Hybrid self-attention network combined with serum Raman spectroscopy for diagnosis of autoimmune diseases

Raman spectroscopy is becoming a potential disease diagnosis method due to its rapidity, high sensitivity, and low water interference. With the development of machine learning, traditional machine learning models and deep learning models represented by convolutional neural networks have been widely used in Raman spectroscopy classification. In view of the limitations of existing models in Raman spectroscopy modeling capabilities, this study designed a hybrid self-attention network with multi-scale convolutional embedding (MCESANet). The model integrates attention networks and multi-layer perceptrons, combines multi-scale convolutional embedding and multi-head self-attention mechanisms to obtain accurate feature representations of Raman spectral sequences, and uses multi-layer perceptrons to effectively transform feature representations into final recognition results. The experimental results of the dataset constructed from serum Raman spectra of patients with various autoimmune diseases and healthy controls showed that the proposed model had the best effect, with all metrics of the model reaching 98.35%, and was stable on unbalanced small datasets, which was superior to traditional machine learning models and popular deep learning methods.