifier functions to interpret context and assign a label. Results: The logical flow is shown in figure A. This approach achieved an 93.73% accuracy in the development set (precision = 57.14%, recall = 69.56%), and 90.48% accuracy (precision = 36.84%, recall = 63.63%) in the test set (figure B). Conclusions: An NLP logic-based approach can identify VT recurrence post ablation from complex clinical notes. NLP can thus be used to accelerate data extraction for clinical purposes or for semi-supervised learning. Future work will determine if logic-based NLP need to be augmented for noisier data, such as the entire EHR system.