Setting limits on neural network's predictive capacity in T1D blood glucose concentration
The ability to anticipate changes in blood glucose (BG) concentration would have a great impact on Type 1 diabetics (T1D). In order to create T1D treatment plans, patients collect a BG concentration time series. It has been demonstrated that various types of recurrent neural networks, such as Long Short Term Memory (LSTM), have success forecasting T1D BG concentrations. However, limited work has been done to characterize the T1D time series or set limits on neural network's predictive capacity. In this work, a T1D patient's 14 day BG concentration time series is studied. First, I test the time series stationarity. Then I use auto-correlation analysis, spectral analysis, and Gaussian process regression to characterize the T1D BG time series. Finally, the LSTM's prediction quality is quantified and interpreted at different prediction intervals. The success or failure of the LSTM's predictions are interpreted using the characterization of the time series.