Figure 1

Reconstruction example for Hematocrit. The mean squared error (MSE) of this reconstruction is 0.23, which is representative of the model, which has a median reconstruction MSE of 0.25. Note how high-frequency movements (e.g., downward spike at around three days) are not reconstructed, but large-scale movements are. This indicates that the model has learned to grasp the “trend” of the time series, making the feature space used for reconstruction more meaningful.