Detection of non-structural protein 1 (NS1) in saliva is a recent finding that is appealing to early non-invasive reporting of NS1 related diseases. It is free from risk of blood infection, an advantage over current detection methods which are mostly serum based. Our approach starts with the acquisition of SERS spectra of saliva and saliva adulterated with low concentration NS1. Each spectrum contains 1801 Raman shift per spot while each sample is impinged with laser 738nm source at 10 spots. A total of 128 spectra are analyzed. Due to this volume of SERS features, dimension reduction is applied prior to classification by a SVM-RBF classifier. Our work here intends to optimize the classifier model with respect to the different PCA criteria. Results show that Scree and CPV criteria makes better classifier model than EOC, for identifying NS1 fingerprint from salivary SERS spectra, through capturing of the most relevant features. In terms of cumulative percentage variance of the dataset, CPV criterion which retains 70 principal components with 90% of cumulative variance is considered the most suitable for the model with a performance of [95.35% 100% 98.41%]. In terms of number of principal components and the associated computational load and time, the best SVM-RBF model is found with Cattel's Scree test criterion, which has a performance of [99.24% 100% 96.97%] with 5 principal components and 34.41% of cumulative variance.