RaSizeNet: An interpretable multimodal deep learning model integrating SERS spectra and inertial microfluidic size distribution for precise detection of complex pathogens

Accurate and rapid identification of pathogens is essential for public health and infectious disease control. Despite notable advances, label-free, high-throughput, and sensitive bacterial detection in complex samples remains a significant challenge. An interpretable multimodal deep learning model, RaSizeNet, was developed by integrating surface-enhanced Raman spectroscopy (SERS) and inertial microfluidic size distribution for high-precision identification of complex bacterial populations. Fe@PCN-Au hybrid particles exhibiting strong SERS activity and magnetic responsiveness were synthesized and embedded into a dual-zone microfluidic chip to enable efficient bacterial enrichment, size-based separation, and signal enhancement. A large-scale multimodal dataset comprising nine representative bacterial species was constructed by combining SERS spectra and outlet distribution data. A size-guided cross-modal attention mechanism was introduced to improve spectral discriminability and model robustness while maintaining independent modality representations. RaSizeNet achieved an average accuracy of 96.22 % in complex samples, significantly outperforming unimodal (SERSNet, 82.23 %) and conventional multimodal (MultiNet, 92.81 %) models. The model also demonstrated superior adaptability and stability under extreme concentrations, varying flow rates, and diverse media conditions. This work establishes a promising framework for integrating microfluidic-SERS sensing with artificial intelligence, providing both theoretical and technical support for real-time multimodal pathogen detection.