Abstract
Objectives
Follicular lymphoma-grade 3 is an aggressive subtype of Follicular lymphoma with a higher recurrence risk and poorer survival outcomes compared to lower-grade Follicular lymphoma. Existing prognostic models often lack accuracy due to disease heterogeneity and insufficient integration of demographic and treatment variables. This study aimed to identify independent prognostic factors and develop a nomogram for predicting OS in Follicular lymphoma-grade 3 patients using the SEER database.
Methods
This study is a retrospective cohort study. Data from 1026 Follicular lymphoma-grade 3 patients diagnosed between 2016 and 2021 were extracted from the SEER database. Patients were randomly divided into training (n = 718) and validation (n = 308) cohorts. Prognostic factors were identified using RSF, LASSO regression, and the Boruta algorithm. A multivariate Cox regression model was used to identify independent prognostic factors, which were incorporated into a nomogram. The model's performance was evaluated using ROC curves, calibration curves, and DCA.
Results
Age, radiotherapy, and liver metastasis were identified as independent prognostic factors for OS. The nomogram demonstrated strong predictive performance with AUC values exceeding 0.7 at 12, 36, and 60 months in both cohorts. Calibration curves confirmed the agreement between predicted and observed OS rates. Risk stratification using the nomogram identified significant survival differences between low-risk and high-risk groups (P < 0.05).
Conclusion
This study developed a validated nomogram for Follicular lymphoma-grade 3 that integrates clinical, demographic, and treatment factors, offering superior predictive accuracy over traditional staging systems. The model provides a reliable tool for individualized prognosis assessment and treatment optimization in clinical practice.
Introduction
Follicular lymphoma (FL) -grade 3 is recognized as one of the most aggressive subtypes of follicular lymphoma. Compared with low-grade FL, it has been associated with a significantly higher risk of recurrence and transformation, and patient survival is often severely affected.1,2 The identification of prognostic factors for FL remains a major focus of clinical research. With the increasing understanding of the pathological mechanisms and clinical characteristics of FL, it has become evident that a single clinical or pathological indicator is insufficient for the comprehensive assessment of patient prognosis. Although the traditional Ann Arbor Stage provides some prognostic information, it falls short in accurately evaluating survival risks in patients with FL-grade 3 due to the high heterogeneity and individual variability of the disease. 3 It has been demonstrated that the inclusion of additional variables, such as demographic and therapeutic factors, can significantly improve the accuracy of survival predictions. Therefore, the identification of key prognostic factors for FL-grade 3 and the development of accurate survival prediction models are of great clinical importance. In recent years, the widespread application of machine learning in the medical field has provided new opportunities. Studies based on data from surveillance, epidemiology, and end results (SEER) registries have enabled further exploration in this area. 4 In the present study, data from the SEER database on patients with grade 3 FL were analyzed using machine learning techniques. An effective multivariate mathematical model was developed to identify key factors influencing patient survival, with the aim of providing evidence-based support for clinicians in formulating personalized treatment strategies.
Method
This study followed the reporting guidelines of the STROBE guidelines. 5 The overall workflow chart was illustrated in Figure 1.
Schematic of the study workflow.
Research subject
This study is a retrospective cohort study based on data extracted from the SEER database. It aims to identify prognostic factors and develop a predictive nomogram for patients with FL-grade 3. We have obtained authorization to collect data from the SEER databases（17 registries research, Nov 2022 Sub, 2000–2021). We used SEER*Stat software (v8.4.3) to extract information on FL-grade 3 patients(excluding those transformed to diffuse large B-cell lymphoma).Inclusion criteria: patients diagnosed with FL-grade 3 (ICD-O-3 code: M9698/3) from 2016 to 2021;Exclusion criteria: 1. Non only one primary tumor; 2. Non pathologically confirmed diagnoses; 3. Race and marital is unknown; 4. Staging information (Ann Arbor Stage, summary stage) is unknown; 5. Surgery is unknown; 6. Distant metastases (bone, brain, liver, and lung) are unknown. The flowchart of patient screening is shown in Figure 2.
Flowchart illustrating patient selection for this study.
