Fusing Multisource Vibration Spectra via a Shared‐Private Network for Cardiovascular Disorder Detection

ABSTRACT

Aortic dissection (AD) and aortic stenosis (AS) are two fatal cardiovascular diseases with overlapping clinical presentations but vastly different treatment strategies, making rapid and accurate differential diagnosis crucial. Existing imaging diagnostic methods have limitations such as radiation exposure, invasiveness, and operator dependence. Vibration spectroscopy (infrared and Raman spectroscopy) enables the noninvasive acquisition of molecular “fingerprint” information from biological samples and thus providing a novel diagnostic approach. However, single‐spectrum modalities offer limited information, and the existing fusion methods struggle to fully exploit the complementary and different features among multisource data. To address this, we propose a multisource vibrational spectroscopy fusion network (MVSFNet) based on shared‐private features. The network extracts private features from infrared and Raman spectra through multiscale convolution modules, generates shared features by a dynamic weight‐sharing mechanism, and designs a bidirectional attention module to realize deep interaction and adaptive fusion of dual‐modal features. Experimental results on clinical AD and AS dried serum spectral datasets demonstrate that MVSFNet achieved 95.00% accuracy with an AUC value of 0.9955 in AD diagnosis tasks, and 91.30% accuracy with an AUC value of 0.9697 in AS severity grading tasks, significantly outperforming various baseline fusion models and single‐modal models. Notably, the model achieved 97.37% accuracy in distinguishing AD from severe AS, showcasing excellent performance and generalization capabilities. Ablation experiments validated the effectiveness of each core module. This study provides a noninvasive auxiliary diagnostic framework for cardiovascular disease detection based on multisource vibrational spectral fusion.