Fairness-Aware EHR Analysis via Structured Missing Pattern Modeling and Adversarial Low-Rank Adaptation

ICLR 2026 Conference Submission16100 Authors

Keywords: Fairness, Electronic Health Records, Missing values, LoRA

TL;DR: We propose a novel two-stage framework to achieve fairness-aware EHR clinical prediction via adversarial LoRA fine-tuning.

Abstract: Deep learning has been widely applied to electronic health record (EHR) analysis, offering strong predictive capabilities for clinical outcome prediction. However, due to inherent disparities across demographic groups (e.g., race, gender), fairness concerns have become increasingly critical. Due to severe instability in jointly optimizing fairness and performance objectives, existing fairness-aware approaches often struggle to balance predictive accuracy and fairness. Moreover, the structured missingness in real-world EHR data further worsens the performance and fairness of predictive models, yet they are frequently overlooked in fairness-aware modeling. In this paper, we propose FEMALA, a novel two-stage EHR analysis framework by explicitly modeling the structured missing patterns within EHR and fairness-aware model adaptation. Particularly, we design a dual-encoder architecture to integrate EHR temporal dynamics and structured missing patterns, thus enhancing performance while improving fairness by handling missingness well. Further, we perform adversarial fine-tuning to decorrelate task and sensitive representations via low-rank adaptation, enabling a better trade-off between fairness and accuracy. Experiments on MIMIC-III/IV datasets demonstrate that our framework achieves state-of-the-art performance in both accuracy and fairness, validating the effectiveness of structured missing pattern modeling and fairness adaptation fine-tuning. The code is anonymously available at https://anonymous.4open.science/r/FEMALA/README.md.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 16100