A Bayesian model with seasonal effects for predicting accrual in clinical trials: Application to HOBIT and BOOST-3 trials for severe traumatic brain injury

Abstract

Background

Accurate accrual prediction is essential for initial planning and ongoing monitoring of clinical trials. Slow accrual can compromise statistical power, increase costs, or lead to premature trial termination. Traditional Bayesian approaches typically assume constant accrual rates and often fail to capture real-world seasonal fluctuations, which can reduce predictive accuracy.

Methods

We developed a Bayesian seasonal accrual model that extends the traditional homogeneous model by incorporating quarter-specific priors to account for seasonal variation. The model combines prior knowledge with observed data up to the monitoring point to obtain accrual predictions using the Bayesian posterior predictive distribution. We applied this approach to quarterly accrual data from two ongoing trials: the Hyperbaric Oxygen Brain Injury Treatment (HOBIT) trial and the Brain Oxygen Optimization in Severe TBI Phase-3 (BOOST-3) trial. Along with the Deviance Information Criterion, model performance was evaluated using RMSE, bias, and standard deviation, calculated from internal predictions of total accruals within observed seasonal quarters. Posterior predictive distributions of accrual after 36 and 30 quarters were also generated.

Results

Both trials exhibited seasonal trends, with the highest accrual rates in summer. The seasonal model yielded lower DIC in both trials. In HOBIT, internal prediction accuracy did not improve, likely due to uniformly low accrual rates. In contrast, the seasonal model outperformed the homogeneous model in BOOST-3 trial, yielding substantially lower RMSE, bias, and SD.

Conclusion

Incorporating seasonal effects into accrual modeling can enhance prediction accuracy, particularly in larger trials with high enrollment, and supports more accurate trial forecasting and resource allocation.