Title : Targeted estimation of heterogeneous treatment effect in observational survival analysis
Medical interventions are often evaluated in terms of their effect on future events such as patient survival, cancer recurrence, or cardiovascular events, which are subject to loss of follow up. The estimation of survival treatment effects is challenging, particularly when the outcome is not directly observable, and randomisation of treatment is not feasible. In this study, we demonstrate a hierarchical framework of heterogeneous treatment effect (HTE) identification and estimation using machine learning techniques for survival data. Our framework allows researchers to discover HTE without predefined subgroups and mitigate the confounding bias using recently developed one-step Targeted maximum likelihood estimation (TMLE). We provide both absolute and relative treatment effect measure (difference in survival probability and hazard ratio at individual, conditional average and average level to assist researchers for better clinical decisions.
Audience Take Away:
- Learn how to estimate individual and conditional survival treatment effect using machine learning techniques.
- Learn how to identify heterogeneous treatment effect based on ITE.
- Learn how to correct for confounding bias for survival outcomes.
- As the treatment effect are measured in absolute terms, researchers and practitioner will find it more interpretable than widely used hazard ratio in clinical practices. The adoption of modern machine learning techniques and confounding correction methods are easily scalable to a large cohort of observational data.