Machine learning techniques for the evaluation of raman spectra to diagnose arthritis in a point-of-care setting
Tom Niessink (TNW-PDT), Tim L Jansen (VieCuri Medisch Centrum), Matthijs Janssen (VieCuri Medisch Centrum), Cees Otto (TNW-PDT)
Abstract
Background
Raman spectroscopy is proposed as a next-generation method for the identification of monosodium urate (MSU) and calcium pyrophosphate (CPP) crystals in synovial fluid. As the interpretation of Raman spectra requires specific expertise, the method is not directly applicable for clinicians. We developed an approach to demonstrate that the identification process can be automated with the use of machine learning techniques. The developed system is tested in a point-of-care-setting at our outpatient rheumatology department.
Method
We collected synovial fluid samples from 246 patients with various rheumatic diseases. We analyzed all samples with our Raman spectroscope. Trained observers classified every Raman spectrum as MSU, CPP or else. We designed two one-against-all classifiers, one for MSU and one for CPP. These classifiers consisted of a principal component analysis model followed by a support vector machine. The resulting model was validated in another 200 consecutive synovial fluid samples.
Findings
The accuracy for classification of CPP with respect to the CPPD classification criteria was 96.0% (95% CI 92.3-98.3), while the accuracy for classification of MSU with respect to the gout classification criteria was 92.5% (95% CI 87.9-95.7). Overall, the accuracy for classification of pathological crystals was 88.0% (95% CI 82.7-92.2). The model was able to discriminate between pathologic crystals, artifacts, and other particles such as microplastics.
Interpretation
We here demonstrate that potentially complex Raman spectra from clinical patient samples can be successfully classified with a machine learning approach, with objective diagnoses independent on the opinion of the medical examiner.