Evaluation of Mortality Data Sources Compared to the National Death Index in the Healthcore Integrated Research Database

Background: Generating fit-for-purpose real-world evidence depends on the underlying data source's quality and information diversity. Mortality ascertainment is a challenge in non-interventional studies given the lack of comprehensive mortality ascertainment in administrative claims or electronic medical records and the under-capture of death events in the Social Security Administration Death Master File (DMF). The National Death Index (NDI) is considered a gold standard for mortality ascertainment in the US. However, this source is time-lagged, and some populations have privacy restrictions preventing NDI linkage. Other death data sources have been developed but currently lack validation

Objectives: To assess the validity of the mortality data sources compared to NDI within the HealthCore Integrated Research DatabaseSM (HIRD).

Methods: This study is a secondary analysis of a standing cohort of individuals in the US diagnosed with advanced cancer in HIRD between January 2010 and December 2018 and with an established NDI linkage. Mortality data sources, including inpatient discharge status, termination reasons, DMF, and online obituary data, all updated monthly in the HIRD, were compared to the NDI. Death status (dead/alive) for each source and the NDI were used to calculate sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and 95% confidence interval (95% CI). Deaths identified in each source 60 days before and 30 days after NDI death were deemed a match

Results: Among 40,692 advanced cancer patients, deaths were identified via NDI linkage for 25,761 (63.3%). Individually, death sources were accurate (all PPV >93%) but had low sensitivity when using the NDI as a gold standard (each with a sensitivity < 50%). The DMF had 32% sensitivity (95% CI: 0.31-0.32), and 98% PPV (95% CI: 97-98). An algorithm using mortality data sources altogether was able to identify most deaths from NDI (Sensitivity: 89%, 95%CI: 88-90) with accuracy (PPV=93%, 95%CI: 93-94). Specificity and NPV had the same magnitude with 90% (95% CI: 89-90) and 82% (95% CI: 82-83), respectively. Other date agreement segments, including ±30, ±15, and ±7 days, had similar results.

Conclusions: Our study demonstrates that creating an algorithm combining DMF with various additional data sources such as inpatient discharge, termination reason, and obituary data provides an accurate and more sensitive assessment of death outcomes among advanced cancer subjects. The algorithm is shown to provide meaningful sensitivity and accuracy compared to NDI. Further research is needed better to understand the performance of this algorithm in other populations.