Lightweight Multimodal CNN for Real-Time Bacterial Classification from Raman Spectroscopy

Delayed pathogen identification is a major driver of inappropriate empirical broad-spectrum antibiotic use and, consequently, antimicrobial resistance (AMR). Conventional culturebased workflows require$24-72$hours, creating a mismatch between diagnostic turnaround and the clinical need for timely, targeted therapy. This paper presents a lightweight multimodal convolutional neural network (CNN) that classifies bacterial species directly from Raman spectra in real time. Unlike existing single-modality approaches, the proposed architecture is the first to jointly process raw one-dimensional spectra and their twodimensional continuous wavelet representations through parallel compact branches, followed by simple feature concatenation. On a three-class bacterial Raman dataset, the network attains 99.13 % test accuracy with only 199 k parameters (0.78 MB) and an average inference time of$\mathbf{1. 5 2 ~ m s}$per sample, yielding a$164 \times$speedup over a support vector machine (SVM) baseline while maintaining comparable accuracy. The exceptionally small memory footprint and sub- 2 ms latency make this the first truly deployable model suitable for portable, resource-constrained point-of-care devices, offering a practical route toward rapid, culture-free bacterial diagnostics to support antibiotic stewardship.