Raman spectroscopy of plasma for biomarker quantification and radiation toxicity prediction in lung cancer

Prupose: Lung cancer is commonly treated using radiation therapy (RT), however treatment effectiveness is limited by the development of radiation-induced lung injury in a subset of patients. Raman spectroscopy is a rapid, non-destructive optical technique that is capable of probing the biochemical composition of a sample. This work investigates the use of Raman spectroscopy on pre-treatment plasma samples to predict the occurrence of clinically significant radiation pneumonitis (RP) in lung cancer patients. Methods: An optimized protocol for Raman spectral acquisition of dried plasma was first developed. Raman spectroscopy was performed on dried plasma from 142 patients: 69 patients presenting RP (grade ≥ 2 post-treatment) and 73 patients as controls (RP grade = 0 post-treatment). Analyses were first conducted to assess whether the acquired Raman spectra captured clinically relevant biochemical variation, using patient lipid panel measurements as references. Spectral analyses were subsequently applied to identify Raman features associated with RP outcome. A final multivariate logistic regression model was developed, combining patient clinical characteristics and Raman features selected using Least Absolute Shrinkage and Selection Operator (LASSO) regularization, and evaluated using a leave-one-patient-out (LOPO) framework. Results: Sampling location within the plasma droplet was identified as a major contributor to overall spectral variance. Meaningful associations between Raman spectral features and measured lipid biomarkers were observed, supporting the ability of the spectra to capture biochemically relevant variation. LOPO cross-validated models using LASSO-regularized feature selection achieved strong predictive performance for triglycerides (r = 0.87) and cholesterol (r = 0.81). A discriminative spectral region between 617−629.8 cm⁻¹ was identified between RP outcome groups, producing a logistic regression model with an area under the curve (AUC) of 0.66. Final LOPO cross-validated logistic regression models combining LASSO-selected Raman spectral features from the inner and outer droplet regions with patient clinical characteristics achieved an AUC of 0.80, showing improved predictive performance from models trained using Raman features alone (AUC = 0.72) or clinical characteristics alone (AUC = 0.69). Conclusion: These preliminary findings suggest that Raman spectroscopy of pre-treatment plasma captures features that may be used to predict whether a patient will develop radiation pneumonitis following radiotherapy.