Data extraction
After eliminating patients with incomplete information, 1026 patients diagnosed with follicular lymphoma-grade 3 were incorporated into the study and randomly assigned in a 7:3 ratio to the training set (n = 718) and the validation set (n = 308). The variables encompassed were age, marital, race, sex, Summary Stage, Ann Arbor Stage, surgery, radiation, chemotherapy, bone metastasis, brain metastasis, liver metastasis, lung metastasis, survival status, and survival time. The primary endpoint was overall survival (OS), defined as the time interval from the date of diagnosis of the specific disease to the date of death, regardless of the cause of death.
Variable definitions
The race is categorized into white, black, and other; the marital is delineated as three classes: married, unmarried, and SDW (divorced, separated, widowed); The Ann Arbor Stage categorized patients into four stages: Stage I, Stage II, Stage III, and Stage IV. The Summary Stage was grouped into three categories: Localized, Regional, and Distant. Yes and No are employed for the classification of the following variables: Bone metastasis, Brain metastasis, Liver metastasis, Lung metastasis, and Surgery. Additionally, radiation therapy and chemotherapy were classified as Yes and No/Unknown.
Statistical analysis
Statistical tools
Ordinal data were analyzed using the rank-sum test, whereas categorical data were evaluated with the χ² test. Survival curves were constructed using the Kaplan-Meier (K-M) analysis, and intergroup survival differences were assessed through the Log-rank test. All data analyses were conducted using R software (version 4.1.3), while statistical analyses were performed with SPSS (version 26.0; IBM, Armonk, NY, USA). p-values ＜0.05 were considered indicative of statistical significance.
Variable selection
Pearson correlation analysis was used to measure the linear relationship between two continuous variables. A Pearson correlation coefficient exceeding 0.8 was considered indicative of strong collinearity between variables. 6 Random survival forest (RSF), a survival analysis method based on the principles of random forest, was applied to compute the importance scores of variables. This method evaluates the significance of each feature and identifies variables with a substantial impact on model predictions. 7 The Boruta algorithm, a feature selection method based on random forests, was employed to identify prognostic factors significantly associated with overall survival (OS). This algorithm iteratively generates shadow features and compares them to the original features to determine their importance. 8 Least absolute shrinkage and selection operator (LASSO) regression analysis, leveraging regularization techniques, was also utilized to perform feature selection by shrinking variable coefficients, thereby identifying significant predictors. 9 Both the Boruta algorithm and LASSO regression were designed to reduce the risk of overfitting. 6 The combination of the Boruta algorithm, Lasso regression analysis, and RSF methods was used to identify variables associated with OS. The intersection of the important variables selected by these methods was included in subsequent multivariate Cox regression analysis to determine independent prognostic factors for OS.
Construction and validation of the nomogram
Independent prognostic factors for OS were identified through multivariate Cox regression analysis and were subsequently used to construct a nomogram for predicting OS. Model performance was evaluated based on discrimination and calibration metrics. The discriminatory ability of the nomogram was assessed using the receiver operating characteristic (ROC) curve and the area under the curve (AUC). Calibration was evaluated using a calibration curve generated from 1000 bootstrap resamples. To assess the clinical utility of the nomogram, decision curve analysis (DCA) was performed to compare the net benefits of different predictive models across varying decision thresholds. The risk score for each patient was calculated based on the nomogram, and the median risk score in the training cohort was used as the cutoff value. Patients in the training and validation cohorts were divided into high-risk and low-risk groups accordingly. K-M survival curves were plotted to visualize risk stratification and compare survival outcomes between the high-risk and low-risk groups. Survival differences between the groups were tested using the log-rank test.
