Quaranta, Patrizia (2025) Causally-Driven Clinical Decision Support: Interventions, Counterfactuals, and What-If Scenarios from Real-World Data. [Tesi di dottorato]

Documento PDF Quaranta_Patrizia_38.pdf Visibile a [TBR] Amministratori dell'archivio Download (6MB) | Richiedi una copia

Abstract

Healthcare research increasingly relies on data-driven methodologies to support clinical decision-making while avoiding expensive, hard-to-implement and sometimes unethical, controlled trials. However, traditional Machine Learning (ML) models are primarily based on statistical correlations and often fail to provide an understanding of the underlying causal relationships. In a context where interventions directly affect human well-being, purely correlational approaches may be insufficient. Causal Reasoning (CR) enables the estimation of intervention effects and counterfactual scenarios from observational data, offering a more interpretable and action-oriented perspective. Nevertheless, its application to real-world clinical data remains challenging due to data heterogeneity, incompleteness, and the lack of prior causal knowledge. This thesis addresses these challenges by proposing a framework for causal discovery and inference from observational clinical data. The framework integrates data pre-processing, feature selection through ML techniques, causal structure discovery, validation using Large Language Models (LLMs), and statistical verification procedures to assess robustness. The resulting causal model supports both interventional and counterfactual queries, allowing the simulation of clinical protocols or therapeutic modifications before their real-world implementation. The methodology was validated through two case studies. The first focuses on diabetes mellitus, using the PIMA Indians Diabetes dataset to evaluate the impact of lifestyle-related interventions. The second applies the framework to occupational health in the maritime sector, using real-world data collected from two shipping companies. In both cases, the inferred causal relationships align with clinical evidence and support what-if scenario analysis. Overall, the results highlight the potential of causal reasoning as a bridge between predictive modeling and clinical interpretability, providing a realistic and interpretable methodology to support safer and more informed medical decisionmaking.

Downloads

Actions (login required)

Modifica documento