OBJECTIVES: Diagnostic algorithms to detect undiagnosed patients with rare diseases have the potential to improve patient health and reduce costs associated with misdiagnosis.  Accurate algorithms are difficult to develop, typically having to overcome Challenges including the tendency for models to I) over-fit, arising from low degrees of freedom / high-dimensionality and II) under-predict the rare disease, arising from the low ratio of confirmed to unconfirmed cases (skewed outcome class distribution). Support Vector Machines (SVMs) are a highly successful class of machine learning algorithms with well-established methods for handling high-dimensionality and skewed outcome class distribution. SVMs therefore represent a promising method to detect patients with undiagnosed rare diseases. This study estimated risk scores for a rare oncology disease using SVMs. The performance of the models was compared to classical methods based on logistic regressions. METHODS: Risk scores for confirmed diagnosis were estimated with logistic regressions (standard, weighted and Firth) and SVMs (regularization, weights and kernel parameters were optimized using internal cross-validation). Patients with high risk scores and without a confirmed diagnosis have a higher probability of being undiagnosed cases.  Model development, validation and testing were carried out on separate random samples from linked primary and secondary care data in the UK (Clinical Practice Research Datalink and Hospital Episode Statistics).  The key performance metric was maximizing Sensitivity at a Positive Predicted Value (PPV) of 10% based on test data. RESULTS: