A Pseudo-Value Approach to Causal Deep Learning of Semi-Competing Risks

Area of Science:

• Biostatistics
• Machine Learning in Medicine
• Cancer Research

Background:

• Cancer studies often prioritize mortality, overlooking non-fatal events like disease recurrence.
• Recurrence is a critical endpoint in lung cancer, influencing treatment options and patient care.
• Causal inference for non-fatal outcomes is complicated by semi-competing risks, where death can prevent recurrence.

Purpose of the Study:

• To develop a robust deep learning approach for estimating the causal effect of treatments on non-fatal cancer outcomes.
• To address challenges in causal inference posed by dependent censoring and complex covariate relationships in semi-competing risks.
• To accurately estimate survival average causal effects for personalized cancer treatment strategies.

Main Methods:

• A three-stage deep learning framework combining Archimedean copula for survival functions and a jackknife pseudo-value approach.
• Estimation of pseudo-survival probabilities at fixed time points to serve as target values for causal estimators.
• Utilization of a deep neural network to link pseudo-outcomes, causal variables, and confounders for direct standardization.

Main Results:

• The proposed method provides consistent causal estimators without requiring proportional hazards assumptions.
• Numerical studies demonstrated the approach's effectiveness in handling dependent censoring and complex confounders.
• Application to the Boston Lung Cancer Study provided insights into the causal effect of surgical resection on recurrence.

Conclusions:

• The deep learning approach offers a powerful tool for causal inference in cancer research, particularly for non-fatal endpoints.
• This method enhances the ability to assess treatment effects on disease recurrence, improving personalized cancer care.
• The study highlights the potential of advanced machine learning techniques to overcome limitations in traditional survival analysis for complex clinical data.