Figure 1

Structure of the normative model based on adversarial autoencoders. In this configuration, the subject data is inputted into the encoder and then mapped to the latent code. This latent code is fed to the decoder with the demographic data, and then the decoder generates a reconstruction of the original data. During the training of the model, the discriminator predicts if its input data came from the latent code or if it was randomly sampled from the chosen prior distribution (e.g. Gaussian distribution). Based on these predictions, the adversarial autoencoder forces the encoder to produce a latent code similar to the prior distribution selected. Since the model is trained on healthy controls data, it is expected that it can reconstruct similar data relatively well, yielding a small reconstruction error. However, the model is expected to generate a high error when processing data affected by unseen underlying mechanisms, e.g. pathological mechanisms.