1. Introduction

Worldwide, cancer is the leading cause of death in every country, and the burden of its morbidity and mortality is increasing rapidly, and the situation of prevention and treatment is still grim [1]. According to estimates of global cancer statistics, there were 19.3 million new cancer cases worldwide in 2020 and nearly 10 million deaths. For women, cervical cancer ranks fourth in both incidence and mortality. In 2020, there were an estimated 604,000 new cases and 342,000 deaths in the world, posing a serious threat to the lives and health of women [1]. Cervical cancer is a malignant tumor that occurs in the female reproductive system. Early detection and accurate diagnosis combined with curative treatment is the key to improving survival [2,3]. In clinical practice, Papanicolaou (Pap) smear, human papillomavirus (HPV) testing, and colposcopy are the most commonly used screening tools. However, these methods have many limitations, such as cost, time, and subjectivity, and often fail to take into account both diagnostic sensitivity and specificity [3,4]. The current gold standard for diagnosing cervical cancer is histopathological examination, but this method is invasive, the patient needs to suffer a lot, and it is not suitable for mass screening [3,5]. Thus, the development of a rapid, non-invasive, and highly accurate cervical cancer screening method has important clinical value and socio-economic significance.

As an inelastic scattering vibration spectroscopy technology, Raman spectroscopy can provide fingerprint type information about the structure and conformation of the sample and has the advantages of fast, non-destructive, and almost no sample preparation [6–9]. However, in reality, the Raman scattering signal is very weak, and there is also a strong fluorescence background interference for biological samples, which hinders the clinical application of this technology [3,6,10]. Fortunately, the discovery of surface-enhanced Raman scattering (SERS) technology can effectively overcome the above-mentioned problems [11]. Thus, SERS has made great progress in the field of cancer and disease detection in recent years [6,7,10,12–16]. In clinical practice, compared with cell or tissue samples, blood samples are easier to obtain and have the advantages of minimal invasiveness, so they are more suitable as preliminary screening materials for diseases [10,17,18]. Both plasma and serum can be obtained from blood samples by centrifugation, and they vary according to whether clotting is allowed [19]. Currently, through the detection of body fluids, here refers to plasma and serum, several studies have been performed on the diagnosis of cervical cancer based on the SERS technique. Feng et al. applied the SERS technique on human plasma to distinguish cervical cancer patients from healthy controls. The diagnostic sensitivity of 96.7% and specificity of 92% were yielded based on principal component analysis (PCA) combined with linear discriminant analysis (LDA) method [3]. Sánchez-Rojo et al. distinguished serum samples from cervical cancer patients and healthy controls based on the SERS technology and also adopted the PCA-LDA method, which achieved high diagnostic sensitivity (88%) and specificity (88%) [20].

Hysteromyoma is a common benign tumor of the uterus. Some of its symptoms are similar to those of cervical cancer. Some patients are likely to confuse the two. Once misjudgment delays the condition, it may cause more serious consequences. Therefore, this study took patients with hysteromyoma into consideration and explored whether serum SERS technology could be used to distinguish hysteromyoma, cervical cancer, and healthy controls. Specifically, firstly, three groups of serum samples were collected, and the corresponding SERS spectra were recorded. Then support vector machine (SVM) algorithm was adopted to establish the three-category diagnosis model. Furthermore, the results of binary classification were also given, that is, the diagnosis results of health and hysteromyoma, health and cervical cancer, and hysteromyoma and cervical cancer were given. To the best of our knowledge, this is the first study to use serum SERS technology to discriminate hysteromyoma and cervical cancer from healthy controls.

2. Materials and methods

2.1 Serum samples

In this research, a total of 66 serum samples from patients were obtained. According to clinical and histopathological diagnosis, 30 cases were hysteromyoma (age range 36.4$\pm$4.3), and the remaining 36 cases were cervical cancer (age range 47.3$\pm$8.9). The serum samples from 30 healthy female volunteers (age range 44.1$\pm$11.9) were also collected as the control group. Among the 36 cervical cancer samples, 17 belonged to stage I, 13 were classified as stage II, 2 were stage III, and 4 were unstaged. All study subjects were from the First Affiliated Hospital of Xinjiang Medical University, and ethical approval was obtained in order to study human serum samples. The serum samples were prepared in accordance with standard hospital procedures and then stored in the refrigerator at −80$^{\circ }\textrm {C}$ until SERS spectra were recorded.

