Methods of causal effect estimation for high-dimensional treatments: A radiotherapy simulation study.

Area of Science:

• Medical Physics
• Radiotherapy
• Causal Inference

Background:

• Radiotherapy treatment outcome analysis is complex due to continuous, spatial, and confounding factors.
• Existing voxel-based estimators may produce biased results by not employing a causal inference framework.

Purpose of the Study:

• Propose a novel Causal Adaptive Lasso (CAL) estimator using sparsity within Pearl's causal framework.
• Address limitations of current voxel-based estimators in radiotherapy dose-outcome analysis.

Main Methods:

• Simulated 2D radiotherapy treatment plans on grids, incorporating organs at risk and target volumes.
• Utilized a directed acyclic graph to model causal relationships and confounding factors.
• Compared CAL against voxel-based regression estimators using simulated planned and delivered doses, evaluating Mean Squared Error (MSE) and bias.

Main Results:

• CAL estimators significantly outperformed voxel-based estimators, achieving up to a four-order-of-magnitude improvement in total MSE.
• CAL demonstrated substantially lower bias, particularly in regions with no dose-response.
• Achieved MSE < 1x10^2 compared to voxel-based MSE ≈ 1x10^6.

Conclusions:

• Sparse causal inference methods enhance the identification of dose-response regions and improve treatment effect estimation.
• Causal inference provides a robust framework to overcome limitations inherent in voxel-based radiotherapy analysis.
• Applying causal inference to clinical data can yield novel insights into treatment complications.