Are you really dead? Validation of death and date of death in patients with COPD in CPRD

Abstract

BACKGROUND: The current Clinical Practice Research Datalink (CPRD) death algorithm identifies a large number of deaths and the date of death (DoD). However, improvements are encouraged. OBJECTIVES: To evaluate a process for identifying and verifying cases of death and the DoD in the CPRD. METHODS: Deaths in a nested case‐control study of all‐cause mortality in a cohort of new users of selected COPD medications (September 2012‐June 2016) were identified in the Hospital Episode Statistics (HES), the Office for National Statistics (ONS), and the CPRD General Practitioner Online Database (GOLD) through an electronic algorithm and classified as confirmed (CONF) or potential (for clinical review). In practices not linkable to HES/ONS (NLP), potential deaths were those with recorded clinical information beyond 15 days after the DoD or with missing transfer‐out date, and CONF deaths were all others. In practices linkable to HES/ONS (LP), CONF deaths were those with matching DoD in ONS, HES, and CPRD GOLD and, if not matching, those with a DoD in ONS ≤3 days before or after the other sources' DoD. Potential deaths were those with non‐matching DoD between ONS, HES, and CPRD GOLD and for which the DoD in ONS was >3 days before or after the DoD in other sources, including patients with missing DoD in ONS or CPRD GOLD. Two physicians blinded to the exposures of interest reviewed the patient profiles of potential deaths to adjudicate the death and DoD. RESULTS: In CPRD GOLD, HES, and ONS, 3822 deaths were identified in 39 788 users of selected COPD medications. Of these, 3610 (94.5%) were CONF deaths through the electronic algorithm, and 212 (5.5%) were potential deaths. Patient profile review of potential deaths confirmed 209 cases (98.6%; 23 in NLP, 186 in LP). In NLP, the DoD changed for 13 cases (56.5%, 6 cases with an error in the year). In LP, the DoD changed for <5 cases; <5 deaths with DoD missing in ONS were not confirmed. Most of the adjudicated deaths in LP had the same DoD in ONS and HES but a later DoD in CPRD GOLD. CONCLUSIONS: Our electronic algorithm confirmed most deaths and DoD. In NLP, we recommend validation of death and DoD for those 122 ABSTRACTS patients with clinical codes beyond 15 days after the DoD or with transfer‐out date missing. In LP, HES/ONS DoDs are usually correct, and most CPRD GOLD DoDs were recorded late. This is likely to occur also in NLP. In LP, we recommend validation of cases with >3 days difference between the DoD in HES and ONS, irrespective of GOLD, or those for which the DoD is missing in ONS.