Proceedings of the Annual Meeting of the Cognitive Science Society

Postoperative delirium is a prevalent acute neurocognitive dysfunction in ICU patients, impairing key cognitive functions and deteriorating clinical prognosis, which is closely related to cognitive science research on neurocognitive regulation. Traditional machine learning prediction methods depend on statistical correlations and suffer from poor interpretability, failing to reflect the inherent causal hierarchy of delirium. To solve this issue, this paper integrates causal inference and proposes a Causal Hierarchy-Guided Feature Reconstruction Model (CDM). It adopts Greedy Equivalence Search and graph neural networks for causal feature reconstruction, followed by Bayesian-optimized XGBoost. Evaluated on the MIMIC-IV dataset, our model achieves superior performance with an AUROC of 0.9054, effectively enhancing model interpretability and offering a novel paradigm for cognitive disorder prediction.