Cross-domain adversarial transformer with soft-label filtering mechanism for unsupervised diagnosis of autoimmune diseases

• Raman spectroscopy provides rapid, non-invasive, and highly sensitive measurement capabilities. The unsupervised learning model developed in this study effectively leverages these features, enabling efficient disease classification without the need for extensive annotations, thereby broadening its clinical applications. • A novel domain adaptation framework was proposed to facilitate knowledge transfer among autoimmune diseases with diagnostic challenges. This framework enables unsupervised training of auxiliary diagnostic models for target diseases. • The proposed model incorporates cross-attention mechanisms for cross-domain feature extraction, introduces an innovative soft label filtering mechanism, and integrates an anti-learning module to mitigate category-biased prediction distributions in the target domain, achieving effective cross-domain feature alignment. With the growing application of artificial intelligence (AI) in medical diagnostics, there is an increasing demand for unsupervised models capable of operating with limited labeled data. Raman spectroscopy, renowned for its rapid and non-destructive measurement capabilities, faces significant challenges in clinical adoption due to the scarcity of reliable labels, particularly for complex conditions such as autoimmune diseases. This study introduces a cross-domain adversarial Transformer equipped with a soft-label filter to address these limitations. The model utilizes serum Raman spectra and incorporates an unsupervised domain adaptation algorithm, employing cross-attention for feature extraction and adversarial training for feature alignment. To mitigate category distribution bias in target domain predictions, a soft-label filter combining smooth labeling and dynamic thresholding is implemented, adaptively enhancing pseudo-label quality. Experimental results demonstrate the model's efficacy, achieving an average accuracy of 83.28 % across multiple autoimmune disease datasets, surpassing the best baseline by 15.42 %. Ablation studies confirm that the soft-label filter improves prediction balance, increasing F1 scores and G-means by 9.71 % and 7.64 %, respectively. This work presents a novel framework for unsupervised training in AI-assisted diagnostics and offers a practical solution to challenges in future research.