Semiology Extraction and Machine Learning-Based Classification of Electronic Health Records for Patients With Epilepsy: Retrospective Analysis.

Background: Obtaining and describing semiology efficiently and classifying seizure types correctly are crucial for the diagnosis and treatment of epilepsy. Nevertheless, there exists an inadequacy in related informatics resources and decision support tools.

Objective: We developed a symptom entity extraction tool and an epilepsy semiology ontology (ESO) and used machine learning to achieve an automated binary classification of epilepsy in this study.

Methods: Using present history data of electronic health records from the Southwest Epilepsy Center in China, we constructed an ESO and a symptom-entity extraction tool to extract seizure duration, seizure symptoms, and seizure frequency from the unstructured text by combining manual annotation with natural language processing techniques. In addition, we achieved automatic classification of patients in the study cohort with high accuracy based on the extracted seizure feature data using multiple machine learning methods.

Results: Data included present history from 10,925 cases between 2010 and 2020. Six annotators labeled a total of 2500 texts to obtain 5844 words of semiology and construct an ESO with 702 terms. Based on the ontology, the extraction tool achieved an accuracy rate of 85% in symptom extraction. Furthermore, we trained a stacking ensemble learning model combining XGBoost and random forest with an F1-score of 75.03%. The random forest model had the highest area under the curve (0.985).

Conclusions: This work demonstrated the feasibility of natural language processing-assisted structural extraction of epilepsy medical record texts and downstream tasks, providing open ontology resources for subsequent related work.