To address the model performance limitations caused by data imbalance in medical insurance fraud detection, we propose the density-aware and particle swarm-optimized Wasserstein generative adversarial network (DPSO-WGAN). First, we employ multi-objective feature selection via particle swarm optimization to extract key features from high-dimensional sparse medical claims. Second, we generate fraud samples using WGAN and introduce a density-diversity dual optimization strategy to enhance the plausibility and representativeness of the synthetic data. Experimental results on real-world medical datasets demonstrate that the proposed method achieves superior performance compared to other baselines across multiple evaluation metrics, with the feature selection and sample screening modules validated for their effectiveness.