Diabetes is a data-intense disease affecting 400 million people worldwide. People with duabetes track glucose, insulin, meals, medications, sleep and activity to make decisions on an event-by-event, hour-by-hour basis. The promise of cognitive computing is to liberate us from the tedious task of data tracking, by contextualizing personalized insights specific to our situation. Medtronic has partnered with IBM Watson Health to advance research into cognitive computing for diabetes and build solutions that leverage data to provide patients with insights that they can use to understand how to better manage their disease. Our cognitive app with IBM Watson will serve as a personal assistant for people with diabetes – by uncovering important patterns and trends using a retrospective analysis of patients’ insulin, continuous glucose monitors and nutritional data – to help people understand how their behavior affects their glucose level in real time.

A key challenge in diabetes management is glucose control – ensuring tight glucose control, avoiding hypo and hyperglycemia (low and high blood sugar, respectively), and reducing complications. The challenge is in correctly estimating the impact of meals, activity, exercise and medication. We retrospectively analyzed anonymous patients’ insulin, CGM and nutritional data (n=10,000) with IBM Watson. Preliminary results suggest that the Watson algorithms are able to accurately predict onset of a potential hypoglycemic within a two-to-four-hour window with AUC of 75-86% of the actual hypoglycemic events. This has enormous potential to help people better understand how their behavior affects their glucose level in real time, and could, in the future, analyze these factors by integrating data from a wearable activity tracker or nutrition app. The algorithm clusters individuals into smaller groups based on anonymous demographic and behavioral profiles, to improve prediction accuracy. We continue to expand our dataset and plan to include auxiliary data (nutrition, fitness, other contextual data) to further improve the prediction performance.