Introduction

Laryngeal carcinoma, which makes up 20% of all malignant tumors of the head and neck, is a prevalent type of these malignancies1,2. Laryngeal cancer is estimated to affect over 1,700,000 people annually, and 123,000 people died from it in 2019—86% of whom were men3,4. The majority of pathological kinds of laryngeal cancer, which include squamous cell carcinoma, are believed to be associated with smoking, alcohol use, human papillomavirus infection, and air pollution. Early detection and treatment are especially crucial for laryngeal cancer because it has an excellent prognosis and quality of life. Advanced laryngeal cancer can be treated with definitive Radiation therapy, combination chemotherapy, or adjuvant radiotherapy and total laryngectomy in addition to surgery, radiotherapy, chemotherapy, and other comprehensive treatments5,6. Different stages, forms, and treatments of laryngeal cancer have varying chances of survival, and the overall patient's 5-year relative survival rate is 64%2. Clinically, we use the conventional TNM staging and the 5-year survival rate to develop an overall understanding of the prognosis of patients, but it is challenging to quantify and depict it7. The conventional Cox proportional hazards (CoxPH) model is a prominent prediction model used to accomplish this goal, but it has limitations since it is based on the presumption that there is a linear relationship between survival outcomes and clinical variables8,9. At the same time, the standard survival analysis is unable to forecast an individual's survival prognosis with any degree of precision. The creation and prediction of tumor-related survival models now uses a large number of sophisticated Machine learning techniques (MLTs). When it comes to predicting the prognosis of tumor patients, MLTs—which include Random Survival Forest (RSF), Gradient Boosting Machine (GBM), eXtreme Gradient Boosting (XGBoost), and deep learning model Deepsurv—have been demonstrated to be more accurate than CoxPH model10,11,12,13,14. Studies on the equivalent machine learning of laryngeal cancer to create a survival model have been conducted concurrently15. Unfortunately, there are two flaws in the way that current research predicts the prognosis and survival of laryngeal cancer. First off, the performance of these models is constrained by the small sample size and homogeneity of the prediction algorithms. Furthermore, the lymph node metastatic rates of supraglottic and subglottic laryngeal carcinoma (non-glottic laryngeal carcinoma), which were 19.9% and 8.0% respectively, were higher. This contributed to the varied survival rates for glottic and non-glottic laryngeal cancer16,17. How to accurately predict the survival of patients with different types of laryngeal cancer has become a key problem. Therefore, in this study, we selected different types of laryngeal cancer patients using SEER database data, and developed survival models for glottic and non-glottic cancer, describing the main factors, to predict the survival of patients with laryngeal cancer more accurately. We develop the survival model using data from the SEER database, and compare the CoxPH model with four widely used machine learning techniques. Last but not least, we apply the model to forecast each person's prognosis, which aligns more with clinical application.

Methods

Data collection

The Surveillance, Epidemiology, and End Results Program (SEER) database was used to gather the study's data (Incidence-Seer Research Plus Data 17 Registries Nov 2021 Sub). Using the SEER*Stat program (version 8.4.2), we retrieved individuals who had been given a larynx carcinoma diagnosis by the third edition of the International Classification of Oncology Diseases (ICD-O-3). The period frame covers instances handled between 2000 and 2019. The following were the inclusion requirements: The behavior was identified as malignant and encoded by position and shape as "larynx".

Data clarity

In total, 54,613 patients with primary laryngeal malignant tumors were included. The median follow-up duration of the sample in this study is 38 months. We used the following exclusion criteria to clean up the data: (1) Patients with limited follow-up information; (2) Patients without T stage (AJCC7), N stage (AJCC7), M stage (AJCC7), or AJCC stage grade information.

Feature selection

We selected variables that were directly related to the clinic, such as age, race, and gender, based on clinical experience. We chose the T stage, N stage, M stage, AJCC stage (AJCC stage 7), tumor size, and pathological categorization to assess the patient's health. Finally, to evaluate the patient's treatment plans, we also included radiation therapy, surgery, and chemotherapy.

Models for survival analysis

A classic model for survival analysis, the Cox proportional hazards (CoxPH) model has been the most commonly applied multifactor analysis technique in survival analysis to date18,19.

