Cross-validation of optimized composites for preclinical Alzheimer's disease

We discuss optimization and validation of composite endpoints for pre-symptomatic Alzheimer's clinical trials. Optimized composites offer hope of substantial gains in statistical power or reduction in sample size. But there is tradeoff between optimization and face validity such that optimization should only be considered if there is a convincing rationale. As with statistically derived regions of interest in neuroimaging, validation on independent datasets is essential.

Using four datasets, we consider the optimized weighting of four components of a cognitive composite which includes measures of (1) global cognition, (2) semantic memory, (3) episodic memory, and (4) executive function. Weights are optimized to either discriminate amyloid positivity or maximize power to detect a treatment effect in an amyloid positive population. We apply repeated 5×3-fold cross-validation to quantify the out-of-sample performance of optimized composite endpoints.

We found the optimized weights varied greatly across the folds of the cross validation with either optimization method. Both optimization methods tend to down-weight the measures of global cognition and executive function. However when these optimized composites were applied to the validation sets, they did not provide consistent improvements in power. In fact, overall, the optimized composites performed worse than those without optimization.

We find that component weight optimization does not yield valid improvements in sensitivity of this composite to detect treatment effects.