Statistical Analysis

The reliability of HR-MAS MR spectroscopic values of the metabolite concentrations (acetate, adipate, alanine [Ala], arginine [Arg], asparagine [Asn], aspartate [Asp], betaine, choline [Cho], creatine [Cr], ethanolamine, fumarate, glucose, glutamate [Glu], glutamine [Gln], glycerol, glycine [Gly], glycerophosphcholine [GPC], histidine [His], isoleucine [Ile], lactate, leucine [Leu], lysine [Lys], methionine [Met], phosphocholine [PC], phosphoethanolamine [PE], phenylalanine [Phe], proline [Pro], serine [Ser], taurine, total choline [tCho, the sum of Cho, PC, and GPC], threonine [Thr], tyrosine [Tyr], uracil, valine [Val], and myo-inositol [m-Ins]) among all 3 specimen types (CNB samples collected in vivo vs central surgical tumor samples collected ex vivo vs peripheral surgical tumor samples collected ex vivo) was assessed using the intraclass correlation coefficient (ICC). Post hoc analysis between the following groups were performed using ICCs with Bonferroni correction: CNB samples collected in vivo versus central surgical tumor samples collected ex vivo, and central surgical tumor samples collected ex vivo versus peripheral surgical tumor samples collected ex vivo. ICC values that did not include 0 in their respective 95% confidence intervals were considered to show statistically significant agreement. ICC values in the following ranges were considered to indicate poor (0–0.2), fair (0.21–0.4), moderate (0.41–0.60), substantial (0.61–0.80), or almost perfect agreement (0.81–1.00).¹⁹ Differences in HR-MAS MR spectroscopic values according to intratumoral location and biospecimen type were analyzed using the paired t test between CNB samples versus central surgical samples and between central surgical samples versus peripheral surgical samples. Statistical analysis was conducted using statistical software (version 20.0; SPSS, Chicago, IL). A 2-tailed P value less than 0.05 indicated a statistically significant difference.

To evaluate whether specimen type affected the metabolic profiling of breast cancer based on multivariate data analysis, we performed multivariate analysis of the spectral data using Matlab (MathWorks, Natick, MA), SIMCA-P 12.0 (Umetrics, Sweden), and Excel (Microsoft, Seattle, WA) programs. Multivariate partial least squares discriminant analysis (PLS-DA) was performed to evaluate whether different specimen types had similar performance in distinguishing patient groups by hormone receptor status (ER, PR, and HER2), for which previous studies have reported clear separation using multivariate models.^4,5,7 To avoid over-fitting of the statistical model, class discrimination models were built until cross-validated predictability values did not increase significantly. Signals contributing to class discrimination were identified by an S-plot, with identification of the corresponding HR-MAS MR spectral data using Chenomx (Spectral database; Edmonton, Alberta, Canada) software and an in-house database.