Introduction: Hypoglycemia is a potentially life-threatening complication of diabetes that can result in seizures, coma, and death, if left untreated. We have built a machine learning system that can predict occurrences of low glucose events from continuous glucose monitoring (CGM) data within 2 to 4 hours before an event.

Methods: IBM Research scientists developed a machine learning system based on a random forest model to predict from data, provided by Medtronic Guardian™ Connect CGM, whether or not a patient’s glucose level will fall below 70 mg/dl, which was defined as a hypoglycemic event. The model was built based on de-identified data collected from voluntary uploads that measured sensor glucose (SG) levels of users every 5 minutes. The system employed 120 million SG data points collected across many years from approximately 10,000 patients. All the patients in the data set were type 1 diabetics who had used a CGM device for at least 11 days and for a median of 272 days. The training features of the ensemble classifier were extracted from the users’ temporal glucose measurements and past event patterns. This was engineered to account for short- and long-term effects of glucose and seasonal elements observed in multiple patients. The random forest method was selected for its robustness, its accuracy, and its ease of maintenance. The data were divided into 80% training and 20% holdout test sets. The model performance was assessed by area under the curve (AUC) of the receiver operating characteristic (ROC) curve and the lead time, which is the time elapsed between the first alert and the time of occurrence of the predicted event.

Results: The model performed with AUC of 0.9, and it is capable of alerting patients more than an hour lead time in 80% of events.

Conclusions: Managing and avoiding hypoglycemia is significant challenge for patients. Systems using modern machine learning techniques have the potential to improve the care of individuals with diabetes and reduce life-threatening complications for these patients.


P. Agrawal: Employee; Self; Medtronic MiniMed, Inc. Stock/Shareholder; Self; Medtronic. L. Cao: None. Y. Chang: None. G.P. Jackson: Employee; Self; IBM. S. Kefayati: None. V. Pavuluri: None. S. Sathe: None. D.S. Turaga: Other Relationship; Self; Medtronic. R. Ngueyep: None. E. Vitkin: Employee; Self; IBM. L. Vu: None. Y. Zhong: None.

Readers may use this article as long as the work is properly cited, the use is educational and not for profit, and the work is not altered. More information is available at