Multimodal serum vibrational spectroscopy with availability-aware deep learning for prediction of pathological complete response in breast cancer.

Accurate assessment of pathological complete response (pCR) after neoadjuvant therapy is clinically important in breast cancer, yet reliable preoperative evaluation remains challenging. Serum vibrational spectroscopy provides a label-free and minimally invasive approach for capturing biochemical variation associated with subsequent treatment response in biofluids. However, clinical application is often constrained by limited sample size, class imbalance, and incomplete acquisition of multimodal spectral data. To address these constraints, we developed an availability-aware multimodal spectroscopy and mask-aware learning framework for serum pCR prediction, combining 532 nm Raman, 785 nm surface-enhanced Raman spectroscopy (SERS), and Fourier transform infrared (FTIR) spectra with CGAN-based augmentation and adaptive multimodal fusion. A conditional generative adversarial network (CGAN) was used to generate synthetic spectra for training augmentation, while MRAM-Net adaptively fused available spectral representations without spectral imputation. Using repeated patient-level cross-validation, the framework achieved an AUC of 94.2% when all modalities were available and maintained robust performance when one modality was unavailable (AUC >88%). In the independent validation cohort, the locked model achieved an AUC of 95.8% and an accuracy of 92.3% using the development threshold. These results indicate that the availability-aware framework can maintain stable performance under clinical data constraints. The proposed framework provides a practical strategy for treatment response assessment in breast cancer and a methodological basis for analyzing incomplete multimodal vibrational spectroscopy datasets in translational applications.