Introduction

Nasopharyngeal carcinoma (NPC) is a malignant tumor originating in the epithelial cells of the nasopharynx with a high incidence in southern China and Southeast Asia. Global cancer statistics estimated 120,416 incidences and 73,476 deaths from NPC worldwide in 20221. Asia accounted for 83.3% of global NPC incidence (100,298 cases) and 83.6% of deaths (61,442 cases)2. China had the highest incidence rate of 51,010 cases with an age-standardized incidence rate of 2.4 cases per 100,000 people2. Early diagnosis of NPC remains challenging due to its deep anatomical location and the lack of overt clinical symptoms in its early stages. Consequently, most patients with NPC are diagnosed at an advanced stage, which increases the risk of cervical lymph node and/or distant metastasis and negatively affects prognosis3. This highlights the urgent need to identify novel biomarkers for the early detection of NPC. Despite recent advances in NPC diagnostics, nasopharyngeal endoscopy combined with histopathological examination remains the gold standard. However, the compliance rate for this procedure is low due to its invasiveness and high costs. The Epstein-Barr virus (EBV) is a significant etiological factor for NPC, and various EBV-related tests, including those for EBV DNA and antibodies, have been developed. Nevertheless, their sensitivity and specificity are insufficient to meet clinical needs4,5. As a result, there is a critical demand for a noninvasive, high-precision method for NPC diagnosis.

S100A9 is a member of the S100 family of calcium-binding proteins and is closely associated with tumorigenesis and progression. It binds Ca2+ and Zn2+ ions with high selectivity and affinity, exhibiting extracellular and intracellular regulatory activities. S100A9 is involved in processes such as cell migration, arachidonic acid metabolism, and bone marrow cell maturation6. Previous studies have shown that S100A9 expression is upregulated in various primary and invasive tumors7,8. Li et al.9 utilized laser capture microdissection to isolate stromal cells from NPC and normal nasopharyngeal epithelial tissues, followed by two-dimensional fluorescence difference gel electrophoresis and mass spectrometry, to demonstrate elevated S100A9 protein levels in NPC stromal cells. However, its expression in the serum of patients with NPC has rarely been reported.

CCL5, a member of the C-C motif chemokine family, binds with high affinity to its receptor CCR5. It is generally recognized that CCL5 attracts T cells and monocytes. CCL5 plays a dual role in cancer: on one hand, it accelerates tumor progression and metastasis10,11; conversely, it enhances anti-tumor immunity by recruiting anti-tumor T cells and dendritic cells to the tumor microenvironment12,13. Overexpression of CCL5 has been observed in several cancers, including breast cancer14,15, esophageal cancer16, gastric cancer17, colorectal carcinoma18,19, prostate cancer20, pancreatic cancer21, and acute lymphocytic leukemia22. However, its role in NPC has not been elucidated.

In this study, we investigated the expression levels of S100A9, CCL5, routine blood indices, and common biochemical indices in the serum of patients with NPC compared with non-NPC controls, including patients with benign nasopharyngeal lesions and healthy individuals. Additionally, we successfully developed models to evaluate the diagnostic and predictive utility of these blood biomarkers for NPC.

Materials and methods

Sample collection

This cross-sectional study recruited 123 patients with NPC and 107 non-NPC controls (34 patients with benign nasopharyngeal conditions such as nasal polyps or rhinitis and 73 healthy subjects) from Xiangya Hospital of Central South University between January 2024 and May 2024. The inclusion criteria were: (a) diagnosis of NPC confirmed by histopathology and (b) availability of complete clinicopathological and baseline data. The exclusion criteria were (a) pregnancy, (b) the presence of other malignancies or prior antitumor treatment, (c) hemolyzed samples, and (d) incomplete patient information. The patient selection process is illustrated in the flowchart in Fig. 1.

Fig. 1
figure 1

Flowchart depicting patient selection.

Full size image

Demographic and clinicopathological data, including sex, age, and cancer stage, were collected. Laboratory examinations included routine blood tests and biomedical analyses. All laboratory data were obtained before initiating treatment. All participants provided informed consent, and this study was approved by the Ethics Committee of Xiangya Hospital of Central South University.

An external validation cohort was also recruited, comprising 19 patients with an initial diagnosis of NPC and 19 non-NPC controls (four with benign nasopharyngeal conditions and 15 healthy individuals) who visited Xiangya Hospital between June and July 2024.

