Cerebrospinal fluid proteomics can identify biomarkers associated with Alzheimer’s disease (AD), but high dimensionality and confounding (e.g., age, APOE4, site effects) can obscure disease- and sex-related signals. We developed an autoencoder-based framework to learn robust, interpretable representations while explicitly modeling sex-associated variation. CSF proteomics from ADNI (N = 515 training, 111 validation, 111 test; 7032 proteins) were adjusted protein-wise for age, APOE4 genotype, and collection site using linear models or generalized additive models, with residualization applied as appropriate. We trained a standard autoencoder (AE) and a conditional variational autoencoder (CVAE) conditioned on sex. Latent representations were evaluated for diagnosis prediction (CN/MCI/AD) and interpretability.