Risk factors associated with overall survival in patients with cervical cancer: a prospective cohort study in Western China comparing random survival forest and Cox proportional hazards models

Abstract

Objective

Cervical cancer (CCa) significantly affects female fertility and quality of life. This study aimed to construct and validate a random survival forest (RSF) model to identify the factors that affect the overall survival (OS) in patients with CCa in China and compare its performance with that of the Cox proportional hazards model (Cox model).

Methods

Data on CCa patients were collected from Chongqing University Cancer Hospital. The performance and discrimination ability of the models were evaluated via the C-index, integrated Brier score (IBS), accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). The Kaplan-Meier (K-M) survival curve was used to analyze the difference in OS between patients with high and low risk predicted by RSF model.

Results

A total of 3,982 patients were included in this study. Comparing to Cox model, the RSF model ranked important variables and identified radiotherapy (RT) as an important treatment measure. A comprehensive analysis of the evaluation indices confirmed that the RSF model outperformed the Cox model (IBS: 0.152 vs. 0.162, C-index: 0.863 vs. 0.764). The RSF model metrics for the validation cohort (VC) were as follows: 1-, 3-, and 5-year AUC (0.908, 0.884, and 0.869), sensitivity (0.746), specificity (0.825), and accuracy (0.808). The OS of low-risk patients predicted by RSF was greater than that of high-risk patients.

Conclusion

The RSF model demonstrated excellent discrimination, calibrated predictions, and stratified risk for CCa patients. Furthermore, it outperformed the Cox model in predicting risks, thus enabling the delivery of personalised treatment and follow-up strategies.

Synopsis

This study developed and validated a random survival forest (RSF) model for cervical cancer overall survival (OS). RSF integrated demographics, therapy, hematological biomarkers and pathogens. RSF effectively stratified OS risk with superior discrimination and calibration compared to the Cox model.

INTRODUCTION

Cervical cancer (CCa) ranks as the fourth most common cancer among women worldwide [1]. In 2016, it accounted for 6.5% of all cancer cases and 4.2% of cancer-related deaths among women in China, highlighting the significant impact of this disease on women’s health [2]. As CCa is both preventable and curable, accurate identification of risk factors is essential for effective prevention and optimal treatment outcomes. Thus, establishing a suitable prognostic model for patients with CCa is of paramount importance.

The Cox proportional hazards model (Cox model), developed by Sir David Cox, is a widely utilized method for analyzing survival data in medicine. It employs hazard ratios (HRs) to identify both risk and protective factors and has been extensively applied in constructing prediction models for cancer prognosis [3]. However, Cox models typically assume linear relationships, while many clinical and pathological features display nonlinear relationships in medical practice [4]. Therefore, there is a need for more broadly applicable models that offer higher prediction accuracy to account for the complexity of prognostic factors [5,6].

The random forest (RF) algorithm is capable of capturing higher-order features and interactions without relying on specific assumptions, which enhances its versatility [7]. The random survival forest (RSF) method, introduced by Ishwaran et al. [8], is a nonparametric decision tree machine learning technique that excels at identifying predictors of events and has demonstrated superior predictive power compared to Cox models [9]. In the context of survival analysis, RSF is frequently employed as a prediction model to identify the most important variables related to time and events, selecting prognostic factors with high predictive performance based on minimal depth and variable importance (VIMP) [10,11]. Due to their high predictive performance and interpretability, RSF are widely utilized in clinical research [12].

Several researchers have explored prognostic risk factors for cancer via novel models. For instance, CCa lymph node metastasis has previously been evaluated by imaging and radiology, but these methods are often subject to subjective interpretation. To address this, 6 machine learning models were constructed using heart rate variability as a characteristic variable, with the RF model emerging as the best predictor of preoperative lymph node metastasis status in CCa patients [13]. Additionally, a nomogram identified pathological type, stage, and chemotherapy as independent prognostic variables for overall survival (OS) in patients with CCa [14]. Furthermore, the RSF model has benn employed to investigate the link between human papillomavirus (HPV) status and prognosis in CCa patients in The Cancer Genome Atlas database [15]. Despite advancements in predictive modeling for CCa survival, a notable gap exists in the literature concerning the development of a comprehensive RSF-based model. This model should incorporate a broad spectrum of variables, including patient demographics, clinical features, treatment strategies, laboratory parameters. Addressing this gap is essential, as it could significantly enhance the accuracy of survival predictions and inform treatment strategies.

