Out-of-distribution (OOD) detection is a crucial step in ensuring robustness and trustworthiness for deep learning (DL) models. This is particularly important for classification tasks in medical applications. Previous work has shown that point-of-care ultrasound (POCUS) imaging combined with a DL classification network could improve access to early breast cancer diagnosis. In this work, we investigate the ability of a conditional latent space variational autoencoder (CLVAE) to detect OOD samples by learning to cluster in-distribution (ID) data. To efficiently cluster the ID data, three types of feature extraction were used as a preprocessing step to the CLVAE. The three different CLVAE models were evaluated on the task of distinguishing four different OOD datasets from the ID data, i.e., breast POCUS images. Three of the OOD datasets were designed to resemble real-world OOD scenarios for breast POCUS imaging: (1) ultrasound images of arteries (non-breast); (2) blurry breast POCUS images; (3) POCUS images with contact artifacts. The best CLVAE performance was achieved by extracting features using an existing ultrasound foundation model, yielding an area under the curve (AUC) of 98.7%, 99.8%, and 88.2% for each of the above-mentioned OOD datasets respectively.