Assessing a Predictive Modeling Technique for Proactive Patient Management of Diabetes

Howard B. Schechter, Walden University
Samir Malkani
Nithyanandam Mathiyazhagan, Walden University

Abstract

Assessing a Predictive Modeling Technique for Proactive Patient Management of Diabetes Patients with type 1 diabetes Patients with type 1 diabetes mellitus (T1DM) may achieve better glucose control when they adjust their insulin based on predicted blood glucose rather than reacting to current glucose levels. Our objective is to develop and assess a computerized model for predicting blood or interstitial glucose in T1DM patients on the insulin pump and continuous glucose sensor.

The approach drew on previous work from the organizational field of knowledge management and extended that work to provide inputs for a patient-oriented inference model. The inputs were used to drive an "Adaptive Neural Fuzzy Inference System" (ANFIS) where the artificial neural network evaluates the inputs, creates rules, and outputs a predicted glucose concentration. The inputs into the model are time of day, mealtime glucose concentration, carbohydrate intake, and insulin intake. The output is the predicted glucose at 5 minute intervals for a 120 minute period following a meal. Data sets from 2 patients over 8 weeks were used to prototype, train, and assess the model. The model was built using the grid partitioning method which produced 125 inference rules.[

The ANFIS learned as patient data were input and the actual output (blue line) vs. predicted graph (red line) became more correlated over a two hour period (Figure 1). The average error of prediction was 31 mg/dl at 30 minutes, 57 mg/dl at 1 hour, and 103 mg/dl at 2 hours. Predictability at 2 hours is less accurate than at 30 minutes because of variability in activity and insulin kinetics. We propose that a rolling prediction model used every 30 minutes will improve the accuracy of the prediction. The utility of this model would be to develop a chip which could be embedded in the pump or sensor to learn from patient data and inform behavior.