Head and neck squamous cell carcinoma prognosis prediction method based on immunohistochemical marker

Abstract

The invention relates to the technical field of crossing of medical artificial intelligence and precision medical treatment, and discloses a head and neck squamous cell carcinoma prognosis prediction method based on an immunohistochemical marker, which comprises the following steps: acquiring clinical variable data and immunohistochemical marker data of a patient, and preprocessing the clinical variable data and the immunohistochemical marker data; screening prognosis prediction factors by adopting regularization regression analysis, constructing a column graph model based on a screening result, and generating a visual prognosis prediction tool; the performance of the visual tool is verified through multiple verification indexes, and the prediction accuracy is guaranteed; on the basis of the verified tool, an online prognosis prediction calculator is developed, and individualized risk stratification and prognosis prediction of the patient are achieved. By integrating multi-dimensional clinical and molecular pathology data and optimizing the model construction and verification process, the accuracy of head and neck squamous cell carcinoma prognosis prediction is improved, and the developed online prognosis tool is convenient to operate, can directly serve clinical diagnosis and treatment decisions, and has high clinical practical value.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of medical artificial intelligence and precision medicine, and particularly relates to a head and neck squamous cell carcinoma prognosis prediction method fusing clinical pathological features and immunohistochemical markers. BACKGROUND

[0002] Head and neck squamous cell carcinoma (hereinafter referred to as "HNSCC") is a malignant tumor with high incidence and poor prognosis worldwide. According to existing research data, the number of new cases per year exceeds 830,000, ranking sixth in the world in terms of high incidence of malignant tumors. More than 60% of patients are in the local advanced or metastatic stage at the time of diagnosis. The recurrence and metastasis rate of locally advanced patients after comprehensive treatment is still as high as 60%, and the median survival time of metastatic patients is less than 1 year. Even after standard treatment, the 5-year survival rate of locally advanced patients is still less than 50%. There is an urgent need for a precise prognosis prediction tool to assist individualized diagnosis and treatment decisions.

[0003] In current clinical practice, the prognosis evaluation and treatment decision of HNSCC mainly rely on the traditional TNM staging system, which is based on tumor size (T), lymph node metastasis status (N) and distant metastasis (M) for anatomical classification, and is the core basis for existing clinical management. At the same time, other related prognosis evaluation schemes have also been developed in the field, including: 1. Single modality prediction model, such as radiomics model relying on image features (for example, public patent CN118675731A), and prediction model based on Ki67 single marker. Although Ki67 is related to the recurrence risk of HNSCC, the prediction performance of single indicator model is limited (AUC<0.7); 2. Prediction model based on public database, such as ATHENA model. Such models rely on public data for construction, but generally lack key clinical variables such as radiotherapy history and smoking status, limiting their applicability in actual populations; 3. Precision prediction model for other solid tumors, such as lung cancer gene mutation deep learning model based on histological images and molecular residual disease prediction model for colorectal cancer based on ctDNA methylation detection. Although the prediction accuracy has been improved by integrating molecular markers and clinical factors, there is still no mature and generalizable scheme established in the field of HNSCC. In addition, some biomarker-related research has also been carried out in the field of HNSCC. For example, it has been reported that the expression level of vimentin is related to lymph node metastasis. However, such research generally has limited sample size and insufficient evidence level, and fails to systematically integrate clinical factors and molecular indicators, making it difficult to be converted into a stable and applicable clinical tool.

[0004] However, the existing technical solutions have significant defects and cannot meet the clinical demand for precise prognosis prediction of HNSCC. The specific deficiencies are as follows: Inherent limitations of traditional TNM staging system: only relying on anatomical information, failing to reflect the molecular biological characteristics of the tumor (such as the epithelial-mesenchymal transition state, HPV infection-related molecular markers), leading to significant differences in prognosis of patients with the same stage. At the same time, this system is a static model, which fails to consider the non-linear relationship between clinical variables such as age, lifestyle (such as smoking) and prognosis (for example, younger patients have more aggressive tumors, and older patients have poorer treatment tolerance), thereby affecting the accuracy of risk assessment.

[0005] Single modality prediction model coverage is not complete: only relying on imaging or a single biological indicator, lacking comprehensive integration of clinical, pathological and molecular level data, it is difficult to fully depict the high heterogeneity of HNSCC, and the prediction ability is limited.

[0006] Insufficient data integration and lack of clinical translation: existing models based on public databases generally lack core clinical variables and have insufficient data integrity; at the same time, most studies only stay at the academic verification level, lack of convenient real-time prediction tools, and difficult to directly support individualized clinical treatment decisions.

[0007] In summary, the existing HNSCC prognosis evaluation technology has the problems of anatomical information limitation, single prediction mode, insufficient data dimension and lack of clinical translation, which makes it difficult to realize individualized risk stratification and accurate prognosis prediction. Therefore, it is urgent to develop a HNSCC prognosis prediction method that can integrate clinical pathological factors and immunohistochemical molecular markers, and consider the prediction accuracy and clinical practicability. SUMMARY

[0008] The purpose of the present application is to provide a head and neck squamous cell carcinoma prognosis prediction method based on immunohistochemical markers to solve the problems of TNM staging limitations, limited prediction efficiency of single modality model and lack of clinical translation tools in the background art.

[0009] To achieve the above-mentioned purpose, the following technical solutions are disclosed: a head and neck squamous cell carcinoma prognosis prediction method based on immunohistochemical markers, the method comprising: Step 1: obtaining the clinical variable data and immunohistochemical marker data of the head and neck squamous cell carcinoma patient, and pre-processing the data to ensure the integrity and consistency of the data; Step 2: screening the prognosis prediction factors by LASSO regression combined with multivariate Cox proportional hazards analysis, constructing nomogram models of overall survival and recurrence-free survival based on the screened prediction factors, and generating a visual prognosis prediction tool; Step 3: verifying the performance of the visual prognosis prediction tool through multiple verification indicators to ensure the prediction accuracy; Step 4: Based on the validated visual prognostic prediction tool, an online prognostic prediction calculator is developed to realize individualized risk stratification and prognosis prediction of patients.

[0010] Preferably, the clinical variable data at least includes age, gender, smoking history, tumor staging information and radiotherapy history of the patient; and the immunohistochemical markers at least include proliferation markers, epithelial-mesenchymal transition markers and virus infection related markers related to the prognosis of head and neck squamous cell carcinoma.

