TY  - RPRT
T1  - Predicting Blood Glucose Levels of Diabetes Patients
A1  - Gülesir, Gizem
Y1  - 2018
M1  - Masters Thesis
T2  - TU Darmstadt
N2  - Time series data is used for modelling, description, forecasting, and control in many fields from
engineering to statistics. Time series forecasting is one of the domains of time series analysis,
which requires regression. Along with the recent developments in deep learning techniques, the
advancement in the technologies personal health care devices are making it possible to apply
deep learning methods on the vast amounts of electronic health data. We aim to provide reliable
blood glucose level prediction for diabetes patients so that the negative effects of the disease
can be minimized. Currently, recurrent neural networks (RNNs), and in particular the long-short
term memory unit (LSTM), are the state-of-the-art in timeseries forecasting. Alternatively, in this
work we employ convolutional neural networks (CNNs) with multiple layers to predict future
blood glucose level of a diabetes type 2 patient. Besides our CNN model, we also investigate
whether our static insulin sensitivity calculation model’s results have a correlation with basal
rate of the patient. We use the static insulin sensitivity data with our prediction model, in
order to find out whether it contributes for a better prediction or not. Our experimental results
demonstrate that calculated static insulin sensitivity values do not have any correlation with
the basal rate. Our convolutional neural network model forecasts multivariate timeseries with
multiple outputs including the blood glucose level with a 1.0729 mean absolute error for the
prediction horizon of 15 minutes.
ER  -