Mitigating Clinical Confounding in AI Models: A Comparative Analysis of Associative and Causal-Fused AI for ICU...

Introduction: In critical environments, patient discharge decisions present significant challenges, particularly regarding the prevention of readmissions. However, traditional AI systems focus on correlations rather than causation, resulting in issues related to explainability and generalization. This study conducts a comparative evaluation between standard predictive and causal-fused models to assess their ability to mitigate clinical confounding on predicting readmissions. Methodology: Utilizing the MIMIC-IV dataset, we predicted 30-day ICU readmissions through a comparative analysis of associative models (XGBoost, LSTM) and their causal-fused equivalents. The Fast Causal Inference (FCI) algorithm mapped latent confounding via a Partial Ancestral Graph (PAG), while counterfactuals were computed using Causal Forests to estimate the Average Treatment Effect (ATE), which was integrated into the predictive models. Performance was evaluated via AUROC, AUPRC, calibration metrics, precision, recall, and F1-score while feature extraction was used to monitor feature realignment. Results: The causal-fused LSTM model maintained a stable AUROC (0.7342 to 0.7357), while the XGBoost AUROC improved (0.6517 to 0.6909). Feature importance extraction revealed a structural realignment; whereas standard models relied heavily on non-actionable frailty proxies such as polypharmacy, causal integration elevated the individualized causal effect of length of stay as a primary predictive driver. The estimated ATE for length of stay was calculated at -0.038. Conclusions: The comparison showed that transitioning to causal-fused AI mathematically resolves a clinical Simpson's Paradox, while also realigning the features based on causal mechanisms without sacrificing predictability. By making this shift, we can mitigate reliance on administrative noise and promote de-confounded interventions rather than passive correlations.