(P-127) Identifying Dipeptidyl peptidase-4 Inhibitor Users with Sufficient Data Completeness in Electronic Health Records

Background: Electronic health records (EHRs) are rich in clinical information; however, patients may seek care in different healthcare delivery systems, which may lead to incomplete data in EHRs and biased results due to misclassification of study variables.

Objectives: To develop an algorithm to identify patients with high EHR data completeness.

Methods: This study was conducted using EHRs from a single center in Taiwan. We selected patients with at least one dipeptidyl peptidase-4 inhibitor (DPP-4i) prescription between December 2011 and December 2014. The first DPP-4i prescription was the index date. Patients were required to have at least one record in consecutive three-month intervals during 6 months before (i.e., baseline) and 24 months after (i.e., follow-up) the index date. Patient-level EHR data were linked to the National Health Insurance Research Database (NHIRD) to calculate the mean proportion of encounters captured (MPEC) by the EHR database. The following variables were selected as candidate predictors for high EHR data completeness: number of emergency room visits, any hospital admission, any weight or glycohemoglobin (HbA1c) test records during baseline and follow-up, and any prescription refill during follow-up. Data were randomly divided into two sets: 75% for training and 25% for validation. Sufficient data completeness was defined as MPEC ≥ 60%. Three models were used to develop the algorithms: lasso regression, stepwise logistic regression and decision tree model.

Results: A total of 4,950 DPP-4i users were identified in the EHR database and linked to the NHIRD; 3,713 of them were in the training set. In the lasso regression, at least one emergency visit, one hospital admission, one refill, ≥ 31 prescribing records, ≥ 12 HbA1c tests, and ≥ 24 outpatient visits during baseline and follow-up were selected; the Spearman correlation was 0.737 between the MPEC in total set and the hat value. In the stepwise logistic regression, at least one emergency visit and one hospital admission during baseline and follow-up, and one refill during follow-up were selected; the area under the curve (AUC) was 0.882 in the training set and 0.888 in the validation set. In the decision tree model, at least one emergency visit and one hospital admission during baseline and follow-up were selected; the AUC was 0.830 in training set and 0.820 in validation set.

Conclusions: All three models suggested that having at least one inpatient admission and one emergency visit during baseline and follow-up period were predictors of data completeness; prescription refill during follow-up may be a secondary predictor. External validation with other hospital EHRs is necessary to assess the generalizability of our prediction algorithms.