Scholarlink

The increasing global prevalence of diabetes represents a significant healthcare challenge, as it carries out a substantial burden due to chronic and acute complications affecting an estimated 415 million individuals worldwide. Predicting these complications accurately remains challenging, particularly for type 1 diabetes (T1D), due to lower prevalence in the community and smaller datasets compared to type 2 diabetes (T2D). This study developed a Deep Transfer Learning model for predicting diabetes complications to enhance the prediction of complications in T1D by incorporating knowledge derived from T2D data. Tested using the UK Biobank (comprising 6,083 T2D and 1,823 T1D cases), we employed a deep neural network trained on T2D genomic data. This training aimed to preserve learned representations crucial to diabetes and its associated complications. Subsequently, transfer learning with width and depth modifications was implemented by adapting this pre-trained model to predict complications in T1D. The study demonstrates that employing transfer learning significantly enhances the accuracy of complication prediction in T1D patients. Despite distinct pathophysiological mechanisms between T1D and T2D, the insights gained from T2D training can be valuable for improving predictions related to T1D complications.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Deep Transfer Learning Approach for Predicting Diabetes Complications Using Genomic Data

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Zelia Soo,
  • Kairui Guo,
  • Mengjia Wu,
  • Coco Huang,
  • Hua Lin,
  • Mark Grosser,
  • Yi Zhang,
  • Jie Lu

摘要

The increasing global prevalence of diabetes represents a significant healthcare challenge, as it carries out a substantial burden due to chronic and acute complications affecting an estimated 415 million individuals worldwide. Predicting these complications accurately remains challenging, particularly for type 1 diabetes (T1D), due to lower prevalence in the community and smaller datasets compared to type 2 diabetes (T2D). This study developed a Deep Transfer Learning model for predicting diabetes complications to enhance the prediction of complications in T1D by incorporating knowledge derived from T2D data. Tested using the UK Biobank (comprising 6,083 T2D and 1,823 T1D cases), we employed a deep neural network trained on T2D genomic data. This training aimed to preserve learned representations crucial to diabetes and its associated complications. Subsequently, transfer learning with width and depth modifications was implemented by adapting this pre-trained model to predict complications in T1D. The study demonstrates that employing transfer learning significantly enhances the accuracy of complication prediction in T1D patients. Despite distinct pathophysiological mechanisms between T1D and T2D, the insights gained from T2D training can be valuable for improving predictions related to T1D complications.