Leveraging Causal AI for Robust Treatment Policy Formulation: An Ensemble Causal Tree Approach with Expert Insights in Critical Care Decision-Making

Personalized healthcare (PHC) aims to tailor optimal treatments for patients based on their heterogeneity. Causality plays a significant role in supporting PHC by providing estimation tools for subpopulation based heterogeneous treatment effect (HTE) estimation. In this study, our goal is to leverage a practical causal inference technique, Causal Trees to estimate HTEs for critical care patients who need ventilation support. Precisely, we aim to identify which treatment strategy - liberal vs. conservative oxygenation is optimal and which covariates are important for different groups of ICU patients based on their HTE estimates using an ensemble of causal trees. From the experimental findings, we can deduct a set of decision rules which further needs to be validated by taking the opinions of clinical researchers and domain experts. Unlike traditional Artificial Intelligence (AI) systems that rely on correlations, our causally-grounded approach ensures that recommendations are based on true cause-and-effect relationships, enhancing both reliability and clinical trust. Our findings contribute to the broader goal of building causally-aware AI systems capable of generalizing to diverse domains, including autonomous systems and emergency response, where managing uncertainty is crucial. By merging interpretability with data-driven, causally-grounded insights, this work aims to foster trust, transparency, and effective human-AI collaboration in critical decision-making.