CoxPH is a statistical technique for survival analysis, which is mainly used to study the relationship between survival time and one or more predictors. The core of the model is the proportional risk hypothesis.

It is expressed as h(t|x) = h0 (t) exp (β|x), h(t|x) is the instantaneous risk function under the given covariable x, h0 (t) is the baseline risk function, on the other hand, exp (β x) represents the multiplicative effect of covariates on risk.

The random survival forest (RSF) model is an extremely efficient integrated learning model that can handle complex data linkages and is made up of numerous decision trees20.

RSF can improve the accuracy and robustness of the prediction, but it does not have a single expression because it is an integrated model consisting of multiple decision trees21. RSF constructs 1000 trees and calculates the importance of variables. To find the optimal model parameters, we adjust three key parameters: the maximum number of features of the tree (mtry), the minimum sample size of each node (nodesize), and the maximum depth of the tree (nodedepth). The values of these parameters are set to mtry from 1 to 10, nodesize from 3 to 30, and nodedepth from 3 to 6. We use a random search strategy (RandomSearch) to optimize the parameters. To evaluate the performance of the model under different parameter configurations, we use tenfold cross-validation and use C-index (ConcordanceIndex) as the evaluation index. The purpose of this process is to find the parameter configuration that can maximize the prediction accuracy of the model through many iterations.

One of the integrated learning methods called Boosting is the gradient boosting machine (GBM) model, which constructs a strong prediction model by combining several weak prediction models (usually decision trees). At each step, GBM adds a new weak learner by minimizing the loss function. The newly added model is trained to reduce the residual generated in the previous step, and the direction is determined by the gradient descent method. It can be expressed as Fm+1(x) = Fm(x) + αmhm(x). Where the Fm(x) is a weak model newly added, and the αm is the learning rate.

XGBoost is an efficient implementation of GBM, especially in optimizing computing speed and efficiency. To reuse the learner with the highest performance, it linearly combines the base learner with various weights22. eXtreme Gradient Boosting (XGBoost) is an optimization of the Gradient Boosting Decision Tree (GBDT), which boosts the algorithm's speed and effectiveness23. The neural network-based multi-task logic regression model developed by Deepsurv outperforms the conventional linear survival model in terms of performance24. DeepSurv uses a deep neural network to simulate the Cox proportional hazard model. Therefore, deepsurv can be expressed as h(t|x) = h0 (t) exp (g(x)), Where the g (x) is the output of the neural network, which represents the linear combination of the covariable x8.

Model training and validation

We categorize five models to adapt to various variable screening techniques used with various models. The RSF, GBM, and XGBoost models are screened using the least absolute shrinkage and selection operator (LASSO) regression analysis, while the CoxPH model is screened using the traditional Univariate and multivariate Cox regression analysis25,26,27.

In contrast, the Deepsurv model can automatically extract features and handle high-dimensional data and nonlinear relationships, so variable screening is not necessary28. We randomly split the data set into t and v datasets (training set and validation set) and test set in the ratio of 9:1 using spss (version 26) to further illustrate the model's dependability. Randomly selected 10% of the data as external verification. Once more, the ratio of 7:3 is used to divide the training set and validation set, and for both splits, the log-rank test is used to evaluate any differences between the two cohorts. The mlr3 package of R (version 4.2.2) uses the grid search approach to fine-tune the hyperparameters in the RSF, GBM, and XGBoost models in the validation set and chooses the most beneficial hyperparameters to build the survival model once the variables have been filtered following the aforementioned stages. Finally, the Deepsurv model is constructed using the Python (version 3.9) sksurv package, and the model is additionally optimized using grid search.

Model evaluation and interpretation

We used the integrated Brier score (IBS), which is appropriate for 1-year, 3-year, and 5-year time points, as the major assessment metric when evaluating the prediction performance of the model in the test set. In addition, the calibration curve is drawn and the conventional time-dependent receiver operating characteristic (ROC) curve as well as the area under the curve (AUC) (1 year, 3 years, and 5 years) are compared. By calculating the clinical net benefit to address the actual needs of clinical decisions, Decision Curve Analysis (DCA), a clinical evaluation prediction model, incorporates the preferences of patients or decision-makers into the analysis. Calculating the various clinicopathological characteristics is also required for the prognosis of contribution. We visualized the survival contribution of several clinicopathological characteristics for 1-year, 3-years, and 5-years using The Shapley Additive Explanations (SHAP) plot.

