3.42 General discussion 199 information for the variables tested: age, creatinine and haemoglobin, whereby cumulative distribution curves of the aggregated data stayed within the 95% confidence interval bands of the original variables. (Schröder et al. – manuscript submitted to BigData) The federated approach to data analysis involves passing lines of analytic codes and summary statistics or regression estimates between data owners and an analytical centre.69 However, present methods for modified ILMA only permits combining studies of similar data or populations69 and were not designed for covariate-adjusted comparative analysis of two or more study populations. To estimate clinical trial generalizability, we intended to compare pooled trial data against registry data while simultaneously accounting for confounding on the individual-level within each data set. Since conventional survival regressions could not be used, we estimated a standardized mortality ratio (SMR) of observed-toexpected event rates for each trial. The expected mortality rate for each trial were calculated based on a predictive model fitted within the registry, which stores data on probability of mortality as a function of age, sex, body mass index, history of diabetes or COPD, LVEF, NYHA class, systolic blood pressure, serum creatinine and smoking status.16,73 If the mortality rate within a trial equals the registry upon prognostic factor adjustment, then the SMR would be 1, i.e. observed mortality=expected mortality. The SMRs were then combined by meta-analysis to obtain a pooled, confounder-adjusted estimate of events in trials relative to registry patients. We have proposed here a straightforward approach to an analytic challenge in federated data analysis that was done sequentially by study partners at the request of the analytic centre. In an established federated database system, a model can be updated directly from the analytic centre, allowing flexibility in the process and timely results. To address gaps in analytic capability, adapted methods for generalized linear models are currently being developed and proposed for federated data sets.74,75 Future perspectives Phase III trials for HF are among the costliest of cardiovascular trials; ranging from USD 142 million for a hypothetical 14500-patient trial in 2001 to an estimated USD 347 million in the 8442-patient trial for sacubitril-valsartan in 2017.76–78 A recurring theme contributing to more than two-fold rise in drug development costs for HF
RkJQdWJsaXNoZXIy MjY0ODMw