跳转至

文章背景与核心概要

电子健康记录（EHR）数据的结构化临床预测由于高维度、异质性、严重的类别不平衡以及跨队列分布偏移而异常困难。尽管表格上下文学习（TICL）和检索增强方法在通用表格基准测试中表现优异，但它们在真实世界临床环境中的行为和可靠性此前仍缺乏系统研究。

本文引入了一个全面的多队列 EHR 基准测试，对比了经典模型、深度表格架构以及 TICL 模型在不同数据规模、特征维度、结果罕见程度和跨队列泛化挑战下的表现。研究结果表明，基于先验数据拟合网络（PFN）的 TICL 模型在低数据状态下具有样本效率优势，但随着数据异质性和类别不平衡的加剧，在朴素的基于距离的检索下性能会下降。

为了克服这一问题，作者提出了 AWARE，这是一个利用监督嵌入学习和轻量级适配器的新型任务对齐检索框架。AWARE 在极端类别不平衡下将 AUPRC 提升了多达 12.2%，这证明了解决检索质量与检索推理对齐问题是将表格基础模型部署到临床工作流中的关键所在。

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

检索对齐的表格基础模型：在真实世界约束下实现电子健康记录的稳健临床风险预测

License icon

  • arXiv ID: arXiv:2604.01841 [cs.AI]
  • Subjects: Artificial Intelligence (cs.AI)
  • Status: Not peer-reviewed. Under revision
  • Submission History:
  • [v1] Thu, 2 Apr 2026
  • [v2] Sun, 31 May 2026
  • [v3] Sat, 22 Aug 2026 (This version)

Authors

作者

Minh-Khoi Pham, Thang-Long Nguyen Ho, Thao Thi Phuong Dao, Tai Tan Mai, Minh-Triet Tran, Marie E. Ward, Una Geary, Rob Brennan, Nick McDonald, Martin Crane, Marija Bezbradica

Executive Summary

执行摘要

Clinical prediction from structured electronic health records (EHRs) is notoriously difficult due to high dimensionality, heterogeneity, severe class imbalance, and distribution shifts across different cohorts. While tabular in-context learning (TICL) and retrieval-augmented methods excel on generic tabular benchmarks, their behavior and reliability in real-world clinical environments remained largely unstudied.

由于高维度、异质性、严重的类别不平衡以及不同队列间的分布偏移，基于结构化电子健康记录（EHR）的临床预测异常困难。尽管表格上下文学习（TICL）和检索增强方法在通用表格基准测试中表现卓越，但它们在真实世界临床环境中的行为和可靠性在很大程度上仍未得到充分研究。

This paper introduces a comprehensive multi-cohort EHR benchmark comparing classical models, deep tabular architectures, and TICL models across varying data scales, feature dimensionalities, outcome rarities, and cross-cohort generalization challenges. The findings reveal that Prior-Data Fitted Networks (PFN)-based TICL models offer sample efficiency in low-data regimes, but struggle and degrade under naive distance-based retrieval as data heterogeneity and class imbalance grow.

本文引入了一个全面的多队列 EHR 基准测试，对比了经典模型、深度表格架构以及 TICL 模型在不同数据规模、特征维度、结果罕见程度和跨队列泛化挑战下的表现。研究结果表明，基于先验数据拟合网络（PFN）的 TICL 模型在低数据状态下具有样本效率，但随着数据异质性和类别不平衡的加剧，在朴素的基于距离的检索下表现不佳并会出现性能退化。

To overcome this, the authors propose AWARE, a novel task-aligned retrieval framework utilizing supervised embedding learning and lightweight adapters. AWARE successfully improves AUPRC by up to 12.2% under extreme class imbalance, demonstrating that solving retrieval quality and retrieval-inference alignment is essential for deploying tabular foundation models in clinical workflows.

为了克服这一问题，作者提出了 AWARE，这是一个利用监督嵌入学习和轻量级适配器的新型任务对齐检索框架。AWARE 在极端类别不平衡下成功将 AUPRC 提升了多达 12.2%，这证明了解决检索质量与检索推理对齐问题是将表格基础模型部署到临床工作流中的核心关键。

Abstract

摘要

Clinical prediction from structured electronic health records (EHRs) is challenging due to high dimensionality, heterogeneity, class imbalance, and distribution shift. While tabular in-context learning (TICL) and retrieval-augmented methods perform well on generic benchmarks, their behavior in clinical settings remains unclear. We present a multi-cohort EHR benchmark comparing classical, deep tabular, and TICL models across varying data scale, feature dimensionality, outcome rarity, and cross-cohort generalization. PFN-based TICL models are sample-efficient in low-data regimes but degrade under naive distance-based retrieval as heterogeneity and imbalance increase. We propose AWARE, a task-aligned retrieval framework using supervised embedding learning and lightweight adapters. AWARE improves AUPRC by up to 12.2% under extreme imbalance, with gains increasing with data complexity. Our results identify retrieval quality and retrieval-inference alignment as key bottlenecks for deploying tabular in-context learning in clinical prediction.

由于高维度、异质性、类别不平衡和分布偏移，基于结构化电子健康记录（EHR）的临床预测面临挑战。虽然表格上下文学习（TICL）和检索增强方法在通用基准测试中表现良好，但它们在临床环境中的具体表现尚不明确。我们提出了一个多队列 EHR 基准测试，用于比较经典模型、深度表格模型以及 TICL 模型在不同数据规模、特征维度、结果罕见程度和跨队列泛化能力方面的表现。基于 PFN 的 TICL 模型在低数据状态下具有样本效率，但随着异质性和不平衡性的增加，在朴素的基于距离的检索下性能会下降。我们提出了 AWARE，这是一个使用监督嵌入学习和轻量级适配器的任务对齐检索框架。在极端不平衡的情况下，AWARE 将 AUPRC 提高了多达 12.2%，并且随着数据复杂性的增加，增益也随之增大。我们的研究结果表明，检索质量和检索推理对齐是部署表格上下文学习用于临床预测的关键瓶颈。

Access the Paper & Resources

访问论文与资源

External References & Citations

外部参考与引用