ChroNet: A multi-task learning based approach for prediction of multiple chronic diseases
- Authors
- Feng, R.; Cao, Y.; Liu, X.; Chen, T.; Chen, J.; Chen, D.Z.; Gao, H.; Wu, J.
- Issue Date
- Dec-2022
- Publisher
- SPRINGER
- Keywords
- Chronic diseases; Convolutional neural networks (CNNs); Electronic medical records (EMR); Multi-task learning; Prediction
- Citation
- Multimedia Tools and Applications, v.81, no.29, pp.41511 - 41525
- Journal Title
- Multimedia Tools and Applications
- Volume
- 81
- Number
- 29
- Start Page
- 41511
- End Page
- 41525
- URI
- https://scholarworks.bwise.kr/gachon/handle/2020.sw.gachon/86195
- DOI
- 10.1007/s11042-020-10482-8
- ISSN
- 1380-7501
- Abstract
- Chronic diseases (such as diabetes, hypertension, etc) are generally of long duration and slow progression. These diseases may be implied in electronic medical records (EMR), and one chronic disease may be accompanied by another. Recently, many methods have been proposed for chronic disease prediction and early detection. However, previous methods mainly focused on predicting one individual disease, thus possibly neglecting potential correlations among multiple diseases. In this paper, we propose a new framework for chronic disease prediction which can take into account possible correlations among multiple chronic diseases, called ChroNet. We propose a Multi-task Learning (MTL) based framework, for multiple chronic disease prediction. First, based on the characteristics of EMR, we introduce a novel approach for data embedding, including Content Embedding and Spatial Embedding. Then, an MTL convolutional neural network (CNN) is designed to perform multiple chronic disease prediction simultaneously. We collect a dataset from 5 local hospitals, involving 48953 patients’ records. Then we conduct abundant experiments for hypertension and type 2 diabetes prediction, based on our dataset. For both hypertension and type 2 diabetes prediction, our proposed framework outperforms known single-task models (with the same CNN layers yet a single branch). Further, our MTL-based framework outperforms several most commonly used traditional machine learning models and convolutional neural networks. Theoretically, our framework can capture general features of different diseases and focus its attention on those features that actually matter for each disease. The performance superiority in experiments indicates that our framework may be able to capture more detailed characteristics of medical structural data after specific embedding, comparing with known single-task models. © 2021, The Author(s), under exclusive licence to Springer Science+Business Media, LLC part of Springer Nature.
- Files in This Item
- There are no files associated with this item.
- Appears in
Collections - ETC > 1. Journal Articles
![qrcode](https://api.qrserver.com/v1/create-qr-code/?size=55x55&data=https://scholarworks.bwise.kr/gachon/handle/2020.sw.gachon/86195)
Items in ScholarWorks are protected by copyright, with all rights reserved, unless otherwise indicated.