Thyroid nodules are more and more widespread, with the frequency of thyroid cancers increasing accordingly. In routine practice, the clinician’s goal is to determine if a nodule is malignant or suspicious, and thus requires surgery. However, nearly one-third of thyroid nodules fall into the ‘indeterminate’ classification segment, often leading to the use of surgery as a diagnostic tool, when it could have been avoided with a better preoperative diagnosis. Molecular tests exist to address this issue, but they are markedly onerous. In this paper, we propose an accurate, low-cost method to predict the malignancy of thyroid nodules using a categorical thyroid nodule medical dataset. The prediction is achieved through the use of a new supervised autoencoder neural network. Results show that the supervised autoencoder outperforms classical classification methods on all evaluation metrics. A proof-of-concept Graphical User Interface was developed to demonstrate the medical usefulness of the latent space of the trained network. Aside from the prediction itself, the supervised autoencoder neural network provides three medically interesting benefits: a prediction confidence score, consistent feature ranking, and an interpretative latent space. All of those benefits are valuable from the clinician’s point of view, as they allow for an efficient, low-cost, and accurate diagnosis.