|
Abstract:
|
Modern statistical methods applied to omics data can produce highly accurate prediction models, if there is sufficient signal in the data, yet accuracy of predictions is only one aspect of having a clinically useful biomarker signature for treatment selection. Importantly, it must be demonstrated that the use of the prediction to guide therapy decisions leads to improved clinical outcomes in comparison to not using the prediction. Observational data and data from completed clinical trials can be used to evaluate the clinical utility of a prediction-based decision rule. We propose to use observational data to emulate a target trial to assess the clinical utility of a prediction-based decision rule. Unlike standard treatment trials, the comparison is not between two treatments but between a prediction-based decision rule and the standard of care. Under certain assumptions, one can estimate or bound the causal effect of using the prediction-based decision rule to guide treatment. This talk will introduce the conceptual framework, highlighting the assumptions and interpretations of the estimands, and illustrate the idea using publicly available data.
|
