Integrated LIBS-Raman spectroscopy coupled with explainable machine learning for biochemical characterization of...

Melanoma remains one of the most aggressive and lethal cutaneous malignancies, underscoring the need for objective and label-free diagnostic methods that complement traditional histopathology. In this study, Laser-Induced Breakdown Spectroscopy (LIBS) and Raman spectroscopy were jointly applied to characterize formalin-fixed paraffin-embedded (FFPE) melanoma and normal human tissues, and the resulting spectra were modeled using an Extreme Learning Machine (ELM) under five activation functions. Model interpretability was achieved through SHapley Additive exPlanations (SHAP) analysis. Among the tested activation functions, the average test accuracies were achieved by sine for LIBS (92.43%) and sigmoid for Raman (74.41%) under stratified spectrum-wise cross-validation, while feature-level fusion achieved 77.88% average test accuracy with sigmoid. SHAP analysis revealed that K (∼766/769 nm) and Ca (∼422 nm) emission lines in LIBS and ∼ 3164 and ∼ 1163 cm^-1 bands in Raman were the dominant discriminative features, corresponding to ionic imbalance, calcium signaling dysregulation, as well as protein conformational changes (amide A/N-H stretching) and protein-lipid-associated C-C/C-N vibrational alterations characteristic of melanoma progression. These results establish a coherent and interpretable framework linking spectral signatures to biochemical mechanisms and demonstrate the potential of compact, multimodal, and mechanism-driven optical diagnostics for precise, transparent cancer assessment.