Glycemic-aware metrics and oversampling techniques for predicting blood glucose levels using machine learning

Abstract

Techniques using machine learning for short term blood glucose level prediction in patients with Type 1 Diabetes are investigated. This problem is significant for the development of effective artificial pancreas technology so accurate alerts (e.g. hypoglycemia alarms) and other forecasts can be generated. It is shown that two factors must be considered when selecting the best machine learning technique for blood glucose level regression: (i) the regression model performance metrics being used to select the model, and (ii) the preprocessing techniques required to account for the imbalanced time spent by patients in different portions of the glycemic range. Using standard benchmark data, it is demonstrated that different regression model/preprocessing technique combinations exhibit different accuracies depending on the glycemic subrange under consideration. Therefore technique selection depends on the type of alert required. Specific findings are that a linear Support Vector Regression-based model, trained with normal as well as polynomial features, is best for blood glucose level forecasting in the normal and hyperglycemic ranges while a Multilayer Perceptron trained on oversampled data is ideal for predictions in the hypoglycemic range.

Fig 1. A sample of a few hours of CGM data from one patient.

The horizontal red line indicates the boundry between normoglycemia and hyperglycemia according to [15]. Note (i) the two gaps in the trace, one shorter and one longer; and (ii) maximum possible sensor reading of 400 mg/dl, even though glucose levels can exceed this amount. This patient experienced hypoglycemia just after 1am followed by severe hyperglycemia later in the morning.

Fig 2. Frequency histogram showing counts of CGM sensor readings for all patients in the training data.

Different colours indicate whether or not the sensor reading is normoglycemic or not.

Fig 3. Illustration of a decision tree used for regression.

Intermediate nodes represent tests of the features and leaf nodes are predictions for ${\widehat{bg}}_{t+30}$.

Fig 4. Illustration of a MLP with a single hidden layer of size five.

Fig 5. Illustration of SMOTE’s artificial example generation technique.

Fig 6. Clarke error grid analysis, reproduced from [28].

Fig 7. Examples of a prediction made by linear SVR for one patient.

The first 120 minutes of the plot (unfilled circles) are the inputs to the model; the last reading at 150 minutes is the prediction (unfilled) and what actually happened (filled). This figure depicts a Type A error.

Fig 8. Similar example to that depicted in Fig 7, but depicting a Type E error.