Lightweight interpretable support vector machine SERS (LiSVM-SERS) learning framework for serological diagnosis of...

Psoriatic arthritis (PsA), a chronic, autoimmune form of arthritis that causes joint inflammation and occurs with the skin condition psoriasis, is difficult to clinically diagnose at an early stage due to insignificant symptoms and lack of biomarkers. Surface-enhanced Raman spectroscopy (SERS) is a powerful tool to identify specific biomolecules for liquid biopsy. Here, we establish a lightweight interpretable support vector machine SERS (LiSVM-SERS) learning framework for serological diagnosis of PsA and identification of potential biomarkers. The fs-laser-nanofabricated Au@mica SERS substrates with a relative standard deviation of 8.80% in intensity homogeneity are employed to obtain reliable serum Raman spectra. The SVM algorithm is performed to train the machine-learning model for PsA classification using only 26 samples (PsA = 13 and healthy controls = 13). The hyperplane coefficient vectors from the trained SVM model are then employed as the filtered weights of Raman shifts, followed by logistic-probability mapping and Welch's t-test to filter highly-significant spectral features, identifying 432 most-significant Raman shifts for 7 potential biomarkers. The trained LiSVM-SERS model demonstrates the outstanding sensitivity and specificity of 98.0% and 96.0%, respectively, validated by an internal independent blind-test cohort (PsA = 30, healthy controls = 30) and a cross disease-control cohort (systemic lupus erythematosus = 15, rheumatoid arthritis = 15, and osteoarthritis = 15). The present work not only provides a simple and lightweight non-invasive technique to diagnose PsA, but also paves a new way to identify the potential biomarkers from the label-free serum SERS spectra for autoimmune disorders.