Purpose NLP increased identification of depression by 35%. NLP may be an efficient way to enhance identification of depression in selected research and quality improvement efforts, especially when high sensitivity is desired. As more medical providers use electronic medical records, NLP likely provide an increasingly useful and efficient tool for both clinical and research purposes. : Administrative data including diagnosis (ICD-9) codes and pharmacy codes are often used to identify depression, but sensitivity is low. Verbal analysis of office notes using Natural Language Processing (NLP) offers an alternative and additional way to identify depression when office notes are available in electronic form. This study compares administrative data versus NLP of office notes to identify diabetes patients with depression.
Methods : For 2,680 adults with diabetes, NLP was used to identify all office notes that included the imbedded term “depress.” Two reviewers (a psychiatrist and a gerontologist) confirmed appropriate identification of depression, excluding inappropriate uses of the term (“denies depression”; “depressed fracture”; “no history of depression”). When in disagreement, the two reviewers looked at the full sequence of notes to determine the appropriate decision. Depression cases identified by administrative data were compared to those identified by NLP in terms of clinical and demographic characteristics.
Results : Among 2,680 incident diabetes patients, 260 (9.7%) were identified as having depression through both methods. An additional 141 (5.3%) were identified with depression through NLP only, and 37 of 260 depression cases identified through administrative codes with no NLP evidence of depression. There were no significant differences in gender, age, comorbidity, or A1c of those identified with depression through NLP alone (N=141) compared to those identified through administrative databases (N=260). However, patients identified through the administrative databases had more primary care visits and had longer continuous enrollment in the health plan.
Conclusions: