Development of an Automated Hierarchical Structurization Processing Method for EEG Reports in Patients with Epilepsy
Abstract number :
2.141
Submission category :
3. Neurophysiology / 3C. Other Clinical EEG
Year :
2024
Submission ID :
221
Source :
www.aesnet.org
Presentation date :
12/8/2024 12:00:00 AM
Published date :
Authors :
Presenting Author: Young Ho Kim, MD – Seoul National University Budang Hospital
Jaeso Cho, MD – Seoul National University Budang Hospital
Hyun Woo Kim, BS – Seoul National University Budang Hospital
Yoon Gi Chung, Ph.D – Seoul National University Budang Hospital
Hunmin Kim, MD – Seoul National University Bundang Hospital
Rationale: In the clinical environment, encephalogram (EEG) plays a crucial role in the diagnosis and treatment of patients with epilepsy. However, the unstructured text results of EEG readings hinder the construction of a structured database for further big data analysis. In this study, we developed a hierarchical structurization processing method using deep learning and natural language processing techniques to convert unstructured EEG readings into a structured format.
Methods: For the structurization of EEG readings, first, we retrospectively collected 12,346 anonymized EEG readings and annotated them as “normal” and “abnormal”. Second, we created bidirectional encoder representations from transformers (BERT)-based deep learning models combined with long short-term memory for automated normal-abnormal binary classification. Third, we employed hierarchical extraction of keywords using regular expressions for natural language processing to identify the presence of abnormal slow background activities and interictal discharges (IEDs) from the abnormal readings. For readings containing IEDs, we further classified them into focal or generalized discharges, with an additional classification of the IED location in cases of focal discharges.
Results: Our approach for EEG readings showed an accuracy of 99.75%, 99.50%, 98.50, and 94.00% for identification of normal or abnormal, presence of seizure, focal or generalized seizure, and presence and location of epileptiform discharges, respectively. Errors were observed in readings formatted differently from the training set and in readings containing typographical errors.
Conclusions: The automated hierarchical structurization processing method for unstructured EEG readings is an effective and useful tool for creating a structured big data analysis database. Further studies are needed to improve the accuracy of the automated system.
Funding: This study was supported by a Korean National Grant from the Ministry of Health and Welfare.
Neurophysiology