The particular prediction

Clinically speaking, various individuals require personalized care. Consequently, it is crucial to estimate the likelihood that a single patient will survive. The survival probability of a certain patient is predicted using the ggh4x package of R (version 4.2.2), along with the contribution of several clinicopathological characteristics to survival. This has major clinical work implications.

Results

Baseline characteristics

The information of 54,613 patients was included. After data cleaning, there were 5953 patients with glottic carcinoma and 4465 patients with non-glottic (supraglottic and subglottic) cancer as a result of the aforementioned exclusion criteria. Figure 1 shows specific cleaning procedures. Table 1 displays the clinical and clinicopathological characteristics of these patients as well as the relevant categorization ratio. In Fig. 2, the survival curve was displayed after patients with glottic and non-glottic cancer were divided into training and validation datasets and testing datasets, respectively.

Figure 1
figure 1

Diagrammatic sketch of study design.

Full size image
Table 1 The information for laryngeal carcinoma patients in the training set and the validation set.
Full size table
Figure 2
figure 2

The t(train), v(validation), and test cohorts' Kaplan–Meier curves. The log-rank test revealed no statistically significant difference between the two cohorts' survival rates (P > 0.05). Unit of time: month. (a) glottic carcinoma, (b) non-glottic carcinoma.

Full size image

Feature selection and model construction

Age, histology, tumor size, RN Eval (Regular Nodes Evaluation), AJCC T, AJCC N, AJCC M, AJCC Stage, surgery, and chemotherapy were the 10 significant mutations identified in the univariate Cox regression analysis for Glottic Carcinoma. Following multivariate Cox regression, age, AJCC T, AJCC N, AJCC Stage, and surgery were the final 5 effective variables to be included. Similarly, the effective variables of univariate Cox regression analysis and multivariate Cox regression analysis of non-glottic laryngeal carcinoma were age, sex, tumor size, RN Eval, AJCC T, AJCC N, AJCC M, AJCC Stage, surgery, radiotherapy and age, sex, AJCC Stage, surgery, radiotherapy. Table S1 displays the outcomes of the univariate and multivariate Cox regression. Machine learning characteristic variables (Fig. 3) were chosen using lasso regression analysis based on the lowest standard. A total of 10 efficient variables were chosen for glottic carcinoma: age, sex, histology, AJCC T, AJCC N, AJCC Stage, RN Eval, radiotherapy, surgery, and tumor size. The 11 effective variables were chosen for non-glottic carcinoma: age, sex, histology, AJCC T, AJCC N, AJCC M, AJCC Stage, chemotherapy, radiotherapy, surgery, and tumor size.

Figure 3
figure 3

The clinicopathological characteristics of the machine learning model were examined using the least absolute shrinkage and selection operator (LASSO) regression. (a) glottic carcinoma, (b) non-glottic carcinoma.

Full size image

Constructing and evaluating survival analysis models

We built the CoxPH model, RSF model, GBM model, and XGBoost model for glottic and non-glottic cancer by the outcomes of multivariate Cox regression analysis and lasso regression analysis, respectively. The Deepsurv model does not require variable screening, hence all 12 variables are used in the model during model development. All survival models are trained to roughly estimate their performance and stability range using Ten-fold cross-validation C-index and IBS. Following the visual examination of the test set, we eventually obtained the following secondary outcomes: 1-year, 3-year, and 5-year ROC curve, calibration curve, C-index, and 1-year, 3-year, and 5-year IBS (Table 2 and Fig. 4). The indicators of the training set are shown in Table S2. All models have IBS that are less than 0.25, which suggests that they can be calibrated well. Figure 5 displays the 1-year, 3-year, and 5-year DCA decision curves for the best model RSF at the same period.

Table 2 Two types of laryngeal cancer are performed using various survival prediction algorithms.
Full size table
Figure 4
figure 4

The chart "RSF Model for predicting 3-and 5-year Survival rates of Laryngeal Cancer patients: calibration Curve and time-dependent ROC Curve" shows the calibration curve and time-dependent ROC curve of RSF model for predicting 3-and 5-year survival rates of laryngeal cancer patients. The calibration curve shows the consistency between the predicted survival rate and the actual survival rate, while the ROC curve provides the performance of the model under different discriminant thresholds. Month is the unit of time. (a) glottic carcinoma, (b) non-glottic carcinoma.