Blood routine examination

Venous blood (3–5 ml) was collected in a K2EDTA anticoagulant tube in the morning, and routine blood examinations were performed using the Beckman Coulter DxH 800 blood cell analyzer. The following counts were measured: WBC, neutrophil, lymphocyte, monocyte, and platelet (PLT). The inflammatory markers were calculated as follows: the neutrophil-lymphocyte ratio (NLR), which is the ratio of neutrophils to lymphocytes; the lymphocyte-monocyte ratio (LMR), which is the ratio of lymphocytes to monocytes; and the PLR (platelet-lymphocyte ratio), which is the ratio of platelets to lymphocytes.

Biomedical test

Venous blood (3–5 ml) was collected in a tube containing separation gel after 10 h of fasting, and the blood was centrifuged at 3000 rpm for 10 min. Biomedical tests were performed using the Beckman AU5800 automatic biochemical analyzer. The following levels were measured: total protein (TP), albumin (ALB), globulin (GLB), total bilirubin (TBIL), direct bilirubin (DBIL), total bile acid (TBA), alanine transaminase, aspartate aminotransferase, triglyceride, total cholesterol, high-density lipoprotein cholesterol, low-density lipoprotein cholesterol (LDLC).

Measurement of S100A9 and CCL5

All serum samples were quickly frozen after centrifugation and stored at − 80 °C until testing. The samples were thawed only once to avoid freeze-thaw cycles. Serum levels of S100A9 and CCL5 were measured using ELISA kits (Beyotime, Nanjing, China) in accordance with the manufacturer’s instructions.

Statistical analysis

Statistical analysis was performed using SPSS software (version 27.0; SPSS Inc., IL, USA). Data normality was assessed using the Kolmogorov–Smirnov test. Continuous variables with a normal distribution are presented as mean ± standard deviation; in contrast, nonnormally distributed data are expressed as medians with interquartile ranges. Group comparisons were conducted using the Mann–Whitney U test for continuous variables and the Chi-square test for categorical variables. Logistic regression analysis was employed to identify risk factors of NPC. Variables with P < 0.05 in the univariate logistic regression analysis were included in the multivariate logistic regression analysis to determine independent risk factors for NPC. Receiver operating characteristic (ROC) curves were generated to assess the diagnostic value of blood indicators for NPC. P < 0.05 was considered statistically significant.

Predictive model construction using machine learning

Feature selection and predictive model construction were performed using Deepwise and the Beckman Coulter DxAI platform v2.0 (https://www.xsmartanalysis.com/). LASSO regression was applied to determine the importance of the influencing factors. Following feature selection, predictive models were developed using four machine-learning algorithms: logistic regression, XGBoost classifier, LightGBM classifier, and random forest classifier. Five-fold cross-validation was used to evaluate the performance of these algorithms. In this method, the dataset was randomly divided into five subsets: four subsets served as the training set for model construction; in contrast, the remaining subset was used as the internal validation set. This process was repeated five times to ensure a robust evaluation of the model performance.

The optimal predictive model was selected based on its performance in the internal validation set. In the training and validation sets, ROC curves and the area under the ROC curve (AUC) were used to assess the model’s performance. Calibration was further evaluated graphically using calibration plots. Decision Curve Analysis (DCA) was employed to evaluate the clinical utility of the model; in contrast, Shapley additive explanation (SHAP) plots were used to illustrate feature rankings. All P values were two-sided, and P < 0.05 was considered statistically significant.

Results

Comparisons of clinical characteristics of participants

A total of 230 participants were enrolled in the study, comprising 123 individuals with NPC and 107 non-NPC participants (34 with benign nasopharyngeal diseases such as nasal polyps or rhinitis and 73 healthy subjects). The clinical characteristics of the study participants are summarized in Table 1. There were no significant differences in age or sex between the two groups (P > 0.050). Among the 123 patients with NPC, 83 (67.5%) were male, and 40 (32.5%) were female, compared with 70 males (65.4%) and 37 females (34.6%) in the non-NPC group. The mean age of the NPC group was 52.26 ± 10.80 years, compared to 49.50 ± 12.11 years in the non-NPC group. The distribution of NPC stages was as follows: 53 patients (12.2%) in stages I + II, 45 patients (36.6%) in stage III, and 64 patients (51.2%) in stage IV.

