Patient identity hides in Raman spectra: adversarial training cuts ID prediction from 68% to 12% while IDH typing holds at 0.8
Synopsis
In glioma Raman spectra, a single spectrum is enough for a model to tell which patient it came from, a sample-specific bias unrelated to biology that can confound tumor typing. On the 46-sample APOLLO dataset, a learnable feature-importance layer is trained adversarially against a two-headed classifier to suppress patient-identity signal while keeping IDH mutation information. Patient-ID balanced accuracy falls from 68% to 12%, while mutant-versus-wild-type classification stays at 0.8 balanced accuracy.
Interpretation
Sample-specific bias in Raman spectra can be measured on its own: a model reading a single spectrum can identify its patient of origin, not just its tumor type. Bias control in Raman work has relied mainly on standard preprocessing such as baseline correction, cosmic-ray removal and denoising, which was not quantitatively assessed against confounders like patient identity. On 300,506 spectra from 46 glioma samples in the APOLLO dataset, the two-headed model reached 68% accuracy for patient identity and 84% accuracy for the IDH1 methylation subtype.
Treating identity as an adversarial target sharply reduces the model's ability to recognize patients while leaving IDH mutation discrimination largely intact. The same model reached 75% AUC for identity on raw data, 69% after RADAR preprocessing and 66% after manual preprocessing, so standard preprocessing only partly weakened the bias. Under stratified 10-fold cross-validation, adversarial training lowered identity balanced accuracy to 12% while mutant-versus-wild-type classification held at 0.8 balanced accuracy.
Perspective
The demonstrated setting is spontaneous Raman spectra from 46 glioma samples in the APOLLO dataset, where data can be separated by patient sample and spectral coverage is broad enough across roughly 300,000 spectra; the authors note stimulated Raman data or datasets with too few spectra may not fit.
The authors note the method may have removed too many features, and the retained features were also weak at capturing IDH status; learned importance is task-dependent, so concentration in the fingerprint region does not mean the high-wavenumber region lacks diagnostic value generally; the method has not been tested on other datasets.