Issue:
A major pharmaceutical client sought to better identify drivers associated with early mortality and late survival among patients evaluated in a large multinational RCT of patients with esophageal cancer.
Solution:
- After completing a comprehensive data pre-processing step, we constructed a series of machine learning (ML) predictive models to determine if treatment regimens in study arms and/or other drivers could predict survival outcomes. Models evaluated included best subset logistic regression, Elastic Net regression, LASSO regression, Classification and Regression Tree (CART), Random Forest (RF) and Gradient-Boosted Machine Learning (GBML) models.
- Our analyses revealed that all models had good predictive ability with AUCs over 0.7. The models also identified other clinical predictors of early mortality and/or late survival. This study led to additional in-depth explorations of these other drivers, with the potential of utilizing these drivers as prognostic markers of treatment success.