[0011] Preferably, the data preprocessing at least includes: supplementing missing data, identifying and screening abnormal data, and standardizing variable data; wherein a non-linear conversion method is used for age data to fit the risk relationship between age and overall survival and recurrence-free survival, wherein the non-linear conversion method is realized by a restricted cubic spline function, and the conversion formula is: , wherein, is a spline node, represents a positive function, is the number of spline nodes, is the age data, is a constant term, , , are the coefficients of the first, second and third order terms of age, respectively, is the kth node coefficient.

[0012] Preferably, the LASSO regression analysis selects the optimal regularization parameter by multi-fold cross-validation to screen out prognostic predictors that have a significant impact on overall survival and recurrence-free survival; The nomogram model is a multivariate Cox proportional hazards model.

[0013] Preferably, the risk function of the multivariate Cox proportional hazards model is represented as: , wherein, is the risk rate at time t, is the baseline risk rate, is the screened prognostic predictor, is the regression coefficient of each prognostic predictor.

[0014] Preferably, the visual prognostic prediction tool is a nomogram, which is generated based on the multivariate Cox proportional hazards model and by the formula The individualized risk score is calculated, and the total survival probability and recurrence-free survival probability of the patient at different time nodes are predicted according to the risk score, and the survival probability calculation formula is: , wherein, is a cumulative baseline risk function.

[0015] Preferably, the plurality of verification indicators include an indicator for evaluating model discrimination and an indicator for evaluating model calibration; wherein the indicator for evaluating model discrimination includes a consistency index and an area under a time-dependent ROC curve; and the indicator for evaluating model calibration includes a calibration curve, a Grønnesby-Borgan test and a Brier score.

[0016] Preferably, the calculation formula of the consistency index is: , wherein, and are survival times of patients i and j, respectively, and are linear prediction values of patients i and j, respectively, is an indicator function; The calculation formula of the Brier score is: , wherein, is the sample size, is the outcome variable, is the survival probability at time t predicted based on the characteristics of patient i.

[0017] Preferably, the online prognosis calculator is realized through a webpage platform; the webpage platform is provided with an interactive interface for receiving input of clinical pathological parameters and immunohistochemical indicators, and a display interface for outputting prognosis prediction results.

[0018] Preferably, the display interface output content of the online prognosis calculator includes the total survival probability and recurrence-free survival probability of the patient at 1 year, 3 years and 5 years time nodes, and the risk stratification result determined based on the individualized risk score; the risk stratification result is determined by a preset risk grade division rule, and at least includes three grades of high risk, medium risk and low risk.

[0019] Beneficial effects: The head and neck squamous cell carcinoma prognosis prediction method based on immunohistochemical markers of the present application realizes the systematic integration of clinical pathological characteristics and immunohistochemical molecular markers in the field of HNSCC for the first time. By obtaining the clinical variable data and immunohistochemical marker data of the head and neck squamous cell carcinoma patient, combining the nonlinear conversion of age in data preprocessing (using a restricted cubic spline function) to capture the complex nonlinear relationship between age and prognosis, and then accurately screening the prognosis prediction factors through LASSO regression combined with multivariate Cox proportional hazards analysis, a multivariate Cox proportional hazards model is constructed based on the screening results and converted into a visual nomogram tool, which effectively overcomes the defects of traditional TNM staging which only relies on anatomical characteristics, ignores molecular biological heterogeneity and nonlinear effects of clinical dynamic factors, and solves the problems of incomplete coverage and insufficient prediction performance (AUC less than 0.7) of existing single modal models.

[0020] Secondly, the head and neck squamous cell carcinoma prognosis prediction method based on immunohistochemical markers of the present application, through the consistency index (C-index), time-dependent ROC curve, calibration curve, Grønnesby-Borgan test and Brier score, etc. Multidimensional indicators are used to systematically verify the model performance, to ensure that the constructed tool has excellent discrimination, good calibration and high robustness, and significantly improves the accuracy of overall survival (OS) and recurrence-free survival (RFS) prediction of HNSCC patients In addition, based on the verified model, the present application further develops an online prognosis calculator, which can be realized through a web platform, can conveniently receive the input of clinical pathological parameters and immunohistochemical indicators, and automatically output individualized 1-year, 3-year and 5-year survival probability and risk stratification results. The risk stratification rules include three levels of high risk, medium risk and low risk. This tool realizes the effective transformation from academic research to clinical application, solves the problems of insufficient data dimension integration and lack of real-time calculation tools that can be directly applied in existing prediction methods, and provides scientific, intuitive and operable technical support for clinical development of individualized risk stratification, formulation of precise treatment plan and optimization of follow-up management of HNSCC patients. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, brief descriptions will be given below to the drawings needed to be used in the embodiments or prior art descriptions. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0022] Figure 1 The flowchart of the head and neck squamous cell carcinoma prognosis prediction method based on immunohistochemical markers provided by the embodiments of the present application; Figure 2 The schematic diagram of the LASSO regression provided by the embodiment of the present application in the prognosis factor screening result in the overall survival and recurrence-free survival model; Figure 3 The schematic diagram of the verification index of the HNSCC overall survival (OS) and recurrence-free survival (RFS) prognosis model provided by the embodiment of the present application; Figure 4 The schematic diagram of the calibration verification result of the OS prognosis model in the HNSCC provided by the embodiment of the present application; Figure 5 The schematic diagram of the calibration verification result of the RFS prognosis model in the HNSCC provided by the embodiment of the present application; Figure 6 The nomogram of the HNSCC overall survival (OS) and recurrence-free survival (RFS) prediction provided by the embodiment of the present application; Figure 7 The schematic diagram of the online prognosis prediction tool of the HNSCC overall survival (OS) and recurrence-free survival (RFS) provided by the embodiment of the present application. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all the other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0024] In this document, the term "comprising" is intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes the elements inherent to such process, method, article or equipment. Without more limitations, the elements defined by the statement "comprising" do not exclude the presence of other identical elements in the process, method, article or equipment including the elements.

