DiT-SFDA: A source-free domain adaptation method for intelligent diagnosis of cardiovascular diseases with limited heart sound samples

In recent years, the application of deep learning in intelligent diagnosis (ID) of cardiovascular diseases (CVDs) has significantly improved diagnostic efficiency and accuracy. However, in practice, owing to data privacy constraints, high labeling cost and specialized medical knowledge, collecting adequate labeled samples continues to present substantial technical difficulties, which makes ID of CVDs under limited samples a challenging issue. In this paper, a novel source-free domain adaptation (SFDA) approach for ID of CVDs, named DiT-SFDA, is proposed by integrating an improved diffusion model based on transformer (DiT) and a semi-supervised domain adaptation network (SDAN). Specifically, the method first converts heart sound (HS) signals into Mel spectrograms that can represent their time–frequency characteristics. Then, more realistic labeled samples are generated through DiT using limited real labeled data, effectively solving training data insufficiency. Subsequently, the generated labeled samples serve as the source domain, while the real samples serve as the limited labeled data in the target domain, and the SDAN based on minimax entropy is employed to further improve the performance of the model. Finally, experimental validation demonstrates that the DiT-SFDA method achieves significantly better diagnostic performance than other methods on two datasets. This innovative approach not only effectively addresses the critical challenge of data scarcity, but also provides an efficient and robust solution for the early screening and precise diagnosis of CVDs.