Summary

In this chapter, we explored how to use neural networks for time series prediction. We used a case study in healthcare: predicting future glucose levels.

Since this is a book on time series analysis and not on deep learning, we adopted a simple neural architecture: a fully connected feedforward neural network with two hidden layers. We used the past 36 values of glucose levels to predict the next six glucose values. Among the next six, we detected the maximum and minimum values and triggered an alarm if one of them exceeded the recommended boundaries.

Before showing the implementation of the neural network in KNIME Analytics Platform, we recapped the basics of neural networks and their original training algorithm: backpropagation.

Then, we looked at how to install and configure KNIME Deep Learning – Keras Integration, which allows us to use the Keras deep learning libraries from the familiar KNIME interface.

Then, we continued with the practical implementation...