SafeImpute: Reliable Clinical Data Imputation via Conformal Selection SafeImpute：通过共形选择实现可靠的临床数据插补

TL;DR

• SafeImpute introduces a novel framework for reliable clinical data imputation that addresses the critical need for error control in high-stakes medical decision-making.
• The method utilizes a two-relation Graph Neural Network (GNN) to capture both intra-patient temporal trajectories and inter-patient clinical similarities from sparse, irregular longitudinal records.
• Reliability is guaranteed through conformal prediction techniques, converting proxy risk scores into p-values and applying the Benjamini-Hochberg procedure to control the False Discovery Rate (FDR) of unacceptable errors.
• Empirical evaluations on Mayo Clinic, MIMIC-III, and MIMIC-IV datasets demonstrate superior performance in both standard imputation accuracy and FDR-controlled selective release compared to existing baselines.

Why It Matters

This research bridges the gap between predictive accuracy and statistical reliability in healthcare AI, addressing a major barrier to the deployment of imputation models in clinical settings where incorrect data can lead to harmful decisions. By providing a mechanism to selectively release only trustworthy imputations with controlled error rates, it enables safer integration of machine learning tools into electronic health record systems and clinical workflows.

Technical Details

• Architecture: Employs a two-relation GNN with adaptive fusion to model complex dependencies in clinical data, regularized by an auxiliary masked reconstruction objective to enhance learning stability.
• Graph Construction: Builds an event graph that simultaneously encodes temporal dynamics within individual patients and clinical similarity metrics across different patients.
• Reliability Mechanism: Implements conformal selection by transforming a proxy risk score into conformal p-values, allowing for rigorous statistical control over the proportion of unreliable imputations released.
• Error Control Strategy: Uses the Benjamini-Hochberg procedure to manage the False Discovery Rate (FDR) of clinically unacceptable errors at a user-defined tolerance level, ensuring that only high-confidence imputations are utilized downstream.

Industry Insight

Healthcare AI developers must prioritize uncertainty quantification and reliability guarantees alongside accuracy metrics when designing models for clinical applications to ensure patient safety and regulatory compliance. The adoption of conformal prediction methods like SafeImpute offers a scalable pathway for integrating probabilistic reasoning into deep learning frameworks, potentially becoming a standard requirement for FDA-approved or clinically validated AI tools. Organizations handling sparse medical data should consider hybrid approaches that combine graph-based representations with statistical error control to maximize the utility and trustworthiness of their predictive systems.

TL;DR

• 提出SafeImpute框架，旨在解决临床纵向记录中稀疏且不规则数据缺失导致的不可靠填补问题。
• 构建事件图以捕捉患者内部时间轨迹和患者间临床相似性，利用双关系GNN进行自适应融合学习。
• 引入共形选择机制，将代理风险分数转化为共形p值，并通过Benjamini-Hochberg过程控制不可接受错误的错误发现率（FDR）。
• 在Mayo Clinic、MIMIC-III和MIMIC-IV数据集上的实验表明，该方法在标准填补精度和FDR控制的 selective-release 评估中均优于基线模型。

为什么值得看

该研究针对高 stakes 的临床决策场景，解决了传统填补方法缺乏可靠性保证的痛点，为医疗AI中的缺失数据处理提供了统计严谨性的新范式。对于关注医疗数据质量、可信赖AI以及临床决策支持系统的研究者而言，其结合图神经网络与共形预测的方法具有重要的参考价值。

技术解析

• 问题定义：聚焦于“可靠临床数据填补”，即在提高平均准确度的同时，能够选择性发布经过统计验证的高置信度结果，严格控制临床不可接受的错误比例。
• 模型架构：SafeImpute 构建了一个事件图，整合了 intra-patient 的时间序列信息和 inter-patient 的相似性信息。使用双关系图神经网络（Two-relation GNN）结合自适应融合策略进行参数学习，并辅以掩码重建目标作为正则化项。
• 可靠性保障机制：采用共形预测（Conformal Prediction）理论，将模型输出的代理风险分数转换为共形 p-values。利用 Benjamini-Hochberg 程序对发布的填补值进行筛选，确保在用户指定的容忍度下，不可接受错误的错误发现率（FDR）得到严格统计控制。
• 实验验证：使用了包含 Mayo Clinic 私有数据以及公开 MIMIC-III 和 MIMIC-IV 数据集的综合评估体系。结果显示 SafeImpute 不仅在常规指标上表现优异，更在基于 FDR 控制的 selective-release 场景中显著领先于其他基线方法。

行业启示

• 从“准确率”到“可靠性”：在医疗等高风险领域，AI模型的评价标准应从单一的预测精度转向包含不确定性量化和错误控制的可信度指标，SafeImpute 展示了如何将统计保证融入深度学习流程。
• 图神经网络在纵向医疗数据中的应用深化：通过显式建模患者间相似性和时间动态，GNN 能有效挖掘稀疏临床数据中的潜在结构，这为处理其他类型的时序异质数据提供了借鉴。
• 合规与临床落地的桥梁：通过 FDR 控制机制，该技术有助于满足医疗监管对数据完整性和安全性的严格要求，为填补算法在实际临床工作流中的部署扫清了部分信任障碍。