ABSTRACT In today's artificial medicine field, combining multiple biosignals for accurate disease diagnosis has gained widespread application. Raman and infrared spectroscopy, emerging noninvasive diagnostic tools, offer significant advantages such as high sensitivity and specificity, demonstrating their unique value as disease markers. However, Raman and infrared spectral data contain significant noise or redundant information. Existing research using only a single modality for diagnosis can result in poor performance and may lead to erroneous feature associations during the fusion process. Furthermore, memory and computational efficiency when processing long sequences and the ability to capture long‐term dependencies are often overlooked. To address these issues, this paper proposes a multiscale fusion model (MSFM) based on the state‐selective space algorithm. This model first extracts features from Raman and infrared spectral data using a shared‐weight encoder. An attention mechanism is then implemented to achieve bidirectional information complementarity between the high‐resolution features of the Raman spectrum and the broadband features of the infrared spectrum. Residual connections preserve the original modality‐specific information while incorporating cross‐modal information, achieving multi‐level fusion from abstract to specific features. In the prediction module, improvements are made to the state‐selective space algorithm, using convolutional layers and the block‐wise processing of the state‐space model to achieve more efficient predictive and classification diagnosis. The model was used to diagnose osteoporosis with an accuracy of 96.25% and an AUC value of 97.19%, providing a new method for disease diagnosis using multimodal spectral data.
Wang et al. (2026) studied this question.