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

Q: What is the clinical evidence from this study?

Population: ICU readmission (n=53639). Intervention: Causal-fused AI models (LSTM, XGBoost) vs. Traditional associative AI models. Primary outcome: AUROC for predicting 30-day ICU readmission (LSTM model).

Q: What does this research mean for the field?

Integrating causal inference into machine learning models realigns feature importance toward actionable causal drivers and improves predictive performance for 30-day ICU readmissions. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

Konstantina Remoundou; Emanuele Koumantakis; Ioanna Roussaki

doi:10.3390/healthcare14142067

July 10, 2026HealthcareOpen Access

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

View Full Paper

Ask AI

Bookmark

Share

Key result

Causal-fused AI preserves ICU readmission prediction while showing extended stay cuts readmissions ~4%.

n=53,639

Figures

Why the study?

Traditional AI systems focus on correlations rather than causation, causing issues with explainability, generalization, and clinical confounding when predicting readmissions.

Does integrating causal inference into AI models improve the prediction and explainability of 30-day ICU readmissions compared to standard associative models in adult ICU patients?

Key Points

To evaluate whether integrating causal discovery and counterfactual estimation into predictive machine learning models mitigates clinical confounding in 30-day ICU readmission forecasts.
Analyzed a retrospective cohort of 53,639 adult ICU patients from the MIMIC-IV dataset, in which 6,162 individuals (11.49%) experienced a readmission within 30 days.
Mapped latent confounding using the Fast Causal Inference (FCI) algorithm to generate a Partial Ancestral Graph (PAG), and derived counterfactuals via Causal Forests to compute the Average Treatment Effect (ATE) for length of stay.
Trained baseline associative models (XGBoost and LSTM) and their causal-fused counterparts, evaluating performance via AUROC, AUPRC, Brier scores, Expected Calibration Error (ECE), and feature importance realignment.
Causal fusion improved XGBoost AUROC from 0.6517 to 0.6909 and AUPRC from 0.2039 to 0.2706, while LSTM AUROC remained stable (0.7342 to 0.7357) alongside AUPRC (0.3255 to 0.3280).
For LSTM, causal integration improved probability calibration by reducing the Brier score from 0.1996 to 0.1936 and ECE from 0.3183 to 0.3087, whereas XGBoost calibration degraded (Brier score rose to 0.2445 and ECE to 0.3798).
Feature importance realigned away from non-actionable frailty proxies toward the individualized causal effect of length of stay (estimated ATE = −0.038), reducing false-positive rates from 9.50% to 6.66% in XGBoost and 22.55% to 20.23% in LSTM.

Structured PICO

Does integrating causal inference into AI models improve the prediction and explainability of 30-day ICU readmissions compared to standard associative models in adult ICU patients?

P

Population

53,639 adult ICU patients from the MIMIC-IV database, who survived their index admission, evaluated for 30-day ICU readmission risk.

I

Intervention

Causal-fused AI models (XGBoost, LSTM) integrating the Average Treatment Effect (ATE) of length of stay via Causal Forests

C

Comparator

Standard associative AI models (XGBoost, LSTM)

O

Outcome

30-day ICU readmission prediction performance (evaluated via AUROC, AUPRC, calibration metrics, precision, recall, and F1-score)

Main Result

Absolute Event Rate: 0.7357% vs 0.7342%

Limitations

Analysis relies exclusively on observational data from a single center, lacking external validation.
Feature space reduction condensed complex medical histories into quantitative proxies, risking unmeasured confounding.
Critical variables such as socioeconomic status, post-discharge medication adherence, and access to outpatient care were absent.
Temporal dynamics of physiological states were simplified to a static discharge snapshot rather than continuous trajectories.
Single-center data (MIMIC-IV) requiring multicenter validation
Tree-based ensembles may struggle to naturally calibrate synthetic counterfactual variables without aggressive hyperparameter tuning
Retrospective observational nature with potential unobserved confounding despite causal discovery algorithms

Implication

Transitioning to causal-fused AI models for predicting ICU readmissions resolves clinical confounding by prioritizing actionable interventions, such as length of stay, over passive correlations like polypharmacy, without sacrificing predictive accuracy.

Abstract

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.

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

Remoundou et al. (2026) studied ICU readmission (n=53,639). Causal-fused AI models (LSTM, XGBoost) vs. Traditional associative AI models was evaluated on AUROC for predicting 30-day ICU readmission (LSTM model). Causal-fused AI models maintained or improved predictive performance for 30-day ICU readmissions (AUROC 0.7357 for LSTM) while identifying that extending length of stay causally reduces readmission risk by 3.8%.

synapsesocial.com/papers/6a53302d4f7abc118aded3f6 https://doi.org/10.3390/healthcare14142067

View Full Paper

Ask AI

Bookmark

Share

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

Figures

Key Points

Structured PICO

Main Result

Limitations

Implication

Abstract

Discussion

Cite This Study

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

Figures

Key Points

Structured PICO

Main Result

Limitations

Implication

Abstract

Discussion

Cite This Study

Keep reading

More related papers

More from Healthcare

Keep reading

More related papers

More from Healthcare