[0025] The embodiment aims to solve the technical problems of the existing head and neck squamous cell carcinoma prognosis evaluation method (such as the traditional TNM staging) which only relies on anatomical features, ignores tumor biological characteristics and clinical key variables, has limited prediction performance and lacks a convenient prognosis tool that can be directly applied in clinical practice, and provides a head and neck squamous cell carcinoma prognosis prediction method based on immunohistochemical markers as shown in Figure 1 The embodiment aims to solve the technical problems of the existing head and neck squamous cell carcinoma prognosis evaluation method (such as the traditional TNM staging) which only relies on anatomical features, ignores tumor biological characteristics and clinical key variables, has limited prediction performance and lacks a convenient prognosis tool that can be directly applied in clinical practice, and provides a head and neck squamous cell carcinoma prognosis prediction method based on immunohistochemical markers as shown in

[0026] Step 1: Obtain the clinical variable data and immunohistochemical marker data of patients with head and neck squamous cell carcinoma, and preprocess the data to ensure the integrity and consistency of the data.

[0027] The data comes from 169 patients with pathologically diagnosed head and neck squamous cell carcinoma who have passed the ethical review (batch number: First Affiliated Hospital of Sun Yat-sen University

[2017] 173). The baseline characteristics of the samples are clear: the median age is 59 years (interquartile range 49-68 years), the proportion of males is 65.09%, the primary site of the tumor is mainly the tongue (62.7%), the clinical stage (using AJCC 8th edition standard) distribution is 24.26% for stage I, 28.40% for stage II, 18.34% for stage III, and 28.99% for stage IV, the median follow-up time is 36.5 months (interquartile range 24.0-45.0 months), and the follow-up outcome includes survival status (alive / dead) and recurrence / metastasis time (from diagnosis or treatment end to first recurrence / metastasis, unit: months), ensuring the clinical representativeness of the samples.

[0028] In this step, the clinical variable data at least includes the patient's age, gender, smoking history, tumor stage information and radiotherapy history. In addition, the clinical variable data can further include drinking history, chewing betel nut history, lymph node metastasis status (whether there is lymph node metastasis), tumor infiltration depth, chemotherapy history (whether receiving chemotherapy), surgical method (such as radical surgery, palliative surgery), radiotherapy dose (such as total radiotherapy dose 60-70Gy), chemotherapy drugs and cycles (such as cisplatin combined with 5-fluorouracil, 1 cycle every 3 weeks, a total of 4-6 cycles), covering key clinical factors affecting the prognosis of head and neck squamous cell carcinoma, avoiding prediction bias caused by missing data dimensions.

[0029] In this step, the immunohistochemical markers at least include the proliferation marker c-myc, the epithelial-mesenchymal transition marker vimentin and the virus infection related marker p16 associated with the prognosis of head and neck squamous cell carcinoma. In addition, in addition to "c-myc (oncogenic proliferation marker), vimentin (epithelial-mesenchymal transition marker), p16 (virus infection related marker)", p53 (tumor suppressor gene, associated with tumor cell apoptosis abnormalities), Ki67 (cell proliferation index, associated with tumor proliferation activity) can also be supplemented to form a multi-dimensional molecular marker system; The expression level is obtained by a standardized immunohistochemical staining process, and the specific steps are as follows: ① Paraffin section deparaffinization: soak in xylene for 3 times, 10 minutes each time, and completely remove the paraffin; ② Gradient ethanol hydration: soak in 100%, 95%, 85%, 75% ethanol for 5 minutes each time, and gradually restore the water content of the tissue; ③ Antigen retrieval: use 0.01M citrate buffer (pH 6.0), 121℃ high pressure heating for 20 minutes, to expose the antigen epitope; ④ Blocking treatment: first incubate in 3% hydrogen peroxide solution at room temperature for 10 minutes to block endogenous peroxidase activity, then incubate in 5% goat serum containing 0.3% Triton X-100 at 37℃ for 1 hour to reduce non-specific binding; ⑤ Antibody incubation: primary antibody (such as c-myc antibody, vimentin antibody) incubate overnight at 4℃, HRP (horseradish peroxidase) labeled secondary antibody incubate for 30 minutes at 37℃; ⑥ Color development and mounting: DAB (3,3'-diaminobenzidine) color development, hematoxylin re-staining of cell nuclei, 1% hydrochloric acid ethanol differentiation to adjust the staining depth, and neutral resin mounting for preservation. The expression level is quantified using the "scoring model", which is implied in claim 3 "staining intensity x percentage of positive cells": staining intensity is graded according to 0-3 points (0 points: no staining, 1 point: weak staining, 2 points: medium staining, 3 points: strong staining), and the percentage of positive cells is the proportion of positive staining tumor cells in the total tumor cells in the field of view (0-100%), with a final score range of 0-300 points; To ensure the accuracy of quantification, a double-blind scoring mechanism is used: two pathologists independently score, and when the difference is ≤10 points, the average value is taken as the final score, and when the difference is >10 points, a third senior pathologist is required to review and determine, to eliminate subjective scoring errors.

[0030] In this step, data preprocessing at least includes: supplementing missing data, identifying and screening abnormal data, and standardizing variable data; wherein the nonlinear conversion method is used for age data to fit the risk relationship between age and overall survival and recurrence-free survival. Specifically, it includes the following aspects: Firstly, for clinical variable data (e.g. missing radiotherapy dose records for some patients), we used "Multiple Imputation (MI)" to supplement the missing values. For immunohistochemical data (e.g. missing scores due to poor staining of individual specimens), we used "re-slicing experiments on archived paraffin-embedded tissue blocks" to supplement the missing values. Based on the complete data, we constructed an imputation model (e.g. using complete variables such as age, tumor stage, and smoking history to predict missing radiotherapy doses), generated five complete data sets, and fitted the model separately. Finally, we combined the results to avoid bias caused by single imputation. Compared with the traditional mean imputation method, multiple imputation can better preserve the original data distribution characteristics and improve data integrity and representativeness.

[0031] Secondly, for abnormal data identification and screening, we used the 3σ principle to determine abnormal data: calculate the mean (μ) and standard deviation (σ) of each variable, and define data outside the range of "μ±3σ" as abnormal data (e.g. age recorded as 150 years old, immunohistochemical score 400 points). For identified outliers, first check the original medical record or test report, and if it is an input error, correct it (e.g. "150 years old" is actually "50 years old"), if it is a real extreme value (e.g. individual patient immunohistochemical score significantly deviates from the regular distribution), it is excluded. This process ensures data authenticity and avoids interference with model estimation by outliers.

