Rapid Modeling Method For Subcutaneous Glucose Concentration Prediction For Artificial Pancreas

Online prediction of glucose concentration has been an important step in blood glucose control for diabetes. Empirical (or "data-driven") modeling techniques have been widely developed and successfully applied for glucose prediction. However, for conventional modeling methods, the work of model identification has to be repeated with sufficient data for each subject which may cause repetitive cost and burden for patients and clinicians and require a lot of modeling efforts. My motivation is how to make full use of the existing models, cost less modelling burden and reduce the amount of training data.Based on the motivations, my research work mainly focuses on the following aspects:1. The events are redefined which improves the sensitivity and specificity index for hyper/hypoglycemic alarms and overcomes the shortcomings of the existing prediction metrics. Then, the effectiveness of the universal model for the early warning of hyper/hypoglycemic episodes is explored.2. First order model migration based on marching iteration is proposed in order to overcome the disadvantages of the conventional modeling methods, which require a lot of data to train the prediction models, and the weakness of the universal model, which doesn’t consider the effects of the exogenous inputs. The proposed algorithm makes full use of the principle of variables and needs a small number of data for model training. It has strong robustness and greatly reduces the cost of modeling.3. Multiple order model migration based on particle swarm optimization is proposed in order to overcome the disadvantages of the first order model migration based on marching iteration, of which the base model is suboptimal and the marching iteration is complex for the adjustment of the exogenous inputs with multiple orders. The algorithm considers the accumulation effect of exogenous inputs and extends the applicability of model migration algorithm. However, with an increasing number of parameters which need identification, it is difficult to use principle of the exogenous inputs, which results in an increasing demand for training data. It reveals that for different data size, different methods may work the best, so an optimal model selection strategy is proposed to select optimal modeling methods in response to different data size.