Development of an explainable machine learning-based multichannel EEG analysis system for cognitive impairment assessment in sleep-related brain disorders

다국어 초록 (Multilingual Abstract)

Machine learning technology, especially deep learning, which has rapidly developed mainly in image and language processing, has been widely applied to medical data analysis. Recently, diagnostic systems that recognize patients with high accuracy have been developed based on various types of medical images, such as radiation and ultrasound imaging. Medical imaging, such as computed tomography and magnetic resonance imaging (MRI), is widely used to diagnose brain disorders. These imaging systems capture the structure of the brain to find structural brain abnormalities. On the other hand, an electroencephalogram (EEG) and functional MRI are mainly used to measure the functional activity of the brain. An EEG is a time-series signal that records the electrical activity of the cerebral cortex using electrodes attached to the scalp. EEGs are mainly used in the field of cognitive neuroscience because they are effective in tracking rapid temporal changes in brain activity when performing cognitive tasks. However, determining which cerebral area exhibits abnormalities is difficult due to the low spatial resolution of an EEG.
This study, therefore, aimed to develop an explainable machine learning-based EEG analysis system. The system estimated cortical activity from an EEG through source localization to resolve poor spatial resolution and generate images. Then, the system was designed to diagnose and examine patients with sleep-related brain disorders based on these images.
The developed system was applied to experimental data for verification. Up to 99% accuracy was achieved in distinguishing patients from normal controls using multichannel EEG recorded during a cognitive task. The abnormal brain activity characteristics of patients were closely related to cognitive impairment. In addition, the developed system classified most brain activity in patients who received medication as normal. The degree of normalization of brain activity was associated with improved symptoms and sleep quality in the patients.
The developed system could be used for diagnosis and treatment monitoring for patients with sleep-related brain disorders. Furthermore, the system might be able to predict treatment response or the possibility of transition in neurodegenerative diseases.

국문 초록 (Abstract)

이미지, 언어처리 등을 중심으로 급격히 발전한 기계학습 기술, 특히 딥러닝은 의료 데이터 분석에도 널리 적용되었다. 최근에는 방사선, 초음파 영상 등 각종 의료영상을 이용하여 높은 정확도로 환자를 진단하는 시스템들이 개발되고 있다. 뇌질환 진단을 위해서는 CT, MRI 등의 의료 영상이 많이 쓰인다. 이러한 영상들은 뇌의 구조를 촬영하여 정상인과 차이를 보이는 환자의 구조적 뇌 이상을 판별한다. 뇌의 기능적 활동을 측정하기 위해서는 뇌파, fMRI 등이 주로 쓰인다. 뇌파는 두피에 부착된 전극에서 대뇌피질의 전기적 활동을 기록하는 시계열 신호이다. 인지과제 수행 시 빠른 뇌활동 변화를 추적하는데 효과적이여서 인지기능 연구에 뇌파가 주로 쓰인다. 그러나, 뇌파는 공간적해상도가 낮아 어떤 대뇌영역에서 뇌활동 이상을 보이는지 판단하기 어렵다.
이에 본 연구에서는 설명가능한 기계학습에 기반한 뇌파 분석 시스템을 개발하고자 하였다. 본 시스템은 낮은 공간적해상도를 해소하는 신호원 국소화를 통해 뇌파로부터 대뇌피질활동을 추정하여 영상으로 가공한다. 가공한 영상을 이용하여 본 시스템은 수면관련뇌질환 환자를 진단 및 조사하도록 설계되었다.
개발된 시스템은 수면관련뇌질환 환자의 실험데이터에 적용되었다. 인지과제 수행 시 기록된 다채널 뇌파를 사용하여 정상대조군으로부터 환자를 구별하는 데 최대 99%의 정확도를 달성했다. 환자의 비정상적 뇌활동 특성은 인지기능 장애와 밀접한 관련이 있었다. 또한 개발된 시스템은 약물치료를 받은 환자의 뇌활동 대다수를 정상으로 판단하였다. 뇌활동의 정상화 정도는 환자의 개선된 증상 및 수면의 질과 관련이 있었다.
이로써 본 연구는 개발한 시스템이 뇌파를 기반으로 하는 환자 진단 및 치료효과 모니터링에 활용 가능함을 보였다. 나아가 개발한 시스템은 환자의 치료효과 예측 또는 신경퇴행성 질환으로의 발전 가능성 예측 등에도 활용 가능할 것으로 보인다.

목차 (Table of Contents)

• Table of Contents
• List of Figures v
• List of Tables viii
• Abstract ix
• Chapter 1. Introduction - 1 -
• 1.1. Evoked potentials in electroencephalogram (EEG) - 2 -
• 1.1.1. EEG - 2 -
• 1.1.2. Evoked potentials - 4 -
• 1.2. Sleep-related brain disorder and cognitive function - 7 -
• 1.2.1. Statistical analysis of ERP - 8 -
• 1.2.2. Sleep-related brain disorder - 10 -
• 1.3. Machine learning-based diagnostic system - 11 -
• 1.3.1. Machine learning - 12 -
• 1.3.2. Explainable machine learning - 16 -
• 1.4. Objectives of the study - 17 -
• Chapter 2. Development of an explainable machine learning-based multichannel EEG analysis system - 19 -
• 2.1. Overview of EEG analysis system - 19 -
• 2.2. Workflow of EEG analysis system - 22 -
• 2.2.1. EEG preprocessing and cortical current source reconstruction - 24 -
• 2.2.2. Machine learning framework - 36 -
• 2.2.3. Spatio-temporal analysis of cortical activity - 48 -
• 2.3. Discussion - 55 -
• Chapter 3. Application case 1 - Discrimination of idiopathic REM sleep behavior disorder (iRBD) using cortical activity during a visuospatial attention task - 58 -
• 3.1. Overview of iRBD - 58 -
• 3.2. Assessment of visuospatial attention - 59 -
• 3.2.1. Experimental procedure - 60 -
• 3.2.2. Dataset - 62 -
• 3.3. Discrimination of iRBD patients based on time-domain analysis of event-related current density time-series - 65 -
• 3.4. Discrimination of iRBD patients based on frequency-domain analysis of event-related current density time-series - 66 -
• 3.4.1. Methods - 67 -
• 3.4.2. Results - 70 -
• 3.5. Characterization of cortical dysfunction in iRBD patients - 71 -
• 3.5.1. Time-domain brain function characteristics - 71 -
• 3.5.2. Frequency-domain brain function characteristics - 76 -
• 3.6. Discussion - 78 -
• Chapter 4. Application case 2 – Discrimination of restless legs syndrome (RLS) patients using cortical activity during a working memory task - 84 -
• 4.1. Overview of RLS - 84 -
• 4.2. Assessment of working memory - 85 -
• 4.2.1. Experimental procedure - 86 -
• 4.2.2. Dataset - 87 -
• 4.3. Discrimination of RLS patients based on time-domain analysis of event-related current density time-series - 89 -
• 4.4. Characterization of cortical dysfunction in RLS patients - 90 -
• 4.5. Discussion - 92 -
• Chapter 5. Application case 3 – Evaluation of working memory improvement in RLS patients following drug treatment - 94 -
• 5.1. Introduction to drug treatment for RLS - 94 -
• 5.2. Methodology for monitoring working memory improvement - 95 -
• 5.3. Analysis of working memory improvement post-treatment - 98 -
• 5.4. Discussion - 102 -
• Chapter 6. Conclusion - 104 -
• References - 106 -
• Abstract (In Korean) - 115 -