Chemically Informed Artificial Intelligence Decoding of Full-Spectrum SERS Fingerprints for pH-Guided Intraoperative Assessment of Freshly Excised Melanoma Specimens

Abstract

Precise intraoperative delineation of melanoma margins remains a clinical challenge; incomplete resection causes recurrence, whereas over-excision compromises aesthetic and functional outcomes. While surface-enhanced Raman spectroscopy (SERS) provides chemically rich vibrational fingerprints, accurate intraoperative quantification is hindered by complex, coupled full-spectrum responses. Herein, we report a chemistry-informed artificial intelligence (AI)-SERS platform designed to support melanoma margin assessment via the detection of tumor-associated metabolic acidification. A pH-responsive HemS–SERS microarray was developed to translate the phenolic protonation/deprotonation of a hemicyanine–xanthene reporter into coordinated intensity redistributions across multiple Raman bands. Guided by this acid–base chemistry, our AI model (CIMPA-pH) encodes these pH-responsive vibrational anchors as learnable directional chemical priors, integrating full-spectrum attention with Henderson–Hasselbalch-guided prediction. This global decoding strategy accurately captures coupled pH responses, reducing prediction errors by approximately 85% and 54–78% compared to manual peak-ratio features and conventional deep-learning baselines, respectively, while a quality-control risk metric further reduces high-error readouts by 29.1%. Providing rapid (10 s) pointwise pH measurements, the platform distinguishes melanomas from normal skin with distinct pigmentation phenotypes in animal models and shows preliminary applicability for assessing melanoma-associated regions, including grossly atypical areas, in human resection specimens. Ultimately, this chemical-prior-guided SERS decoding strategy provides an interpretable framework with potential for metabolism-guided intraoperative margin assessment.