Abstract
INTRODUCTION: To improve the utilization of continuous- and flash glucose monitoring (CGM/FGM) data we have tested the hypothesis that a machine learning (ML) model can be trained to identify the most likely root causes for hypoglycemic events. METHODS: CGM/FGM data were collected from 449 patients with type 1 diabetes. Of the 42,120 identified hypoglycemic events, 5041 were randomly selected for classification by two clinicians. Three causes of hypoglycemia were deemed possible to interpret and later validate by insulin and carbohydrate recordings: (1) overestimated bolus (27%), (2) overcorrection of hyperglycemia (29%) and (3) excessive basal insulin presure (44%). The dataset was split into a training (n = 4026 events, 304 patients) and an internal validation dataset (n = 1015 events, 145 patients). A number of ML model architectures were applied and evaluated. A separate dataset was generated from 22 patients (13 'known' and 9 'unknown') with insulin and carbohydrate recordings. Hypoglycemic events from this dataset were also interpreted by five clinicians independently. RESULTS: Of the evaluated ML models, a purpose-built convolutional neural network (HypoCNN) performed best. Masking the time series, adding time features and using class weights improved the performance of this model, resulting in an average area under the curve (AUC) of 0.921 in the original train/test split. In the dataset validated by insulin and carbohydrate recordings (n = 435 events), i.e. 'ground truth,' our HypoCNN model achieved an AUC of 0.917. CONCLUSIONS: The findings support the notion that ML models can be trained to interpret CGM/FGM data. Our HypoCNN model provides a robust and accurate method to identify root causes of hypoglycemic events.