Patient-Level Reassessment of Breast Raman Spectroscopy: Data Leakage and Limits of Subtype Classification

Reliable interpretation of Raman spectra from breast tissue requires evaluation on patients absent from model training. We reanalysed a public ex vivo dataset comprising 641 spectra, 80 tissue samples and 71 donors to assess the effect of validation design on breast cancer classification. In a GNU Octave reproduction of the deposited analysis, an indexing statement caused an average of 56% of testfold spectra also to be selected for training. With the original and corrected indexing operations, four-class support-vector-machine accuracy was 87.6% and 59.1%, respectively; the reproduction used an approximate kernel-scale setting rather than the original MATLAB heuristic. Under donor-disjoint nested validation, five classifiers achieved 55–60% four-class accuracy, with invasive lobular carcinoma sensitivity of only 6–15%. Cancer-versus-normal classification remained more robust (support vector machine accuracy 90.0%; donor-bootstrap 95% confidence interval 83.8–95.3%). A fixed-pipeline analysis gave four-class accuracies of 81.7%, 77.9%, 60.9% and 62.1% when folds were formed at the spectrum, measurement-point, sample and donor levels, respectively. A bandresolved variance decomposition indicated that donor identity accounted for the largest share of within-class variance at the carotenoid bands (1157 and 1525 cm-1; 58–63%), whereas class information was concentrated at the phenylalanine band (1004 cm-1). Donor-level band ratios differed between cancer and normal tissue but not between DCIS and invasive lobular carcinoma. Adding the highwavenumber region, varying preprocessing, or increasing the number of lobular carcinoma training donors from about 2 to 8 did not raise lobular carcinoma sensitivity above 25%. In a simulation of the measurement chain with real fluorescence backgrounds and Poisson shot noise, donor-shared backgrounds and collection gains produced little optimism after preprocessing, whereas donor-specific tissue chemistry did. Donor-separated conformal calibration yielded 91.0% empirical marginal coverage for the four-class task at a nominal 90% target, but coverage was lower for invasive subtypes; class-conditional calibration equalised coverage across classes but returned large prediction sets. These findings identify substantial validation sensitivity in this dataset and underline the need for patient-level evaluation, reproducible code and classspecific uncertainty reporting.