Table 1.
Data Groups Characteristics.
Figure 1.
Schematic configuration of the experimental setup.
The system consists of three infusion syringe pumps for [3-3H] glucose, insulin and variable glucose respectively. Arterial catheter is connected to the infusion pumps, and venous catheter is used for manual blood sampling. A closed-loop, computer controlled system is proposed for maintaining plasma glucose concentration within the desired level during HEGC.
Figure 2.
Hyperinsulinemic-euglycemic glucose clamp experiment (HEGC) protocol – Schematic illustration.
Representation of the experimental design of the clamp study: The animals were studied under basal conditions for the first 2 hours and under hyperinsulinemic conditions over the last 2 hours. Period–I is characterized by rapid changes in glucose concentration. Period-II exhibits a near steady state behavior of insulin and plasma glucose concentrations. Circles represent times at which blood samples were taken.
Table 2.
Network Parameters and Ranges.
Figure 3.
Glucose pump controller design stage block diagram.
The output of this stage is an ensemble of 50 sets of Artificial Neural Network (ANN) connection weights, created using the Test set of data and the best fit parameters vector.
Table 3.
Animals Characteristics.
Figure 4.
Real time glucose pump controller block diagram.
In each time slot in a real time experiment, an input vector f(Pi,Gi) is calculated where Pi is the pump setting and Gi is the blood glucose level of step i. The controller’s prediction output Pi+1 is derived as the median of 50 predictions.
Table 4.
Optimized Parameters and Best Performance.
Figure 5.
Artificial neural networks model predictive performance.
Plasma glucose concentration and experimental glucose infusion rate during hyperinsulinemic-euglycemic glucose clamp phase in comparison to target values (data presented for animal TG11).
Figure 6.
Regression analysis between the predicted and desired values of the ANN glucose pump controller.
Performance results of the Test set simulation. A: ANN trained using Levenberg-Marquardt (LM) optimization algorithm, B: ANN trained using Gradient-Descent with momentum and adaptive learning rate algorithm.
Figure 7.
Evaluation of the error in the prediction of glucose infusion rate over different levels of random noise.
Random Gaussian density function noise with zero mean, and variance corresponding to signal to noise ratios (SNR) of 5 dB to 35 dB was added to the input data. The prediction error is expressed in mean ± SEM over 100 simulations.
Figure 8.
Regression analysis between the predicted and desired values calculated using feedback control algorithm.
The feedback control algorithm of DeFronzo et al. [5] was used for performance comparison, with a sampling rate of 10 minutes interval.