The growing use of artificial intelligence in modern medicine poses significant challenges due to the lack and imbalance of high-quality clinical data, which arises from ethical, logistical, and technical limitations. We propose a framework that addresses this challenge by utilizing diffusion models, a state-of-the-art generative technique, to provide high-fidelity and diverse synthetic features rather than images. The framework implements a complete pipeline that combines feature extraction with pre-trained convolutional networks and a U-Net-based diffusion model to generate new clinical representations in scenarios with strong imbalance, such as melanoma diagnosis. Through rigorous experiments on two real datasets, we demonstrated that this technique consistently improves critical standard performance metrics compared to classical augmentation methods such as replication or Gaussian noise, positioning it as an effective and reproducible solution for data augmentation. The document describes the method and empirical results that validate this approach, presenting a practical and replicable method to enhance the robustness of models used in medical applications. Our results suggest that synthetic feature generation with diffusion not only enhances classification in limited clinical contexts but also paves the way for future research in domains where data access remains a persistent bottleneck.