Label-free serum SERS combined with RFE-GBDT algorithm for non-invasive screening of liver cancer

Early diagnosis of liver cancer is crucial for developing clinical treatment strategies and improving patient survival rates. However, current diagnostic methods are often invasive, complex, and time-consuming, making them unsuitable for early screening in practical settings. Therefore, there is an urgent need to develop efficient and convenient non-invasive diagnostic techniques. This study presents a non-invasive optical diagnostic approach based on surface-enhanced Raman spectroscopy (SERS) and a deep learning algorithm for liver cancer staging identification and auxiliary screening. We systematically collected high-quality SERS spectral data from serum samples of patients with different stages of liver cancer (T1, T2, T3), hepatitis B (HBV), and healthy controls (Normal). Recursive feature elimination (RFE) was employed for feature selection, eliminating redundant spectral bands and retaining features highly relevant to classification, which significantly enhanced the model's discriminative ability. The selected features were input then into a gradient boosting decision tree (GBDT) model. Through residual iterative optimization, the model effectively captured nonlinear feature interactions, and key spectral bands were interpreted using the local interpretable model-agnostic explanations (LIME) algorithm. Compared to other commonly used classifiers such as logistic regression (LR) and random forest (RF), the RFE-GBDT model demonstrated superior performance in liver cancer staging tasks, achieving an accuracy of 92.68% in the five-class classification. The results indicate that the integration of SERS technology with the RFE-GBDT algorithm holds promise as an efficient and non-invasive auxiliary tool for the early diagnosis of liver cancer.