Breaking data barriers in medical diagnosis with MSDGD framework based on Gaussian Diffusion Generation

Domain knowledge gaps and scarcity of data on rare medical conditions pose significant challenges to leveraging Medical Service Data (MSD) effectively, resulting in compromised research quality and uninformed medical decision-making. To address these limitations, we propose a novel Gaussian diffusion-based MSD generation framework, MSDGD. This framework includes embedding modules such as the Parents-Children Numerical encoding scheme (ParentChild), which encodes column pair interactions, the Random Column Rearrangement algorithm (RamCol), which uncovers hidden multi-column relationships, and a Spatial Dimensional Transformation strategy (DimTrans) for optimal multi-row feature extraction. Additionally, we develop a new UNet model (MSDUNet) and a column relationship predictive process to enhance MSDGD optimization. We perform three sub-experiments to evaluate the effectiveness of MSDGD in generating MSD, and to assess the framework across five metrics, we use (1) a Simulation Table without Column Relationship, (2) a Simulation Table with Column Relationship, and even (3) a complex relationship table derived from real-world U.S. Chronic Disease Indicators datasets. This table contains more than 40 columns and 8x4 categories. The experimental results indicate that MSDGD has a high potential for improving medical service quality, achieving a 97.97% reduction in errors rate, thus promoting research by generating dependable and diversified generation of medical services data.