Table 1 Demographic and clinical characteristics of the study subjects.
Full size table

The NPC group showed significantly higher levels of S100A9, CCL5, TP, GLB, LDLC, neutrophil, PLT, NLR, and PLR compared to the non-NPC group (P < 0.05). Conversely, levels of ALB, TBIL, DBIL, TBA, lymphocytes, and LMR were significantly lower in the NPC group (P < 0.05) compared to controls.

Comparison of clinical characteristics of patients with different stages of NPC

As shown in Table 2, the majority of patients with NPC (87.8%) were diagnosed with advanced stage III/IV disease. Notably, levels of S100A9, ALB, GLB, WBC, neutrophil, and NLR varied significantly among different NPC stages (P < 0.05). The levels of S100A9, GLB, WBC, neutrophils, and NLR were significantly higher in patients with more advanced stages than in those with earlier stages; in contrast, ALB levels were significantly lower in patients with advanced stages than in those with earlier stages.

Table 2 Comparison of clinical characteristics across different clinical stages of NPC.
Full size table

Prediction factors for NPC analyzed by logistic regression

In this study, univariate logistic regression was used to analyze 24 potential risk factors to determine their association with the occurrence of NPC (Table 3). Eleven predictive factors were identified as being significantly associated with the incidence of NPC: S100A9, CCL5, TP, ALB, GLB, DBIL, LDLC, neutrophils, PLT, NLR, and PLR. Subsequently, multivariate logistic regression analysis was performed on these 11 factors, revealing that S100A9, CCL5, TP, and ALB (P < 0.05) were independent predictive factors for the incidence of NPC (Table 3).

Table 3 Logistic regression analyses for identifying prediction factors for NPC.
Full size table

LASSO regression for prediction factors

This study collected 24 features of patients newly diagnosed with NPC. After excluding irrelevant and redundant features, 15 features (S100A9, CCL5, TP, ALB, GLB, TBIL, DBIL, TBA, LDLC, neutrophils, lymphocytes, PLT, PLR, NLR, and LMP) were retained for LASSO regression analysis. LASSO regression was then applied to identify factors associated with NPC. The results indicated that S100A9, CCL5, ALB, GLB, TBA, LDLC, NLR, and PLR were the most relevant factors linked to NPC (Fig. 2).

Fig. 2
figure 2

Predictor selection using LASSO regression analysis. (A) Bias selection of the tuning parameter (λ) in LASSO regression based on the minimum standard (left dashed line) and 1-SE standard (right dashed line). (B) Tenfold cross-validation for tuning parameter selection in LASSO regression, with a joint plot based on log-likelihood. The 1-SE standard (right dashed line) was used to select predictive factors for this study. SE: standard error.

Full size image

Diagnostic value of S100A9, CCL5, and other blood biomarker for NPC

The diagnostic values of S100A9, CCL5, and other blood markers for NPC were assessed using univariate ROC analysis. The results showed that S100A9 had the highest diagnostic efficacy, with an AUC value of 0.803 (Fig. 3A; Table 4). Based on the ROC curve coordinates, the cutoff value for S100A9 concentration was determined to be 40.396 ng/mL, with a sensitivity of 77.2% and specificity of 73.8%. Other biomarkers, including GLB (AUC = 0.789), NLR (AUC = 0.754), PLR (AUC = 0.750), and CCL5 (AUC = 0.731) also showed good diagnostic performance. In contrast, TBA (AUC = 0.606), ALB (AUC = 0.597), and LDLC (AUC = 0.594) demonstrated poor diagnostic efficacy.

Fig. 3
figure 3

Detection efficiencies analyzed using receiver operating characteristic (ROC) curves. (A) ROC curves depicting the predictive performance of individual biomarkers for NPC. (B) ROC curves comparing predictions from four models: XGBoost, Logistic Regression, LightGBM and Random Forest.

Full size image
Table 4 Diagnostic efficiency of S100A9, CCL5, and other blood biomarker for NPC.
Full size table

Comparison of machine learning algorithms and identification of the optimal predictive model

To improve the diagnostic value of serum S100A9 levels in distinguishing between patients with NPC and non-NPC, machine learning models were constructed that combined S100A9 levels with other variables to evaluate their predictive reliability. Four machine learning algorithms were utilized: logistic regression, XGBoost, LightGBM, and random forest classifiers. The optimal predictive model was identified based on its performance in the internal validation set. In the validation cohort, all four models achieved an AUC exceeding 0.85. The logistic regression model exhibited the best performance, with an AUC of 0.868; in contrast, the AUC of XGBoost, Random Forest, and LightGBM were 0.860, 0.856, and 0.854, respectively (Fig. 3B; Table 5).

