Comparison of machine learning models for the label-free detection of viruses via surface-enhanced Raman spectroscopy

Early identification of viral pathogens is essential for patient care, spread reduction, and epidemiological monitoring. Common detection techniques include Polymerase Chain Reaction (PCR), viral antigen detection, and serology. While powerful, each of these are targeted approaches suffer from the need of prior knowledge and test for only one to a few viruses at a time. Moreover, targeted (labeled) approaches may miss instances of co-infection and fail when faced with novel strains or when a virus presents atypically. Optical strategies have the potential to overcome these limitations, detecting viruses in a label-free manner, particularly when coupled with machine learning. However, there are many machine learning models, each with their own benefits and limitations. Moreover, the success of each model is highly dependent upon the data provided and the problem at hand. Here, we investigate the ability for a multitude of machine learning models to identify viruses, including both a deep model (Convolutional Neural Network) and traditional models (Naïve Bayes, Logistic Regression, Support Vector Machine, Random Forest, and XGBoost). Structurally, viruses are classified into two major groups: enveloped and non-enveloped. In this study, we investigated each model's ability to identify both an enveloped virus (influenza A H1N1) and non-enveloped virus (human polyomavirus 2, JCPyV). We compare both the performance of each model as well as the favorable and unfavorable characteristics of each. Ultimately, a cross-validated AUC greater than 0.99 was achieved for JCPyV detection and 0.95 for influenza A detection following model optimization, demonstrating the ability of machine learning models to detect virus in a label-free manner.