| 摘 要: 近年来,糖尿病的患病率持续上升,增加了并发症和死亡风险。因此,准确预测血糖对糖尿病患者的健康管理至关重要。针对血糖数据的复杂性和不稳定性,提出了一种基于堆叠卷积神经网络(Stacked Convolutional Neural Network,CNNS)、自注意力蒸馏机制(Self Attention Distillation，SAD)以及自适应通道(Squeeze-and-Excitation，SE)和Transformer的预测模型(CNNS-SAD-SETransformer)。模型结合了不同时间粒度的血糖数据(过去48h和2h),利用克拉克误差网格分析工具对未来30min和60min血糖预测结果进行误差分析。实验结果
表明,该模型在均方根误差(Root Mean Square Error,RMSE)等多个评估指标上的表现均优于其他预测模型,具有更好的预测性能。 |
| 关键词: 血糖预测 多粒度 自注意力蒸馏 Transformer 通道注意力 克拉克误差网格 |
|
中图分类号:
文献标识码: A
|
| 基金项目: 山东省自然科学基金项目(ZR022MF272);山东省重点研发计划项目(2019GGX101068) |
|
| Research on Blood Glucose Predicti on Based on Self-Attention Distillation and Channel Adaptive Mechanism |
|
ZHOU Wen1, XUE Jian1, HAO Fanchang1, CAO Huaimin2, HU Yu3
|
(1.School of Computer and Artificial Intelligence, Shandong Jianzhu University, Jinan 250101, China;
2.The People’s Hospital of Gaotang, Shandong Province, Gaotang 252800, China;
3.Beijing Intelligent Decision Medical Technology Co., Ltd., Beijing 100028, China)
zouzon2022@163.com; xuejian464@163.com; haofanchang18@sdjzu.edu.cn; cao.hm@163.com; 15210358424@163.com
|
| Abstract: In recent years, the prevalence of diabetes has continued to rise, increasing the risk of complications and mortality. Accurate blood glucose prediction is therefore crucial for health management in diabetic patients. To
address the complexity and instability of blood glucose data, this study proposes a novel prediction model (CNNs-SADSETransformer) integrating Stacked Convolutional Neural Networks (CNNs), a Sel-f Attention Distillation (SAD)
mechanism, and an adaptive channel mechanism (Squeeze-and-Excitation, SE) with Transformer. The model combines blood glucose data of different temporal granularities (past 48 hours and 2 hours) and employs the Clarke Error Grid
Analysis tool to evaluate prediction errors for 30 minutes and 60 minutes future blood glucose levels. Experimental results demonstrate that the model outperforms other prediction models across multiple evaluation metrics, including
Root Mean Square Error (RMSE), exhibiting superior predictive performance. |
| Keywords: blood glucose prediction multi-granularity self-attention distillation Transformer channel attention Clarke error grid |