Table 5 Diagnostic efficacy of four classifiers in the training and testing cohorts.
Full size table

The Logistic Regression model demonstrated the highest predictive performance among the tested algorithms. Using feature ranking within the logistic regression model, the top five features—GLB, ALB, CCL5, S100A9, and PLR—were selected for final model construction. As shown in Table 6; Fig. 4, the model exhibited strong discriminatory ability in distinguishing NPC from non-NPC. In the training and validation cohorts, the AUC of the model exceeded 80% (Fig. 4A and B). Additionally, the calibration curve revealed a strong correlation between the predicted and actual outcomes, indicating excellent calibration. The DCA curves (Fig. 4C and D) further demonstrated that the model provided substantial clinical benefit.

Table 6 Diagnostic efficiency of the logistic regression model in the training and validation cohorts for fivefold cross-validation.
Full size table
Fig. 4
figure 4

Performance of the prediction model. (A) ROC curve for the training cohort. (B) ROC curve for the validation cohort. (C) Calibration curve analysis demonstrating model calibration. (D) Decision curve analysis demonstrating clinical utility.

Full size image

As illustrated in Fig. 5A and B, the SHAP algorithm was used to interpret feature importance in the logistic regression model. GLB, PLR, CCL5, ALB, and S100A9 were identified as the most influential features for predicting NPC outcomes. A larger mean absolute Shapley value indicated a greater impact on the model’s predictions.

Fig. 5
figure 5

SHAP-based interpretation of the logistic regression model. (A) SHAP values illustrate the influence of each feature on the outcome. Each point represents one sample. Higher feature values are shown in red; in contrast, lower feature values are shown in blue. (B) Ranking of feature importance based on SHAP analysis. (C) SHAP force plot for patients with NPC, visualizing the contribution of individual features to predictions.

Full size image

Figure 5C provides individual force plots for NPC and non-NPC predictions. For patients with NPC, the probability prediction value was 0.84. Positive contribution features, shown in red, pushed the prediction score higher; in contrast, negative contribution features, shown in blue, reduced the score. The length of each arrow represents the magnitude of a feature’s impact on the prediction, with longer arrows indicating greater influence.

A web-based tool was developed using the Deepwise and Beckman Coulter DxAI platform v2.0. This tool allows users to input parameter values to predict the risk of NPC. Patients are categorized as NPC and non-NPC based on calculated probabilities. http://www.xsmartanalysis.com/model/list/predict/model/html?mid=19297&symbol=6anUc17318ru57DY5035.

External validation of the logistic regression model

An external validation cohort comprising 38 patients (19 with an initial diagnosis of NPC and 19 without NPC) was recruited from Xiangya Hospital between June and July 2024. As shown in Fig. 6, the AUC of the model for the external validation cohort was 0.817, confirming its predictive reliability.

Fig. 6
figure 6

Performance of external validation for the prediction model. (A) ROC curve for the external validation cohort. (B) Calibration curve analysis for the external validation cohort. (C) Decision curve analysis for the external validation cohort.

Full size image

Discussion

The use of readily available blood tests to screen high-risk individuals offers a potential alternative to relatively hard-to-obtain or expensive imaging modalities and invasive nasopharyngeal endoscopy. After evaluating clinical symptoms, if a patient exhibits signs suggestive of NPC, clinicians can further use predictive models based on blood markers to assess disease risk to guide the next steps. This approach has the potential to reduce unnecessary patient expenditure on invasive tests. In this study, S100A9, CCL5, and other common hematological markers were combined to evaluate their predictive accuracy for NPC. Furthermore, LASSO regression was employed to reduce dimensionality and avoid model overfitting, making it a widely used technique for variable selection and prediction model optimization23.

