Figure 9.

This figure shows the effects of differentially privacy respecting (DP) Variational Autoencoders for Heterogeneous and Incomplete Data (HI-VAE) training on the HI-VAE step of the model. (Left) reconstruction loss change between DP and conventional model training for laboratory data at visit 1 for the SP513 study; (middle) epsilon plotted against reconstruction loss for different delta values; (right) epsilon over 500 epochs, given different deltas. A noise multiplier of 1.1, norm clipping at 1.6, and a learning rate of 0.01 were used. Further plots can be found in Section E of the Supplementary Materials.