Alternative per-protocol estimates: secondary analyses of data from the Balanced randomised controlled trial

Background: The Balanced trial was designed to answer the question of whether anaesthetic depth affects postoperative mortality when a vulnerable patient undergoes major surgery. Patients were recruited between 2012 and 2017 at 73 centres in seven countries. In the intention-to-treat analysis (n=6644), there was no significant difference in one year mortality between patients randomised to surgery under deep (target BIS 35) or light anaesthesia (target BIS 50). However the separation between randomised groups was only 8.4 BIS units and the trial was criticised for being underpowered. Methods In a secondary analysis of this trial's data, we made alternative per-protocol estimates designed to improve the power of the trial. We added an additional covariate - each patient's deviation from target BIS - to the original analysis, statistically recreating the desired separation of 15 BIS units between randomised groups. We used multiple imputation to recover missing BIS values. We also assessed whether a proportional hazards Cox model was appropriate for the analysis of one year mortality. Results Our alternative per-protocol estimates did not differ materially from the original per-protocol estimate. The gain in precision through using all intent-to-treat patients for our per-protocol estimates was offset by the additional variance introduced when modelling missing BIS values. When modelling missing BIS, we found regional differences: in China, the separation between randomised groups was far higher (13.6 BIS units) than in any other region. Estimates and plots assessing proportional hazards suggested increasing late mortality under deep anaesthesia, most notably in China. Conclusion Our hypothesis is that deep anaesthesia in the Balance trial led to higher postoperative delirium, which in turn led to an increase in late mortality. In future trials, patients should be followed for more than a year and cause of death recorded.