[0032] Thirdly, for variable data standardization, considering the dimensional differences of different variables (e.g. age "years", immunohistochemical score "0-300 points", radiotherapy dose "Gy"), we used Z-score standardization processing, the formula is: Through this method, all variables are converted to a unified scale with a mean of 0 and a standard deviation of 1, avoiding bias in the model due to dimensional differences, and ensuring fair weighting of different predictors in the model.

[0033] In addition, for the key variable of age, we used a restricted cubic spline function (RCS) to achieve nonlinear transformation to fit the "U-shaped" risk relationship between age and overall survival and recurrence-free survival (patients ≤35 years old and >65 years old have significantly higher risk of prognosis). Specifically, we set 4 spline nodes, the node position was determined by the 5%, 35%, 65%, and 95% quantiles of the age data (i.e. calculated by the formula "quantile(x, prob)", where "x" is the patient age data (unit: years), "prob" is the parameter used to specify the quantile position to be calculated). is the patient age data (unit: years) to be processed;​​​ representing the 5%, 35%, 65%, 95% quantile of the age data, the corresponding values of the four quantiles are the node positions of the restricted cubic spline), the nonlinear transformation of the age data uses a restricted cubic spline function, and the transformation formula is: , Correspondingly, when the number of nodes k = 4, the transformation formula is: wherein, is the constant term, that is, the model intercept, , , are the coefficients of the age linear term, quadratic term, and cubic term (obtained by model fitting), respectively, is the kth node coefficient (controls the degree of curvature of the curve at the node), is the kth node value (determined by the 5%, 35%, 65%, 95% quantiles of the age data, such as the 42, 52, 65, and 72-year-old nodes of the 169 patients), is the positive part function (that is, when , take , otherwise take 0), and k is the number of spline nodes (k = 4 in this method). This transformation breaks through the limitations of traditional linear models and can accurately capture the high-risk interval of young and old patients, significantly improving the explanatory power and accuracy of model prediction. The restricted cubic spline function is implemented in R language by the "rcs(age, 4)" function, and the "U-shaped" curve of age and prognosis is fitted, which clearly shows that ≤35 years old (hazard ratio HR = 1.82) and >65 years old (HR = 1.67) are high-risk age intervals, verifying the defect of traditional linear assumption that underestimates the risk of extreme age.

[0034] The innovation of this step is that by checking the ethics compliance and sample representativeness, integrating clinical variables and immunohistochemical multi-indexes, systematically processing missing and abnormal data, combining variable standardization and nonlinear modeling, the integrity, authenticity, standardization and precision of input data are significantly improved. Compared with existing single modal data (such as imageomics or single marker model) dependent, this multi-dimensional data preprocessing scheme more comprehensively describes the anatomical and molecular biological characteristics of HNSCC, providing high-quality input for subsequent factor screening and model construction. Second, multiple imputation of missing data and 3σ screening of abnormal data solve the common problem of incomplete and noisy clinical data, reducing the interference of data quality on model performance; variable standardization eliminates the dimension difference, avoids the bias of the model to high-value variables, and ensures that each predictor participates in model fitting fairly; nonlinear conversion of age accurately identifies the U-shaped risk interval, breaking through the limitations of traditional linear assumptions, enabling the model to more truly reflect the association between age and prognosis, and the preprocessing operation optimizes input data from four dimensions of data integrity, authenticity, standardization and precision, which forms a synergy with data acquisition, lays a solid foundation for subsequent regularization regression screening (step 2), model validation (step 3), application development (step 4), and ultimately improves the accuracy and robustness of the overall prognosis prediction method. Among them, the U-shaped risk refers to the nonlinear association between a factor (such as age) and the risk of disease prognosis (such as the risk of death or recurrence of head and neck squamous cell carcinoma patients): the value of the factor in the two end intervals (such as age ≤ 35 years and > 65 years), the prognosis risk of patients is significantly increased; while in the middle interval (such as age 36-65 years), the risk is relatively low, and its risk trend shows a similar "U" curve distribution in the coordinate system. This relationship shows that the effect of the factor on prognosis is not simply linearly increasing or decreasing, but there are two risk peaks, which need to be accurately captured by a nonlinear model (such as a restricted cubic spline).

[0035] Step 2: Screen the prognostic predictors by LASSO regression combined with multivariate Cox proportional hazards analysis, construct nomogram model based on the screened predictors, and generate visual prognostic prediction tool for predicting overall survival (OS) and recurrence-free survival (RFS) of patients.

[0036] In this step, LASSO regression combined with multivariate Cox proportional hazards analysis is used for factor screening and modeling. Specifically, LASSO regression analysis selects the optimal regularization parameter through multiple-fold cross-validation to screen out prognostic predictors that have a significant impact on OS and RFS; then a nomogram model is constructed based on the screening results, and a visual prognostic prediction tool is generated. The nomogram model of the present embodiment is a multivariate Cox proportional hazards model. Secondly, the key factors that have a significant impact on prognosis are variables that have statistically significant effects on both OS (time from diagnosis / treatment end to death) and RFS (time from diagnosis / treatment end to first tumor recurrence / metastasis) and have clear clinical significance. LASSO regression uses a regularization penalty mechanism to compress the coefficients of redundant variables to 0, and finally only retains factors that can substantially change the survival risk of patients (such as nonlinearly converted age variables, high expression of vimentin indicating enhanced tumor invasiveness, history of radiotherapy associated with reduced risk of death, etc.).

[0037] In terms of implementation details, in the LASSO regression analysis, all patient data is used for model construction; LASSO regression analysis is implemented through the glmnet package of R language, and 10-fold cross-validation is used to select the optimal regularization parameter λ to obtain stable variable screening results. In the modeling process of OS and RFS, LASSO regression selects clinical factors and immunohistochemical indicators with predictive value, and further incorporates them into the multivariate Cox proportional hazards model. Finally, the constructed model is used for subsequent nomogram development and verification.

