The main goal of this poster is to explore statistical methods that can be used to identify clinical trial patients who are likely to respond to placebo, known as placebo responders. We consider a Phase III program for a novel antipsychotic drug with two Phase III trials. The information from the first trial could be used to build predictive models for the second trial.
Neural nets can be used to create regression models for modeling complex relationships among clinical trial variables. A neural net model was developed using the Keras package in Python for the first clinical trial. Standard cross-validation methods were used to select a model architecture with the best validation score. The resulting model was applied to the second trial to predict patient responses. The discrepancies between the actual and predicted responses in the second trial were analyzed using tree-based regression methods to identify patients with a disproportionate response to placebo. Information on key characteristics of placebo responders is crucial for selecting appropriate patient populations and neural net models can help improve signal detection in future neuroscience trials.