Biomedical Fragmented Attention Spectral Transformer Network for Infrared Spectra Analysis

Rapid and non-invasive molecular characterization of biological samples using Fourier Transform Infrared (FTIR) spectroscopy is increasingly important in clinical diagnostics, tissue analysis, and microbiology. However, existing approaches rely on conventional chemometric and machine learning pipelines that require extensive preprocessing and lack real-time capability. In this work, we propose BioFASTNet (Biomedical Fragmented Attention Spectral Transformer Network), an end-to-end deep learning framework for real-time spectral analysis. The model integrates a Multi-resolution Convolutional Feature Extractor (MCFE) to capture multi-scale spectral features and a chemistry-informed Fragment-wise Attention Module (FAM) to model localized, functional group–specific regions. This design enables effective learning of both global spectral patterns and localized vibrational signatures. BioFASTNet is tailored for imbalanced multilabel classification, addressing challenges arising from overlapping molecular attributes and skewed data distributions. Experimental results demonstrate superior performance on a functional group prediction dataset, achieving 97.81% accuracy and 91.97% F1-score, outperforming state-of-the-art (SOTA) models such as CIRSNet and AggMapNet. On an additional odor prediction dataset, the model achieves a SOTA F1-score of 36.68%. Furthermore, interpretability analyses using fragment-wise attention and Grad-CAM++ reveal that the model focuses on chemically meaningful spectral regions. These results highlight BioFASTNet as an efficient, accurate, and interpretable solution for real-time FTIR-based biomedical analysis. Source code and pre-trained models are available at: https://github.com/durgesgameta/biofastnet