Short-term prediction method of blood glucose based on temporal multi-head attention mechanism for diabetic patients. (April 2023)

• Record Type: Journal Article
• Authors: Yang, Guanci; Liu, Saisai; Li, Yang; He, Ling
• Is Part Of: Biomedical signal processing and control. Volume 82(2023)
• Publication Date: 2023-04
• Subjects: Blood glucose prediction -- Multi-head attention mechanism -- Spatiotemporal features -- Sequence data; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104552
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Temporal multi-head attention model to obtain global and local spatiotemporal features. Unsupervised outlier detection algorithm for abnormal blood glucose without prior knowledge. Imputation algorithm based on feature engineering to fill missing blood glucose values. The proposed method can attain comparatively good performance. Abstract: The hyperglycemic state of people with diabetes can lead to metabolic and healthy disturbances in the body. Diabetes is mainly treated clinically by conservative treatment, which requires frequent and continuous measurement of blood glucose concentration. Accurate blood glucose prediction plays an important role in the future blood glucose management of patients. To improve the accuracy of blood glucose prediction, this paper proposes a short-term prediction method of blood glucose based on temporal multi-head attention mechanism for diabetic patients (PBGTAM). Firstly, a detection algorithm of abnormal blood glucose based on Adaptive Density-Based Spatial Clustering of Applications with Noise is proposed by using an autonomous neighborhood parameter selection method. Secondly, the imputation algorithm based on feature engineering is proposed to fill the missing blood glucose values. Thirdly, we propose temporal multi-head attention model to obtain global and local spatiotemporal features from sequence data, in which a temporal series features module is designed to insight the detail feature so as to keep the useful information contained in the details, and an intensity correlation module is used to obtain global features of sequence data and then to achieve strong correlation between blood glucose data, and a sequence features module is employed to learn temporal sequence features between sequence data to guarantee continuous prediction of future blood glucose values. Finally, the comparative results show that PBGTAM outperforms six state-of-the-art methods in terms of the overall prediction root mean square error, forecasting accuracy for warning, and clinical consistency with 20.57, 84.35% and 85.18%, respectively.