An interpretable CNN model for NMR-based whole metabolomic profiling of sepsis

• Proposing a novel interpretable CNN architecture (NmCNN) for direct analysis of raw NMR metabolomics profiles. • Validated on dual sepsis cohorts (whole blood & platelets) with high sensitivity outperforming conventional methods. • Integrated SHapley Additive exPlanations(SHAP) analysis reveal biomarkers undetectable by traditional quantitative analysis. • CNN-SHAP framework enables big data readiness, scalable with growing clinical NMR datasets. • Establishes protocol for AI-driven metabolic pattern recognition with minimal preprocessing requirements. NMR metabolomics faces critical sensitivity limitations when analyzing whole spectral profiles, particularly in clinical biomarker discovery where weak metabolite signals often elude detection. To address this challenge, we present NMR metabolomics convolutional neural network(NmCNN) as an interpretable 1D-convolutional neural network framework that innovatively integrates spectral translation invariance with explainable AI. Our approach presents a novel concept distinct from traditional machine learning methods, and the incorporation of an interpreter exhibits a higher sensitivity in detecting differential metabolites. The model incorporates translation invariance from CNN, facilitating efficient processing of spectra and effectively addressing issues pertaining to deviations in spectral peak alignment. Validated on sepsis biomarker discovery for whole blood and blood platelet samples, NmCNN achieved 90–95 % sensitivity, 80 %-85 % specificity and area under curve 0.83–0.96 The interpretability framework with SHapley Additive exPlanations (SHAP) identified previously undetected metabolic features, demonstrating sensitivity improvement over conventional quantification methods through its whole-spectrum utilization. Overall, our NmCNN model offers a distinctive approach that enhances the utilization of raw NMR-based spectral metabolomics profiles in AI-driven clinical research, thereby improving interpretability and sensitivity in metabolomics analysis.