Our findings revealed that serum levels of S100A9 and CCL5 were significantly elevated in patients with NPC. The elevated serum S100A9 levels were consistent with previous findings9, indicating high expression of S100A9 in NPC tissues. In this study, S100A9 demonstrated the best diagnostic efficacy for NPC, with an AUC of 0.803, sensitivity of 77.2%, and specificity of 73.8%. Additionally, S100A9 levels were significantly higher in patients with advanced-stage disease than in those with early-stage disease. S100A9 is a potent proinflammatory mediator involved in acute and chronic inflammation and plays a pivotal role in inflammation-associated cancers7. Elevated S100A9 levels have been reported in various cancers24,25. As an inflammatory factor, S100A9 has strong chemotactic effects on leukocyte aggregation, adhesion, and migration, amplifying local proinflammatory microenvironments26. Studies27,28 have shown that S100A9 promotes tumor growth and progression by binding to cell surface receptors, increasing NF-κB-dependent transcriptional activity, and modulating inflammatory and immune responses. It activates cytokines such as TNFα, IL-6, IL-8, and IL-1β through the NF-κB pathway29 and creates a feedback loop by stimulating the synthesis of proinflammatory cytokines and chemokines30. Guo et al.31 demonstrated that the PI3K/Akt pathway is crucial for NPC progression, and a recent study32 suggested that S100A9 may influence NPC proliferation via this pathway.

CCL5, a member of the CC chemokine subfamily, binds to CCR1, CCR3, CCR4, and CCR5, with the highest affinity for CCR533. Recent research has explored the role of the CCL5/CCR5 axis in various cancers, including gastric, pancreatic, breast, and head and neck cancers33. CCL5/CCR5 activates pathways such as PI3K/AKT, NF-κB, and ERK/MEK, driving abnormal tumor cell proliferation. Additionally, these pathways regulate MMP and inflammatory factors, facilitating tumor metastasis and invasion34.

Among blood abnormalities observed in patients with NPC, increased levels of TP, GLB, LDLC, neutrophils, PLT, NLR, and PLR were notable. Conversely, ALB, TBIL, DBIL, TBA, lymphocytes, and LMP levels were significantly lower compared to controls (P < 0.05). Chronic inflammation is closely linked to cancer development. Peripheral blood markers such as neutrophil, lymphocyte, monocyte, and platelet counts, as well as derived indices like NLR and PLR, are cost-effective and noninvasive inflammatory markers widely used in cancer research. Elevated serum GLB levels, induced by proinflammatory cytokines such as IL-6, IL-1β, and TNFα, reflect systemic inflammation35. Previous studies36,37 have linked high GLB levels not only to inflammatory responses but also to tumor initiation, recurrence, and outcomes. ALB, indicative of nutritional status, has been associated with tumor progression and malignancy38.

Univariate and multivariate logistic regression analyses identified S100A9, CCL5, TP, and ALB as independent factors influencing NPC risk.

These markers, along with GLB and PLR, were used to develop a predictive logistic regression model, which demonstrated strong performance and clinical applicability. The model exhibited excellent discrimination (AUC = 0.877) and calibration in the internal validation cohort, with similar results (AUC = 0.817) in the external validation cohorts. These findings align with prior studies, such as that conducted by Wu et al.39, however, our study uniquely incorporated routine blood parameters.

Clinically, we developed a predictive model for NPC using S100A9, CCL5, ALB, GLB, and PLR levels. This model may enhance early diagnosis of NPC and provides a foundation for future research. However, this study has some limitations. First, this study was conducted in a single center, with both the training cohort and the external validation cohort being data sources from the same hospital, and thus may have suffered from selection bias. Second, the study lacked prospective data, had a small external validation sample size. While using the existing methodology, we also used 10-fold cross-validation and Bootstrap simulation to similarly validate the reliable diagnostic accuracy and clinical utility of the predictive model40,41 (As shown in Supplementary material). Nevertheless, a larger, well-powered study might be necessary to validate the findings. Future studies will be aim to increase the sample size and conduct multicenter studies in collaboration with multiple hospitals for larger and more diverse patient populations. Third, our analysis of blood markers such as S100A9 and CCL5 was limited to expression levels; further investigations are needed to identify the signaling pathways influenced by these molecules and to elucidate their roles in NPC pathogenesis and progression. In the future, we will thoroughly explore whether treatment or coexisting health issues affect levels of S100A9, CCL5, or the other biomarkers, and take a closer look at potential confounding factors.

Conclusion

Our findings demonstrated that serum levels of S100A9, CCL5, and other blood indicators, including GLB, ALB, and PLR, hold diagnostic value for NPC. The logistic regression model based on these biomarkers showed strong predictive performance and clinical utility. However, large clinical trials are needed to confirm these results.