Full size image
Figure 5
figure 5

The 3-year and 5-year decision-making curves are based on the RSF model. The decision curve shows the net benefit of the model prediction under different patient risk thresholds. By comparing the net income of model prediction with and without model prediction under a specific risk threshold, the application value of the model in clinical decision-making can be evaluated. In the figure, the horizontal axis represents the decision threshold and the vertical axis represents the net income. The points on the curve represent the relative net income that the model can predict under a given risk threshold. Ideally, the higher the curve, the greater the net income provided by the model within a wider threshold range, that is, the higher the value of the model in clinical application. (a) glottic carcinoma, (b) non-glottic carcinoma.

Full size image

Visualization of the optimal model's evaluation indexes

The RSF model is the most effective one for both glottic and non-glottic carcinomas, and its C-index in the test set is 0.687 for glottic and 0.657 for non-glottic, respectively. Their 1-year, 3-year, and 5-year IBS were 0.116, 0.182, and 0.195 for glottic carcinomas, and 0.130, 0.215, and 0.220 for non-glottic carcinomas, respectively. Figure 6 depicts the impact of several clinicopathological characteristics on patient survival for the RSF model of two subtypes of laryngeal cancer. AJCC Stage, age, and AJCC T are the first three factors that have the greatest impact on glottic carcinoma. And it is AJCC Stage, age, surgery for non-glottic cancer.

Figure 6
figure 6

The SHAP plot of the RSF model. The vertical axis lists many clinical characteristics, while the horizontal axis shows how the variable affected the outcomes. The likelihood of dying increases with a feature's SHAP value. The picture reflects the Shap values predicted by 1-year,3-year, and 5 years. (a) glottic carcinoma, (b) non-glottic carcinoma.

Full size image

The particular forecast

Two patients with glottic carcinoma and two individuals with non-glottic carcinoma were chosen at random. Their clinicopathological data is listed below.

Glottic cancer: Patient 1 is a male, aged 70 to 79 years, with non-squamous cell carcinoma, T3N0, AJCC III, undergoing surgery and radiotherapy; tumor size is unknown; Patient 2 is a male, aged 60 to 69 years, with squamous cell carcinoma, T3N2c, AJCC IVA, not undergoing surgery and undergoing radiotherapy; tumor size is less than 1 cm.

Non-glottic laryngeal carcinoma: Patient 1 is a male, 50–59 years old, squamous cell cancer, T1N0M0, AJCC I, surgery, radiation, chemotherapy, tumor size less than 1 cm; Patient 2: Male, 50–59 years of age, squamous cell carcinoma, T2N3M0, AJCCIVB, surgery, radiation, no chemotherapy, tumor size less than 1 cm. Figure 7 depicts their unique forecasting chart.

Figure 7
figure 7

Survival prediction of individual patients. The horizontal axis indicates various ages (1-year, 3-year, 5-year), the vertical axis shows the contribution to survival, while the lines of various hues represent various clinical parameters. Less than 0 indicates a detrimental contribution to survival, whereas more than 0 indicates a beneficial one. (a) glottic carcinoma, (b) non-glottic carcinoma.

Full size image

Discussion

In otorhinolaryngology, head and neck surgery, laryngeal carcinoma is a common malignant tumor. Early laryngeal cancer has an occult quality. As examination and treatment techniques advance, the fibrolaryngoscope, for instance, is being used more frequently in clinics to play a significant role in early screening for laryngeal cancer. Nevertheless, since many patients ignore early symptoms such as hoarseness and throat discomfort, many people will mistake them for chronic diseases including chronic pharyngitis, leading to delayed diagnosis and treatment. More than 60% of patients are diagnosed with advanced cancer, based on studies, which significantly lowers the efficiency of laryngeal cancer treatment2.