Results
Baseline characteristics of patients
The optimal age cutoff values were determined using X-tile software (version 3.6.1) and categorized into three groups: ≤ 60 years, 61–74 years, and ≥75 years (Figure 3A-C). The complete clinicopathological and demographic characteristics of patients with FL-grade 3 are summarized in Table 1. In the overall population, the majority of patients were aged ≤60 years (43.37%), White (85.58%), and married (63.16%). Early-stage disease (Ann Arbor Stage I + II) accounted for 40.65% of cases, while advanced-stage disease (Stage III + IV) comprised 59.35%. According to the Summary Stage, the distant stage was the most common (58.58%). Among metastatic sites, bone metastasis was the most frequently observed (6.92%). A minority of patients underwent surgery (24.46%) or radiotherapy (16.18%), while a larger proportion received chemotherapy (67.64%). There were no statistically significant differences in other variables between the training and validation cohorts.
The optimal cut-off values for age were determined by X-tile software (A-C).
Demographic and treatment characteristics of patients with follicular lymphoma-grade3.
SDW: Separated, divorced, widowed
Identification of independent prognostic factors
Pearson correlation analysis was conducted to calculate the Pearson correlation coefficients between all variables (Figure 4). The results showed that none of the coefficients exceeded 0.8, indicating the absence of multicollinearity among the variables. Therefore, all variables were included in the subsequent feature selection process.
The linear relationship between two continuous variables is measured by Pearson correlation analysis.
For the LASSO regression analysis, the coefficient path plot (Figure 5A) revealed that as the regularization parameter -lambda increased, the regression coefficients of the variables gradually shrank and eventually approached zero. The optimal lambda value was determined using ten-fold cross-validation (Figure 5B). At lambda min, 14 variables were retained, while lambda 1se resulted in no variables being selected. To balance model complexity with the risk of overfitting, the penalty coefficient at lambda min was chosen, ultimately retaining 14 key variables with non-zero coefficients (Figure 5C).The Boruta algorithm was applied to evaluate the importance of all variables on the outcome variable. Variables were categorized as “important” (green icons), “unimportant” (red icons), or “undetermined” (yellow icons). In the final analysis, only variables explicitly marked as “important” were retained, while “unimportant” variables were excluded, ensuring model simplicity and predictive stability. Using the RSF method, the importance score of each variable was calculated to quantify its contribution to survival outcomes. The results indicated that, except for sex and chemotherapy, all other variables had positive importance scores, suggesting significant contributions to survival outcomes. Based on the ranking of importance scores, the top 10 variables were selected for further analysis to optimize the model's interpretability and predictive performance. By integrating the results from the three feature selection methods, six intersecting variables were identified (age, marital status, radiotherapy, liver metastasis, Summary Stage, and Ann Arbor Stage) and included in the multivariate Cox regression analysis. The final multivariate Cox analysis revealed that age, radiotherapy, and liver metastasis were independent prognostic factors for OS (P < 0.05) (Table 2).
The Lasso regression analysis (A-C), Boruta algorithm (D), and RSF (E) were used to determine the important variables for OS (F).
Multivariate Cox regression analysis for OS in training set.
SDW: Separated, Divorced, Widowed
Construction and validation of the nomogram
A nomogram for predicting OS was constructed based on the three independent prognostic factors identified through multivariate regression analysis: age, radiotherapy, and liver metastasis (Figure 6). The nomogram assigns a weighted score to each prognostic factor based on its relative contribution to OS prediction. For individual patients, specific variable values are used to calculate the corresponding scores for each factor. These individual scores are then summed to yield a total score, which is mapped to a specific survival probability. This approach enables personalized survival prediction for OS in individual patients.
Nomogram for prediction of 12 month, 36 month, 60 month OS based on independent risk factors for follicular lymphoma-grade 3 patients. In the nomogram, each level of each variable means a score on the ‘points' scale. Add up each score and draw a straight line down to the 12month,36month,60month scale to get corresponding OS.
Validation of the nomogram
The performance of the OS nomogram was assessed using ROC curves at 12, 36, and 60 months in both the training and validation cohorts. High AUC values were observed for both cohorts, with the training cohort showing values of 0.77 (95% CI, 0.701–0.84), 0.774 (95% CI, 0.714–0.834), and 0.803 (95% CI, 0.735–0.871), respectively, and the validation cohort showing values of 0.718 (95% CI, 0.605–0.831), 0.718 (95% CI, 0.614–0.823), and 0.718 (95% CI, 0.59–0.847) (Figure 7). AUC values exceeding 0.7 in both the training and validation cohorts indicated that the constructed model exhibited good predictive ability. In addition, the calibration curve revealed that the predicted values closely matched the observed values along the 45° ideal line in both the training and validation cohorts, demonstrating the excellent calibration ability of the model (Figure 8). DCA was conducted to further evaluate the clinical utility of the model. It was found that, across most threshold probability ranges, the nomogram provided higher net benefits compared to the Ann Arbor Stage model (Figure 9).
