Non-invasive screening for ovarian cancer by combining serum SERS with interpretable machine learning models
Early identification of malignant ovarian tumors is critical for informing treatment decisions and enhancing patients' quality of life. As the third most prevalent gynecologic cancer globally, ovarian cancer remains challenging to diagnose due to the high cost, limited accessibility, and radiation exposure associated with current screening techniques. This study integrates surface-enhanced Raman spectroscopy (SERS) with feature selection techniques and deep learning frameworks to construct a diagnostic model for detecting ovarian cancer based on serum component analysis. The goal is to realize efficient and precise non-invasive screening for the disease. High-quality SERS spectra were first collected from serum samples of patients with clinically confirmed ovarian cancer, healthy individuals, and those with ovarian endometrioma. Subsequently, the Light Gradient Boosting Machine (LightGBM) algorithm was employed as the base classifier to perform two-stage feature selection, utilizing both the model's intrinsic feature importance scores and SHapley Additive exPlanation (SHAP) values. Finally, a Deep Neural Network (DNN) was incorporated and trained via backpropagation to optimize the weights and biases of neuronal connections, thereby improving the predictive performance of the overall network model. After feature selection, the DNN algorithm achieved an accuracy rate of 92.03% in the five-fold cross-validation for the three types of recognition - healthy individuals, ovarian cancer, and potentially malignant ovarian endometrioma. In the evaluation of the independent test set, the accuracy rate still reached as high as 86.96%. In addition, compared with traditional machine learning algorithms, the classification performance of DNN is also the best. The findings above demonstrate that the integration of serum SERS with the robust LightGBM-DNN algorithm offers a promising strategy for clinical ovarian cancer screening.