Despite postoperative adjuvant radiotherapy and chemotherapy do not have a favorable prognosis, patients with advanced laryngeal cancer with lymph node metastases sometimes undergo partial laryngectomy or even total laryngectomy. Patients with early laryngeal cancer have a decent prognosis, and their quality of life will significantly diminish as a consequence of total laryngectomy29. The physical foundation and developmental base of the larynx are unique. Glottic type, supraglottic type, and subglottic type are three subtypes of laryngeal cancer. The glottic area and subglottic region's structure is derived from the storage trachea germ base, whereas the supraglottic region's structure is derived from the oropharynx germ. As a result, these two regions have different fibrous fascia and lymphatic drainage systems. Clinically, glottic and non-glottic cancer have significantly different risks of lymph nodes and distant metastases. The likelihood of survival varies significantly between various forms of laryngeal cancer, too.

There is some research on the likelihood of surviving laryngeal cancer, however, the majority of them use the outdated Kaplan–Meier estimator survival model (Kaplan and Meier, 1958), which is unable to incorporate the patient's variables. The obvious drawback of the KM survival model is that diverse clinicopathological variables influence how the tumor develops and evolves5,30.

The CoxPH model, which can handle censored and censored data as well as continuous and sub-type variables, is suggested as a solution to the problem of covariable fitting. The most used model for predicting survival, the CoxPH model, measures the effect of covariables on survival time using a partial regression coefficient and risk ratio. CoxPH model's assumptions that the risk rate is constant and that the logarithm of the risk rate is a linear function of the covariable are limiting. If the hypothesis is incorrect, the prediction will be biased8,31.

Machine learning-based survival analysis has been increasingly used in recent years to forecast the survival of tumor patients. Machine learning can handle complex, nonlinear data, extract relevant features and information, and enhance the model's generalizability and accuracy. Machine learning does not need to make as many assumptions about data distribution or risk functions as the CoxPH model requires. More significantly, we can estimate each patient's survival after the development of a machine-learning model, which has enormous clinical importance. Different machine learning algorithms can be used to handle various types of data and have varying properties. Based on the CoxPH model's survival analysis, RSF, XGBoost, GBM model, and deep learning model Deepsurv are added in this work.

Many academics have discovered that RSF is a survival analysis model with good performance in earlier investigations. It creates several decision trees by self-sampling and combines the outcomes of each tree's predictions by voting or averaging32,33.

Of course, similar machine learning studies on the survival analysis of head and neck malignancies, such as laryngeal cancer, hypopharyngeal carcinoma, oropharyngeal carcinoma, and nasopharyngeal carcinoma, have been conducted by various researchers34,35. However, as previously mentioned, different embryonic sources and fibrous fascia tissues cause laryngeal carcinoma to be divided into different subtypes. Because previous studies have not distinguished between different subtypes, there will inevitably be a discrepancy between the expected results and the actual results. Based on this, we created the five survival analysis models mentioned above, one for glottic carcinoma and the other for non-glottic carcinoma. RSF is an excellent model, to sum up. The C-index of two separate subtypes of laryngeal carcinoma RSF reached 0.687 and 0.657, respectively, in the final test set. The integrated Brier score (IBS) of their 1-year, 3-year, and 5-year time points is, respectively, 0.116, 0.182, 0.195 (glottic type), and 0.130, 0.215, 0.220 (non-glottic type). This demonstrates the RSF model's high degree of reliability and strengthens our conclusion. The SHAP plot can also more easily convey how risk factors affect specific survival outcomes when compared to the conventional CoxPH analysis nomogram plot. Furthermore, using the RSF machine learning model, we can build the individual survival probability curve for any patient and display their survival prognosis in a more precise manner. This raises the study's clinical relevance even further.

This study has some limitations. First of all, glottic carcinoma and supraglottic carcinoma account for the vast majority of subtypes of laryngeal carcinoma, and because there is a dearth of data on subglottic type, we are unable to develop a survival analysis model for subglottic carcinoma alone. It can only be split into glottic type and non-glottic type as a result. Theoretically, a more precise division results in a more precise forecast. Second, while not terrible, our model C-index still has to be refined by academics.

In conclusion, we compared the prognostic value of patients with various subtypes of laryngeal cancer using five survival prediction model algorithms, and we selected the best RSF algorithm based on which we established survival prognosis prediction for patients with two subtypes of laryngeal cancer, model it and depict it. To advance customized medicine, we also give professionals a tailored patient prognosis prediction model at the same time. Our research demonstrates that the RSF algorithm offers promising therapeutic potential for the prognostic prediction of laryngeal cancer.