In light of this, our study aims to develop and validate a prognostic model utilizing the RSF approach specifically tailored for CCa. Our primary objective is to evaluate the comparative predictive performance of this model against the traditional Cox model. By doing so, we seek to identify key clinical features that are most predictive of patient prognosis. This comparison is anticipated to refine our capacity to guide medical diagnosis and treatment, devise appropriate follow-up strategies, and ultimately contribute to improve survival outcomes for patients with CCa.

MATERIALS AND METHODS

1. Study population and data source

This prospective cohort study utilised a dataset from 4,525 patients with CCa from the tumour database established at Chongqing University Cancer Hospital (CUCH) between May 1, 2015, and December 12, 2019. This study adhered to the principles of the Declaration of Helsinki and obtained approval from the Ethics Committee of CUCH. All participants provided written informed consent. Fig. 1 illustrates the rigorous and systematic inclusion and exclusion process based on predefined criteria, resulting in the final analysis of 3,982 patients, who were randomly assigned to a training cohort (TC) (n=2,788) and a validation cohort (VC) (n=1,194) for model development and assessment, respectively. The inclusion criteria for patients were as follows: 1) completion of the entire treatment course, including surgery, radiotherapy (RT), and chemotherapy (CT); 2) age of 18 years or older; 3) treatment received at CUCH; 4) histological confirmation of primary CCa; 5) availability of complete baseline and follow-up information. The exclusion process was as detailed below: 1) patients without complete follow-up records were excluded (n=132); 2) patients lacking imaging confirmation of primary CCa were also excluded (n=155); 3) additionally, patients who did not receive standard treatment were excluded (n=256).

Fig. 1. Flowchart detailing the patient selection process, including inclusion and exclusion criteria, along with the random assignment to TC and VC.

CCa, cervical cancer; TC, training cohort; VC, validation cohort.

Patient demographics, including age, medical insurance (urban employee-based basic medical insurance/urban resident-based basic medical insurance, UEBMI/URBMI), marital status, and ethnicity, were considered in this study. Clinical features, such as clinical stage classified according to the International Federation of Gynaecology and Obstetrics (FIGO 2018) and pathological type (squamous cell carcinoma, adenocarcinoma, and adenosquamous carcinoma), were collected. The treatment strategies used by patients, including surgery, RT, and CT, were recorded. Laboratory parameters, including high-risk human papillomavirus (hrHPV, only HPV-16/HPV-18/HPV-52/HPV-58) status, β2-microglobulin, the serum albumin to globulin ratio (A/G), the neutrophil-lymphocyte ratio (NLR), the platelet-lymphocyte ratio (PLR), the neutrophil (N) count, the lymphocyte (LYM) count, and the albumin (Alb) count, were also retrieved. The cut-off points for continuous variables were determined via X-tile. The continuous variables were transformed into categorical variables according to the following values: β2-microglobulin (≤2.5, >2.5), A/G (≤1.8, >1.80), NLR (≤4.5, >4.5), PLR (≤216.52, >216.52), N (≤6.46, >6.46), LYM (≤0.89, >0.89), and Alb (≤66.67, >66.67). The primary endpoint of interest was the OS of CCa patients, calculated from the date of CCa diagnosis to the date of patient death or the last follow-up, whichever occurred first.

2. Model construction

Cox model

The TC and VC were randomly assigned. Univariate Cox regression analysis was performed to verify the risk factors for CCa. Variables with p-values <0.05 were entered into the multivariate Cox regression model to analyse the associations. Comprehensive analysis of univariate Cox regression and stepwise Cox model was used to construct a Cox model that considered factors such as patient demographics, clinical features, treatment strategies, laboratory parameters. The Cox model was developed through TC and was created based on the final Cox regression model calculated through a stepwise process. Multivariate Cox regression results were used to construct Cox models of 1-, 3-, and 5-year CCa.

