Development and internal validation of an XGBoost-based prognostic prediction model and risk stratification system for overall survival in pediatric neuroblastoma: a SEER-based study

Introduction

Neuroblastoma (NB) is a heterogeneous malignancy arising from the developing sympathetic nervous system and represents the most common extracranial solid tumor in children (1,2). Although NB accounts for approximately 6–10% of all pediatric cancers, it is responsible for nearly 15% of childhood cancer-related deaths owing to its highly variable biological behavior and propensity for metastatic spread (3-5). Clinical outcomes range from spontaneous regression in some low-risk patients to rapid disease progression and treatment failure in high-risk cases. Despite substantial advances in multimodal treatment strategies, including chemotherapy, surgical resection, radiotherapy, autologous stem cell transplantation, immunotherapy, and targeted therapies, the prognosis of high-risk NB remains unsatisfactory, with long-term survival rates still below 50% in many cohorts (6-8). Accurate risk assessment is therefore essential for treatment selection, prognostic counseling, and follow-up planning. Currently, prognostic evaluation of NB primarily relies on established risk stratification systems, including the International Neuroblastoma Staging System (INSS) and the International Neuroblastoma Risk Group Staging System (INRGSS) (9-11). These systems incorporate important clinical and biological factors and play a central role in therapeutic decision-making. However, they mainly classify patients into broad risk categories and may not provide individualized estimates of survival probability. Furthermore, interactions among demographic, tumor-related, and treatment-related variables are often complex and nonlinear, making individualized prognostic assessment challenging using conventional risk classification approaches alone.

Several prognostic models for NB have been reported in recent years, including nomogram-based models and regression-based survival prediction tools developed from institutional and population-based datasets (12,13). Although these models have demonstrated reasonable predictive performance, many were developed using conventional statistical methods, relied on relatively limited predictor sets, or lacked comprehensive evaluation across different modeling approaches. In addition, few studies have systematically compared multiple machine-learning algorithms for individualized survival prediction in large population-based cohorts of NB patients.

Machine learning has emerged as a promising approach for clinical prediction because of its ability to model complex nonlinear relationships and high-order interactions among predictors (14-17). Compared with traditional regression-based methods, machine-learning algorithms may offer improved flexibility when handling heterogeneous clinical data and have shown encouraging results in prognostic modeling across multiple malignancies (18-21). Among these algorithms, extreme gradient boosting (XGBoost) has attracted considerable attention due to its regularization mechanisms, computational efficiency, and strong predictive performance in oncology-related prediction tasks (18-21). Nevertheless, evidence regarding its utility for individualized survival prediction and risk stratification in NB remains limited.

Therefore, the present study aimed to develop and internally validate prognostic prediction models for overall survival (OS) in patients with NB using a large population-based cohort from the Surveillance, Epidemiology, and End Results (SEER) database. Six machine-learning algorithms, including XGBoost, logistic regression (LR), support vector machine (SVM), random forest (RF), k-nearest neighbor (KNN), and decision tree ID3, were compared to identify the optimal predictive model. In addition, we sought to establish an XGBoost-based risk stratification system to facilitate individualized prognostic assessment. The proposed model is intended as a complementary prognostic tool based on routinely available clinical variables and requires further external validation before clinical implementation. We present this article in accordance with the TRIPOD reporting checklist (available at https://tp.amegroups.com/article/view/10.21037/tp-2026-0337/rc).

Methods

Data collection and study population

This prognostic model development study was conducted using data extracted from the SEER database of the National Cancer Institute. The SEER program collects population-based cancer incidence, treatment, and survival data from multiple cancer registries and currently covers approximately 48% of the U.S. population. Data quality is ensured through standardized coding procedures, routine quality control audits, and systematic follow-up mechanisms. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

SEER*Stat software (version 8.4.4) was used to identify patients diagnosed with NB (ICD-O-3 code: 9500/3) or ganglioneuroblastoma (ICD-O-3 code: 9490/3) between January 2000 and December 2021.

Eligible patients met the following criteria: (I) age ≤18 years; and (II) histopathological confirmation of NB or GNB. Exclusion criteria included: (I) missing tumor size or SEER stage information; (II) unknown surgical treatment status; and (III) survival time <1 month or unknown survival duration.

After applying the eligibility criteria, 2,427 patients were included in the final analysis (Figure 1).

Figure 1 Flowchart of study design and patient selection. AUC, area under the curve; ML, machine learning; ROC, receiving operating characteristic; SEER, Surveillance, Epidemiology, and End Results.

