A novel IagPLS baseline correction method for glioma identification using Raman spectroscopy.

As the most aggressive malignant tumours of the central nervous system, gliomas urgently require real-time intraoperative molecular diagnostic techniques to overcome the invasive limitations of conventional pathology and the uncertainties of imaging. Although Raman spectroscopy can provide non-invasive biomolecular fingerprinting information, its signal is susceptible to interference from baseline drift caused by tissue autofluorescence, resulting in masking of the effective information of the spectrum. Existing baseline correction methods (, polynomial fitting and asymmetric least squares) struggle to balance the challenges of noise suppression, feature preservation, and adaptation to heterogeneous tissue spectra. In this study, we propose an improved adaptive gradient-derived penalized least squares (IagPLS) method that integrates three innovative mechanisms: curvature-driven dynamic regularization, which dynamically adjusts the smoothing intensity through a gradient-sensitive penalty term and protects biomarker-rich regions while suppressing high-frequency noise; SHAP algorithm-guided feature protection, which identifies and diagnoses key Raman peaks and constructs region-specific weight constraints to avoid oversmoothing; and quantum-inspired global optimization, which models the weight update as a tunnelling potential well model and uses a Monte Carlo simulated annealing strategy to jump out of the local optimum. Based on the validation of 423 clinical Raman spectra (157 normal tissues/266 glioma tissues), IagPLS showed a significant advantage: the glioma identification accuracy of its corrected spectra reached 96.1% (tumour F1 score: 0.97) after random forest classification, which was significantly better than that of airPLS (89.4%) and agdPLS (87.0%). The key indicators show that the feature peak prominence of the spectra during IagPLS processing is improved by 82.05% compared to agdPLS, the negative residual area is reduced by 89.79% compared to airPLS, and the processing speed is improved by 43.64% compared to airPLS. SHAP interpretability analysis confirmed that the protected biomarker regions contributed 1.07-fold to classification and were highly compatible with glioma-specific spectral features. The algorithm takes less than 0.1 seconds for a single correction, combining biological interpretability with superior spectral correction to provide a reliable pre-processing tool for intraoperative optical biopsy systems. Its algorithmic framework can be extended to multimodal biomedical spectral analysis, such as near-infrared and mid-infrared spectroscopy, to promote the innovation of complex spectral pre-processing technology in precision medicine.