RSF model

The RSF model was constructed via the following procedures: 1) A sample subset was created by randomly selecting samples from the eligible data with replacement by bootstrapping (nonparametric). Approximately 33% of the data from each sample were excluded as out-of-bag (OOB) data. 2) In each sample, a survival tree was constructed using randomly selected covariates. 3) Binary splitting was performed via log-rank (LR) splitting rules. 4) The survival tree was expanded until it reached its terminal nodes. Additionally, the mean of the cumulative hazard functions (CHFs) for each tree was evaluated to obtain the ensemble of CHFs. 5) OOB data were used to evaluate the prediction error rate. The optimal number of trees is determined based on the minimum error rate, as a lower error rate indicates a better model fit. Moreover, RSF can be used to identify risk factors for variable selection, ranking the importance of variables by VIMP or minimal depth. In the RSF model, 18 attributes related to CCa were selected based on their VIMP via the minimal depth method. These attributes included ethnicity, A/G, marriage status, medical insurance, Alb, N, pathology, LYM, occupation, surgery, RT, β2-microglobulin, CT, PLR, NLR, age, hrHPV, and stage.

3. Assessment of model performance

In this study, we utilised the C-index and integrated Brier score (IBS) to evaluate the performance of the two models. The C-index, which assesses the discriminatory ability of a survival model, quantifies the proportion of patient pairs for which the model accurately predicts the order of outcomes, thereby reflecting the model’s predictive accuracy. A C-index value of 0.5 signifies that the model lacks discriminatory ability, while a value of 1 indicates perfect discriminatory ability, demonstrating the model’s capacity to accurately predict the sequence of events. Consequently, the closer the C-index is to 1, the stronger the model’s discriminatory ability [16]. Additionally, the Brier score, calculated as the mean square error (MSE) of the probability predictions, was used to assess the overall performance of the models. In survival models, the Brier score is commonly integrated over time to accommodate the time-varying nature of survival data, resulting in the IBS. The Brier score ranges from 0 to 1, where a score of 0 signifies perfect prediction, indicating no discrepancy between predicted probabilities and actual outcomes, while a score of 1 represents the worst possible prediction, characterized by maximum error [17]. A lower Brier score indicates better predictive calibration for a given dataset. Therefore, we assessed the performance of the Cox and RSF models by integrating the C-index, IBS, accuracy, sensitivity, specificity and area under the receiver operating characteristic (ROC) curve (AUC).

4. Statistical analysis

The data were analysed via R version 4.2.1 (Institute for Statistics and Mathematics, Vienna, Austria). A variety of R packages were utilized for the development and evaluation of models. The survival package (version 3.3-1) was employed to conduct survival analyses, which included fitting Cox models and generating Kaplan-Meier (K-M) survival curves. The foreign package (version 0.8-82) facilitated the reading and writing of data files from other statistical software, such as SPSS, SAS, and Stata. For regression modeling strategies, particularly in the calibration and validation of Cox models, we utilized the rms package (version 6.3-1). The timeROC package (version 0.4) was used for computing and plotting time-dependent ROC curves. The randomForest package (version 4.6-14) was employed to construct random forest models, including RSF. Finally, the pec package (version 2022.03.06) was utilized for calculating and plotting prediction error curves [18,19]. The statistical significance level was set at p≤0.05 for 2-sided tests. The K-M survival curve was used to analyze the difference in OS between patients with high and low risk predicted by RSF model.

RESULTS

1. Clinical characteristics of patients

This study analyzed a total of 4,525 patients diagnosed with CCa between 2015 and 2019 using data from the CHCU. Among the 543 patients who did not meet the inclusion criteria, they were excluded from the study. A total of 2788 patients (approximately 70% of the data) were assigned to the TC, and the remaining 1,194 patients (approximately 30% of the data) were assigned to the VC. The median survival time was 33.64 months. Table S1 summarises the characteristics and treatment methods of patients in the TC and VC groups. The mean age at diagnosis was 52.840 years (range: 18–91), and most patients were married (n=3,613, 90.73%). Squamous cell carcinoma was the predominant tumour pathology (n=3,468, 87.09%). A subset of patients received RT (n=1,389, 34.88%) and CT (n=1,393, 34.98%). Approximately half of the patients underwent surgery as part of their treatment (n=2,021, 50.75%). Notably, there were no significant differences observed between the TC and VC.