2.2 Preparation of silver nanoparticles

Silver (Ag) colloids were provided by Simple & Smart Instrument Equipment Co., Ltd (Nanjing, China). Briefly, 200 ml of silver nitrate solution with a concentration of 1.0 mM was first heated to boil, and then 5.0 ml of 1% trisodium citrate was added dropwise with vigorous stirring and continued to boil for 1 h until the solution turned gray. After the solution was cooled, add distilled water to maintain the volume of 200 ml [21]. The prepared Ag colloid was then characterized by transmission electron microscopy (TEM) (JEM-2100, JEOL Ltd., Japan) and ultraviolet (UV)-visible absorption spectroscopy (AvaSpec-ULS2048CL-EVO, Avantes Inc., The Netherlands).

2.3 SERS measurements

Before SERS measurement, serum samples were thawed at room temperature, and then 10 $\mathrm{\mu}$l Ag colloid was mixed with 10 $\mathrm{\mu }$l blood serum in 1:1 proportion. After incubating for about 1 h at 4 $^{\circ }\textrm {C}$, a drop of the resulting mixture was transferred onto an aluminum foil and naturally dried for SERS collection. A Raman micro-spectrometer (ATR8300MP, Optosky Photonics Inc., China) equipped with a 785 nm diode laser was used for the measurement of serum SERS spectra in the range of 400–1800 cm$^{-1}$. A 20 $\times$ microscope objective lens (NA = 0.40) was used, and the laser power and integration time were 1 mW and 3 s, respectively. Four spectra were recorded for each sample, and the average spectrum was utilized for further processing and analysis. Thus, a total of 96 serum SERS spectra were obtained, including 30 spectra of healthy controls, 30 spectra of hysteromyoma, and 36 spectra of cervical cancers.

2.4 SERS data processing and SVM classification

The raw SERS spectra were firstly smoothed by the Savitzky-Golay algorithm (polynomial order 5, frame length 7) to reduce the noise, and the baseline was subsequently removed by the adaptive iteratively reweighted penalized least squares (airPLS) algorithm [22]. Finally, each spectrum was processed by vector normalization. Specifically, each Raman intensity was divided by the ’norm’ to obtain the normalized spectrum [23]. The spectral data were then submitted to the supervised classification algorithm, i.e., SVM, to build the diagnostic models. The C-SVM model based on LIBSVM package (version 3.22) was adopted for SVM classifications [24]. Three commonly used kernel functions, namely linear, polynomial, and Gaussian radial basis function (RBF) were used to build classification models. Each model was evaluated by the leave-one-out cross-validation (LOOCV) method [10]. All programs were implemented in MATLAB language (R2020a, MathWorks Inc., USA).

