Evaluation of chemical factor analysis model validation techniques for FTIR spectroscopy cancer data matrices

摘要

Abstract: Fourier transform infrared fiberoptic evanescent wave (FTIR- FEW) spectra, in the middle infrared (MIR) region, of human normal skin tissue, and melanoma and benign tumors were analyzed using chemical factor analysis (CFA). The first step in CFA is to determine the rank of the factor space and in this study several model validation techniques were employed. In particular we compare results obtained from Complete Cross Validation (CCV), Binary Cross Validation (BCV), Fisher variance ratios (F-Tests), Malinowiski indicator function (IND) and significance level (%SL). All methods' results were in agreement expect for F-Tests which differed with the other methods for normal skin tissue and melanoma tumors' rank. Using the two highest ranking eigenvectors for normal skin tissue as a basis set for cluster analysis, normal skin and melanoma tumors' clusters were well separated and localized in the two dimensional factor space. The projection of benign tumors' spectral points in this factor space revealed that some of the benign tumors had already regressed into either cancerous and other abnormal skin states. Furthermore the angular disparity between normal skin tissue and melanoma tumors' eigenvectors was successfully used to discriminate between the two skin states. !21