Validation of an International Classification of Disease (ICD) 10 Algorithm for Hypoglycemia in Administrative Claims Data

Background: Hypoglycemia is a common adverse outcome in diabetes. The ability to accurately identify severe hypoglycemia in administrative claims data can facilitate the comparative safety analysis of antidiabetics. A validated algorithm for hypoglycemia based on ICD-9 diagnosis codes exists, but the performance of an ICD-10 version of the algorithm is unknown.

Objectives: To determine the positive predictive value (PPV) of the ICD-10 version of a validated ICD-9 hypoglycemia algorithm.

Methods: This validation study was conducted as part of a multi-center retrospective cohort study initiated to compare the effectiveness and safety of oral antidiabetic drugs among patients with renal impairment. We assembled a cohort of Medicare-enrolled diabetes patients with linked electronic health record (EHR) data in the University of North Carolina (UNC) Health System from 10/2015 to 12/2017, who had ≥6 months of continuous Medicare coverage and were new users of an oral antidiabetic that met the inclusion and exclusion criteria. We mapped a validated ICD-9 hypoglycemia algorithm to ICD-10 codes and identified potential hypoglycemia events from the Medicare claims. For patients with multiple potential hypoglycemia events, we retained the first event for analysis. We confirmed hypoglycemia events by EHR review. Potential hypoglycemia events with a charted diagnosis or evidence of hypoglycemia in the EHR were validated as hypoglycemia events. We calculated the PPV and 95% confidence intervals (CI) of the ICD-10 hypoglycemia algorithm and a kappa statistic between two investigators for a subset of potential hypoglycemia events to evaluate inter-rater agreement of the EHR reviews.

Results: We reviewed 586 potential hypoglycemia events based on claims data. Health encounters associated with 267 potential hypoglycemia events (45.7%) occurred outside of the UNC Health System; therefore, charts were not available for review. Of the remaining potential hypoglycemia events (n=319), we confirmed 264 (82.8%) as true events and 55 (17.2%) as false positives. Thus, the ICD-10 hypoglycemia algorithm had a PPV of 82.8% (95% CI: [78.1%, 86.7%]). Two investigators independently adjudicated 181 potential hypoglycemia events. EHR review for hypoglycemia events had a high inter-rater agreement (kappa = 95.3%, 95% CI: [91.4%, 99.4%]).

Conclusions: The ICD-10 hypoglycemia algorithm performed well and could facilitate future research on hypoglycemia using administrative claims data and implementation of quantitative bias analysis. Further validation studies are required to evaluate the performance of the algorithm in other populations.