2. Construction and verification of the Cox model

In our analysis, we performed both univariate and multivariate assessments of the impact of various risk factors with TC on CCa outcomes. As outlined in Table 1, several factors were found to be significantly associated with the prognosis of CCa. Medical insurance (only UEBMI), age, pathological type (only adenosquamous carcinoma), stage (including stage II/III/IV), hrHPV, surgery, RT, CT, β2-microglobulin, NLR, PLR, LYM, Alb, and N were significant predictors of OS in patients with CCa in the univariable analysis. The multivariable analysis revealed that OS in patients with CCa is significantly influenced by various prognostic factors. Notably, advancing age was found to be negatively correlated with OS, indicating that older patients face an increased risk of adverse outcomes. Additionally, the pathological type of the tumor emerged as a significant predictor of survival, with adenocarcinoma and adenosquamous carcinoma associated with reduced OS compared to squamous cell carcinoma. Furthermore, the pathological stage of the tumor was identified as a robust indicator of survival, with more advanced disease stages correlating with decreased OS. Moreover, elevated serum levels of β2-microglobulin, a biomarker frequently indicative of disease progression, were linked to a decline in OS. The NLR and the PLR, both recognized markers of systemic inflammation, exhibited negative correlations with survival. Elevated levels of NLR and PLR suggest a more aggressive disease state and are associated with poorer prognosis. These findings highlight the potential role of increased systemic inflammation, as reflected by NLR and PLR, in disease progression and its contribution to elevated mortality risk. In addition, we found that medical insurance (only UEBMI), surgery, CT, and Alb were independent predictors. The ROC curves for the Cox model are shown in Fig. 2. The AUC values for TC at 1, 3, and 5 years were 0.806, 0.754, and 0.774, respectively, whereas those for VC were 0.816, 0.777, and 0.772, respectively. The sensitivity of the Cox model for VC was 0.762; the specificity was 0.697, and the accuracy was 0.71. The performance of the Cox model is summarized in Table 2, which displays the C-index values of 0.764 (95% confidence interval [CI]=0.74–0.78) for TC and 0.782 (95% CI=0.75–0.81) for VC.

Table 1. Univariate and multivariate analysis of the risk factors for cervical cancer in the training sets.

Variables		Univariable analysis		Multivariable analysis
		HR (95% CI)	p-value	HR (95% CI)	p-value
Medical insurance
	URBMI	1.00		1.00
	UEBMI	0.66 (0.54–0.81)	<0.001	0.70 (0.57–0.86)	0.001
Age (yr)		1.03 (1.02–1.04)	<0.001	1.02 (1.01–1.02)	<0.001
Marital status
	Married	1.00
	Unmarried/divorced/other	1.20 (0.91–1.60)	0.203
Ethnic
	Han	1.00
	Minority	1.18 (0.72–1.94)	0.515
Pathological type
	Squamous carcinoma	1.00		1.00
	Adenocarcinoma	1.03 (0.78–1.37)	0.832	1.36 (1.01–1.82)	0.040
	Adenosquamous carcinoma	1.55 (1.03–2.34)	0.037	2.44 (1.60–3.70)	<0.001
Stage
	I	1.00		1.00
	II	2.27 (1.72–2.98)	<0.001	1.80 (1.35–2.39)	<0.001
	III	3.99 (3.06–5.20)	<0.001	3.31 (2.50–4.39)	<0.001
	IV	12.23 (9.09–16.47)	<0.001	9.42 (6.87–12.91)	<0.001
hrHPV
	Negative	1.00
	Positive	0.81 (0.68–0.96)	0.017
Surgery
	No	1.00		1.00
	Yes	0.48 (0.40–0.58)	<0.001	0.81 (0.67–0.98)	0.027
Radiotherapy
	No	1.00
	Yes	0.44 (0.35–0.56)	<0.001
Chemotherapy
	No	1.00		1.00
	Yes	0.42 (0.33–0.54)	<0.001	0.32 (0.25–0.41)	<0.001
β2-microglobulin
	≤2.5	1.00		1.00
	>2.5	2.09 (1.75–2.50)	<0.001	1.50 (1.24–1.82)	<0.001
A/G
	≤1.8	1.00
	>1.8	0.87 (0.65–1.17)	0.362
NLR
	≤4.5	1.00		1.00
	>4.5	2.64 (2.19–3.18)	<0.001	1.61 (1.29–2.01)	<0.001
PLR
	≤216.52	1.00		1.00
	>216.52	2.34 (1.96–2.80)	<0.001	1.44 (1.17–1.78)	0.001
Lymphocyte count
	≤0.89	1.00
	>0.89	0.42 (0.34–0.52)	<0.001
Albumin count
	≤66.67	1.00		1.00
	>66.67	0.43 (0.29–0.64)	<0.001	0.59 (0.39–0.90)	0.013
Neutrophil count
	≤6.46	1.00
	>6.46	1.88 (1.50–2.37)	<0.001

