Detection of amelanotic melanoma on the basis of NIR autofluorescence features

Raman spectroscopy in chronic heart failure diagnosis based on human skin analysis

“…The experimental spectral set with 272 predictors (the number of discrete wavenumbers in the analyzed spectral regions) was reduced to a certain number of Latent Variables (LVs) that are the best to capture the covariance between the spectral data and the assigned classes. The number of LVs for the constructed PLS-DA models was chosen according to several criteria [19]. First, we took into account the criterion of the local minimum of the root mean square error (RMSE) of the performed 10-fold cross-validation.…”

“…of LVs is a key concern when building a correct and valid statistical model to analyze the experimental dataset to avoid model overfitting or underfitting. Some recent papers [19] lack the description of the rule applied to selecting the optimal number of LVs or PCs (principal components) at all. In other manuscripts [20][21][22][23][24], the methodology of selecting the correct number of PCs/LVs raises doubts and leads to a significant overestimation of the classification models.…”

“…To check the obtained model performance by means of a new part of spectral data analysis, a true diagnostic capability of the PLS-DA model based on an independent test set was required. This portion of data is not included into the calibration set [19], therefore such model evaluation provides an opportunity to find a real performance of the proposed model. Once the PLS-DA classification model has been built and trained using the CV algorithm on the calibration set, we verified its capability by means of the test set (20% of the total spectral set).…”

“…Chemometrics methods have become widespread for RS data analysis, which is clearly reflected in a large number of manuscripts in the last few years [17,18]. However, several studies demonstrated the results that are characterized by significant overestimation of the proposed approaches based on the RS clinical effectiveness due to some flaws of applying the chemometric methods to analyze the RS data [17,[19][20][21][22][23][24][25]. In particular, using an unreasonably large number of principal components is the most common flaw in PLS that usually leads to overoptimistic results, for example, when a portable RS tool could provide the diagnostic result with an excellent accuracy [22,26].…”