3. Results and discussion

3.1 Enhancement effect of Ag colloid and SERS spectra analysis

Figure 1(A) shows the absorption spectra of the pure Ag colloid and the mixture of Ag colloid and serum. The absorption maximum of pure Ag colloid is located at 415 nm, and a small red shift occurs in the absorption peak after the addition of serum, indicating that the mixture was aggregated to some extent [25]. To explore the detection sensitivity of the prepared Ag colloid, we mixed different concentrations of rhodamine 6G (R6G) aqueous solution with the Ag colloid in 1:1 proportion and carried out SERS collection after they were naturally dried, and the results are shown in Fig. 6(A). According to the equation [26], the enhancement factor of Ag colloid is approximately 2.42 $\times$ 10$^{6}$. In order to assess the enhancement effect of Ag colloid on serum samples, we compared the spectra before and after Raman scattering enhancement in Fig. 1(B). In Fig. 1(B), (a) and (b) are SERS spectrum and regular Raman spectrum of the same serum sample under the same experimental parameters. It can be seen that some basic biochemical substances in the serum, such as proteins, nucleic acids, etc., are adsorbed on the Ag nanoparticles, resulting in a significant increase in the Raman signal, while the Raman peak is not observed in the serum sample without Ag colloid [7]. In (c), by increasing the integration time and laser power, some weak Raman peaks of the serum sample can be obtained, but at this time, the spectral noise is larger, and the spectral peaks are less. (d) is the Raman spectrum of the Ag colloid placed on the aluminum foil. Compared with (a), the Raman signal of Ag nanoparticles is very weak and has no effect on the SERS spectrum of serum. Further, we measured the spectra of different proportions of Ag colloid mixed with serum (Fig. 6(B)). It can be seen that even a small amount of Ag colloid mixed with serum (0.1:1) can also detect the spectral signal. When the two were mixed in a ratio of 1:1, a better serum SERS spectrum could be obtained.

 

Fig. 1. (A) The UV/visible absorption spectra of the Ag colloid and the mixture of Ag colloid and serum. The inset picture in the upper right corner shows the TEM micrograph of pure Ag colloid. (B) Comparison of (a) SERS spectrum of the serum sample from a cervical cancer patient, (b) the regular Raman spectrum of the same serum sample without the Ag nanoparticles, (c) the regular Raman spectrum of the same serum sample without the Ag nanoparticles under different experimental parameters, and (d) the Raman signal of the Ag colloid. The experimental parameters of (a), (b), and (d) are the same, the integration time is 3 s, and the laser power is 1 mW. However, the integration time of (c) is 20 s, and the laser power is 100 mW.

Download Full Size | PDF

Figure 2(A) displays the normalized mean SERS spectra after pre-processing from cervical cancer, hysteromyoma, and healthy serum samples in the range of 400-1800 cm$^{-1}$, respectively. The light gray shaded area represents the standard deviations of the means. The SERS peaks located at 489, 631, and 722 cm$^{-1}$ have the strongest signals and can be obviously observed in all three groups. In addition, other relatively weak peaks at 583, 807, 884, 956, 1001, 1095, 1130, 1205, 1326, 1440, 1583, and 1654 cm$^{-1}$ can also be seen. These peaks in the serum SERS spectra correspond to different biochemicals and vibration modes. According to previously published literature [3,7,10,17,25,27], Table 1 lists the tentative assignments of the observed SERS bands. The mean SERS spectra of the three groups of serum samples have similar spectral profiles, such as the position and bandwidth of the Raman peaks, which indicates that there is little difference between the serum biochemical components of patients with hysteromyoma and cervical cancer and those of healthy controls. Figure 2(B) shows the histogram of the mean intensities values and the corresponding standard deviations of the three groups of serum SERS peaks. The significantly different peaks among the three groups were determined by one-way ANOVA with the definition of p-value < 0.05 [28]. Although the spectral profiles are relatively similar between the different groups, there were still statistical differences in the intensity values of some spectral peaks. These subtle differences make it possible to screen hysteromyoma and cervical cancer with serum SERS technology combined with the SVM algorithm.

 

Fig. 2. (A) Comparison of normalized mean SERS spectra of healthy control group, hysteromyoma, and cervical cancer, and shaded area represents the standard deviations. (B) Histogram of the mean intensities and standard deviations of SERS peaks among three groups. * corresponds to p < 0.05.

