A Hybrid Dynamic Wavelet-Based Modeling Method for Blood Glucose Concentration Prediction in Type 1 Diabetes

Area of Science:

• Biomedical Engineering
• Computational Intelligence
• Endocrinology

Background:

• Diabetes Mellitus (DM) is a significant public health concern.
• Accurate blood glucose concentration (BGC) prediction is crucial for optimizing type 1 DM (T1DM) therapy.

Purpose of the Study:

• To develop and evaluate a novel hybrid modeling approach for BGC prediction in T1DM.
• To address the risks associated with hyper- and hypoglycemia through improved glucose monitoring.

Main Methods:

• A dynamic wavelet neural network (WNN) model with heuristic input selection was developed.
• Two hybrid models, HDWNN and HDFWNN, were proposed, utilizing genetic algorithms and fuzzy rule induction.
• Models were validated using real patient data and a FDA-approved T1DM simulator.

Main Results:

• The HDFWNN and HDWNN models achieved high accuracy, with gFIT = 0.97 ± 0.01 and gR² = 0.88 ± 0.07.
• Prediction errors (RMSE) for HDFWNN and HDWNN were significantly lower (11.23 ± 2.77 and 10.79 ± 3.86 mg/dl) compared to a jump NN method.
• Novel glucose-based assessment metrics (gFIT, gESODn, gR²) were introduced for performance evaluation.

Conclusions:

• The proposed hybrid models demonstrate superior performance in BGC prediction compared to existing methods.
• Fuzzy rule induction in the HDFWNN model represents a key innovation in wavelet modeling for diabetes management.
• These findings support the clinical utility of advanced predictive modeling for T1DM patients.