Variational autoencoder generating synthetic subject FC and timeseries based on various datasets (PNC and BSNIP)
- On-line demo available at https://aorliche.github.io/DemoVAE/
- Download the models from a link in the above app
- See server directory for how to use model to generate samples
We condition a variational autoencoder on demographic data, and at the same time train it to decorrelate latent features from these demographics
This is important because many datasets are skewed with respect to demographics.
Prediction based on fMRI may be predicting the phenotype of interest OR predicting the demographic (age, sex, race, etc.) and inferring the phenotype based on the inherent demographic bias in the dataset.
In some cases, decorrelating latents from demographic information destroys predictive ability.
We find that most clinical and computerized battery fields in the PNC and BSNIP datasets are biased by demographics, and once latent features are decorrelated by DemoVAE, are no longer significantly correlated with the fMRI data.
An exception to this finding are Antipsychotic medication use in the BSNIP dataset as well as PANSS scores for severity of schizophrenia symptoms, which are still significantly correlated after demographic decorrelation.
We find the DemoVAE creates fMRI functional connectivity (FC) samples that are high in quality and represent the full distribution of subject fMRI.
Additionally, it correctly captures the group differences of the demographic features it is trained on.
Prediction using models trained on synthetic DemoVAE data is almost as good as prediction using models trained on real data.
Preprint manuscript available at: (coming soon).
Submitted to IEEE journal.
Personal website: https://aorliche.github.io/
Lab website: https://www2.tulane.edu/~wyp/
Email me