The 12-month (A), 36-month (B) and 60-month ROC (C) curves in the training set. The 12-month (D), 36-month (E) and 60-month (F) ROC curves in the validation set.
The 12-month, 36-month and 60-month of calibration curve for the training set (A). The 12-month, 36-month and 60-month of calibration curve for the validation set (B).
The 12-month (A), 36-month (B) and 60-month (C) DCA curves in the training set. The 12-month (D), 36-month (E) and 60-month (F) DCA curves in the validation set.
Risk stratification
The total score for each patient was calculated using the OS nomogram. The optimal cutoff value for the risk score was determined based on the median score, categorizing patients into a low-risk group (total score ≤100) and a high-risk group (total score >100). Kaplan-Meier survival curves (Figure 10) demonstrated that overall survival was significantly better in the low-risk group compared to the high-risk group in both the training and validation cohorts (P < 0.05).
The Kaplan-Meier survival curve of training set (A) and validation set (B).
Discussion
Although novel therapeutic strategies have significantly improved the survival outcomes of patients with FL, the prognostic differences across various pathological grades still require further evaluation. FL-grade 3, in particular, poses substantial clinical challenges due to its higher aggressiveness and recurrence rates, making precise survival prediction for this specific subgroup especially critical and urgent. 10 Unfortunately, no accurate OS prediction model tailored specifically for FL-grade 3 is currently available. Most existing models rely on single clinical or pathological indicators, which fail to comprehensively capture the complexity of survival risks. To address this gap, our study introduced a novel approach that integrates machine learning with Cox regression analysis. Machine learning was employed to handle high-dimensional data and complex nonlinear relationships, while Cox regression preserved the statistical interpretability of the model. This combined method enabled the development of a more accurate and clinically practical predictive model. The proposed model provides a robust tool for risk stratification and individualized treatment planning for patients with FL-grade 3. 11
Through a comparative analysis of multiple predictive models, we found that prognostic models incorporating a broader range of variables demonstrated higher accuracy and reliability in predicting OS in oncologic diseases, compared to traditional staging systems. The SEER database, a large cancer registry covering approximately 30% of the U.S. population across 22 regions, provides rich multidimensional data, including tumor anatomy, pathological characteristics, demographic factors, and treatment modalities. Its large sample size offers an ideal platform for accurate survival prognosis of specific diseases.
This study extracted data from 1026 patients with FL-grade 3 from the SEER database for variable selection. During the variable selection process, a single method may not comprehensively capture all potential variables related to survival outcomes. Therefore, to overcome this limitation, three feature selection methods were employed: RSF, LASSO and Boruta algorithm. RSF, with its built-in feature importance scoring function, can identify variables significantly associated with survival outcomes and is particularly robust against overfitting, making it suitable for handling complex datasets. LASSO, effective for high-dimensional data, reduces redundancy by retaining only the most relevant features, thus improving the model's predictive accuracy and interpretability. The Boruta algorithm identifies features with significant impact on the target variable by comparing the importance of actual features with randomly generated features. Each method has its unique advantages and complements the others.12–14 By combining RSF, LASSO, and Boruta for feature selection, this study effectively enhanced the accuracy and reliability of variable selection, successfully identifying variables significantly associated with survival outcomes. Based on these selected variables, a Cox regression risk model was constructed, and a user-friendly nomogram for clinical application was generated. Ultimately, the study results indicated that age, radiotherapy, and liver metastasis were independent prognostic factors for overall survival.