Predictor variables and outcome definition

Candidate predictors included demographic variables (age, sex, race), tumor-related characteristics (histological subtype, primary tumor site, tumor size, tumor grade, distant metastasis, SEER stage, and regional lymph node involvement), and treatment-related variables (surgery, regional lymph node surgery, chemotherapy, and radiotherapy).

The primary outcome was OS, defined as the time from diagnosis to death from any cause or last follow-up. Survival time was obtained from the SEER variable “Survival Months”. Patients alive at the last follow-up were treated as censored observations.

For prognostic prediction, 1-, 3-, and 5-year OS outcomes were evaluated.

Data preprocessing and missing data handling

Data preprocessing followed standard SEER procedures. Patients with missing information regarding tumor size, SEER stage, surgical treatment status, or survival duration were excluded from the analysis. Therefore, a complete-case analysis was performed.

Continuous variables were categorized using X-tile software (version 3.6.1) to determine optimal prognostic cut-off values. Age was categorized into <1, 1 to <2, and ≥2 years, while tumor size was categorized into ≤60, >60–95, and >95 mm.

Model development and internal validation

The study was designed as a prognostic model development study with internal validation. Patients were randomly assigned to a training cohort (80%, n=1,941) and an internal validation cohort (20%, n=486) using a fixed random seed to ensure reproducibility. Feature selection was performed using least absolute shrinkage and selection operator (LASSO)-Cox regression. The optimal penalty parameter (λ) was determined using five-fold cross-validation.

Six machine learning algorithms were implemented to develop predictive models for 1-, 3-, and 5-year OS in NB patients, including XGBoost, LR, SVM, RF, KNN, and ID3. XGBoost is an optimized version of gradient boosting decision trees (GBDT) that incorporates L1 and L2 regularization to prevent overfitting, employs a quantile approximation algorithm for accelerated feature discretization, and enhances efficiency through parallel computing and feature subsampling (22). LR is a generalized linear model widely used for binary classification. It maps linear combinations of input variables onto a probability space using the logistic function, estimating the likelihood of a given sample belonging to a particular class (23). SVM is a supervised learning algorithm that constructs a hyperplane to maximize the margin between different classes. It is particularly effective for small-sample, nonlinear, and high-dimensional data classification, demonstrating strong performance in medical data analysis (24,25). RF is an ensemble learning method that aggregates multiple decision trees through a voting mechanism, enhancing model generalization and stability. It is capable of handling high-dimensional input features and is less prone to overfitting. KNN is an instance-based algorithm that classifies samples by computing their similarity to training samples. It assigns a class label based on a majority vote from the k-nearest neighbors, making it suitable for multi-class classification and non-linear decision boundary problems (26). ID3 constructs decision trees using information gain as the splitting criterion. It recursively selects features with the highest information gain to partition data, producing an interpretable decision process.

Hyperparameter optimization was performed within the training cohort using five-fold cross-validation combined with grid search. Model performance was subsequently evaluated in the independent internal validation cohort. Hyperparameters including n_estimators [100, 200], max_depth [3, 5, 7], learning_rate [0.01, 0.1, 0.2], subsample [0.8, 1.0], and colsample_bytree [0.8, 1.0] were optimized using grid search. The optimal parameter combination identified during cross-validation was subsequently used to construct the final model.

Model performance assessment

Model discrimination was evaluated using time-dependent receiver operating characteristic (ROC) curves, area under the ROC curve (AUC), and concordance index (C-index). Prediction accuracy was assessed using the Brier score. Model performance was evaluated in both the training and internal validation cohorts. Higher AUC and C-index values indicate better discrimination, whereas lower Brier scores indicate improved prediction accuracy. Calibration analysis was performed to assess the agreement between predicted and observed 5-year OS probabilities in the internal validation cohort.

Risk stratification

To further evaluate the discriminatory power of the predictive model, a risk score was computed for each patient. The optimal cut-off value for the risk score was determined, enabling risk stratification into three prognostic groups. Kaplan-Meier survival analysis and log-rank tests were performed to assess differences among risk groups, validating prognostic distinctions between them.

Statistical analysis

Continuous variables (e.g., age, tumor size) were converted into categorical variables using X-tile software (version 3.6.1) to optimize their statistical efficiency (Figure 2). Specifically: age was categorized into three groups: <1, 1 to <2, and ≥2 years. Tumor size was stratified into ≤60, >60–95, and >95 mm. Categorical variables were presented as frequencies (percentages), and comparisons between groups were performed using the chi-square test. A P value <0.05 was considered statistically significant. All data analyses and machine learning modeling were conducted using Python (version 3.12.9).

