BackgroundClinical practice guidelines define Clostridium difficile infections (CDI) as diarrhea (≥3 unformed stools in 24 h) with either a positive C difficile stool test or detection of pseudomembranous colitis. Diagnostic modalities such as toxigenic culture and nucleic acid amplification testing can identify the presence of toxigenic C difficile in stools. But these tests are confounded by the presence of asymptomatic colonization of toxigenic C difficile and lead to overdiagnosis of CDI. The presence of two large toxins, toxin A and B (TcdA and TcdB) is necessary for pathogenicity. Detection of toxins using toxin enzyme immunoassay is difficult as it has low sensitivity and moderate specificity. Raman spectroscopy (RS) is a novel technology that is used to detect bacteria and their toxins. RS does not require any reagents for detection such as antibodies, enzymes, primers, or stains. We hypothesize that RS is a sensitive method to detect C difficile toxins in stool and will solve the problem of overdiagnosis of CDI. Materials and methodsCDI negative stool samples were spiked with concentrations (1 ng/mL, 100 pg/mL, 1 pg/mL, and 0.1 pg/mL) of TcdA and TcdB. RS was performed on air-dried smeared samples of stool supernatant on a mirror-polished stainless-steel slide. As RS of feces is difficult because of confounding background material and autofluorescence, samples were photo-bleached before spectral acquisition to reduce autofluorescence. Raman spectra were obtained, background corrected, and vector normalized. The data were split into training (70%) and test (30%) datasets. The machine learning methods used on the training data set were Support Vector Machine with Linear and Radial Kernels, Random Forest, Stochastic Gradient Boosting Machine, and Principle Component Analysis—Linear Discriminant Analysis. Results were validated using a test data set. The best model was chosen, and its accuracy, sensitivity, and specificity were determined. ResultsIn our preliminary results, at all concentrations (1 ng/mL, 100 pg/mL, 1 pg/mL, and 0.1 pg/mL), TcdA or TcdB spiked stool was distinguished from unspiked stool by all models with accuracies ranging from 64% to 77%. Gradient Boosting Machine, Principle Component Analysis—Linear Discriminant Analysis, and Support Vector Machine Linear Kernel performed best with sensitivities ranging from 69% to 90% and specificities ranging from 43% to 78%. ConclusionsUsing RS, we successfully detected TcdA and TcdB in stool samples albeit with moderate-to-high sensitivity and low-to-moderate specificity. Sensitivity and specificity could be further increased with the implementation of deep learning methods, which require large sample sizes. In terms of sensitivity, RS performs better than toxin enzyme immunoassay and has the potential to rapidly detect C difficile toxins in stool at clinically relevant concentrations and thereby help mitigate overdiagnosis of CDI.
Read full abstract