Abstract 12039: Assessing the Applicability of Three Heart Failure Randomised Controlled Trials by Calibration to Scottish Registry

Abstract

Intro: Randomised controlled trials are the gold standard method for assessing treatment effects. However, those who participate in trials of heart failure (HF) are usually highly selected (by the protocol, investigators and patient-choice) and, on average, younger and with fewer comorbidities than those encountered in clinical practice. Trials can be calibrated to better reflect routine clinical practice by re-estimating treatment effects after adjusting for differences in characteristics between trial and target populations, whilst maintaining the principle of randomisation. Methods: Using individual-level participant data, we calibrated 2 trials of HF with predominantly reduced left ventricular ejection fraction (HFrEF) - COMET (100% HFrEF) and DIG (87% HFrEF) and 1 trial of HF with preserved left ventricular ejection fraction TOPCAT - to 8,012 HF patients from a Scottish HF register using two methods - one regression-based and one using inverse propensity score weighting (IPSW). Results: Compared to register patients, trial participants had lower furosemide doses and were younger (mean (standard deviation) age 73 (12) years vs 62 (11), 64 (11) and 69 (10) years) with higher eGFR (59 (23) vs 67 (21), 62 (21) and 68 (18) ml/min/1.73m 2 for register vs COMET, DIG and TOPCAT respectively). For each trial, effect estimates from uncalibrated and calibrated analyses were similar (Figure 1; e.g. odds ratios in COMET for all-cause mortality were 0.83 (0.74, 0.93), 0.71 (0.58, 0.88) and 0.97 (0.72, 1.27) in the uncalibrated, IPSW and regression-based analyses). The calibrated estimates for all three trials were at most 1.5-fold less precise (i.e., 1.5-fold wider 95% confidence intervals) than the standard estimates. Conclusions: Calibration may be used to estimate the applicability of trial findings to populations typically encountered in clinical settings, with only small reductions in the precision of the effect estimates, and without breaking randomisation.