OBJECTIVES : In health outcomes research using real-world data, the lack of definitive variables for discontinuation of treatment may create bias in the estimation of treatment duration. This study aimed to develop an algorithm to identify patients who are on treatment at the end of the database using electronic medical records (EMR).

METHODS : A retrospective analysis of EMR data (Advanced Gastric Cancer Cohort, Flatiron Health, NY) was conducted among patients initiating intravenous (IV) cancer therapy from 1/1/2013 through 3/31/2017. Patients were defined as being on treatment who were alive and whose last IV drug was administrated within a time window prior to the study end date. Different time windows (from 15 days through 60 days by 5-day increments) were evaluated and validated using refreshed data with longer follow up (by 05/30/2017). Cohen’s Kappa was used for agreement against the refreshed data using SAS 9.4.

RESULTS : Among 1407 eligible patients, 700 received treatment with a mean age of 67.5 (SD=12.0); 70.1% were male. The proportion of patients on treatment ranged from 16.1-22.4% depending on the time window applied to the dataset. The 25-day time window was identified to have the highest Kappa, in which 133 out of 700 (19.0%) patients were correctly classified as being on treatment. The mean (SD) and median of treatment duration for all 700 patients were 96.3 (112.4) and 64 days. After excluding patients using the 25-day rule, the mean (SD) and median were 86.7 (93.5) and 57 days.

CONCLUSIONS : In this study, 25-day prior to the end of follow up was found to be the best time window to determine patients who remained on treatment at the end of the database. The accuracy of determining discontinuation of therapy will improve the accuracy of treatment duration estimates and reduce bias in health outcomes research.