Fig. 5.

Overview of our machine reading pipeline. Top: Each patient’s EHR is processed to extract the date of primary hip replacement surgery, any coded record of revision surgeries, and all clinical and operative notes. Bottom: From the patient’s coded data and primary hip replacement operative report, we tagged all mentions of implants, complications, and anatomical locations. We defined pairs of relation candidates from these sets of entities, and labeled them using data programming via the Snorkel framework. These labeled data were then used to train a deep learning model. When applied to unseen data, the final model’s final output consists of timestamped, structured attribute data for implant systems, implant-related complications, and mentions of pain