[0038] The objective function of LASSO regression is: , wherein, is the regression coefficient vector of each prediction variable, is the sum of squares of residuals, is the L1 norm of the regression coefficient (i.e., the sum of the absolute values of the coefficients), and N is the sample size (N=118 in the present method), is the N-dimensional outcome variable vector (total survival / relapse-free survival event indicator variable, 1=event occurred, 0=censored), is the Nxp-dimensional prediction variable matrix (each row represents 1 patient, and each column represents 1 prediction variable). is the regularization parameter (controls the strength of the penalty, the larger the penalty, the stronger the penalty, and more coefficients are compressed to 0), that is, by penalizing the regression coefficients, the coefficients of unimportant variables are compressed to 0, and finally the variables with non-zero regression coefficients are retained as prognostic predictors.

[0039] In addition, the screened prognostic predictors include: OS included prognostic predictors: age (transformed by restricted cubic spline), TNM-N stage, radiation therapy, c-myc, vimentin, p16; RFS included prognostic predictors: age (transformed by restricted cubic spline), smoking history, clinical stage, radiation therapy, vimentin, p16. Among them, radiation therapy is a protective factor (regression coefficient is negative), vimentin is a risk factor (regression coefficient is positive), which is consistent with clinical cognition.

[0040] As shown in Figure 2 , in the figure, part A is the ten-fold cross-validation plot of OS, which shows the optimal lambda value selected by λ_min (minimum error) and λ_1se (1 standard error criterion). Part B is the LASSO coefficient trajectory plot of OS, λ_1se (λ=0.018) selects 6 non-zero predictors: age (nonlinear cubic spline), TNM-N stage, radiation therapy, c-myc, vimentin and p16. Part C is the ten-fold cross-validation plot of RFS, which determines the optimal lambda value by λ_min and λ_1se to balance the model complexity. Part D is the LASSO coefficient trajectory plot of RFS, which retains 6 non-zero predictors: age (nonlinear cubic spline), smoking, clinical stage, radiation therapy, vimentin and p16.

[0041] Further, the risk function of the multivariate Cox proportional hazards model is represented as: , where, is the hazard rate at time t; is the baseline hazard rate, is the selected prognostic predictors, is the regression coefficient of prognostic predictors.

[0042] In this step, the Cox proportional hazards model is implemented by the R language rms package, and the model structure is as follows: overall survival model: , recurrence-free survival model: , where, is the survival time (unit: month); is the overall survival outcome variable (1 indicates death, 0 indicates survival or censoring); is the recurrence-free survival outcome variable (1 indicates recurrence or metastasis, 0 indicates no recurrence / metastasis or censoring); Age variable after transformation by 4-node restricted cubic spline; N stage in TNM stage (representing regional lymph node metastasis, divided into N0, N1, N2, N3, respectively assigned 0, 1, 2, 3); Radiotherapy history (1 indicates having received radiotherapy, 0 indicates not having received radiotherapy); Immunohistochemical score of c-myc protein (0-300 points); Immunohistochemical score of vimentin protein (0-300 points); Immunohistochemical score of p16 protein (0-300 points); Smoking history (1 indicates having a smoking history, 0 indicates no smoking history); Clinical stage (using AJCC 8th stage standard, divided into I, II, III, IV, respectively assigned 1, 2, 3, 4). After model construction, Schoenfeld residual test is required to verify the proportional risk assumption (P>0.05), and variables that do not meet the assumption (such as individual subgroups) are stratified to ensure that the model meets the core assumptions of Cox model.

[0043] The regression coefficients β of each predictor in the risk function are obtained by model fitting, and the positive and negative signs represent the direction of influence, and the absolute value represents the strength of influence: the specific risk function of overall survival model: , Where, is the instantaneous risk rate of the patient at time t given the covariate X; is the baseline risk rate; is the age value after transformation by restricted cubic spline; is assigned 1 when TNM-N stage is N3, otherwise 0; Immunohistochemical score of c-myc protein (here refers to the standardized value); Immunohistochemical score of vimentin protein (here refers to the standardized value); Radiotherapy history (1 indicates having received radiotherapy, 0 indicates not having received); Immunohistochemical score of p16 protein (here refers to the standardized value); 0.98, 1.32, 0.87, 0.92, -1.15, -0.95: the regression coefficients of each variable.

[0044] The specific risk function of recurrence-free survival model: , Where, 1 if the patient is in stage IV, otherwise 0; 0.85, 0.55, 1.21, 0.89, -0.75, -0.88: regression coefficients of each variable.

[0045] Application example: a 60-year-old patient (TNM-N0 stage, receiving radiotherapy, cmyc=100 points, vimentin=50 points, p16=30 points) is substituted into the overall survival model, and the 5-year survival probability is calculated to be about 75.2%, which is consistent with the clinical follow-up results.

[0046] In this step, LASSO regression uses 10-fold cross-validation and L1 norm penalty to solve the problem of high-dimensional variable collinearity while accurately screening key predictors and avoiding redundant variable interference; the Cox proportional hazards model and its risk function quantify the degree and direction of the influence of each factor on survival risk, and convert risk from an abstract concept to a calculable mathematical index; the combination of the two ensures the simplicity of the model (only important factors are retained) and the scientificity of the prediction (based on the gold standard model of survival analysis), which improves the prediction accuracy compared with traditional empirical staging.

[0047] Step 3: Verify the performance of the visualized prognosis prediction tool through various verification indicators to ensure prediction accuracy.

[0048] In this step, a multi-dimensional verification index system is used to systematically evaluate the performance of the visualized prognosis prediction tool to ensure the accuracy and stability of the prediction. The verification indicators include indicators for evaluating the discrimination of the model and indicators for evaluating the calibration of the model. Among them, the indicators for evaluating discrimination include consistency index (C-index) and area under the curve (AUC) of time-dependent ROC curve (t-ROC); the indicators for evaluating the calibration of the model include calibration curve, Grønnesby-Borgan test and Brier score.

[0049] The calculation formula of the consistency index (C-index) is: , wherein, and are the survival times of patients i and j, respectively, and are the linear prediction values of patients i and j, respectively, is an indicator function defined as "1 if the condition is true, otherwise 0".

[0050] And the calculation formula of the Brier score at time t is: , wherein, is the sample size, Yi is the outcome variable (1 for event occurrence, 0 for censoring), Pi is the survival probability at time t predicted based on the characteristics of patient i.

