A unified machine learning framework for oral cancer classification using Raman spectra from tissue, serum, and urine

Raman spectroscopy has emerged as a promising label-free technique for cancer detection; however, variability across biological sample types and the limited availability of complete multi-source data per patient present challenges for developing generalisable classification models. This study proposes a unified machine learning framework for oral cancer classification using Raman spectra acquired from heterogeneous biological samples, including tissue, serum, and urine. Each spectrum was treated as an independent observation while maintaining patient-level data integrity during model evaluation. A multilayer perceptron (MLP) incorporating sample-type information was evaluated alongside classical machine learning models, including support vector machine (SVM) and random forest (RF). Spectral preprocessing consisted of baseline correction, vector normalisation, and controlled smoothing within the fingerprint region (600-1800 ^-1). Model performance was assessed using patient-level group K-Fold cross-validation, with classification thresholds optimised using a macro-averaged Youden's index. The proposed framework demonstrated strong diagnostic performance across heterogeneous biosamples. The MLP achieved an overall area under the curve (AUC) of 0.905 ± 0.042, while SVM and RF achieved AUCs of 0.964 ± 0.025 and 0.948 ± 0.026, respectively. Cohort comparability analysis confirmed no significant differences in age or sex distribution across the three sample types. Raman spectral analysis revealed distinct biochemical differences between normal and oral cancer samples, including sample-dependent variations in bands associated with proteins, lipids, amino acids, and metabolites. Notably, no single Raman peak exhibited a consistent intensity trend across tissue, serum, and urine, highlighting the influence of biological medium on spectral signatures. These findings provide proof-of-concept evidence that a single domain-conditioned architecture can learn cancer-associated spectral representations that transfer across biochemically distinct biological matrices, establishing a methodological foundation for future multi-modal spectroscopic studies using matched patient cohorts.