Multi-modal fusion model combines SERS spectroscopy and clinicopathological features to predict neoadjuvant therapy response in breast cancer.

Breast cancer remains a significant global health threat to women, with neoadjuvant therapy (NAT) playing a critical role in treatment. Early prediction of NAT efficacy is essential for personalizing therapy and improving patient outcomes. The Miller-Payne (MP) grading system is a widely accepted standard for evaluating treatment response, categorizing patients as non-major histologic responders (MP1∼MP3) or major histologic responders (MP4∼MP5). This study developed a multi-modal fusion model integrating clinicopathological features and pre-treatment serum surface-enhanced Raman spectroscopy (SERS) data to predict NAT response in breast cancer patients. Leveraging Principal Component Analysis (PCA) for spectral dimensionality reduction and a Transformer architecture for feature extraction, the model achieved an accuracy of 92.6 % on the training cohort, significantly outperforming single-modal models using only SERS or clinicopathological features. Double-blind validation on an independent cohort confirmed the model's generalizability with an accuracy of 90 % and an area under the receiver operating characteristic curve (AUC) of 93 %. SERS analysis revealed significant spectral differences related to uric acid, tryptophan, phospholipids, and collagen, which have potential as biomarkers for NAT efficacy prediction. This study innovatively combined serum SERS data with clinicopathological features to predict NAT response in breast cancer patients. The multi-modal fusion model, enhanced by PCA and a Transformer architecture, captured biomolecular and clinical information, improving prediction accuracy and robustness. This non-invasive, cost-effective tool enables clinicians to avoid ineffective NAT, optimize treatment strategies, and improve patient outcomes.