(437e) Sub-Classifying Type 2 Diabetes by Combining Nonlinear Dynamic Modeling and Multivariate Classification Methods

Type 2 diabetes is one of the biggest health problems in the United States today. This condition inhibits the body from properly converting glucose into energy. This can cause glucose levels to either run very high, which destroys capillaries and can lead to macular degeneration, kidney failure, even the amputation of a limb, or very low, which leads to diabetic coma and possibly death. According to the American Diabetes Association's website, 23.6 million Americans (or roughly 1 in 12) have this condition, and it has risen 13.5% in just the last three years.

To help people with their diabetes, we would like to be able to build a model that predicts their blood glucose at a given point in time. For this, we have chosen a block-oriented Wiener model due to its great flexibility. For this model, each input xi is passed through a dynamic linear block. This block is based on a second-order-plus-dead-time-plus-lead differential equation. The resulting vi's are then passed through a static nonlinear block, which for our model is a quadratic regression model with interactions. Our full model has 111 parameters.

The input variables used in our modeling can be broadly classified into three major groups: food, movement, and physiological. For food, we used the carbohydrates, proteins and fats consumed. For movement, we used transverse acceleration (both its peaks and its mean absolute deviation), average longitudinal acceleration, and energy expenditure. For physiological, we used heat flux, galvanic skin response, and near-body temperature. In addition to these variables, time of day, which varied from 0 (at midnight) to 1339 (at 11:59 pm) in our model, was used. The movement and physiological variables were measured using a SenseWear® Pro 3 Body Monitoring System made by BodyMedia, Inc.

At Iowa State, we performed four weeks of data collection, modeling and analysis on twenty-four type 2 diabetic subjects. K-means clustering was used to sub-classify these 24 subjects into groups based on the parameter estimates of our modeling. For this purpose, we considered the full set of 111 parameters as well as the residence times for each input. The results of this work show that diabetic subjects have a very wide range of parameters (indicative of differences in blood glucose behavior) and thus may benefit from different diagnostic, treatment, and glucose management approaches.