A/G, serum albumin to globulin ratio; CI, confidence interval; HR, hazard ratio; hrHPV, high-risk human papillomavirus; NLR, neutrophil-lymphocyte ratio; PLR, platelet-lymphocyte ratio; UEBMI, urban employee-based basic medical insurance; URBMI, urban resident-based basic medical insurance.

Fig. 2. Receiver operating characteristic curve of Cox model for OS in TC and VC. (A) TC, (B) VC.

AUC, area under the receiver operating characteristic curve; OS, overall survival; TC, training cohort; VC, validation cohort.

Table 2. Evaluation indication of models used.

Models	TC			VC
	C-index	Error rate	IBS	C-index	Error rate	IBS
Cox model	0.764	0.236	0.162	0.782	0.218	0.167
RSF (log. rank)	0.863	0.238	0.152	0.787	0.229	0.129

IBS, integrated Brier score; RSF, random survival forest; TC, training cohort; VC, validation cohort.

3. Development and validation of the RSF model

Error rates were calculated for the RSF model with fewer than 500 survival trees, and the most important variables were evaluated, as depicted in Fig. 3. The results revealed that the error rate remained constant when the number of surviving trees exceeded 350. The results of the study revealed that among the top 10 feature variables, clinical stage was the most significant, followed closely by age, CT, RT, pathology, NLR, PLR, Alb, β2-microglobulin, LYM, and N. The predictive performance of the RSF model is illustrated in Fig. 4, which showcases the ROC curves for OS in both TC and VC. The AUC values of the RSF model for estimating 1-, 3-, and 5-year OS in the TC were 0.904, 0.876, and 0.883, respectively, whereas in the VC, they were 0.908, 0.884, and 0.869, respectively. The sensitivity of the RSF model for the VC was 0.746; the specificity was 0.825; and the accuracy was 0.808. Table 2 presents the C-index values for the RSF model, which were 0.863 (95% CI=0.85–0.88) for TC and 0.787 (95% CI=0.76–0.81) for VC.

Fig. 3. OOB VIMP and error rate of the RSF model.

A/G, serum albumin to globulin ratio; Alb, albumin; HPV, human papillomavirus; LYM, lymphocyte; N, neutrophil; NLR, neutrophil-lymphocyte ratio; OOB, out-of-bag; PLR, platelet-lymphocyte ratio; RSF, random survival forest; VIMP, variable importance.

Fig. 4. Receiver operating characteristic curve of RSF model for OS in TC and VC. (A) TC, (B) VC.

AUC, area under the receiver operating characteristic curve; OS, overall survival; RSF, random survival forest; TC, training cohort; VC, validation cohort.

4. Comparison of the Cox and RSF models

The optimal model segmentation rule was selected based on the minimised prediction error. Fig. S1 shows that the RSF model had the lowest prediction error compared with the Cox models. As shown in Fig. S2, the predicted survival probability and observed survival frequencies showed a high level of concordance, confirming the accurate predictive ability of the RSF model. As presented in Table 2, when the IBS criterion was used, the RSF model achieved the lowest value of 0.15 compared with the Cox model. These results indicate that the RSF model has excellent overall performance.

