A predictive modeling engine using neural networks: Diabetes management from sensor and activity data

Abstract

Diabetes is a common but serious chronic disease. Nearly 8% of Americans who are aged 65 and older (about 10.9 million) suffer from this deadly disease. Self-management of this disease is possible, yet the older population lack knowledge, have denial and often lack motivation to do so. Recently we have demonstrated sensor-based network architecture within the home to monitor daily activities and biological vital parameters [25]. The data is mined to find patterns and abnormal values. Through daily text messages that are sent to the subjects, we have achieved to influence behavior change using persuasive principles. In this paper, we analyze the daily data and demonstrate that a model to profile the subject's daily behavior is possible using Artificial Neural Networks (ANN). Such a profiling has the advantage of knowing the situations, when the subject's daily activity deviates from its “normal profile”, which may be a possible indication of an onset of some health condition or disease. Lastly we develop an ANN based model to predict blood sugar level based on previous day's activity and diet intake. Such a model can be used to help a subject with high blood sugar to adjust daily activity to reach a target blood sugar level and also gives a care-giver advance notice to intervene in adverse situations.