Classification of Salivary Adulterated NS1 SERS Spectra Using PCA-Cosine-KNN

Abstract

Review of literature on dengue fever (DF) reveals the most popular biomarkers and diagnostic medium are IgG/IgM and blood plasma respectively. As such, the current diagnostic methods are prone to blood borne infection. Presence of nonstructural protein 1 (NS1), another biomarker of DF, was detected in saliva of DF infected subjects using Enzyme-Linked Immunosorbent Assay (ELISA), but of low sensitivity. Our previous work has found Surface Enhanced Raman Spectroscopy (SERS), a confluent of photonic and nano-technology, is able to detect and produce a molecular fingerprint of NS1 from its salivary spectra. This implies an early, non-invasive, blood borne infection free detection method for DF, with the many associated advantages. Since K-nearest neighbor (KNN) is known for its strength in pattern recognition of signals and images, it is chosen to classify between NS1 positively and negatively adulterated NS1 samples here. Our work here intends to investigate the effect of number of nearest neighbours (k-value), classifier rules on KNN classifier with Cosine distance rule, subjected to three termination criteria of Principal Component Analysis (PCA). Healthy and adulterated NS1 samples from our UiTM-NMRR-12-1278-12868-NS1-DENV database were first analyzed with SERS. After pre-processing to remove undesired features, performance of the different KNN classifiers with Cosine distance rule as k-value and classifier rules were varied, with optimized features sets derived from the termination criteria of PCA, were evaluated and compared, in terms of sensitivity, specificity, precision and accuracy. From the results, it is observed that all the classifier models attained the highest performance of 100% in accuracy, precision and ROC performance, except for the Scree-Cosine-KNN models with Consensus classifier rule. And the CPV- KNN models with k-value of 1, 3 or 5 are the best in view of trade-off between computation load and performance, for all classifier rules, when Cosine distance rule is used.