(139) An Evaluation of Patient Activity Definitions and Informative Censoring in Analyses of Real-World Overall Survival in Electronic Health Records (EHR) in the United States

Background: In a time-to-event (TTE) analysis of real-world overall survival (rwOS), some patients (pts) may not have a recorded date of death (DoD) during the follow-up period. These pts are censored at last confirmed activity date (e.g., last known date at which the patient was verifiably alive). Common methods in TTE analysis assume non-informative censoring, meaning loss-to-follow-up provides no information on the patient’s subsequent likelihood of survival; however, in real-world data (RWD), this assumption may not always be true.

Objectives: To contextualize the potential for bias due to informative censoring in a US-nationwide, oncology EHR-derived database, we aimed: (1) to evaluate the impact of abstracted and structured data vs. structured data only upon pts’ last confirmed activity date; (2) to assess the abstracted variables used to define censor date for prognostic value with respect to survival.

Methods: Analyses were completed for three tumor types: small cell lung (SCLC), advanced head & neck (advHNC), and metastatic pancreatic (mPanc), cancers. For inclusion into the study, pts’ date of initial (SCLC), advanced (advHNC), or metastatic diagnosis (mPanc) was required to be on or after Jan 1 2013, 2011, and 2014, respectively. Data through the end of 2017 were included. Among pts without a DoD (i.e., censored), two confirmed patient activity definitions (D) were evaluated using Flatiron Health EHR-derived, de-identified RWD: D1) last confirmed structured activity date (LCSAD), including the date of vitals, medication administration, or a reported laboratory test and D2) latest of LCSAD or last abstracted date (e.g., biomarker test date, abstracted treatment date). The number and percentage of censored pts with a change in their censor date using D2 vs. D1 was summarized. Among pts experiencing a change in censor date, the abstracted data type associated with the change in censor date was described.

Results: Among censored pts with SCLC (n = 1,614), advHNC (n = 2,078), and mPanc (n = 1,633), there were 162 (10%), 44 (2%), and 136 (8%) pts, respectively, who experienced a change in censor date between D2 vs. D1. Among those pts, the most common abstracted data censor date source was last clinic note date (70%) in SCLC, date of disease recurrence (98%) in advHNC, and date of hospice referral (99%) in mPanc.

Conclusions: While the addition of abstracted data impacted ≤10% of censored pts in each tumor type, the abstracted data types associated w/ D2 in advHNC and mPanc include sources of information highly likely to provide prognostic information for survival in these cohorts. These results signal the potential for informative censoring in these cohorts; further research can more fully diagnose the presence and impact of this bias.