CASL-VAE Learns Latent Variables from Unpaired Data for Disease Analysis

In scientific and clinical research, quantifying variability between target and reference populations, such as diseased versus healthy individuals, is a fundamental challenge. Existing methods often struggle with this, especially when paired data is unavailable and target populations exhibit diverse variations. This paper proposes CASL-VAE, a novel deep contrastive latent variable model designed to address these limitations. CASL-VAE excels at learning structured latent generative factors directly from unpaired data. Its key innovation lies in factorizing variation into two distinct types: continuous common latent factors, which are shared across different populations, and hierarchical salient latent factors. These salient factors specifically model target-specific heterogeneity, breaking it down into discrete subtypes and continuous variation within each subtype. Through variational inference, the model optimizes approximate joint likelihood across both reference and target domains using unpaired data, providing a robust framework for generating paired samples and conducting cross-domain analysis. Validation on semi-synthetic neuroimaging data demonstrated CASL-VAE's superior performance in recovering subtypes and generating paired samples compared to existing clustering and generative models. Furthermore, its application to Alzheimer's disease data successfully revealed biologically plausible heterogeneity, highlighting its potential for uncovering complex disease mechanisms.