Related Experiment Video
Updated: Jun 20, 2026

A Method for Manipulating Blood Glucose and Measuring Resulting Changes in Cognitive Accessibility of Target Stimuli
Published on: August 12, 2016
GAT-BiGRU: explainable multi-task temporal graph learning for glucose forecasting, hypoglycemia risk, and
Muhammad Abdullah Sarwar1, Sarmad Maqsood1, Egle Belousovienė2
1Department of Software Engineering, Faculty of Informatics, Kaunas University of Technology, Kaunas, LT-51386, Lithuania.
This study introduces an explainable AI framework for type 1 diabetes management, accurately predicting glucose levels and hypoglycemia risk while offering personalized insulin adjustment recommendations.
Area of Science:
- Artificial Intelligence in Medicine
- Biomedical Data Science
- Diabetes Technology
Background:
- Type 1 diabetes (T1DM) management involves complex glucose variability influenced by insulin, meals, activity, and circadian rhythms.
- Existing deep learning models often treat glucose forecasting and hypoglycemia detection separately, lacking transparency.
- There is a need for integrated, explainable AI solutions for T1DM glucose management.
Purpose of the Study:
- To develop an explainable, multi-task temporal graph framework for joint glucose trajectory prediction and hypoglycemia risk assessment.
- To provide bounded, patient-specific insulin adjustment recommendations.
- To enhance transparency in AI-driven diabetes management.
Main Methods:
- Utilized a Temporal Graph Attention-BiGRU (GAT-BiGRU) framework transforming continuous glucose monitoring (CGM) data into a temporal graph.
- Employed a graph-attention encoder and BiGRU for multi-head message passing and long-term dependency capture.
- Integrated a prediction-driven counterfactual module for generating insulin adjustments (basal/bolus) using patient-specific Insulin-to-Carbohydrate Ratio (ICR) and Insulin Sensitivity Factor (ISF), with GNNExplainer for interpretability.
Main Results:
- Achieved high accuracy in glucose prediction (e.g., MAE 9.40 mg/dL on OhioT1DM) and hypoglycemia detection (PR-AUC 0.93 on OhioT1DM).
- Demonstrated strong performance on the BrisT1D dataset (PR-AUC 0.98, MAE 9.36 mg/dL).
- The insulin module provided personalized recommendations, including ~10% basal adjustments and individualized meal-bolus updates.
Conclusions:
- The Temporal GAT-BiGRU framework offers accurate glucose prediction and hypoglycemia risk assessment through temporal graph reasoning and sequence modeling.
- The system provides interpretable explanations, supporting personalized decision-making for diabetes management.
- This approach facilitates advanced personalized decision support and advancements in closed-loop glucose management systems.
Related Concept Videos
Hypoglycemia and Glucagon
Hypoglycemia
Glucose Homeostasis: Regulation of Blood Glucose
During fasting, when blood glucose levels are low, the pancreas secretes glucagon. it...
Hyperglycemia
Hormones Regulating Blood Glucose
In addition to accelerating glucose uptake and utilization, insulin has...
Glucose Homeostasis: Pancreatic Islets and Insulin Secretion
Insulin and C-peptide are co-secreted in...
