Deep transfer learning: a novel glucose prediction framework for new subjects with type 2 diabetes

• DOI: 10.1007/s40747-021-00360-7
• Publication Date: 2021-04-07T19:02:44Z
• Authors: Xia Yu; Tao Yang; Jingyi Lu; Yun Shen; Wei Lu; Wei Zhu; Yuqian Bao; Hongru Li; Jian Zhou
• Keywords: Transfer of learning; Similarity (geometry); Dynamic Time Warping

Abstract Blood glucose (BG) prediction is an effective approach to avoid hyper- and hypoglycemia, achieve intelligent management for patients with type 1 or serious 2 diabetes. Recent studies have tended adopt deep learning networks obtain improved models more accurate results, which often required significant quantities of historical continuous glucose-monitoring (CGM) data. However, new limited dataset, it becomes difficult establish acceptable network prediction. Consequently, the goal this study was design a novel framework instance-based network-based transfer cross-subject based on segmented CGM time series. Taking effects biodiversity into consideration, dynamic warping (DTW) applied determine proper source domain dataset that shared greatest degree similarity subjects. After that, method designed cross-domain personalized model combined generalization capability. In case study, clinical demonstrated additional from other subjects, proposed achieved predictions subjects