5. Stratification analysis of the RSF model

In order to observe the relationship between OS and risk level in patients with CCa, Fig. S3 shows the K-M survival curves generated by the RSF model, which effectively differentiated the OS in patients with CCa in the 2 risk groups. The OS rate of low-risk patients was significantly greater than that of high-risk patients, with increasing survival time in both the TC and the VC.

DISCUSSION

In this study, Cox and RSF models were developed using patient demographics, clinical features, treatment strategies, and laboratory parameters from the follow-up database of CUCH. The RSF model, which uses the log-rank splitting rule, was compared with the Cox model in terms of IBS values, C-indices, and AUC values, successfully distinguishing between low- and high-risk patients. Based on the aforementioned results, it is evident that the RSF model demonstrated a C-index of 0.863 in the TC and 0.787 in the VC. In contrast, the Cox model exhibited a C-index of 0.764 in the TC and 0.782 in the VC. These findings indicate that the predictive accuracy of the RSF model surpasses that of the Cox model. In a study focusing on prognostic prediction models for postoperative CCa patients, the C-index values for the Cox and RSF models were reported as 0.731 and 0.753, respectively, further corroborating the RSF model’s superior predictive accuracy [16]. Additionally, the RSF model demonstrated higher accuracy (0.808) and specificity (0.825) in the VC compared to the Cox model’s accuracy (0.710) and specificity (0.697). The sensitivity of both models was found to be comparable, with no significant differences observed. This finding aligns with another study that employed multiple machine learning models to predict the prognosis of breast cancer patients, which also indicated that the RSF model’s accuracy and specificity exceeded those of other machine learning models, including the Cox model [20]. The AUC values for the RSF model for 1-, 3-, and 5-year OS in the VC were 0.908, 0.884, and 0.869, respectively, all of which surpass those of the Cox model, thereby further substantiating the RSF model’s superior discriminatory capacity. Furthermore, the IBS for the RSF model was 0.129, compared to 0.167 for the Cox model, indicating better predictive calibration for the RSF model. Similarly, in a study predicting OS in lung transplant patients, it was demonstrated that the RSF model’s predictive ability was superior to that of the Cox model, with the Cox model’s AUC and IBS being 0.658 and 0.205, respectively, while the RSF model’s AUC and IBS were 0.879 and 0.130, respectively [21]. The differences between the two models can be understood from several perspectives. The RSF model is a non-parametric machine learning technique that does not rely on specific distributional assumptions and is capable of handling non-linear relationships as well as high-dimensional data. In contrast, the Cox model is a semi-parametric model that assumes constant HR, which may not always align with clinical data. Importantly, we selected a comprehensive set of variables, including patient demographics (age, ethnicity, marital status), clinical features (stage, pathology), treatment strategies (surgery, RT, CT), and laboratory parameters (β2-microglobulin, NLR, PLR, Alb). The selection of these variables was based on their relevance to survival outcomes and the availability of high-quality data. However, the Cox model’s inability to accommodate variables with non-linear relationships may result in the loss of critical information regarding certain important variables, which could adversely affect the specificity and accuracy of OS predictions.

