Cardiovascular and bleeding endpoints in clinical trials: adjudication compared to linked hospital ICD10 codes

Background Cardiovascular trials can be operationally complex and costly, particularly when long-term follow-up and detailed event adjudication are planned. Administrative, routinely collected data, such as hospital admissions and mortality records, offer an alternative to traditional trial-specific data collection methods. Purpose To examine whether trials with cardiovascular endpoints should use administrative health data for endpoint ascertainment, in place of traditional clinical adjudication. Methods Using data from a large aspirin trial conducted in Australia, we examined agreement between adjudicated endpoints and endpoints defined by ICD-10 codes from linked administrative datasets, namely the National Hospital Morbidity Dataset and National Death Index. Endpoints included in this study were major adverse cardiovascular events (MACE defined as fatal or non-fatal ischaemic stroke, myocardial infarction, or coronary heart disease death) and major haemorrhagic events (intracranial, gastrointestinal, or other bleeding). Results Of 16,544 Australian participants, 11,873 had at least one linked hospitalisation record, and 1,659 had linked cause of death. Table 1 shows that fatal and non-fatal myocardial infarction from linked data had substantial agreement with adjudicated events (kappa = 0.74; sensitivity 73.3%). Similar results were found for ischaemic stroke (kappa = 0.75; sensitivity 75.1%) and coronary heart disease death (kappa = 0.67; sensitivity 62.4%). Intracranial haemorrhage had moderate agreement (kappa = 0.45) and sensitivity 78.3%, while serious gastrointestinal bleeding had moderate agreement (kappa = 0.43) and sensitivity 44.0%. Other major bleeding demonstrated fair agreement (kappa = 0.32; sensitivity 24.4%). Hazard ratios for aspirin effects on major haemorrhage were similar using adjudicated endpoints (1.47, 95% CI 1.25 to 1.75) and linkage (1.59, 95% CI 1.31 to 1.95). For MACE, aspirin effects with adjudicated endpoints (hazard ratio 0.89, 95% CI 0.84 to 1.04) and linkage (0.83, 95% CI 0.71 to 0.98) were similar but statistical significance at p<0.05 differed. Conclusions Using administrative data produced substantial agreement with adjudicated cardiovascular endpoints. Administrative data agreed less well with adjudicated that rely more heavily on clinical judgement, such as major bleeding. Linked administrative data may have utility advantages that can enhance outcome capture for trials but is an inadequate replacement for adjudication where clinical nuance or coding limitations exist.