[0051] The verification results are as follows: For the indicators for evaluating the discriminability of the model: Consistency index (C-index): the OS prognosis model C=0.79 (95% CI: 0.72-0.83), and the RFS prognosis model C=0.73 (0.68-0.78), both of which are significantly better than the traditional TNM staging (C=0.65-0.70); Time-dependent ROC curve: the OS model 1-year AUC=0.95, 3-year AUC=0.821, and 5-year AUC=0.823; the RFS model 1-year AUC=0.79, 3-year AUC=0.805, and 5-year AUC=0.813, indicating that the model has excellent discriminability for both short-term and long-term prognosis.

[0052] As shown in Figure 3 , in the figure, part A is the Brier score curve of the OS prognosis model, showing that it maintains stable prediction performance in short-term, medium-term, and long-term follow-up. Part B is the Brier score curve of the RFS prognosis model, indicating that the prediction error is low over time. Part C is the time-dependent ROC curve of overall survival, reflecting strong discriminability for survival outcome. Part D is the time-dependent ROC curve of recurrence-free survival, showing optimized long-term recurrence risk stratification ability.

[0053] For the indicators for evaluating the calibration degree of the model: Calibration curve: generated by 1000 times bootstrap sampling, the predicted 1 / 3 / 5-year survival probability is highly consistent with the actual observed value in the ±5% interval of the 45° reference line, such as the patient group with a predicted 5-year survival rate of 70%, the actual survival rate is 68.3%; Grønnesby-Borgan test: P>0.05 at all time points (such as P=0.32 for the 5-year OS model), indicating that there is no statistical difference between the predicted value and the actual observed value; Brier score: the smaller the value, the more accurate the prediction, the OS model 1-year =0.074, 3-year =0.140, and 5-year =0.141; the RFS model 1-year =0.113, 3-year =0.162, and 5-year =0.159, all ≤0.2, at a low error level (average error sum of squares accounted for <15%).

[0054] As shown in Figure 4As shown in the figure, Part A presents the calibration curves for predicting 1-year (A), 3-year (B), and 5-year (C) OS using 1000 self-sampling iterations, demonstrating a high degree of agreement between the predicted and observed probabilities (the curves are close to the ideal 45° reference line). Part B shows the Grønnesby-Borgan test results for the OS prognostic model calibration, confirming no significant deviation from the ideal agreement.

[0055] like Figure 5 As shown in the figure, Part A presents the calibration curves for predicting 1-year (A), 3-year (B), and 5-year (C) OS using 1000 self-sampling iterations, demonstrating a high degree of agreement between the predicted and observed probabilities (the curves are close to the ideal 45° reference line). Part B shows the Grønnesby-Borgan test results for the RFS prognostic model calibration, confirming no significant deviation from the ideal agreement.

[0056] In summary, the multi-dimensional validation index system in this step comprehensively demonstrated the reliability of the model: the high C-index and AUC value indicate that the model can effectively distinguish patients with different risks; the calibration curve and Grønnesby-Borgan test verified the consistency between the predicted values ​​and the actual outcomes; the low Brier score proves that the prediction error is small; compared with the existing single validation index method, this system ensures the model performance from three dimensions: discrimination ability, calibration accuracy, and error control, significantly enhancing its scientific rigor and robustness, and providing sufficient scientific basis for clinical application.

[0057] Step 4: Based on the validated model, develop a visual nomogram tool and an online prognostic calculator to achieve individualized risk stratification and prognostic prediction for patients.

[0058] In this step, a visual nomogram tool and an online prognostic calculator are developed based on the validated model to achieve individualized risk stratification and prognostic prediction for patients. The nomogram is visualized using a formula... Calculate the individualized risk score, where Score is the individualized risk score (range 0–300), βi is the regression coefficient of the i-th predictor, and Xi is the value of that factor. For example, the regression coefficient of vimentin in the overall survival model is 0.92. When a patient's vimentin score is 100, the corresponding risk score for that factor is 92. A final total score ≥150 indicates high risk, and <80 indicates low risk.

[0059] Based on the risk score, the overall survival (OS) and recurrence-free survival (RFS) probabilities of patients at different time points are predicted. The survival probability is calculated using the following formula: , in, This is the cumulative baseline risk function.

[0060] It is feasible that nomogram is generated by R language rms package and nomogram function, for example, survfit(cox_model, newdata, t=60) can directly output 5-year survival rate. The core conversion logic is: , Among them, is the fitted Cox proportional hazards model object, including model regression coefficients, baseline risk rate, variable information and other core parameters; is a logical function, that is, a function for converting linear prediction values into 0-1 interval survival probability, , for converting linear prediction values into survival probability labels between 0 and 1; is the model centralization parameter, used to adjust the baseline of linear prediction values; is the function name, used to convert the Cox proportional hazards model into a clinically readable point scale visualization chart; is a function list, which is a set of custom functions for calculating survival probability, and in this method it is a single-element list containing only one logical function based on ; is the independent variable input into , specifically the linear prediction value output by the Cox model; is the label text, the label name of the survival probability result in the nomogram, which is "surv" (i.e. survival probability) in this method, used for clinical personnel to identify the meaning of the result.

[0061] Among them, the nomogram is presented as a clinically readable point scale chart, with each prediction factor corresponding to an independent scale line (such as vimentin score 0-300 corresponding to 0-100 points), and the total risk score is calculated by the formula: , Among them, is the individualized risk score of the patient (range 0-300 points); is the regression coefficient of the i-th prediction factor in the Cox model (such as the regression coefficient of vimentin in the total survival model is 0.92); is the actual value of the i-th prediction factor (such as a patient's vimentin score is 100 points). The score range is 0-300 points, ≥150 points is high risk, and <80 points is low risk.

[0062] Secondly, the survival probability calculation formula is realized by the Survival function in R language, and the code logic is: , Among them, is the survival function extracted from the Cox model;​ is the fitted Cox proportional hazards model object; is the function for calculating 5-year survival probability; is the linear predictor (i.e. ) and 60 represents 60 months, i.e. 5 years.

[0063] The survival probability calculation formula is: , wherein, is the survival probability of the patient at time t (e.g. 5-year survival rate = 0.68) given the covariates X; is the cumulative baseline hazard function (i.e. the integral of the baseline hazard rate from 0 to t); is the regression coefficient of each predictor; is the p prognostic predictors after screening; and t is a time variable (unit: month).

[0064] Example: the risk score of a patient is 120 points, and the 5-year overall survival probability is 0.68 (68%) obtained by substituting the formula, which is highly consistent with the actual follow-up of 67.5%, verifying the accuracy of the formula.