Figure 2 The optimal cutoff values of tumor size identified by X-tile. (A) The frequency distribution map of age. (B) The Kaplan-Meier curves for the subgroups of age (<1, 1 to <2, ≥2 years) for overall survival. (C) The frequency distribution map of tumor size. (D) The Kaplan-Meier curves for the subgroups of tumor size (≤60, >60–95, >95 mm) for overall survival.

Results

Patient characteristics

A total of 2,427 patients diagnosed with NB or GNB were included in this study. Among them, 53.23% were male, and 76.1% were White. Regarding age distribution, 47.80% of patients were aged ≥2 years. The adrenal gland was the most common primary tumor site, accounting for 50.56% of cases. At the time of diagnosis, 46.97% of patients presented with distant metastasis. In terms of treatment, 80.84% of patients underwent surgery (either radical or local resection), 65.47% received chemotherapy, and 24.43% underwent radiotherapy. All cases were randomly divided into a training set (n=1,941) and a validation set (n=486) in an 8:2 ratio. There were no significant differences in baseline characteristics between the two groups (P>0.05). Table 1 summarizes the detailed characteristics of the patient cohorts.

LASSO regression and variable selection

Feature selection was performed using LASSO regression. Figure 3A illustrates the coefficient paths of predictor variables at different log(λ) values. As λ increased, the coefficients of non-essential variables were gradually shrunk to zero, indicating a reduced contribution to the model. Figure 3B shows that the partial likelihood deviance increased as λ decreased, suggesting a risk of overfitting. Based on the λ value corresponding to the minimum partial likelihood deviance, six key prognostic variables were ultimately selected: age, primary tumor site, tumor size, distant metastasis, SEER stage, and chemotherapy.

Figure 3 The results of the LASSO regression. (A) LASSO coefficient profiles; (B) cross-validation curve of LASSO regression. LASSO, least absolute shrinkage and selection operator.

Predictive models development and performance evaluation

The time-dependent AUC of the six machine learning models in the training and validation sets is presented in Figure 4, with detailed predictive performance metrics summarized in Table 2. During the training phase, the RF model achieved the highest AUC values at all time points and demonstrated a lower Brier score, indicating strong discrimination between events and non-events and relatively accurate probability estimation. Additionally, its C-Index [95% confidence interval (CI)] was 0.806 (0.787–0.825), further confirming its strong predictive ability. However, in the independent validation set, the XGBoost model exhibited superior stability and generalization capability, achieving AUCs of 0.767, 0.796, and 0.807 for 1-, 3-, and 5-year survival predictions, respectively. The C-Index was 0.784 (95% CI: 0.746–0.821), significantly outperforming the other models. This indicates that XGBoost not only fit the training data well but also maintained robust predictive performance on unseen data. To further evaluate the reliability of predicted survival probabilities, calibration analyses were performed for the XGBoost model. As shown in Figure 5, the calibration curves for 5-year overall survival prediction demonstrated good agreement between the predicted probabilities and observed outcomes in both the training cohort (Brier score =0.1056) and the internal validation cohort (Brier score =0.1172). The calibration plots indicated that the XGBoost model provided reasonably accurate probability estimates, supporting its potential utility for individualized prognostic assessment.

Figure 4 ROC curves and AUCs demonstrating the predictions of the six models: XGBoost, LR, SVM, RF, KNN, and ID3. (A,C,E,G,I,K) The training cohort. (B,D,F,H,J,L) The validation cohort. AUC, area under the curve; ID3, Iterative Dichotomiser 3; KNN, K-nearest neighbor; LR, logistic regression; RF, random forest; ROC, receiving operating characteristic; SVM, support vector machine; XGBoost, extreme gradient boosting.

Figure 5 Calibration curve of the XGBoost model for predicting 5-year overall survival in the training cohort (A) and validation cohort (B). OS, overall survival; XGBoost, extreme gradient boosting.

Considering model complexity, interpretability, and practical applicability, although the RF model demonstrated a slight advantage in the training phase, XGBoost was ultimately recommended as the optimal predictive model in this study due to its better generalization ability and stability.

Feature importance of the XGBoost model

The feature importance analysis of the XGBoost model (Figure 6) revealed that “metastatic status” contributed the most to survival prediction, followed by “SEER stage”, “age”, “chemotherapy”, “tumor size”, and “primary tumor site”, in decreasing order of importance. This finding aligns with the clinical consensus that metastatic disease is an independent adverse prognostic factor.

Figure 6 The XGBoost model was used to calculate the importance of six features. SEER, Surveillance, Epidemiology, and End Results; XGBoost, extreme gradient boosting.