Download Full Size | PDF

Table 1. The peak positions and tentative assignments of the primary Raman bands.^a

View Table | View all tables in this article

3.2 Classification of serum spectral data based on SVM algorithm

In this study, the SVM algorithm was adopted to develop diagnostic models for classifying the serum SERS spectra of the three groups. Three common kernel functions, namely linear, polynomial, and RBF kernel functions, were used for comparison. For the linear kernel SVM algorithm, there is a penalty parameter C that needs to be optimized, which can make the best trade-off between training error and generalization ability [10]. Unlike the linear kernel, the RBF kernel can non-linearly map the data to a higher-dimensional space so that it can handle the case where the relationship between the class label and the attribute is non-linear [29]. The RBF kernel has two parameters that need to be optimized, namely the penalty parameter C and the Gaussian radial width $\gamma$. In addition to linear and RBF kernels, the polynomial kernel is also a commonly used basic kernel. However, the polynomial kernel has more hyperparameters to be considered, which increases the complexity of the model [29]. For the polynomial kernel function, the penalty parameter C, the kernel parameter $\gamma$, and the order d of the polynomial were optimized. The study shows that trying exponentially growing sequences is a practical method to identify good parameters in parameter optimization [29]. Thus, for the linear kernel SVM algorithm, by maximizing the classification accuracy as the optimization criterion to obtain the optimal C, the search was performed in the range of 2$^{-5}$ to 2$^{20}$ with a step of the power of two. In Fig. 3, (a) shows the LOOCV accuracy as a function of parameter log$_2$C. It can be seen that the penalty parameter C has a great influence on the diagnostic accuracy of the model. When log$_2$C = 0, the cross-validation only has a diagnostic accuracy of 24.0%. By changing the parameter C, when C = 2$^6$, the overall accuracy for classifying the three groups can reach 84.4%. For the two hyperparameters C and $\gamma$ of the RBF kernel SVM model, the grid search method was used to optimize. The range of C was set from 2$^{-5}$ to 2$^{15}$, and $\gamma$ from 2$^{-5}$ to 2$^{10}$, increasing in steps of powers of two. For different parameter combinations, the pair with the best cross-validation accuracy was selected. In Fig. 4, (a) shows the classification accuracy of the RBF kernel SVM model to classify the three groups under different combinations of parameters C and $\gamma$. The darker the color indicates, the higher the classification accuracy. When C = 2$^7$ and $\gamma$ = 2$^0$, the classification accuracy of cross-validation reaches the highest, 85.4%. For the polynomial kernel SVM model, this study considered three parameters: C, $\gamma$, and the polynomial order d. Because there are three parameters that need to be optimized, this process is extremely time-consuming. Therefore, this research first fixed the order of the polynomial to optimize the penalty parameter C and the kernel parameter $\gamma$. The optimization range and process of the parameters C and $\gamma$ were the same as the RBF kernel SVM model. Then changed the order of the polynomial to find the optimal order d. It can be seen from Fig. 5(a) that when the order of the polynomial d = 5, the model achieves the highest classification accuracy, and the corresponding accuracy is 86.5%. For classifying the three groups, the LOOCV accuracy of linear, RBF, and polynomial kernels are 84.4%, 85.4%, and 86.5%, respectively. The classification performances of the three kernel functions are similar. Among them, the SVM model based on the polynomial kernel achieves the best results. Table 2 shows the corresponding confusion matrix and correct rates. However, if compared with the time complexity of each model, the parameter optimization time of the linear kernel SVM model is the shortest, followed by the RBF kernel, and the model based on the polynomial kernel is the most time-consuming. When the classification performance of the three kernel functions is close, it is a good choice to use a linear kernel to build the diagnostic model.

 

Fig. 3. LOOCV accuracy of the linear kernel SVM models as a function of parameter C. (a) represents the classification results of the three groups, and (b) to (d) represent the binary classification results of healthy and hysteromyoma, healthy and cervical cancer, and hysteromyoma and cervical cancer, respectively.

Download Full Size | PDF

 

Fig. 4. LOOCV classification performance as a function of parameter C and $\gamma$ using the RBF kernel SVM models. The darker the color indicates the higher the classification accuracy. The right side of each image gives the highest and lowest classification accuracy of each model, and also gives the optimal parameter combination corresponding to the model when the highest accuracy is achieved. The models represented by (a) to (d) are the same as (a) to (d) in Fig. 3.

Download Full Size | PDF

 

Fig. 5. LOOCV accuracy of the polynomial kernel SVM models as a function of polynomial order d. The models represented by (a) to (d) are the same as (a) to (d) in Fig. 3.

