CARES-Net: a channel-attention residual network for multi-disease classification in small-sample ¹H NMR metabolomics...

NMR-based metabolomics is widely used in disease diagnosis due to its non-destructive and quantitative advantages. However, its analytical performance is often limited by the high-dimensional feature space derived from segmented 1D ¹H NMR spectra, excessive feature variables, and small sample sizes. These characteristics impede accurate correlation analysis, reduce model robustness, and frequently cause overfitting and unstable classification results. Existing statistical and machine learning methods struggle to effectively capture nonlinear metabolic patterns under such conditions. Therefore, there is a clear need for a more robust and intelligent analytical framework to improve disease classification accuracy in small-sample NMR-based metabolomics.

We propose CARES-Net, a channel-attention residual network specifically designed for ¹H NMR-based small-sample metabolomics disease classification. The model integrates residual connections with a channel attention mechanism to adaptively reweight spectral features, thereby enhancing informative metabolite signals while suppressing noise and redundant variables. Serum ¹H NMR datasets from children with malnutrition and patients with pancreatic ductal adenocarcinoma (PDAC) were used to evaluate the model. To address sample imbalance and feature redundancy, the SMOTE algorithm was applied for data augmentation and XGBoost was employed for feature selection. CARES-Net achieved classification accuracies of 86.25% for malnutrition and 92.50% for PDAC, significantly outperforming traditional machine learning models and representative deep learning approaches, demonstrating superior feature learning and generalization capabilities under small-sample conditions.

This study demonstrate that CARES-Net provides an effective solution to the challenges of high dimensionality and limited sample sizes in NMR-based metabolomics. By improving feature representation and classification stability, the proposed method enhances diagnostic reliability and accuracy. This work offers a promising analytical framework for small-sample metabolomics studies and supports the development of intelligent diagnostic systems.