Raman spectra files (in .SPC format) were processed using the Rametrix™ LITE Toolbox in MATLAB r2018a (MathWorks; Natick, MA) as described previously (Fisher et al., 2018 (link)). Briefly, spectra were (i) truncated to include a Raman shift range of 400–1,800 cm−1, (ii) baselined using the Goldindec algorithm (Liu et al., 2015 (link)) (baseline polynomial order = 3; estimated peak ratio = 0.5; smoothing window size = 5), (iii) vector normalized, and (iv) scan replicates averaged for each patient. PCA and DAPC models were also built using the Rametrix™ LITE Toolbox. Multiple DAPC models were produced by varying the number of principal components (PCs) used in model construction.
The Rametrix™ PRO Toolbox v1.0 was used to perform leave-one-out analysis on all DAPC models. Spectra classification for each left-out spectrum (i.e., “healthy” or “unhealthy”) was predicted and compared to the actual classification. The averaged spectrum from each healthy individual or CKD patient was excluded from model construction and predicted in the leave-one-out routine. Thus, the leave-one-out validation was done with respect to individual specimens and individuals, not according to scan replicates. Model accuracy was calculated as the percentage of spectra where classification was predicted correctly. Sensitivity (i.e., the true-positive rate) and specificity (i.e., the true-negative rate) were also calculated and reported as percentages.
Rametrix™ PRO also has the capability to calculate “random chance” values of prediction accuracy, sensitivity, and specificity for any dataset. While this may be obvious for datasets with only two possible classifications (i.e., “healthy” or “unhealthy”), it is less obvious for datasets with multiple potential classifications with unequal representation. In these cases, the calculated accuracy, sensitivity, and specificity of leave-one-out validation routines are best presented relative to their random chance values.
The Rametrix™ PRO Toolbox v1.0 was used to perform leave-one-out analysis on all DAPC models. Spectra classification for each left-out spectrum (i.e., “healthy” or “unhealthy”) was predicted and compared to the actual classification. The averaged spectrum from each healthy individual or CKD patient was excluded from model construction and predicted in the leave-one-out routine. Thus, the leave-one-out validation was done with respect to individual specimens and individuals, not according to scan replicates. Model accuracy was calculated as the percentage of spectra where classification was predicted correctly. Sensitivity (i.e., the true-positive rate) and specificity (i.e., the true-negative rate) were also calculated and reported as percentages.
Rametrix™ PRO also has the capability to calculate “random chance” values of prediction accuracy, sensitivity, and specificity for any dataset. While this may be obvious for datasets with only two possible classifications (i.e., “healthy” or “unhealthy”), it is less obvious for datasets with multiple potential classifications with unequal representation. In these cases, the calculated accuracy, sensitivity, and specificity of leave-one-out validation routines are best presented relative to their random chance values.
Free full text: Click here