Download Full Size | PDF

Table 2. Confusion matrix for classifying the three groups using polynomial kernel SVM algorithm.

View Table | View all tables in this article

Furthermore, we evaluated the performance of using the serum SERS technology combined with the SVM algorithm in the binary classification: the distinction between health and hysteromyoma, health and cervical cancer, and hysteromyoma and cervical cancer. The parameter optimization process of binary classification is the same as above. It can be seen from Figs. 3–5 that the result of binary classification is better than the classification of three groups. Especially for the distinction between health and hysteromyoma, health and cervical cancer, the classification accuracy of the three kernel functions is higher than 90%, and the results are satisfactory. For the classification between hysteromyoma and cervical cancer, in addition to the linear kernel SVM model, the classification accuracy is also more than 90%. The performance of the SVM algorithm based on the linear kernel is slightly worse, and the classification accuracy is 87.9%. The confusion matrixes of the binary classification results based on the SVM algorithm are detailed in Tables 3–5.

Table 3. Confusion matrix to distinguish between healthy and hysteromyoma with different algorithms.

View Table | View all tables in this article

Table 4. Confusion matrix to distinguish between healthy and cervical cancer with different algorithms.

View Table | View all tables in this article

Table 5. Confusion matrix to distinguish between hysteromyoma and cervical cancer with different algorithms.

View Table | View all tables in this article

The performance of the above classification models was all evaluated using the LOOCV method, but the diagnostic ability of the classification algorithm can also be evaluated by the prediction accuracy of unknown test samples [30]. Thus, we further selected 10 samples from 96 samples for blind testing, including 3 healthy cases, 3 hysteromyoma cases, and 4 cervical cancer cases. The remaining 86 samples were used to build the classification model. Specifically, the five-fold cross-validation method was first used to find the optimal parameters, then the model was trained and finally used to predict the unknown 10 blind samples. For the linear kernel SVM model, the sample prediction is wrong in 2 out of 10 cases, while the RBF kernel SVM model is only wrong in 1 case. Based on the polynomial kernel model, by selecting the appropriate order, the same result as the RBF kernel can also be obtained. Table 6 shows the blind test results for three groups using the RBF kernel-based SVM algorithm. As can be seen from Table 6, 3 cases of hysteromyoma and 4 cases of cervical cancer are predicted correctly. Among the 3 healthy samples, only 1 case is misclassified as cervical cancer. Overall, the results obtained based on SVM algorithm above are very interesting and promising.

Table 6. Blind test results for three groups using RBF kernel based SVM algorithm. Health, hysteromyoma, and cervical cancer are represented by 0, 1, and 2, respectively.

View Table | View all tables in this article

3.3 Comparison with principal component analysis combined with linear discriminant analysis

SVM is a machine learning method based on statistical learning theory. Due to the existence of the kernel functions, it can not only deal with linearly separable problems but also can be used for problems where the data are not linearly separable, so it is considered to be superior to traditional linear classification approaches, such as principal component analysis (PCA)-linear discriminant analysis (LDA) [10,31]. For our spectral data, to verify whether the SVM-based performances are better than the traditional linear method, we compared the LOOCV results of PCA-LDA with the above SVM classification results. For the classification of the three groups, one-way ANOVA was used to determine the most diagnostically significant PC scores in distinguishing hysteromyoma, cervical cancer, and healthy control group using the definition of p < 0.05 [28]. As a result, seven PC scores, accounting for 91.75% of the total variance, were selected and input to LDA to construct the diagnostic model. Table 7 shows the results of classifying the three groups using the PCA-LDA algorithm. The cross-validation accuracy based on PCA-LDA is 82.3%, which is lower than the SVM models. For the distinction between healthy and hysteromyoma, three diagnostically significant PC scores (PC1 < 0.05, PC3 < 0.05, PC6 < 0.05) were obtained, and the corresponding confusion matrix is shown in Table 3. Similarly, Tables 4 and 5 show the classification results of health and cervical cancer, hysteromyoma and cervical cancer, respectively. In general, the classification results of SVM are slightly better than PCA-LDA, especially for distinguishing three classes.

