Diagnostic performance of attenuated total reflectance Fourier transform infrared (ATR-FTIR) spectroscopy for detecting osteoarthritis and rheumatoid arthritis from blood serum

Abstract

Osteoarthritis (OA) and rheumatoid arthritis (RA) are the two most common rheumatic diseases worldwide, causing pain and disability. Both conditions are highly heterogeneous, and their onset occurs insidiously with non-specific symptoms, so they are not always distinguishable from other arthritis during the initial stages. This makes early diagnosis difficult and resource-demanding in clinical environments. Here, we estimated its diagnostic performance in classifying ATR-FTIR spectra obtained from serum samples from OA patients, RA patients, and healthy controls. Altogether, 334 serum samples were obtained from 100 OA patients, 134 RA patients, and 100 healthy controls. The infrared spectral acquisition was performed on air-dried 1 µl of serum with a diamond-ATR-FTIR spectrometer. Machine learning models combining Partial Least Squares Discriminant Analysis (PLS-DA) and Support Vector Machine (SVM) were trained to binary classify preprocessed ATR-FTIR spectra as follows: controls vs. OA, controls vs. RA, and OA vs. RA. For a separated test dataset and the validation dataset, the overall model performance was better in classifying OA and RA patients, followed by the RA and controls, and lastly, between OA and controls, with corresponding AUC-ROC values: 0.72 (0.05; standard deviation for 100 iterations), 0.67 (0.04; standard deviation for 100 iterations), and 0.61 (0.06; standard deviation for 100 iterations) (test dataset) and 0.87 (0.02; standard deviation for 100 iterations), 0.87 (0.02; standard deviation for 100 iterations), 0.70 (0.07; standard deviation for 100 iterations) (validation dataset). In conclusion, this study reports robust binary classifier models to differentiate the two most common arthritic diseases from blood serum, showing the potential of ATR-FTIR as an effective aid in arthritic disease classification.