Baiying Lei

Automated medical audio diagnosis supports disease detection, especially in resource-limited settings. However, scarce and imbalanced audio data limit model performance, requiring effective data augmentation. Most existing augmentations use fixed policies and ignore acoustic differences among samples, making personalized augmentation difficult.To address this, we propose a FairAug, which is one of the first fairness-aware adaptive augmentation paradigms to formulate augmentation selection as a Markov Decision Process (MDP). It connects the diagnostic model and augmentation agent in five steps: state construction, action modeling, Hybrid Batch-Sample (HBS) execution, reward computation, and network optimization. HBS first collects sample statistics via efficient batch-level exploration, then switches to personalized sample-level control after performance stabilizes.When HBS enters the sample-level phase, fairness under class imbalance requires two things. First, the difficulty scores should be more comparable across classes. Second, the grouping boundaries should adapt to each class. Therefore, we introduce Fair Confidence Measurement (FCM) and Adaptive Decision Boundary (ADB). FCM maps sample features to fixed Equiangular Tight Frame (ETF) anchors and derives difficulty scores from the separation between targets and competitors. ADB updates class-wise boundaries using these scores and operation-level feedback. This allows policy selection to follow relative sample difficulty and action effects.Experiments on five benchmarks and four backbones show gains across respiratory, cardiac, and cough sound tasks. FairAug achieves the best reported scores on all four event-level evaluations of the SPRSound dataset, with an average improvement of 5.49%. Across four backbones on the HF_Lung_V1 dataset, it further improves average recall, accuracy, and macro-F1 over standard fine-tuning by an average of 2.54%, 1.72%, and 2.36%, respectively. FairAug is a universal data augmentation tool for scarce and imbalanced medical audio data with no inference overhead, greatly enhancing the value of limited data.