Predicting glucose concentration in type 1 diabetes patients using artificial neural networks

Abstract

Focus: Artificial neural networks may have the potential to predict blood glucose concentration in type 1 diabetes patients who are on an intensive insulin regimen, on the basis of relevant data that patients can provide themselves.

Methods: A feed-forward back-propagation ANN was constructed and trained with 308 records from a single patient, containing data about time, previous blood glucose values, carbohydrate ingestion, insulin intake, stress and physical activity. It was run multiple times to measure average performance for some feasible parameter choices.

Findings: An ANN containing a single layer of 8 neurons reported a root-mean-square error ( ± standard deviation) of rmse = 2.156 ± 0.131 mmol/l, a mean absolute error percentage ( ± SD) of mae = 19.4% ± 1.3 pp, and a correlation coefficient ( ± SD) of r = 0.662 ± 0.042. However, the data set does not generalise well.

Conclusion: While limited predictive success has been achieved, much work remains to be done before it can find practical application.