DiGAN breakthrough: advancing diabetic data analysis with innovative GAN-based imbalance correction techniques

Zhao, Puyang, Liu, Xinhui, Yue, Zhiyi, Zhao, Qianyu, Liu, Xinzhi, Deng, Yuhui and Wu, Jingjin (2024) DiGAN breakthrough: advancing diabetic data analysis with innovative GAN-based imbalance correction techniques. Computer Methods and Programs in Biomedicine Update, 5. ISSN 2666-9900

Text (Liu_digan-breakthrough--published) - Published Version Available under License Creative Commons Attribution Non-commercial. Download (1MB)

• Scopus publication

Abstract

In the rapidly evolving field of medical diagnostics, the challenge of imbalanced datasets, particularly in diabetes classification, calls for innovative solutions. The study introduces DiGAN, a groundbreaking approach that leverages the power of Generative Adversarial Networks (GAN) to revolutionize diabetes data analysis. Marking a significant departure from traditional methods, DiGAN applies GANs, typically seen in image processing, to the realm of diabetes data. This novel application is complemented by integrating the unsupervised Laplacian Score for sophisticated feature selection. The pioneering approach not only surpasses the limitations of existing techniques but also sets a new benchmark in classification accuracy with a 90% weighted F1-score, achieving a remarkable improvement of over 20% compared to conventional methods. Additionally, DiGAN demonstrates superior performance over popular SMOTE-based methods in handling extremely imbalanced datasets. This research, focusing on the integrated use of Laplacian Score, GAN, and Random Forest, stands at the forefront of diabetic classification, offering a uniquely effective and innovative solution to the long-standing data imbalance issue in medical diagnostics.

Item Type:	Article
Official URL:	https://www.sciencedirect.com/journal/computer-met...
Additional Information:	© 2024 The Authors
Divisions:	Statistics
Subjects:	R Medicine Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Date Deposited:	29 Apr 2024 12:57
Last Modified:	15 Nov 2024 07:21
URI:	http://eprints.lse.ac.uk/id/eprint/122841

Actions (login required)

View Item