Get started
Work

Development and implementation of an end-to-end causal inference workflow integrating large language and causal generative models

Year: 2026
Type: dissertation
Abstract: Causal inference, the process of estimating the effect of one variable on another under different interventions, has been a fundamental challenge in recent decades. In biomedicine, determining which interventions actually improve health is crucial for millions of patients. However, identifying reliable causal relationships from complex, mostly observational data remains challenging. It requires the integration of expert knowledge, probabilistic modeling, and computational methods. A key step often missing from effective causal inference is causal discovery, which involves constructing causal graphs — directed acyclic graphs representing cause-effect relationships. Without well-defined causal graphs, researchers often rely on predefined diagrams or lack explicit causal structures. Existing approaches for causal discovery range from purely statistical methods to expert-driven manual constructions, which are frequently infeasible. Large Language Models (LLMs) offer a promising opportunity to automate and scale causal discovery by leveraging their embedded and native domain knowledge. Despite their demonstrated reasoning and generalization capabilities, their potential for producing causal graphs that enable causal inference remains under-explored. The aim of this thesis is to build probabilistic causal graphs using LLMs and to evaluate their potential for causal inference in biomedical applications. This work proposes methodologies for LLM-based causal graph construction, beginning with a baseline approach that relies on the general, pre-trained knowledge of a foundational LLM. Next, we will advance to a Retrieval-Augmented Generation (RAG) framework that incorporates domain-specific context from external knowledge bases to mitigate known limitations, such as hallucinations and outdated information. To more closely model real-world scenarios, the framework introduces uncertainty by generating probabilistic edges and quantifying confidence and hallucination levels numerically. To perform causal inference, we will integrate the LLM-estimated graphs with causal generative models. These models go beyond modeling observational distributions, enabling intervention modeling (estimating population distributions under hypothetical changes to a variable) and counterfactual estimation (predicting individual outcomes under alternative interventions). After implementation, the entire pipeline is validated. Causal graph construction is validated against established ground truths, including a general cause-effect benchmark, while the practical utility of the LLM-generated graphs is assessed through their application to causal inference on real-world biomedical data with quantitative evaluations By combining LLM-driven graph discovery with causal generative modeling, this work develops a complete end-to-end framework for automated causal inference in biomedical domains.
Source: Hispana
Language: English
Open Access status: green
OpenAlex ID: W7134150827
Cites:
Cited by:
Related to: