Diabetes Phenotyping Using the Electronic Health Record
- PDF / 162,876 Bytes
- 3 Pages / 595.276 x 790.866 pts Page_size
- 3 Downloads / 207 Views
J Gen Intern Med DOI: 10.1007/s11606-020-06231-0 © Society of General Internal Medicine 2020
INTRODUCTION
Accurate electronic health record (EHR) identification of diabetes is necessary for promoting best practices, clinical decision support, and population health. However, simple definitions have limitations: laboratory-based definitions may miss patients with well-controlled diabetes, medications would miss diet-controlled patients or erroneously capture those taking a “diabetes medication” for another indication, and International Classification of Diseases (ICD) codes require accurate clinician coding. While prior work has examined laboratory work, diabetes medications, and ICD-9 codes,1–3 they have not incorporated newer ICD-10 coding. This study evaluated whether combination of EHR criteria could create a diabetes patient registry that would be useful to clinical practices. Pitfalls from false positives and negatives were clarified with chart review.
METHODS
We conducted a retrospective cohort study of patients aged 18–85 with inpatient or outpatient encounters from January 1, 2014, to November 15, 2016, to the NYU Langone Health System. We created six clinical phenotype definitions using Epic EHR data: [1] type 2 diabetes ICD-10 code listed at least once in either the problem list, medical history, or encounter; [2] laboratory work (hemoglobin A1C ≥ 6.5); [3] active or discontinued outpatient diabetes medication; [4] presence of any of the above; [5] presence of either [1] or [2]; and [6] presence of either [1] or [3]. The EHR pull identified 62,661 patients who met the ICD-10 code definition, 36,151 patients who met the laboratory work definition, and 30,855 patients who met the diabetes medication definition. Together, these represented 76,114 unique patients. Of these, we sampled 106 charts. We also identified 1000 random charts that met none of the above diabetes definitions; we sampled 99 charts from this latter group. Four physicians manually reviewed the total 200 charts to validate each diabetes definition. Reviews were Prior Presentation Parts of this study were presented in abstract form at the SGIM 2018 Annual Meeting, April 12, 2018, Denver, CO. Received August 19, 2019 Accepted September 10, 2020
completed independently, with an overlap of 51 medical records that demonstrated 88% agreement (κ = 0.43). Differences were adjudicated by H.M.W. To estimate clinical phenotype definition accuracy using chart sampling, we used the methods described by Rosenman and colleagues4 to calculate sensitivity, specificity, and positive predictive value (PPV). Physician review was the gold standard in these calculations. The sensitivity of ICD-10 codes alone was compared to each of the other EHR-based phenotypes using the delta process4. Qualitative reviews of false positives and negatives were performed with chart review.
RESULTS
Of the 205 charts selected for review, 46% (94 out of 205) were adjudicated as having diabetes. This included 3 of the 99 charts from the group meeting none of the diabetes defin
Data Loading...