4. Conclusion and outlook

In summary, this paper investigated the use of serum SERS technology to distinguish patients with hysteromyoma and cervical cancer from healthy controls. High-quality SERS data can be obtained by mixing a small volume of serum with the Ag colloid. The SVM algorithm was adopted to establish the diagnosis models, and the classification results based on the three kernel functions were compared and analyzed. In addition, the results based on PCA-LDA were also given, which demonstrates the superiority of the SVM algorithm. The research indicates that the combination of serum SERS and SVM algorithm has great application potential in screening hysteromyoma and cervical cancer. Compared with clinical diagnosis methods, the method proposed in this study has the advantages of being rapid, non-invasive, and label-free.

However, this study also has some flaws. For example, the sample size of each class is not enough, and more serum samples need to be collected to further evaluate the utility of this method in clinical detection. At present, cervical cancer samples are mainly I and II, and it is necessary to incorporate into more stages III and IV and to explore whether the serum-based SERS method can be used for the diagnosis of cervical cancer at various stages. In addition, as the sample size increases, different machine learning and deep learning algorithms can be employed and compared to find the best model for screening hysteromyoma and cervical cancer.

Appendix

 

Fig. 6. (A) SERS spectra of R6G with the different concentrations (from 10$^{-4}$ M to 10$^{-9}$ M). (B) Spectra of Ag colloid mixed with serum in different ratios.

Download Full Size | PDF

Table 7. Confusion matrix for classifying the three groups using PCA-LDA algorithm.

View Table | View all tables in this article

Funding

National Natural Science Foundation of China (62071059, 61801042); BUPT Excellent Ph.D. Students Foundation (CX2020113); State Key Laboratory of Pathogenesis, Prevention and Treatment of Central Asian High Incidence Diseases Fund (SKL-HIDCA-2019-5); Science and Technology Department of Xinjiang Uyghur Autonomous Region (2019E0282).

Disclosures

The authors declare no conflicts of interest.

Data availability

Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.

References

1. H. Sung, J. Ferlay, R. L. Siegel, M. Laversanne, I. Soerjomataram, A. Jemal, and F. Bray, “Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries,” Ca-Cancer J. Clin. 71(3), 209–249 (2021). [CrossRef]  

2. E. Weiderpass, D. Hashim, and F. Labrèche, “Malignant tumors of the female reproductive system,” in Occupational Cancers, (Springer, 2020), pp. 439–453.

3. S. Feng, D. Lin, J. Lin, B. Li, Z. Huang, G. Chen, W. Zhang, L. Wang, J. Pan, R. Chen, and H. Zeng, “Blood plasma surface-enhanced Raman spectroscopy for non-invasive optical detection of cervical cancer,” Analyst 138(14), 3967–3974 (2013). [CrossRef]  

4. K. Aljakouch, Z. Hilal, I. Daho, M. Schuler, S. D. Krauß, H. K. Yosef, J. Dierks, A. Mosig, K. Gerwert, and S. F. El-Mashtoly, “Fast and noninvasive diagnosis of cervical cancer by coherent anti-stokes Raman scattering,” Anal. Chem. 91(21), 13900–13906 (2019). [CrossRef]  

5. J. Mo, W. Zheng, J. J. Low, J. Ng, A. Ilancheran, and Z. Huang, “High wavenumber Raman spectroscopy for in vivo detection of cervical dysplasia,” Anal. Chem. 81(21), 8908–8915 (2009). [CrossRef]  

6. S. Feng, R. Chen, J. Lin, J. Pan, G. Chen, Y. Li, M. Cheng, Z. Huang, J. Chen, and H. Zeng, “Nasopharyngeal cancer detection based on blood plasma surface-enhanced Raman spectroscopy and multivariate analysis,” Biosens. Bioelectron. 25(11), 2414–2419 (2010). [CrossRef]  

