FairGRPO: Towards Fair Reasoning Foundation Models for Clinical Diagnosis

Medical artificial intelligence systems have achieved remarkable diagnostic capabilities, yet they consistently exhibit performance disparities across demographic groups. While recent multimodal large language models (MLLMs) have advanced clinical diagnosis through integrated analysis of diverse medical data, they inherit and often amplify biases present in training datasets dominated by majority populations. We introduce Fairness-aware Group Relative Policy Optimization (FairGRPO), a hierarchical reinforcement learning approach that promotes equitable learning across heterogeneous clinical populations. FairGRPO employs adaptive importance weighting of advantages based on both representation, task difficulty, and data source, automatically discovering latent demographic groups through unsupervised clustering when labels are unavailable. Through comprehensive experiments across 7 clinical datasets spanning 4 clinical modalities, we demonstrate that FairGRPO reduces the standard deviation of F1 scores across demographic groups by up to 28.9% compared to GRPO, while improving overall F1 score by 3.8%. Notably, our training dynamics analysis reveals that FairGRPO progressively improves fairness throughout optimization, while baseline RL methods exhibit deteriorating fairness as training progresses. Based on FairGRPO, we release FairMedGemma-4B, a fairness-aware clinical MLLM that achieves state-of-the-art performance while demonstrating significantly reduced disparities across demographic groups. Our code, models, and fairness evaluation framework are publicly available at this anonymous link: https://anonymous.4open.science/r/fairness_submission-D923/.