Figure 4.

Analysis Workflow: (1, 2) We start by creating the patient timelines and generating drug-disease association frequencies. (3) Next, we filter drug-disease pairs to those having at least one thousand common patients. (4) This count of 492 thousand pairs roughly doubles when we aggregate up the SNOMED-CT hierarchy. (5) Then, we apply regression techniques to generate features. (6, 7) These features then feed into an SVM that classifies whether the disease in a given drug disease pair is an indication or adverse event.