Negative results for the prediction of postprandial hypoglycemias from insulin intakes and carbohydrates : analysis and comparison with simulated data

Dubosson, Fabien (University of Applied Sciences and Arts Western Switzerland (HES-SO Valais-Wallis)) ; Mordanyuk, Natalia (eXiT group, University of Girona) ; López, Beatriz (eXiT group, University of Girona) ; Schumacher, Michael (University of Applied Sciences and Arts Western Switzerland (HES-SO Valais-Wallis))

Diabetic patients usually take insulin bolus right before eat-ing a meal. A wrong dosage of insulin may lead to a hypoglycemia. Be-ing able to anticipate such insulin-induced, postprandial hypoglycemias would enable warning of the patients about the risk associated with the quantity of insulin they are planning to take. In this work, we explore the feasibility of predicting these postprandial hypoglycemias by using information available at pre-meal time, such as glucose levels, planned insulin intakes and carbohydrates estimations. First, an experiment has been done on a dataset acquired on real patients, for which several classes of machine learning algorithms have been tried. The obtained results do not offer predictions that are useful enough to consider any usage in real-life applications. These kinds of datasets — acquired on real patients — suffer heavily from missing data and incorrect carbohydrates estimations though. In order to analyse the impact of these flaws on the obtained results, the same experiment has been run on a simulated dataset. Re-sults support that even with the simulated dataset, which does not have missing data and which has precise carbohydrates intake, these features alone are not able to predict postprandial hypoglycemia. Therefore, im-proving the quality of patients annotations is not enough to solve the problem, and using these features without further features engineering does not offer good results.

Conference Type:
full paper
Economie et Services
Institut Informatique de gestion
Vienna, Austria, 24 June 2017
Vienna, Austria
24 June 2017
5 p.
Published in:
Proceedings of the Second Workshop on Artificial Intelligence for Diabetes
Appears in Collection:

Note: The status of this file is: restricted

 Record created 2017-11-19, last modified 2020-10-27

Download fulltext

Rate this document:

Rate this document:
(Not yet reviewed)