7. S. Chen, S. Zhu, X. Cui, W. Xu, C. Kong, Z. Zhang, and W. Qian, “Identifying non-muscle-invasive and muscle-invasive bladder cancer based on blood serum surface-enhanced Raman spectroscopy,” Biomed. Opt. Express 10(7), 3533–3544 (2019). [CrossRef]  

8. J. González-Solís, J. Rodríguez-López, J. C. Martínez-Espinosa, C. Frausto-Reyes, L. Jave-Suárez, A. Aguilar-Lemarroy, H. Vargas-Rodríguez, and E. Martínez-Cano, “Detection of cervical cancer analyzing blood samples with Raman spectroscopy and multivariate analysis,” in AIP Conference Proceedings, vol. 1226 (American Institute of Physics, 2010), pp. 91–95.

9. A. Huefner, W. L. Kuan, S. L. Mason, S. Mahajan, and R. A. Barker, “Serum Raman spectroscopy as a diagnostic tool in patients with Huntington’s disease,” Chem. Sci. 11(2), 525–533 (2020). [CrossRef]  

10. S. Li, Y. Zhang, J. Xu, L. Li, Q. Zeng, L. Lin, Z. Guo, Z. Liu, H. Xiong, and S. Liu, “Noninvasive prostate cancer screening based on serum surface-enhanced Raman spectroscopy and support vector machine,” Appl. Phys. Lett. 105(9), 091104 (2014). [CrossRef]  

11. M. Fleischmann, P. J. Hendra, and A. J. McQuillan, “Raman spectra of pyridine adsorbed at a silver electrode,” Chem. Phys. Lett. 26(2), 163–166 (1974). [CrossRef]  

12. X. Li, T. Yang, S. Li, D. Wang, Y. Song, and K. Yu, “Different classification algorithms and serum surface enhanced Raman spectroscopy for noninvasive discrimination of gastric diseases,” J. Raman Spectrosc. 47(8), 917–925 (2016). [CrossRef]  

13. G. Cennamo, D. Montorio, V. B. Morra, C. Criscuolo, R. Lanzillo, E. Salvatore, C. Camerlingo, M. Lisitskiy, I. Delfino, M. Portaccio, and M. Lepore, “Surface-enhanced Raman spectroscopy of tears: toward a diagnostic tool for neurodegenerative disease identification,” J. Biomed. Opt. 25(08), 1 (2020). [CrossRef]  

14. D. Lin, S. Feng, J. Pan, Y. Chen, J. Lin, G. Chen, S. Xie, H. Zeng, and R. Chen, “Colorectal cancer detection by gold nanoparticle based surface-enhanced Raman spectroscopy of blood serum and statistical analysis,” Opt. Express 19(14), 13565–13577 (2011). [CrossRef]  

15. D. Lin, Y. Wang, T. Wang, Y. Zhu, X. Lin, Y. Lin, and S. Feng, “Metabolite profiling of human blood by surface-enhanced Raman spectroscopy for surgery assessment and tumor screening in breast cancer,” Anal. Bioanal. Chem. 412(7), 1611–1618 (2020). [CrossRef]  

16. X. Lin, D. Lin, Y. Chen, J. Lin, S. Weng, J. Song, and S. Feng, “High throughput blood analysis based on deep learning algorithm and self-positioning super-hydrophobic sers platform for non-invasive multi-disease screening,” Adv. Funct. Mater. 31(51), 2103382 (2021). [CrossRef]  

17. Q. Wu, S. Qiu, Y. Yu, W. Chen, H. Lin, D. Lin, S. Feng, and R. Chen, “Assessment of the radiotherapy effect for nasopharyngeal cancer using plasma surface-enhanced Raman spectroscopy technology,” Biomed. Opt. Express 9(7), 3413–3423 (2018). [CrossRef]  

18. A. A. Bunaciu, Ş. Fleschin, V. D. Hoang, and H. Y. Aboul-Enein, “Vibrational spectroscopy in body fluids analysis,” Crit. Rev. Anal. Chem. 47(1), 67–75 (2017). [CrossRef]  

