Scholarlink

<p>This study addresses the challenges of performance degradation and decision unfairness in multi-ethnic electronic health record data for early readmission risk prediction in diabetes, where racial distribution shifts and sample imbalance often lead to inconsistent generalisation across demographic groups. To tackle these issues, a unified framework is proposed that integrates the MedFair Diffusion Block with Causal-Invariant Domain Generalisation to jointly improve robustness and fairness in cross-racial prediction. The MedFair Diffusion Block is designed as a source-side, fairness-oriented augmentation module trained only on the training split of the designated source domain. It generates fairness-enhanced samples to alleviate source-domain imbalance and underrepresentation without accessing target-domain data, thereby avoiding information leakage during cross-domain evaluation. On this basis, the Causal-Invariant Domain Generalisation module maps both original and fairness-enhanced source samples into a shared latent space, disentangles relatively stable predictive factors from group-sensitive variations, and strengthens structural alignment through invariance constraints to improve cross-group transfer stability. Comprehensive experiments on the multi-ethnic Diabetes 130-US hospitals 1999–2008 dataset show that, compared with several representative baselines, the proposed model consistently improves predictive performance and fairness across four source-domain settings. In particular, it achieves average gains of approximately 2.0 percentage points in AUC and 1.5 percentage points in accuracy, while reducing Demographic Parity Difference and Equalised Odds Difference by more than 25% on average. These results indicate that the proposed framework provides a more balanced and stable trade-off between predictive performance and racial fairness under single-source cross-racial generalisation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Fair and robust early readmission risk prediction from electronic health records via diffusion-based data augmentation and causal-invariant representation learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月27日

查看全文
  • Haodong Lu,
  • Jian Xu

摘要

This study addresses the challenges of performance degradation and decision unfairness in multi-ethnic electronic health record data for early readmission risk prediction in diabetes, where racial distribution shifts and sample imbalance often lead to inconsistent generalisation across demographic groups. To tackle these issues, a unified framework is proposed that integrates the MedFair Diffusion Block with Causal-Invariant Domain Generalisation to jointly improve robustness and fairness in cross-racial prediction. The MedFair Diffusion Block is designed as a source-side, fairness-oriented augmentation module trained only on the training split of the designated source domain. It generates fairness-enhanced samples to alleviate source-domain imbalance and underrepresentation without accessing target-domain data, thereby avoiding information leakage during cross-domain evaluation. On this basis, the Causal-Invariant Domain Generalisation module maps both original and fairness-enhanced source samples into a shared latent space, disentangles relatively stable predictive factors from group-sensitive variations, and strengthens structural alignment through invariance constraints to improve cross-group transfer stability. Comprehensive experiments on the multi-ethnic Diabetes 130-US hospitals 1999–2008 dataset show that, compared with several representative baselines, the proposed model consistently improves predictive performance and fairness across four source-domain settings. In particular, it achieves average gains of approximately 2.0 percentage points in AUC and 1.5 percentage points in accuracy, while reducing Demographic Parity Difference and Equalised Odds Difference by more than 25% on average. These results indicate that the proposed framework provides a more balanced and stable trade-off between predictive performance and racial fairness under single-source cross-racial generalisation.