CHAPTER 2.1 18 causes were treated as competing events. For first HF hospitalization, all-cause deaths were treated as competing events. Then, standardised mortality ratios (SMRs) were used to compare adjusted mortality rates between the trials and the SwedeHF registry population. First, we fitted a Poisson model with 11 prognostic indicators from a validated MAGGIC HF risk score (age, sex, LVEF, NYHA class, serum creatinine, chronic obstructive pulmonary disease (COPD), diabetes, systolic blood pressure, body mass index (BMI), HF duration, smoking status) in a stepwise manner to the trial-eligible SwedeHF patients’ data.20,21 Next, the model with the derived β coefficients was applied to each trial to estimate each individual’s expected mortality, which was then summed across all participants to derive total expected mortality counts. The observed mortality count for each trial was divided by the expected mortality count to give the SMRs. An SMR value > 1 indicated that the observed risk of mortality in a trial was higher than the risk predicted based on SwedeHF patients as the reference population. The SMR was risk-adjusted for 11 prognostic factors to address heterogeneity between the trials. This was considered sufficient adjustment to pool the trials using fixed effect meta-analysis without introducing partial pooling. The corresponding 95% CI was determined using methods described by Breslow and Day.22 SMRs were not estimated for HF hospitalization because its existing risk prediction models do not have adequate discriminative performance compared to those designed to predict mortality.23 For cardiovascular causes of mortality, the Poisson model has taken into account competing risk from other causes of death as every patient’s follow-up duration was included in the estimation of the number of events. Rather than predicting cumulative probabilities, the Poisson model gives a prediction of the number of events for each individual which can be summed to obtain the total expected number of events in a trial. Missing data was multiply imputed by chained equations using the mice package in R.24 The number of imputations was set at 20.25 Statistical significance was set at 0.05. Statistical analysis was performed using the R statistical software version 3.6.1 (R Core Team, 2019) and Stata SE Version 15 (StataCorp LP, College Station, TX).26,27 The largest RCTs (BEAUTIFUL and SHIFT) in this analysis only included patients who were in sinus rhythm and the BEAUTIFUL study included a population who had CAD; therefore, sensitivity analyses were conducted in subsets of registry patients who were (i) in sinus rhythm or (ii) diagnosed with CAD. The fully-adjusted
RkJQdWJsaXNoZXIy MjY0ODMw