Diagnosis of uterine diseases by label-free serum SERS fingerprints with machine learning

Early detection of uterine diseases is critically important for women's reproductive health. Here, we propose a novel and robust serum-based SERS analysis platform that integrates machine learning algorithms. This is the first application of it in the diagnosis and classification of various uterine diseases. We firstly collected serum samples from a cohort consisting of 27 healthy volunteers and 61 patients with clinically diagnosed uterine diseases. By simply mixing serum samples with silver nanoparticles (Ag NPs), we can obtain the stable and repeatable SERS spectra within 20 s. The optimized 1D-CNN model achieved outstanding performance in classifying the three uterine diseases compared with the other seven methods, attaining a comprehensive accuracy of 93.32 %, precision of 93.73 %, recall of 92.99 % and F1 score of 93.32 %. These results underscore the model's strong discriminative capability and reliability for clinical diagnostic applications. Furthermore, we enhanced the interpretability of the classification results of the 1D-CNN model by implementing Grad-CAM algorithm, thereby significantly improving the reliability and clinical applicability of the classification results. The integration of label-free SERS technology with advanced machine learning algorithms has demonstrated significant potential in uterine disease diagnostics, offering a comprehensive solution that combines high accuracy, exceptional sensitivity, operational simplicity, non-invasive and low-cost.