Primary Submission Category: Bayesian Causal Inference
Optimizing Clinical Trial Design with Causal Learning: Insights and Predictions for Improved Success Rates
Authors: Shaurya Gaur, Lucia Pagani, Lorenzo Rigolli,
Presenting Author: Shaurya Gaur*
Clinical trials, crucial for evaluating new treatments, are becoming increasingly expensive and time-consuming with a substantial number of trials failing due to intrinsic design challenges. In this paper, we propose an optimization tool for adjusting clinical trial design to increase their probability of success. Our tool utilizes a weighted logistic regression predictive model trained on a dataset of 41,269 trials from clinicaltrials.gov (CTGov), with labels obtained from TrialTrove.
The weights enable us to decorrelate features and obtain a causal interpretation of their impact on trial outcomes. Given the parametric nature of logistic regression, we can easily maximize the probability of trial success within a specific domain, allowing us to identify and implement optimized designs for proposed trials.
The AUROC score obtained by the logistic regression predictive model is 0.623. Furthermore, we collected evidence on the improved quality of optimized designs by comparing scores obtained from non-linear models. The best-performing model achieved a score of 0.695; however, it is less interpretable compared to our model.
The practical implications of this work are significant for pharmaceutical companies aiming to enhance their trial success rates. By employing our tool, they can optimize trial designs, potentially saving resources and increasing the likelihood of successful outcomes.
Keywords: Clinical trials, Weighted Logistic Regression, Causal learning, Optimization