Osteoarthritis (OA) is a prevalent and progressive joint disease whose diagnosis from radiographs often requires expert-labeled data, which is expensive and time-consuming to obtain. Variational Autoencoders (VAEs) offer a way to learn compact, unsupervised representations that may be reused for downstream classification in low-label scenarios. In this work, we assess whether a VAE can learn latent features from hip radiographs that support OA classification with minimal supervision. We evaluate the model’s reconstruction quality, latent space structure, and diagnostic utility under label scarcity and label noise. Results show that VAE-derived features outperform raw pixel and random baselines, suggesting the latent space captures diseaserelevant structure. These findings underscore the potential of VAEs as scalable, label-efficient tools for clinical imaging tasks like OA diagnosis.