Presented are systems and methods for using expert knowledge to generate, train, and use a medical data model that uses medical data from a number of sources to generate likelihoods that a given set of symptoms is caused or related to one or more illnesses. Various embodiments accomplish this by parsing medical and non-medical data into keywords and target words to learn, e.g., based on a characteristic of the parsed words, an association between keywords and target words. Based on the learned associations, likelihood scores are then generated that represent, for example, a relationship between a set of symptoms and an illness, a treatment, and an outcome.