Retrieval-aligned Tabular Foundation Models Enable Robust Clinical Risk Prediction in Electronic Health Records Under Real-world Constraints

Researchers present AWARE, a retrieval-aligned framework for improving clinical risk prediction in electronic health records using tabular foundation models. The method addresses limitations of naive retrieval-augmented approaches in clinical settings, achieving up to 12.2% improvement in AUPRC under extreme class imbalance while maintaining robustness across varying data complexity.

This research addresses a critical gap in applying modern machine learning techniques to clinical prediction tasks. Electronic health records present unique challenges—extreme class imbalance, high-dimensional feature spaces, and significant distribution shifts across patient populations—that generic tabular learning benchmarks fail to capture. The paper demonstrates that while retrieval-augmented in-context learning methods show promise on standard benchmarks, their naive application to clinical data degrades performance substantially as complexity increases.

The healthcare industry has increasingly sought to leverage advanced machine learning for risk stratification and early intervention. Traditional classical and deep learning approaches have proven insufficient for EHR prediction tasks due to their inability to handle sparse, heterogeneous clinical data effectively. The emergence of tabular foundation models and in-context learning represented a potential breakthrough, yet their clinical applicability remained empirically unvalidated under real-world constraints.

AWARE's contribution centers on alignment between the retrieval mechanism and the downstream prediction task through supervised embedding learning and lightweight adapters. This targeted approach directly addresses why naive distance-based retrieval fails: it doesn't account for task-specific relevance in clinically heterogeneous data. The framework's performance gains scaling with data complexity suggest its viability for large-scale EHR deployments across diverse healthcare systems.

The implications extend beyond academic validation. Healthcare organizations seeking to implement AI-driven clinical prediction systems now have evidence-based guidance on which architectures perform reliably at scale. The identification of retrieval quality as a bottleneck provides a concrete optimization target for practitioners deploying these systems in production environments with limited resources.

• →AWARE achieves 12.2% AUPRC improvement under extreme class imbalance by aligning retrieval mechanisms with prediction tasks.
• →Tabular in-context learning models show sample efficiency in low-data regimes but degrade under naive retrieval as data heterogeneity increases.
• →Supervised embedding learning and lightweight adapters effectively address the retrieval-inference alignment gap in clinical prediction.
• →Multi-cohort evaluation reveals that retrieval quality directly impacts clinical model robustness across varying data scales and outcome rarity.
• →Task-aligned retrieval frameworks enable robust deployment of foundation models in real-world EHR systems with extreme class imbalance.