19. D. R. Parachalil, J. McIntyre, and H. J. Byrne, “Potential of Raman spectroscopy for the analysis of plasma/serum in the liquid state: recent advances,” Anal. Bioanal. Chem. 412(9), 1993–2007 (2020). [CrossRef]  

20. S. Sánchez-Rojo, B. Martínez-Zérega, E. Velázquez-Pedroza, J. Martínez-Espinosa, L. Torres-González, A. Aguilar-Lemarroy, L. Jave-Suárez, P. Palomares-Anda, and J. González-Solís, “Cervical cancer detection based on serum sample surface enhanced Raman spectroscopy,” Rev. Mex. Fis. 62, 213–218 (2016).

21. M. Kashif, M. I. Majeed, M. A. Hanif, and A. ur Rehman, “Surface enhanced Raman spectroscopy of the serum samples for the diagnosis of Hepatitis C and prediction of the viral loads,” Spectrochim. Acta, Part A 242, 118729 (2020). [CrossRef]  

22. Z. M. Zhang, S. Chen, and Y. Z. Liang, “Baseline correction using adaptive iteratively reweighted penalized least squares,” Analyst 135(5), 1138–1146 (2010). [CrossRef]  

23. R. Gautam, S. Vanga, F. Ariese, and S. Umapathy, “Review of multidimensional data processing approaches for Raman and infrared spectroscopy,” EPJ Tech. Instrum. 2(1), 8–38 (2015). [CrossRef]  

24. C. C. Chang and C. J. Lin, “Libsvm: a library for support vector machines,” ACM T. Intel. Syst. Tec. 2(3), 1–27 (2011). [CrossRef]  

25. K. Zhang, X. Liu, B. Man, C. Yang, C. Zhang, M. Liu, Y. Zhang, L. Liu, and C. Chen, “Label-free and stable serum analysis based on Ag-NPs/PSi surface-enhanced Raman scattering for noninvasive lung cancer detection,” Biomed. Opt. Express 9(9), 4345–4358 (2018). [CrossRef]  

26. N. Cheng, D. Chen, B. Lou, J. Fu, and H. Wang, “A biosensing method for the direct serological detection of liver diseases by integrating a SERS-based sensor and a CNN classifier,” Biosens. Bioelectron. 186, 113246 (2021). [CrossRef]  

27. S. Feng, J. Pan, Y. Wu, D. Lin, Y. Chen, G. Xi, J. Lin, and R. Chen, “Study on gastric cancer blood plasma based on surface-enhanced Raman spectroscopy combined with multivariate analysis,” Sci. China Life Sci. 54(9), 828–834 (2011). [CrossRef]  

28. J. Lin, Z. Huang, X. Lin, Q. Wu, K. Quan, Y. Cheng, M. Zheng, J. Xu, Y. Dai, H. Qiu, D. Lin, and S. Feng, “Rapid and label-free urine test based on surface-enhanced Raman spectroscopy for the non-invasive detection of colorectal cancer at different stages,” Biomed. Opt. Express 11(12), 7109–7119 (2020). [CrossRef]  

29. C. W. Hsu, C. C. Chang, and C. J. Lin, “A practical guide to support vector classification,” (2003).

30. Y. Yu, Y. Lin, C. Xu, K. Lin, Q. Ye, X. Wang, S. Xie, R. Chen, and J. Lin, “Label-free detection of nasopharyngeal and liver cancer using surface-enhanced Raman spectroscopy and partial lease squares combined with support vector machine,” Biomed. Opt. Express 9(12), 6053–6066 (2018). [CrossRef]  

31. A. Maguire, I. Vega-Carrascal, J. Bryant, L. White, O. Howe, F. Lyng, and A. Meade, “Competitive evaluation of data mining algorithms for use in classification of leukocyte subtypes with Raman microspectroscopy,” Analyst 140(7), 2473–2481 (2015). [CrossRef]