Emotion classification using physiological data with machine learning: case studies on healthy people and Alzheimer’s disease patients with dementia
- 주제(키워드) Emotion classification , Sensor , Physiological data , EEG , GSR , Machine learning , Artificial intelligence , Brain , Alzheimer’s disease , Dementia
- 발행기관 아주대학교
- 지도교수 손경아
- 발행년도 2021
- 학위수여년월 2021. 2
- 학위명 박사
- 학과 및 전공 일반대학원 컴퓨터공학과
- 실제URI http://www.dcollection.net/handler/ajou/000000030542
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
In recent years, studies, products, and services that used artificial intelligence and physiological sensor s have been published. One of the trends is classifying the user's context, such as their emotional state, using physiological data. One of the motivations for classifying emotion classification studies is that emotions influence human behaviors, decision making, health, learning efficiency. Many emotion classifications studies have been conducted; however, none of the studies classified boredom using electroencephalogram (EEG) and galvanic skin response (GSR) data. Furthermore, studies that targeted Alzheimer's disease (AD) patients with dementia for classifying their emotions did not exist. In this study, we did a literature review related to emotion classification based on physiological sensors and set healthy people (13 males and 15 females, mean age 23.62) and AD patients with dementia (30 females, mean age 83.9) as target groups. Then, we designed a data collection protocol for each target group and collected physiological data (healthy group: EEG and GSR, and AD patients with dementia group: EEG data) when exposed to video and image stimuli designed to evoke the target emotions. Using the data, I trained emotion classification models with conventional machine learning and deep learning algorithms. As a result, the model trained with multilayer perceptron (MLP) showed 79.98% mean accuracy from 1,000 iterations of five-fold cross-validation. For the model of AD patients with dementia, the ensemble model that consisted of MLP and convolution neural network showed 73.33% accuracy from leave-one-out cross-validation. Additionally, we analyzed the correlations between boredom state and collected data and indicated that healthy people's approach to classifying emotion is possible for AD patients with dementia. These results can be utilizing for affective computing systems and understanding correlations between emotional states and physiological responses.
more목차
Introduction 1
1.1. Emotion classification 1
1.2. Physiological data 2
1.3. Alzheimer's disease 4
1.4. Thesis statement 5
1.5. Summary of contributions 5
Background 7
2.1. Emotion model 7
2.2. Alzheimer's disease 8
2.3. Emotion classification studies 10
2.3.1. For healthy people 10
2.3.2. For neurological disorders 13
Boredom classification for healthy people 17
3.1. Overview 17
3.2. Data collection 18
3.2.1. Participants 18
3.2.2. Sensors 19
3.2.3. Protocol 21
3.3. Model training 24
3.3.1. Window size 24
3.3.2. Features 24
3.3.3. Conventional machine learning algorithms 26
3.3.4. Feature refinement 28
3.4. Results 28
3.4.1. Questionnaire 28
3.4.2. Conventional machine learning algorithm selection 30
3.4.3. Hyperparameter Tuning 31
3.4.4. Final Performance Analysis 36
3.5. Discussion 41
Emotion classification for Alzheimer's disease patients 44
4.1. Overview 44
4.2. Data collection 45
4.2.1. Participants 45
4.2.2. Target emotion 46
4.2.3. Stimuli 47
4.2.4. Sensor 48
4.2.5. Protocol 51
4.3. Model training 52
4.3.1. Conventional machine learning algorithms 53
4.3.2. RNN model 55
4.3.3. CNN model 58
4.3.4. Ensemble models 62
4.4. Results 63
4.4.1. Interview 63
4.4.2. Conventional machine learning algorithms 64
4.4.3. RNN models 69
4.4.4. CNN models 73
4.4.5. Ensemble results 75
4.5. Discussion 77
Conclusion 81
Reference 84
