Answers are grounded in this paper and cite their sources.
On unseen patients, Raman spectroscopy separates breast cancer from normal tissue at about 91% accuracy, but separates DCIS, ILC and IDC at only about 59%.We re-evaluated machine-learning classifiers on the public Imperial College ex vivo breast Raman dataset (641 spectra, 80 samples, 71 donors) using donor-disjoint validation, with preprocessing and classifier settings re-implemented from the deposited pipeline.The number of principal components was chosen by nested cross-validation inside training donors only.Over 20 random seeds, a support vector machine reached 90.8% accuracy (seed SD 0.6) for cancer versus normal and 59.3% (SD 1.3) for four-class subtype classification, with ILC sensitivity of 17.9%.Published figures of 97.7% and 91.0% were not reproduced under donor-disjoint validation; a literal reading of the cross-validation step in the deposited MATLAB scripts reproduces them to within about 3 points.Split-conformal prediction calibrated on separate donors gave 89.1% coverage at a 90% target for cancer versus normal, but only 83.8% coverage and 31% single-label predictions for four classes.In simulations calibrated to the real variance structure, spectrum-level validation overestimated accuracy on new donors by 12 to 23 points, while donor-level validation was nearly unbiased.We recommend donor-level splits, nested model selection and calibrated abstention for Raman tissue classifiers.
No takes yet. Share an insight, caveat, or question.
khoshnami et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: