I am trying to fit a model which given symptoms predicts the possible disease. We have a dataset with around 450000 rows, with patients, their vital signs, a textual description of the symptoms as the independent Variables and the diagnosis as the dependent variable. we mapped the textual complaint to a list of symptoms in UMLS codes and expanded the list to dummy variables for each accuring ymptom in the dataset and tried to train with that but with very poor results. any better approach you may propose?