[0065] In this step, the restricted cubic spline function accurately captures the non-linear relationship between age and prognosis, solving the limitations of traditional linear assumptions; nomograms convert complex models into intuitive tools, and risk scores and survival probability formulas provide quantitative standards, allowing clinical staff to quickly obtain individualized prognostic information without statistical background; compared with traditional staging reports, this method reduces the prediction error of 5-year survival rate for patients by doctors, improving the accuracy of clinical decision-making.

[0066] As shown in FIG. 1, Figure 6 Part A of the figure is an OS nomogram, which integrates age, TNM-N staging, radiotherapy status and immunohistochemical markers (c-myc, vimentin, p16), calculates an individualized risk score by accumulating the weighted scores of each variable, and calculates the corresponding 1-year, 3-year and 5-year OS probability according to the total score. Part B is an RFS nomogram, which combines age, smoking history, clinical stage, radiotherapy status and immunohistochemical markers (vimentin, p16). The model generates a personalized recurrence risk score by accumulating weighted variables, and can predict 1-year, 3-year and 5-year RFS probability.

[0067] In addition, the online prognosis calculator is realized through a web platform; the web platform is provided with an interactive interface for receiving input of clinical pathological parameters and immunohistochemical indexes, and a display interface for outputting prognosis prediction results.

[0068] ​Furthermore, the online prognosis calculator's display interface outputs the following: the patient's overall survival probability and relapse-free survival probability at 1-year, 3-year, and 5-year time points, as well as the risk stratification results determined based on individualized risk scores; the risk stratification results are determined by preset risk level classification rules, including at least three levels: high risk, intermediate risk, and low risk.

[0069] It is feasible to develop the online prognosis calculator based on the Shiny framework and use a "sidebarLayout" layout: Interactive interface (left side): Continuous parameters (age, immunohistochemistry score) are entered via slider input (age 20-90 years, step size 1 year; score 0-300 points, step size 1 point); categorical parameters (TNM stage, smoking history, etc.) are selected via radio buttons, with clear option labels (e.g., "TNM-N stage: N0 / N1 / N2 / N3").

[0070] Display interface (right side): ① Survival probability table (1 / 3 / 5-year overall survival and relapse-free survival probability, retained to 3 decimal places and marked with 95% confidence interval); ② Dynamic survival curve (patient curve is red, population baseline curve is blue, marked with probability at key time points); ③ Risk stratification labels (low risk / intermediate risk / high risk, corresponding to green / yellow / red).

[0071] Hierarchical rules are based on " " Low risk (annual check-up recommended) Medium risk (quarterly follow-up examination recommended). This is considered high-risk (follow-up examination every 2 months + intensive treatment is recommended). Among them, This represents the difference between the predicted survival rate of patients and the baseline survival rate of the population (measuring the relative risk of patients); the predicted survival rate of patients is the individualized survival rate calculated using the survival probability formula (e.g., a patient's 5-year predicted survival rate is 68%); the baseline survival rate of the population is the overall survival rate of 169 patients (calculated using the Kaplan-Meier method, e.g., a 5-year baseline survival rate of 42.3%). The baseline survival rate of the population is calculated using the Kaplan-Meier method for 169 patients (1-year 78.2%, 3-year 55.1%, 5-year 42.3%). Local deployment requirements: R ≥ 4.0.3, installation of shiny, rms, and other dependent packages, memory ≥ 4GB, startup command "Re"shiny::runApp('HNSCC_OS',port=3838)"", supports use in environments without network access.

[0072] like Figure 7As shown in the figure, part A is the overall survival prediction tool interface (https: / / dongxiaotang.shinyapps.io / HNSCC_OS / ), which can generate a personalized 1-year, 3-year and 5-year overall survival probability report containing a dynamic calibration curve by inputting clinical pathological characteristics and immunohistochemical scores. Part B is the recurrence-free survival prediction tool interface (https: / / dongxiaotang.shinyapps.io / HNSCC_RFS / ), which realizes personalized recurrence risk assessment by integrating patient-specific variables and visualizing prediction accuracy.

[0073] In this step, the online prognosis calculator realizes the transformation of the model from academic research to a clinical tool: the interactive interface reduces the operation threshold, the display interface makes the results intuitive and easy to understand, and the risk stratification and clinical recommendations form a prediction-decision closed loop; compared with traditional manual calculation, the prediction time is greatly shortened, and the local deployment supports data security and improves clinical practicability.

[0074] In summary, the head and neck squamous cell carcinoma prognosis prediction method based on immunohistochemical markers of the embodiment integrates the clinical variable data and immunohistochemical marker data of head and neck squamous cell carcinoma patients, optimizes the data quality after data preprocessing, uses LASSO regression to screen key factors that have a significant impact on prognosis, constructs a Cox proportional hazards model based on the screening results and converts it into a intuitive nomogram, verifies the reliability of the model through multiple indexes such as consistency index and calibration curve, and finally develops an online prognosis calculator to realize precise prognosis prediction from multi-dimensional data input to individualized survival probability and risk stratification result output.

[0075] Feasibly, the embodiment also provides a system applying the head and neck squamous cell carcinoma prognosis prediction method based on immunohistochemical markers as described above, which comprises: A data acquisition module for acquiring patient clinical variable data and immunohistochemical marker data; A preprocessing module for multiple imputation of missing values, outlier screening and standardization of data; A model construction module for screening prediction factors using LASSO regression, constructing a Cox proportional hazards model and a nomogram; A verification module for verifying the performance of the model through consistency index, time-dependent ROC curve, calibration curve and Brier score; A prediction module for calculating individualized survival probability and risk stratification based on the model and generating a visual report.

[0076] Secondly, the embodiment also provides a device corresponding to the head and neck squamous cell carcinoma prognosis prediction system based on immunohistochemical markers, which comprises: a memory storing a computer program and patient clinical variable data, immunohistochemical marker data, model parameters; a processor realizing the following functions when executing the computer program: (1) calling a data preprocessing module to clean and standardize the input data; (2) calling a model construction module to generate a prognosis prediction model through LASSO regression and Cox proportional hazards model; (3) calling a verification module to evaluate the performance of the model; (4) calling a prediction module to output 1-year, 3-year, and 5-year overall survival (OS) / recurrence-free survival (RFS) probability and risk stratification results.

