842-P: Analysis of Primary Care Provider (PCP) EHR Notes for Discussions of Prediabetes Using Natural Language Processing (NLP) Methods

Background: Prior studies using structured EHR data suggest that patients with prediabetes are not receiving evidence-based care, but prediabetes may be discussed in unstructured data. We developed and validated an NLP tool to identify discussions about prediabetes in EHR notes. Methods: We included adult patients without diabetes with an in-person office visit at any general medicine clinic at a single academic center and at least one HbA1c 5.7-6.4% between 7/1/2016 and 12/31/2018. In phase 1, we devised an initial keyword search strategy based on clinical experience. We extracted notes with at least one of the keywords and manually annotated them to determine whether they represented clinical discussions of prediabetes. We used the annotated notes to train and evaluate multiple machine learning and deep learning classifiers to replicate human annotation. In phase 2, we applied a similar annotation process and machine learning method on notes from a different group of clinic practices. We analyzed notes from phase 2 to describe the content of prediabetes discussions: labs ordered or reviewed (HbA1c or fasting glucose); lifestyle counseling; diabetes prevention program (DPP) discussion/referral; nutrition referral/discussion; and metformin discussion or ordering/continuation. Results: We identified 269 patients with prediabetes discussions in phase 2. Most commonly, PCPs provided lifestyle counseling (80%), reviewed current labs (63%) and ordered follow-up labs (60%). PCPs rarely discussed/referred to a nutritionist (4%). We did not find any discussions/referrals to a DPP. Metformin was discussed, ordered or continued in