Actinic keratosis labeling and single-band instability in near-infrared spectroscopy of skin cancer: a nested, patient-level re-analysis

Background. Portable near-infrared (NIR) spectroscopy has been proposed as an adjunct for skin-lesion triage, and the public NIR-SC-UFES dataset has made in vivo NIR spectra of skin lesions available for machine-learning benchmarks. Two questions that bear on how such benchmarks should be read have received little attention: how results depend on the labeling of actinic keratosis (ACK), a pre-malignant lesion that the dataset convention places in the non-cancer group, and whether the wavelengths that classifiers rank as important are stable enough to support spectroscopic interpretation. Methods. We analyzed the original NIR-SC-UFES file (714 in vivo spectra, 331 patients, 908-1676 nm, 125 points; MicroNIR Pro 1700). After standard normal variate (SNV) normalization, five classifiers (RBF support vector machine, logistic regression, random forest, PLS-DA and a one-dimensional convolutional neural network) were evaluated with nested cross-validation in which every patient was confined to a single outer fold (5 outer x 3 inner folds; hyperparameters selected on inner folds only). We compared four labelings of ACK and two pairwise tasks, measured the stability of band importance over 60 patient-level bootstrap resamples, and tested the effect of removing top-ranked, random or contiguous bands. Results. On the standard binary task (647 spectra, 319 patients) the SVM reached macro-F1 0.764 +/- 0.005, balanced accuracy 0.772 +/- 0.006 and ROC-AUC 0.842 +/- 0.006 (mean +/- SD over five repetitions of the outer split); sensitivity was 0.72 and specificity 0.83. Logistic regression, random forest and the 1D-CNN were within the fold-to-fold variability of the SVM (paired Wilcoxon test, SVM vs. CNN, p = 0.63; patient-bootstrap 95% CI of the AUC difference -0.006 to 0.064). The labeling of ACK changed performance more than the choice of classifier did: macro-F1 was 0.764 with ACK as non-cancer, 0.743 with ACK excluded and 0.677 with ACK counted as positive. ACK was more separable from malignant lesions (AUC 0.876) than benign lesions were (AUC 0.813), while benign lesions and ACK were only moderately separable (AUC 0.794). Band importance was unstable at the level of single bands (mean Jaccard overlap of the top-10 bands 0.34; mean Spearman correlation 0.51), yet the top-ranked band fell within 908-1051 nm in 59 of 60 resamples. Removing the top 5-20 ranked bands lowered macro-F1 by at most 0.016, and removing 20 contiguous bands around the top band by 0.028; removing random bands had no effect. Conclusions. On this cohort, classical and deep classifiers perform similarly under nested patient-level validation, while the definition of the non-cancer class has a larger effect. The informative spectral region is reproducible, but individual band rankings are not, so single-wavelength chemical assignments should be avoided. Studies on NIR-SC-UFES should state how ACK is labeled and should report band importance across resamples.