Compared to the methods employed in this study, existing models have certain limitations in several aspects. For example, in 2023, Hu JC et al. constructed a 24-month disease progression (POD24) prediction model for FL, which provided initial prognostic assessment for FL patients. 15 However, this model relies solely on pathological type and clinical indicators, lacking treatment information and demographic factors, which limits its ability to fully reflect the survival risk of patients. Similarly, prognostic models proposed by Lu YX and Wan X et al primarily focus on single clinical indicators, failing to capture the complexity of the disease and its impact on survival risk.16,17 In contrast, the prognostic model in this study integrates disease staging, clinical characteristics, treatment methods, and demographic factors (such as age, marital status, race, and gender), providing a more comprehensive assessment of prognosis. This approach improves both the accuracy and clinical applicability of the model.
In addition, the OS prognostic model for FL patients, developed by Li XL et al. in 2024, provides predictive guidance for patients with primary gastrointestinal FL. 18 However, the study was limited to patients with primary gastrointestinal involvement and did not include stratified analysis based on the pathological grade of FL. This limitation makes the model less comprehensive and accurate when evaluating the prognosis of FL patients across different pathological grades. Interestingly, they identified radiotherapy as an independent risk factor for OS in FL, which aligns with the findings of our study. Our results are also consistent with several retrospective studies, one of which analyzed 404 early-stage FL patients and found that, compared to chemotherapy, radiotherapy did not show a shorter OS and even demonstrated better progression-free survival. In our study, 50% of patients receiving radiotherapy also received chemotherapy, which may suggest that the combination of radiotherapy and chemotherapy may play a positive role in disease control compared to chemotherapy alone. In 2021, Chen et al. analyzed the survival outcomes of patients with peripheral T-cell lymphoma-not otherwise specified (PTCL-NOS) using the SEER database, and found that patients in stages I-II who received combined chemotherapy and radiotherapy had a higher overall survival (OS) compared to those who received chemotherapy alone. 19 This finding is consistent with our conclusions . However, for patients with high tumor burden, chemotherapy is still considered the preferred treatment, which may be related to subgroup differences in the study population. Currently, there is no unified standard for the optimal treatment strategy for FL- grade 3, and the choice of treatment regimen remains controversial. Therefore, we believe that the impact of radiotherapy and chemotherapy on survival in FL-grade 3 patients needs to be further validated through large-scale retrospective studies.
Based on this, we further conducted an importance analysis, and found that age is the most significant predictor of OS in FL-grade 3 patients, consistent with previous studies. Increasing age is a strong adverse prognostic factor for FL patients. A Swedish registry study on follicular lymphoma reported that the 10-year survival rate for patients aged 18–49 years was 92%, but it declined to 83% and 78% in the 50–59 and 60–69 age groups, respectively. For patients over 70 years old, the 10-year survival expectation further dropped to 64%. 20 This may be related to physiological aging, an increase in chronic comorbidities, poor treatment tolerance in elderly patients, and the higher incidence of adverse tumor-related biological characteristics in this population.
Notably, this study also identified liver metastasis as an independent prognostic factor for OS in FL-grade 3, which has important clinical implications. This may be related to the liver's role as a key metabolic and detoxifying organ. Liver metastasis often indicates advanced disease with more aggressive characteristics. Patients with liver metastasis are typically at higher risk for complications, including liver failure and immune suppression, which significantly shorten survival. 21 Additionally, due to the complexity of the disease, treatment options for these patients are more limited, further reducing survival rates. In contrast, while bone, brain, and lung metastasis are also considered distant organ metastases, they did not emerge as independent predictors of OS in FL-grade 3. Several large-scale retrospective studies have shown that bone marrow biopsy positivity has little impact on treatment evaluation for FL, less than 1%. 22 The results for brain and lung metastasis may be limited by insufficient sample sizes, leading to statistical bias. Currently, there is a lack of sufficient evidence regarding the impact of distant organ metastasis on the survival prognosis of FL-grade 3, and further validation through large-scale retrospective studies is needed.
