Information-Guided Fusion of Multimodal Vibrational Spectroscopy for Disease Diagnosis Based on Symbiotic Attention Decoupled Contrastive Learning

Vibrational spectroscopy has gained significant attention in medical diagnosis due to its high sensitivity and nondestructive nature. Raman spectroscopy and infrared spectroscopy complement each other in their selection rules, vibration responses, and wavenumber coverage. Combining these two techniques can overcome the limitations of individual spectra, enhancing the accuracy of molecular structure identification. However, existing deep learning fusion methods often overlook the diagnostic advantages of different modalities, leading to overreliance on strong modalities or interference from weak modality noise, resulting in unstable fusion and imbalanced information flow. We propose a Symbiotic Attention Fusion Decoupled Network (SAFDN) to effectively model the information guidance mechanism. In the prefusion stage, we combine multilayer perceptrons and convolutional neural networks for intramodal encoding, laying the foundation for cross-modal fusion. Then, we design Symbiotic Attention Fusion (SAF) and Parasitic Attention Fusion (PAF) mechanisms to simulate biological symbiosis and parasitism, achieving a differentiated information enhancement. Finally, a supervised multimodal contrastive learning decoupling network is introduced to balance cross-modal consistency and intramodal cohesion, improving feature decoupling and semantic fusion. Experiments on cancer, autoimmune diseases, and cardiovascular disease data sets show that SAFDN outperforms existing methods, achieving accuracy and AUC values of 90.49%/0.9649, 95.48%/0.9866, and 96.67%/0.9934, respectively. SAFDN validates the advantages of the symbiotic effect in vibrational spectroscopy disease classification tasks through an in-depth comparison and analysis of fusion and loss mode ratios. This model provides an efficient solution for rapid, noninvasive precision medical diagnosis, improving the accuracy and interpretability of disease classification.