Several independent prognostic factors have been identified, aligning with previous research findings. Specifically, age, pathology, advanced disease stage, high expression of β2-microglobulin, the NLR, the PLR and N are risk factors for OS in patients with CCa. Conversely, factors such as medical insurance, surgical intervention, radiotherapy, chemotherapy, and lower levels of LYM and Alb are considered protective factors. Medical insurance has been identified as a significant prognostic factor for CCa, likely due to the financial burden associated with treatment. The high costs of targeted therapy and immunotherapy for advanced CCa increases the risk of mortality in Southwest China, where the out-of-pocket payment rate is relatively low [22]. However, opinions regarding the significance derived from the benefits of diagnosis and treatment for CCa patients covered by Medicaid differ, indicating a continued need for long-term follow-up in CCa healthcare [23,24]. Surgery remains the primary treatment option for patients with early-stage CCa, while chemotherapy serves as an adjunct therapy for those with advanced, recurrent, or metastatic CCa following surgery and radiotherapy. Patients who undergo both surgery and chemotherapy experience a reduced risk of death from CCa [25,26]. The combination of surgery and radiotherapy constitutes the main treatment approach for CCa [27,28]. A meta-analysis indicated that postoperative radiotherapy and concurrent chemoradiotherapy yield similar prognostic effects on OS and progression-free survival [29]. Notably, a multinational group of radiation oncologists has examined the current challenges associated with radiotherapy and recommended that concurrent chemoradiotherapy be combined with immunotherapy, targeted therapy, or other innovative treatment modalities [30]. The RSF model revealed the importance of variables such as RT, which were not deemed significant in the multivariate analysis. This finding emphasizes the predictive value of the RSF model in identifying relevant clinical factors. Furthermore, haematological biomarkers including NLR, PLR, and Alb, along with β2-microglobulin, were validated as independent prognostic factors [25,31,32,33]. Although this study does not establish hrHPV as a crucial prognostic factor for CCa, HPV infection is acknowledged as a necessary condition for the development and progression of CCa [34]. There is also evidence indicating that vaccination can reduce the incidence of HPV infection [35]. Vaccination is not only effective in preventing CCa but is also vital for the prevention of HPV-related lower genital tract lesions [36]. Moreover, in patients requiring cervical excision, preoperative HPV vaccination has been shown to decrease the recurrence rate of local cervical intraepithelial neoplasia, while postoperative vaccination helps prevent lower genital tract lesions [37,38]. Surprisingly, HPV-negative and HPV-positive cell lines, as well mouse models are being developed to provide translational preclinical models for investigating the mechanisms of HPV-associated disease thereby accelerating preclinical research on HPV-related conditions [39].

The RSF model has consistently demonstrated superior predictive performance across various studies, including those involving patients with CCa. By addressing the limitations associated with the Cox model’s assumption, the RSF model exhibits enhanced performance as the number of events increases. Previous research has effectively utilized the RSF model for predicting patient prognosis, underscoring its efficacy. Consequently, the RSF model serves as a valuable risk classification tool enabling clinicians to assess patient prognosis and inform treatment decisions. The strengths of this study are rooted in the RSF model’s effective risk stratification performance and its comparative analysis with the Cox model to investigate prognostic factors in patients with CCa.

In this study, we employed stringent inclusion and exclusion criteria to ensure the homogeneity of the research sample. While these criteria have enabled us to concentrate on a specific target population, they may have inadvertently introduced selection bias. The utilization of data from a single institution may also limit the external generalizability of the findings. Consequently, further well-designed prospective studies are essential to validate these results. We encourage future research to adopt broader inclusion criteria to enhance the representativeness of the sample and the generalizability of the findings. Additionally, this study did not include computed tomography parameters, specific types of surgery, or cancer grade. Our focus was solely on the decision to undergo surgery and the stage of cancer. Future studies should consider incorporating these factors, including computed tomography parameters, type of surgery, and cancer grade.

The RSF model provided a more accurate and robust prediction of OS in patients with CCa compared to the Cox model. While both the RSF and Cox models demonstrated varying levels of prediction accuracy, they identified similar risk factors. Additionally, the RSF model highlighted RT and ethnicity as significant variables, effectively distinguishing between low-risk and high-risk patients, thereby enhancing its clinical utility in risk stratification. Consequently, in Western China, it is essential to emphasize the critical role of RT in the management of CCa. Effort should be made to improve the expertise of radiotherapy specialists and to optimize the utilization of RT equipment and resources within hospitals, while also establishing standardized protocols for the application of RT.

SUPPLEMENTARY MATERIALS

Table S1

Demographics and clinicopathologic characteristics of the training and validation sets

Fig. S1

Prediction error curves for the Cox and RSF model in TC and VC. (A) TC, (B) VC.

Fig. S2

Calibration of RSF model in TC and VC. (A) TC, (B) VC.

Fig. S3

Kaplan-Meier curves of different risk levels to estimate the OS of patients in TC and VC. (A) TC, (B) VC.