2019 Fiscal Year Annual Research Report
強縦断的な生体医学信号の深層学習と健康関連の人工知能応用
| Project/Area Number |
19F19081
|
|---|
| Research Institution | The University of Tokyo |
|---|
Principal Investigator |
山本 義春 東京大学, 大学院教育学研究科(教育学部), 教授 (60251427)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
QIAN KUN 東京大学, 教育学研究科(研究院), 外国人特別研究員
|
|---|
| Project Period (FY) |
2019-10-11 – 2022-03-31
|
| Keywords | Signal Processing / Internet of Things / Artificial Intelligence |
|---|
| Outline of Annual Research Achievements |
We have built an open access heart sound database, i.e., heart sounds Shenzhen corpus (HSS1.0), which is the largest open access heart sound database collected from single medical centre. In addition, we made a benchmark work on using both the classic machine learning (ML) models and the state-of-the-art deep learning (DL) models. A pilot study based on HSS1.0 by using wavelet analysis combining with deep recurrent neural networks (RNNs) were presented. We also investigated the capacity of ML and/or DL for new tasks in other interdisciplinary fields including robotic control, materials science, security surveillance, ecology, and smart grid. A perspective study on using affective computing technologies to facilitate the mental health monitoring and therapy was published. We proposed a novel transfer learning models pre-trained by audio data for heart sound classification task, which was demonstrated to be superior to the representations extracted by models pre-trained by the widely used image data. We gave a brief opinion work on computer audition (CA) for healthcare applications. Moreover, we summarised and conducted a comprehensive review study on snore sound classification using CA methods. Based on the aforementioned studies, we found that: First, a standard open access database is the prerequisite for a sustainable research of artificial intelligence (AI) based healthcare applications. Second, we can collaborate more with researchers from other fields to commonly contribute to innovations.
|
| Current Status of Research Progress |
Current Status of Research Progress
1: Research has progressed more than it was originally planned.
Reason
We have implemented an AI-enabled system for automatically analysing the elderly’s behaviour by their daily life usage data of the smart appliances, e.g., television or refrigerator. To the best of our knowledge, this is a first study on using smart appliances’ data to analyse the elderly’s behaviour. The results are encouraging to demonstrate the feasibility to use this framework to monitor the elderly’s symptom status when they are living alone. Motivated by the success of HSS1.0, we re-defined the task, and made modifications of the original audio recordings, to make a new database, i.e., HSS1.5. We are now working towards to release both HSS1.0 and HSS1.5 in the near future workshops and/or challenges. Then, we can make a fair comparison of the participants’ contributions, and summarise the state-of-the-art in current heart sound classification studies. In addition, we have a collaborative work with other AI experts on the multi-label learning algorithms. With an international collaboration between Japan, China, Germany, and UK, we have made a series pilot studies/reviews on COVID-19. We are now working on the development of an open source deep learning toolkit, i.e., deepSELF, which has a combination of the state-of-the-art deep learning, transfer learning, and end-to-end learning frameworks. By using this powerful toolkit, we have successfully beat the official baseline in the ComParE 2020 Challenge on the Mask Task. We also contribute two other collaborative papers on music and education AI based applications.
|
| Strategy for Future Research Activity |
The current achievements of this research have already been much more than original plan. In future work, we will investigate more potential valuable fields. First, we will continuously investigate the ML and DL methods for analysis of some types of intensive longitudinal biomedical signals, e.g., spontaneous physical activity. We would like to understand some fundamental knowledge about the features and models for the classification and/or regression tasks. Second, we will make an intensive study on leveraging CA based methods for analysing the speech of COVID-19 patients. We are already collaborating with the colleagues from a background of medicine in Wuhan, China, and the AI researchers in Germany and UK. We would like to present a benchmark work of the database and a series of studies on using the state-of-the-art signal processing and ML methods for analysis of the database. We are also planning to organise several challenges and/or workshops on the topic of AI for fighting against the COVID-19. Third, we will develop another open source toolkit for data augmentation, which is essential to improve the generalisation of DL models, specifically, for data scarcity scenarios. Fourth, we will investigate the ML strategies, like semi-supervised learning, active learning, and cooperative learning, for their capacity in reducing the human annotation works. Fifth, we will start some specific AI based healthcare applications for elderly people who are living alone.
|
-
-
-
-
[Journal Article] Computer Audition for Healthcare: Opportunities and Challenges2020
Author(s)
Kun Qian, Xiao Li, Haifeng Li, Shengchen Li, Wei Li, Zuoliang Ning, Shuai Yu, Limin Hou, Gang Tang, Jing Lu, Feng Li, Shufei Duan, Chengcheng Du, Yao Cheng, Yujun Wang, Lin Gan, Yoshiharu Yamamoto, and Bjoern W. Schuller
-
Journal Title
Frontiers in Digital Health
Volume: -
Pages: in press
Peer Reviewed / Int'l Joint Research
-
-
[Journal Article] Deep Wavelets for Heart Sound Classification2019
Author(s)
Kun Qian, Zhao Ren, Fengquan Dong, Wen-Hsing Lai, Bjoern W. Schuller, and Yoshiharu Yamamoto
-
Journal Title
Proceedings of the International Symposium on Intelligent Signal Processing and Communication Systems (ISPACS)
Volume: -
Pages: 1-2
Peer Reviewed / Int'l Joint Research
-