Nomograms have been widely used for individualized survival prediction. In this study, we constructed a nomogram based on the aforementioned prognostic factors to predict the OS of FL-grade 3 at 12, 36, and 60 months. ROC curve analysis showed that the AUC values of the nomogram for predicting OS at 12, 36, and 60 months were all above 0.7, indicating good performance in terms of both sensitivity and specificity. The traditional Ann Arbor Stage is primarily used to describe the anatomical distribution and extent of disease; however, its application in individualized survival assessment has certain limitations. In contrast, the nomogram, as a novel graphical prognostic tool, demonstrated superior net benefit over the Ann Arbor Stage, as shown by decision curve analysis. Furthermore, calibration curves indicated good agreement between the predicted and actual survival rates. The nomogram constructed in this study demonstrated strong discrimination and calibration abilities in both the training and validation cohorts. This further supports the reliability and clinical applicability of the nomogram in predicting survival in FL-grade 3. Additionally, risk stratification analysis can help clinicians identify high-risk populations, enabling early intervention and personalized treatment.
Through the constructed nomogram, we visually demonstrated the combined impact of key prognostic factors on patient outcomes, helping clinicians quantify the relative contributions of each variable, thus enabling individualized risk assessment and providing a scientific basis for the formulation of precision treatment strategies. Although the findings of this study are of significant clinical relevance, certain limitations exist. Firstly, the SEER database is based on retrospective data, which may introduce selection and information biases, affecting the generalizability of the results. Due to the lack of further differentiation between FL-3A and FL-3B subtypes in the SEER database for patients with FL-grade 3, potential differences in clinical presentation and prognosis between FL-3A and FL-3B may impact the clinical applicability of the model. Additionally, studies have shown that gene mutations such as KMT2D, CREBBP, and EZH2 are closely associated with the prognosis of follicular lymphoma. 23 However, the SEER database does not include molecular biological variables, preventing their inclusion in our analysis. The SEER database lacks precise documentation of treatment regimens, including critical details such as specific chemotherapy protocols and dosages (R-CHOP or R-Bendamustine), which limits our ability to conduct more refined stratification analyses regarding the relationship between treatment modalities and survival outcomes. Similarly, the SEER database does not include these clinical variables (such as performance status and comorbidities) in patient records, preventing us from incorporating these factors for further refined analysis in our study. The SEER database used in this study covers 22 regions in the United States, representing approximately 30% of the population, and is designed to provide representative epidemiological data. However, White patients accounted for 85% of our study cohort, which may partly reflect the overall White population proportion in the U.S. but may also be influenced by racial disparities in disease incidence, healthcare accessibility, and data collection. This imbalance in racial distribution may limit the generalizability of our findings to other racial groups. Therefore, future studies should further investigate survival outcomes across different geographic regions and racial groups to ensure the broader applicability of the findings. These limitations warrant further exploration, such as conducting large-scale, multi-center prospective studies or integrating molecular biological data into databases, to further validate and optimize the model developed in this study and improve its clinical applicability and predictive accuracy.
Conclusion
By analyzing a large population dataset from the SEER database, this study developed a nomogram for predicting the prognosis of FL- grade 3. The model integrates multiple clinical and demographic variables, enabling more accurate predictions of individual patient survival rates and providing a basis for personalized management and treatment strategies. Our study aims to provide clinicians with a reliable tool to support precise prognostic assessments for FL-grade 3 facilitate individualized treatment, and offer stronger guidance for patient management.
Footnotes
Acknowledgements
We are very grateful to the SEER program for approving the registration and the SEER database.
Ethics considerations
The authors state that this article does not contain any studies with human participants or animals so exempt from institutional review board approval. Informed consent from study participants was not required as this was a retrospective analysis of an existing database.
Authors’ contributions/CRediT
Ruilan Zhong performed the data extraction, statistical analysis, literature investigation, and drafted the manuscript.Ruifu Wei and Chunyu Zhang performed the statistical analysis and revised the manuscript. Xunxiu Ji and Lihua Hu assisted in data acquisition and validation. Zhenxin Mei and Limei Li supervised the literature investigation, statistical analysis, and reviewed the manuscript. All authors read and approved the final manuscript.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study was supported by the Hainan Provincial Natural Science Foundation of China (No. 825RC885).
Conflicting interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.