The Influence of NegEx on ICD-10 Code Prediction in Swedish: How is the Performance of BERT and SVM Models Affected by Negations?

Abstract

Clinical text contains many negated concepts since the physician excludes irrelevant symptoms when reasoning and concluding about the diagnosis. This study investigates the machine interpretation of negated symptoms and diagnoses using a rule-based negation detector and its influence on downstream text classification task. The study focuses on the effect of negated concepts and NegEx preprocessing on classifier performance for predicting ICD-10 gastro surgical codes assigned to discharge summaries. Based on the experiments, NegEx preprocessing resulted in a slight performance improvement for traditional machine learning model (SVM) and had no effect on the performance of the deep learning model KB/BERT.