[0077] In addition, the embodiment also provides a computer readable storage medium storing a computer program, the program being executed by a processor to realize the following steps: a) obtaining age, TNM stage, radiotherapy history, and c-myc, vimentin, and p16 expression level data of a HNSCC patient; b) using a restricted cubic spline to convert the age variable, and factorizing the classification variable; c) screening variables through LASSO regression, and constructing a Cox proportional hazards model; d) calculating individualized risk scores and survival probabilities, and generating a prediction report including a nomogram and a survival curve.

[0078] In the embodiments provided by the present application, it should be understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, code, or any appropriate combination thereof. For hardware implementation, the processor can be implemented in one or more of the following: application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), processors, controllers, micro-controllers, microprocessors, other electronic units designed to perform the functions described herein, or a combination thereof. For software implementation, the procedures described herein can be implemented with associated hardware to perform the procedures described herein. In implementation, the procedures described above can be stored in a computer readable storage medium or transmitted as one or more instructions or code on a computer readable storage medium. The computer readable storage medium includes a computer storage medium and a communication medium, and the communication medium includes any medium that facilitates transfer of a computer program from one place to another. The storage medium can be any available medium that can be accessed by a computer. The computer readable storage medium can include, but is not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage medium, or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer.

[0079] Finally, it should be noted that the above-described embodiments are merely exemplary of the application and should not be used in a limiting sense. Rather, the description is intended to cover any alternatives, modifications, and equivalents, which can be made to the embodiments described herein, by those who are skilled in the art in light of the above teachings, without departing from the spirit and scope of the application.

Claims

1. A method for predicting the prognosis of head and neck squamous cell carcinoma based on immunohistochemical markers, characterized in that, The method includes: Step 1: Obtain clinical variable data and immunohistochemical marker data from patients with head and neck squamous cell carcinoma, and preprocess the data to ensure its integrity and consistency; Step 2: Screen prognostic predictors by combining LASSO regression with multivariate Cox proportional hazards analysis, construct nomogram models of overall survival and relapse-free survival based on the screened predictors, and generate a visual prognostic prediction tool. Step 3: Validate the performance of the visualized prognostic prediction tool using multiple validation metrics to ensure prediction accuracy; Step 4: Develop an online prognostic prediction calculator based on the validated visual prognostic prediction tool to achieve individualized risk stratification and prognostic prediction for patients.

2. The method for predicting the prognosis of head and neck squamous cell carcinoma based on immunohistochemical markers according to claim 1, characterized in that, The clinical variable data include at least the patient's age, sex, smoking history, tumor stage information, and radiotherapy history; the immunohistochemical markers include at least proliferation markers, epithelial-mesenchymal transition markers, and virus infection-related markers associated with the prognosis of head and neck squamous cell carcinoma.

3. The method for predicting the prognosis of head and neck squamous cell carcinoma based on immunohistochemical markers according to claim 2, characterized in that, The data preprocessing includes at least: supplementing missing data, identifying and filtering outlier data, and standardizing variable data; wherein, a nonlinear transformation method is used for age data to fit the risk relationship between age and overall survival and relapse-free survival, wherein the nonlinear transformation method is implemented through a restricted cubic spline function, and the transformation formula is: ， in, For spline nodes, Denotes the positive part function, The number of spline nodes. For age data, For constant terms, For age data, , , These are the coefficients of the linear, quadratic, and cubic terms of age, respectively. Let be the coefficient of the k-th node.

4. The method for predicting the prognosis of head and neck squamous cell carcinoma based on immunohistochemical markers according to claim 1, characterized in that, The LASSO regression analysis selected the optimal regularization parameter through multi-fold cross-validation to screen out prognostic predictors that have a significant impact on overall survival and recurrence-free survival. The nomogram model is a multivariate Cox proportional hazards model.

5. The method for predicting the prognosis of head and neck squamous cell carcinoma based on immunohistochemical markers according to claim 4, characterized in that, The risk function of the multivariate Cox proportional hazards model is expressed as follows: ， in, Let be the risk rate at time t. As the benchmark risk rate, These are the screened prognostic predictors. These are the regression coefficients of each prognostic predictor.

6. The method for predicting the prognosis of head and neck squamous cell carcinoma based on immunohistochemical markers according to claim 5, characterized in that, The visualized prognostic prediction tool is a nomogram, which is generated based on the multivariate Cox proportional hazards model and expressed using a formula. Calculate an individualized risk score, and predict the patient's overall survival probability and relapse-free survival probability at different time points based on the risk score. The survival probability calculation formula is as follows: ， in, This is the cumulative baseline risk function.

7. The method for predicting the prognosis of head and neck squamous cell carcinoma based on immunohistochemical markers according to claim 1, characterized in that, The various validation metrics include metrics for evaluating model discrimination and metrics for evaluating model calibration; wherein, the metrics for evaluating model discrimination include the consistency index and the area under the time-dependent ROC curve; and the metrics for evaluating model calibration include the calibration curve, the Grønnesby-Borgan test, and the Brier score.

8. The method for predicting the prognosis of head and neck squamous cell carcinoma based on immunohistochemical markers according to claim 7, characterized in that, The formula for calculating the consistency index is as follows: ， in, and These are the survival times of patients i and j, respectively. and Let be the linear predicted values ​​for patients i and j, respectively. For indicator functions; The formula for calculating the Brier score is as follows: ， in, For sample size, As the outcome variable, The survival probability at time t is predicted based on the characteristics of patient i.

9. The method for predicting the prognosis of head and neck squamous cell carcinoma based on immunohistochemical markers according to claim 1, characterized in that, The online prognostic calculator is implemented through a web platform; the web platform is equipped with an interactive interface for receiving clinical pathological parameters and immunohistochemical indicators, and a display interface for outputting prognostic prediction results.

10. The method for predicting the prognosis of head and neck squamous cell carcinoma based on immunohistochemical markers according to claim 9, characterized in that, The online prognosis calculator's display interface outputs the following: the patient's overall survival probability and relapse-free survival probability at 1-year, 3-year, and 5-year time points, as well as the risk stratification results determined based on individualized risk scores; the risk stratification results are determined by preset risk level classification rules and include at least three levels: high risk, intermediate risk, and low risk.

Images