Project: Type 1 Diabetes (T1D) using Metabolic Chamber data
Goal: Building a multilevel perceptron for predicting temporal need for insulin towards maintaining glucose levels in Type 1 Diabetes patient.
Data: Metabolic chamber (a controlled environment) data on activity levels, food intake, insulin bolus, glucose
Research involved: Application of existing framework NARX-ANN (Nonlinear Auto Regressive Exogenous – Artificial Neural Network) in Matlab in the research domain of T1D.
Method: Recurrent Neural Nets using Feed Forward Architecture with back propagation and gradient descent used for training. The ANN output layer has a neuron that estimates the patient Glucose levels for future time samples. The neurons of the ANN input layer receive Insulin boluses, exercise levels, dietary intake and Glucose level values that are measured from previous time samples. This way, after an appropriate training process the NARX-ANN can learn the dynamics system. Mathematically, an estimate of Glucose value ( ) , for each sample time k, is calculated as a function of an exogenous input vector , with multi-dimensional input vector , glucose as an output vector , and generated non-linear hidden layer neuron outputs , j:{1…J}, with J hidden layers.
Motivation: Doable within 2 months due to existing framework. An experimentally controlled environment will give an opportunity to study the true dynamics among the variables under study. Assess performance of this approach among other approaches that our lab is focusing on.
Reference: Iriyogen et al. 2013. A NARX neural network model for enhancing cardiovascular rehabilitation therapies. Neurocomputing 109 (2013) 9–15