Diagnosis and activity prediction of SLE based on serum Raman spectroscopy combined with a two-branch Bayesian...

This study aims to examine the impact of systemic lupus erythematosus (SLE) on various organs and tissues throughout the body. SLE is a chronic autoimmune disease that, if left untreated, can lead to irreversible damage to these organs. In severe cases, it can even be life-threatening. It has been demonstrated that prompt diagnosis and treatment are crucial for improving patient outcomes. However, applying spectral data in the classification and activity assessment of SLE reveals a high degree of spectral overlap and significant challenges in feature extraction. Consequently, this paper presents a rapid and accurate method for disease diagnosis and activity assessment, which has significant clinical implications for achieving early diagnosis of the disease and improving patient prognosis.

In this study, a two-branch Bayesian network (DBayesNet) based on Raman spectroscopy was developed for the rapid identification of SLE. Serum Raman spectra samples were collected from 80 patients with SLE and 81 controls, including those with dry syndrome, undifferentiated connective tissue disease, aortitis, and healthy individuals. Following the pre-processing of the raw spectra, the serum Raman spectral data of SLE were classified using the deep learning model DBayes. DBayesNet is primarily composed of a two-branch structure, with features at different levels extracted by the Bayesian Convolution (BayConv) module, Attention module, and finally, feature fusion performed by Concate, which is performed by the Bayesian Linear Layer (BayLinear) output to obtain the result of the classification prediction.

The two sets of Raman spectral data were measured in the spectral wave number interval from 500 to 2000 cm-1. The characteristic peaks of serum Raman spectra were observed to be primarily located at 1653 cm^-1 (amide I), 1432 cm^-1 (lipid), 1320 cm^-1 (protein), 1246 cm^-1 (amide III, proline), and 1048 cm^-1 (glycogen). The following peaks were identified: 1653 cm^-1 (amide), 1432 cm^-1 (lipid), 1320 cm^-1 (protein), 1246 cm^-1 (amide III, proline), and 1048 cm^-1 (glycogen). A comparison was made between the proposed DBayesNet classification model and traditional machine and deep learning algorithms, including KNN, SVM, RF, LDA, ANN, AlexNet, ResNet, LSTM, and ResNet. The results demonstrated that the DBayesNet model achieved an accuracy of 85.9%. The diagnostic performance of the model was evaluated using three metrics: precision (82.3%), sensitivity (91.6%), and specificity (80.0%). These values demonstrate the model's ability to accurately diagnose SLE patients. Additionally, the model's efficacy in classifying SLE disease activity was assessed.

This study demonstrates the feasibility of Raman spectroscopy combined with deep learning algorithms to differentiate between SLE and non-SLE. The model's potential for clinical applications and research value in early diagnosis and activity assessment of SLE is significant.