Scholarlink

Diabetes disrupts the regulation of blood glucose levels, causing hyperglycemia (high glucose levels) and hypoglycemia (low glucose levels). Hypoglycemia, in particular, can result in cognitive impairment, seizures, and loss of consciousness, making its prediction and avoidance crucial for enhancing patients’ quality of life. With the help of pattern recognition techniques and the relatively widespread use of continuous glucose monitoring (CGM) devices, intelligent systems can be developed, which in turn can predict hypoglycemic events in advance. In the present paper we propose to use a Long Short-Term Memory (LSTM) model and a Graph Attention Network (GAT) for this task and we evaluate both models on a real-world CGM dataset of 37 type 1 diabetes persons. In an empirical evaluation, we observe that the LSTM model excels in short-term predictions (15 min), whereas the GAT model performs better for longer horizons (30 and 60 min). Additionally, the GAT model demonstrates more stable performance across all horizons and lower variability across different datasets compared to the LSTM.

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

LSTM Networks and Graph Neural Networks for Predicting Events of Hypoglycemia

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Fabian Hüni,
  • Jose Garcia-Tirado,
  • Kaspar Riesen

摘要

Diabetes disrupts the regulation of blood glucose levels, causing hyperglycemia (high glucose levels) and hypoglycemia (low glucose levels). Hypoglycemia, in particular, can result in cognitive impairment, seizures, and loss of consciousness, making its prediction and avoidance crucial for enhancing patients’ quality of life. With the help of pattern recognition techniques and the relatively widespread use of continuous glucose monitoring (CGM) devices, intelligent systems can be developed, which in turn can predict hypoglycemic events in advance. In the present paper we propose to use a Long Short-Term Memory (LSTM) model and a Graph Attention Network (GAT) for this task and we evaluate both models on a real-world CGM dataset of 37 type 1 diabetes persons. In an empirical evaluation, we observe that the LSTM model excels in short-term predictions (15 min), whereas the GAT model performs better for longer horizons (30 and 60 min). Additionally, the GAT model demonstrates more stable performance across all horizons and lower variability across different datasets compared to the LSTM.