Heterogeneous temporal representation for diabetic blood glucose prediction

Area of Science:

• Biomedical Informatics
• Machine Learning
• Data Science

Background:

• Blood glucose prediction (BGP) aids diabetic patient monitoring and clinical decision-making.
• Multi-patient BGP is challenging due to heterogeneous and uncertain continuous glucose monitoring (CGM) data.

Purpose of the Study:

• To propose the first graph-based Heterogeneous Temporal Representation (HETER) network for multi-patient BGP.
• To address data heterogeneity and uncertainty in CGM data for improved BGP.

Main Methods:

• HETER utilizes a flexible subsequence repetition (SSR) method for data alignment.
• It constructs and learns relationships between samples as a graph to capture global temporal characteristics.
• Incorporates a temporally-enhanced mechanism and linear residual fusion for local dependencies.

Main Results:

• HETER was validated on real-world data from 112 patients across two hospitals.
• Achieved maximal improvements of 31.42% (MAE), 27.18% (MAPE), and 34.85% (RMSE) over the second-best method.
• Demonstrated robustness and accuracy in multi-patient BGP scenarios.

Conclusions:

• HETER effectively integrates global and local temporal information to mitigate heterogeneity and uncertainty in multi-patient CGM data.
• The approach shows potential for extension to other clinical tasks involving structured medical data.