How to improve statistical power in a trial with SCA2 patients using natural history data

Randomized, double-blind, placebo-controlled trials are the gold standard for evaluating treatment effects, as they minimize biases. However, they raise ethical concerns regarding participant exposure to potentially ineffective interventions. For rare neurodegenerative diseases, these trials are further challenged by the limited number of patients. Our goal is to leverage recent advancements in statistical methodologies to enhance the statistical power of these trials. We aim to test these methods to the ATRIL study, which evaluated the efficacy of riluzole in patients with dominant cerebellar ataxia SCA2. We trained a progression model on natural history data to capture the progression of the disease, using Disease Course Mapping. We explored the prediction-powered inference for clinical trials method (PPCT), which leverages the prognostic scores predicted by this model, to offer an unbiased comparison between the observed progression in treated patients and the predicted progression if they had not been treated. Then, we compared PPCT to the prognostic covariate adjustment and the Hybrid Augmented Inverse Probability Weighting (H-AIPW) methods, which respectively incorporate the prognostic score to the analysis of covariance or to the Augmented Inverse Probability Weighting (AIPW) method. These methods were applied to the ATRIL trial, which included treated and placebo SCA2 patients, using a model trained on patients from US and European cohorts (EUROSCA, SPATAX, CRC-SCA). Disease Course Mapping models accurately forecasted the prognostic scores, namely the predicted one-year SARA score progression, with correlations around 0.15. The treatment effect estimator had a lower variance using the PPCT (0.342), prognostic covariate adjustment (0.344), and H-AIPW (0.348) methods compared to the classical variance (0.407). Notably, PPCT would reduce the sample size by 14.5% (6 patients). Using the PPCT estimator of the average treatment effect, the same variance reported in the ATRIL paper could have been achieved with 39 patients instead of 45. Our study addresses the challenge of statistical power in clinical trials and explores new methodological advancements. Applying the PPCT, prognostic covariate adjustment, or H-AIPW methods helps reduce confidence intervals, enhance statistical power, or decrease the required sample size for the trial by leveraging information from external data. Although the ATRIL's trial conclusion on riluzole remains unchanged, these methods could have enabled the selection of fewer patients, which is crucial for rare neurodegenerative diseases.