Bruno T Scodari, Roland Brown, Xiaotong Jiang, Changyu Shen, Kyle Ferber, Shuang Wu, Feng Gao, Szofia Bullain, Brian Millen
A machine learning PS achieved a Pearson correlation of 0.48 between predicted and observed CDR-SB changes in an internal test set (N = 398). Adjusting for this PS in the held-out trial increased power from 80% to 87.9% (95% confidence interval [CI]: 85.5%-90.2%) with the original sample size. Alternatively, this approach could reduce the required sample size by 19.7% (95% CI: 13.7%-25.7%) while maintaining 80% power.
INTRODUCTION: A prognostic score (PS) summarizes a patient's expected disease progression and can increase the statistical efficiency of clinical trials when included as an analysis covariate.
METHODS: We pooled patient data from observational studies and randomized trials for early Alzheimer's disease (AD) and trained PS candidates to predict 18-month changes in the Clinical Dementia Rating Scale - Sum of Boxes (CDR-SB) score. The efficiency gains achieved through covariate adjustment were evaluated in a held-out trial (N = 650).
RESULTS: A machine learning PS achieved a Pearson correlation of 0.48 between predicted and observed CDR-SB changes in an internal test set (N = 398). Adjusting for this PS in the held-out trial increased power from 80% to 87.9% (95% confidence interval [CI]: 85.5%-90.2%) with the original sample size. Alternatively, this approach could reduce the required sample size by 19.7% (95% CI: 13.7%-25.7%) while maintaining 80% power.
DISCUSSION: Our findings support the use of PS adjustment for enhancing the efficiency of early AD trials.