Determination of latent tuberculosis infection from plasma samples via label-free SERS sensors and machine learning

Effective diagnostic tools for screening of latent tuberculosis infection (LTBI) are lacking. We aim to investigate the performance of LTBI diagnostic approaches using label-free surface-enhanced Raman spectroscopy (SERS). We used 1000 plasma samples from Northeast Thailand. Fifty percent of the samples had tested positive in the interferon-gamma release assay (IGRA) and 50 % negative. The SERS investigations were performed on individually prepared protein specimens using the Raman-mapping technique over a 7 x 7 grid area under measurement conditions that took under 10 min to complete. The machine-learning analysis approaches were optimized for the best diagnostic performance. We found that the SERS sensors provide 81 % accuracy according to train-test split analysis and 75 % for LOOCV analysis from all samples, regardless of the batch-to-batch variation of the sample sets and SERS chip. The accuracy increased to 93 % when the logistic regression model was used to analyze the last three batches of samples, following optimization of the sample collection, SERS chips, and database. We demonstrated that SERS analysis with machine learning is a potential diagnostic tool for LTBI screening.