Using Machine Learning and Silver Nanoparticle-Based Surface-Enhanced Raman Spectroscopy for Classification of Cardiovascular Disease Biomarkers

Characterizing complex biofluids using surface-enhancedRaman spectroscopy(SERS) coupled with machine learning (ML) has been proposed as a powerfultool for point-of-care detection of clinical disease. ML is well-suitedto categorizing otherwise uninterpretable, patient-derived SERS spectrathat contain a multitude of low concentration, disease-specific molecularbiomarkers among a dense spectral background of biological molecules.However, ML can generate false, non-generalizable models when datasets used for model training are inadequate. It is thus critical todetermine how different SERS experimental methodologies and workflowparameters can potentially impact ML disease classification of clinicalsamples. In this study, a label-free, broadband, Ag nanoparticle-basedSERS platform was coupled with ML to assess simulated clinical samplesfor cardiovascular disease (CVD), containing randomized combinationsof five key CVD biomarkers at clinically relevant concentrations inserum. Raman spectra obtained at 532, 633, and 785 nm from up to 300unique samples were classified into physiological and pathologicalcategories using two standard ML models. Label-free SERS and ML couldcorrectly classify randomized CVD samples with high accuracies ofup to 90.0% at 532 nm using as few as 200 training samples. Spectraobtained at 532 nm produced the highest accuracies with no significantincrease achieved using multiwavelength SERS. Sample preparation andmeasurement methodologies (e.g., different SERS substrate lots, samplevolumes, sample sizes, and known variations in randomization and experimentalhandling) were shown to strongly influence the ML classification andcould artificially increase classification accuracies by as much as27%. This detailed investigation into the proper application of MLtechniques for CVD classification can lead to improved data set acquisitionrequired for the SERS community, such that ML on labeled and robustSERS data sets can be practically applied for future point-of-caretesting in patients.