DERIVATION OF A MODEL THAT ACCURATELY PREDICTS CARDIOVASCULAR FROM NON-CARDIOVASCULAR CAUSE OF DEATH USING ADMINISTRATIVE HEALTH DATA SOURCES

Abstract

Cardiovascular (CV) mortality remains an important endpoint in CV outcome studies. Traditionally, cause of death (COD) is determined with expert adjudication following clinical notes and attestations from study investigators; however, this approach is cumbersome and expensive. Alternatively, COD is available within routinely collected administrative healthcare data, but there is typically a prolonged latency in the collecting and reporting of this information which may be ineffective for prospective research studies. An innovative and cost-effective method to supplant the delay is to derive and validate an accurate prediction model of COD analyzing routine healthcare data leading up to a fatal event. Predicting CV mortality in this manner provides a real-time practical solution for pragmatic CV clinical trials studying this outcome. The model was developed using all deaths that occurred between 2008-2012 in the CArdiovascular HEalth in Ambulatory Care Research Team (CANHEART) cohort of almost the entire Ontario adult population in 2008 created using linkage of 17+ health-related databases. Patients 40 years and older at the time of death with COD information available in the Ontario Vital Statistics Database were included (n = 362,778); COD was categorized into CV (n = 104, 590) or non-CV (n = 258,188) causes according to standard ICD-10 diagnostic codes. Candidate sociodemographic factors, past medical history and healthcare utilization variables that could be ascertained from health administrative data sources were considered as potential predictors. A backwards elimination model building strategy using logistic regression was used to ascertain the final set of predictors in the model. Model diagnostics including discrimination and calibration were computed to assess model performance. A single model for males and females combined was determined to be most appropriate. The model contained 34 predictors and had good discrimination (c-statistic = 0.831) and calibration (HL χ2 = 1394.6, p < 0.0001). Estimated differences between observed versus predicted probabilities of CV death was < 10% across all risk decile groups (Figure). Strong predictors for CV death included recent stroke (OR, 2.16; 95% CI, 2.11-2.21), heart failure (OR, 1.44; 95% CI 1.41-1.47), and any cardiovascular (OR, 3.94; 95% CI 3.83-4.04) or ophthalmologic (OR, 1.43; 95% CI 1.14-1.79) related hospitalization 1 month before death. The model will be validated in independent primary and secondary prevention cohorts. An administrative healthcare data model had good discrimination and calibration in determining whether a death was attributed to CV or non-CV causes suggesting potential use in pragmatic clinical trials.