Use of decision support systems in the hospital to manage glucose
Implementation of clinical decision support tools
Biosensing, diabetes data science and decision support
Abstract
Background and Aims
Real-time data sources are abundant and doubling every two years, with many based on biosensors and relevant to healthcare. This trend is particularly prominent in diabetes, where continuous glucose monitoring (CGM) revolutionized the assessment of key clinical parameters and opened a new research and training field–Diabetes Data Science.
Methods
The high temporal density of CGM data requires new metrics and analytics that take advantage of contemporary methods, including artificial intelligence, machine learning, and data farming. To be relevant, any contemporary CGM-based decision support system (DSS) needs to utilize these types of approaches.
Results
This presentation continues the review of methods enabling automated DSS that we initiated a year ago, and discusses updates, such as the use of a stochastic-process approximation of the day-to-day variation in human physiology and behavior to individualize and optimize diabetes control. Pattern recognition, clustering, and classification are at the core of this approach, in which the collection of all possible daily patterns, is defined by machine-learning classification of CGM profiles into clusters derived from a library of over 50,000 CGM daily profiles accumulated by recent clinical trials.
Conclusions
CGM daily profiles can be classified into identifiable clusters and within each cluster the individual CGM profiles are similar. Over time, each person transitions through a sequence of clusters, which contribute differently to glycemic control. Treatment adaptation is assisted by presenting pattern/risk predictions, with the goal to maximize the probabilities for favorable daily transitions.