Scholarlink

<p>Diabetes mellitus is an endocrine disorder and one of the most fatal chronic diseases. Diabetes patients’ inability to control blood glucose (BG) at normal levels can damage blood vessels and nerves, leaving irreparable consequences. Therefore, continuous BG monitoring is essential for these patients. Various studies on BG level prediction reveal its necessity and can help patients receive timely treatment. Deep learning-based techniques have significantly advanced BG prediction, among which long-short-term memory (LSTM) and temporal convolution networks (TCNs) have shown exemplary performance. However, LSTM suffers from the vanishing gradient problem over long-term sequences, and TCNs are not ideal for long-term dependency recognition. To address these problems, we incorporate a self-attention method with TCN into the feature extraction process to capture dependencies more efficiently. This method targets two sets of clinical and simulated type 1 diabetes (T1D) data at prediction horizons (PHs) of 30 and 45&#xa0;min. Our experiments show that the model, relying on root mean square error (RMSE), mean absolute relative difference (MARD), and coefficient of determination (<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="13042_2025_2758_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="21" /> </InlineMediaObject> <EquationSource Format="TEX">\(R^2\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation>) for PH 30 and 45&#xa0;min, outperforms recent state-of-the-art approaches. By combining the TCN and self-attention, we improve BG prediction accuracy by capturing both local and global temporal correlations in time-series data.</p>

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

Blood glucose prediction for type 1 diabetes based on self attention-TCN module

  • 期刊论文

  • 发表日期：

    2025年8月1日

查看全文
  • Leila Talebi Jouneghani,
  • Mohammad Ahangarkiasari,
  • Amirhossein Nikoofard

摘要

Diabetes mellitus is an endocrine disorder and one of the most fatal chronic diseases. Diabetes patients’ inability to control blood glucose (BG) at normal levels can damage blood vessels and nerves, leaving irreparable consequences. Therefore, continuous BG monitoring is essential for these patients. Various studies on BG level prediction reveal its necessity and can help patients receive timely treatment. Deep learning-based techniques have significantly advanced BG prediction, among which long-short-term memory (LSTM) and temporal convolution networks (TCNs) have shown exemplary performance. However, LSTM suffers from the vanishing gradient problem over long-term sequences, and TCNs are not ideal for long-term dependency recognition. To address these problems, we incorporate a self-attention method with TCN into the feature extraction process to capture dependencies more efficiently. This method targets two sets of clinical and simulated type 1 diabetes (T1D) data at prediction horizons (PHs) of 30 and 45 min. Our experiments show that the model, relying on root mean square error (RMSE), mean absolute relative difference (MARD), and coefficient of determination ( \(R^2\) R 2 ) for PH 30 and 45 min, outperforms recent state-of-the-art approaches. By combining the TCN and self-attention, we improve BG prediction accuracy by capturing both local and global temporal correlations in time-series data.