Integrated spectral-ML framework for the high-fidelity discrimination of cancer cells in liquid biopsy

Abstract

The clinical adoption of liquid biopsy for early cancer detection is often hindered by the need for complex sample preparation and the difficulty of identifying rare malignant cells within a dominant background of normal blood cells. While surface-enhanced Raman spectroscopy (SERS) offers distinct analytical advantages as a label-free, non-invasive tool, its practical implementation faces challenges due to limited training samples in early stage, weak signals and the necessity for expert spectral interpretation, which introduces undesirable operator-dependent variability. To address these challenges, this study proposes an integrated SERS, machine learning (Spectral-ML) framework to automatically extract features, improving diagnostic consistency and efficiency. This study presents a systematic evaluation of three nonlinear ML classifiers, including random forest (RF), extreme gradient boosting (XGB), and support vector machine (SVM), while integrating advanced spectral preprocessing methods such as modified Vancouver Raman algorithm, multiframe non-local means (MNLM) denoising with principal component analysis (PCA) for classification. We validated this framework across two datasets consisting of lung cancer cell lines: Dataset 1 (A549/WBC) and Dataset 2 (H1299/WBC). The results demonstrated that the combination of SVM, PCA, and MNLM achieved the optimal performance, yielding a high accuracy when distinguishing cancer cells from a white blood cell background, effectively simulating a post-erythrocytelysis clinical scenario. PCA-based feature analysis further provided biological interpretability by mapping discriminative bands to specific nucleic acid and protein markers.