Risk stratification

The XGBoost model was further employed for risk stratification, categorizing patients into high-risk, intermediate-risk, and low-risk groups. Based on optimal survival differences, two risk score thresholds were identified at 0.32 and 0.16 (Figure 7A), which were used to classify patients in both the training and validation cohorts. Kaplan-Meier survival analysis demonstrated that 5-year survival rates in the high-risk group were significantly lower than those in the intermediate- and low-risk groups. The log-rank test confirmed that the differences among the risk groups were statistically significant (Figure 7B,7C).

Figure 7 The XGBoost model-based risk stratification. Risk score distribution (A); Kaplan-Meier analysis was performed on high-, medium- and low-risk groups within both the training (B) and testing (C) cohorts. XGBoost, extreme gradient boosting.

Discussion

In this study, we identified six readily available clinical variables associated with OS in patients with NB, including metastatic status, SEER stage, age, chemotherapy, tumor size, and primary tumor site. Based on these variables, six machine-learning algorithms were developed and compared for prognostic prediction. Among the evaluated models, XGBoost achieved the best overall predictive performance in both the development and internal validation cohorts. Feature importance analysis further identified metastatic status as the most influential predictor. Using the predicted risk scores generated by the XGBoost model, we established a risk stratification system that categorized patients into low-, intermediate-, and high-risk groups. These findings suggest that routinely available clinical variables may provide useful prognostic information and support individualized survival assessment in pediatric NB.

Over the past decade, improving prognostic assessment to refine clinical decision-making for cancer patients has been a major research focus (27). However, traditional methods for predicting survival outcomes based on single prognostic factors face challenges such as limited accuracy and the complexity of integrating multiple factors. Given the significant heterogeneity in NB prognosis and the inadequacy of current tumor staging systems in providing precise, personalized predictions, developing widely applicable individualized models is of great clinical significance. Although Zhuo et al. (12) developed a nomogram model and risk stratification system for predicting OS in NB patients, and Amano et al. (13) integrated urinary catecholamine metabolites (HVA and VMA) to construct a risk assessment system for NB diagnosis and pre-treatment evaluation, these models primarily rely on traditional statistical methods and suffer from difficulties in clinical feature collection and poor generalizability. In recent years, machine learning-based artificial intelligence models have gained increasing attention in clinical practice, emerging as valuable tools for developing predictive models and significantly enhancing disease outcome predictions. Recent studies have examined XGBoost’s application in predicting cancer survival and validated its robust predictive performance across various cancers, such as elderly intrahepatic cholangiocarcinoma and breast cancer with brain metastases (28). In this study, among the six machine learning models, XGBoost demonstrated the most outstanding prognostic prediction performance. Nevertheless, although the model achieved acceptable predictive performance, its results should be interpreted in the context of internal validation only, and further external validation is required before broader clinical implementation can be considered.

To systematically evaluate XGBoost’s feature importance and enhance model interpretability, we ranked the predictive power of individual features. The results showed that metastatic status was the most predictive feature, followed by SEER stage, age, chemotherapy, tumor size, and primary tumor site. These findings align with previous research. Metastatic status is widely recognized as an independent prognostic factor, significantly negatively correlated with OS and cancer-specific survival (CSS) (29). Clinical data suggest that adrenal-origin NBs exhibit a higher tendency for metastasis, with bone and liver metastases being the most common single-organ metastatic patterns (30,31). However, current treatment strategies provide limited improvements in 5-year survival rates for patients with distant metastasis (32). leading to a poorer prognosis for these patients. SEER staging is a classification method used in the SEER program to describe the extent and severity of cancer, categorizing it into localized, regional, and distant stages. It aids clinicians in predicting patient survival and disease progression. In our study, SEER stage also demonstrated high predictive utility for NB prognosis within the XGBoost model. Age is another critical prognostic factor in NB. For stage 3 and 4 MYCN non-amplified tumors, patients younger than 18 months exhibit better event-free survival (EFS) than those older than 18 months (33). Chemotherapy was also identified as a prognostic factor through LASSO regression, potentially due to treatment-related mortality events (12). Traditionally, prognostic factors in NB are combined to stratify patients’ risk levels and guide treatment decisions. However, major classification systems such as the Children’s Oncology Group (COG) and the International Neuroblastoma Risk Group (INRG) do not include tumor size as a prognostic factor. Nevertheless, recent studies suggest that tumor size plays a crucial role in NB prognosis (34). Additionally, previous studies have indicated that primary tumor site significantly influences prognosis. Adrenal-origin tumors are associated with lower survival rates, whereas thoracic-origin tumors show higher survival rates (31).

To further explore the clinical relevance of the proposed model, we developed an XGBoost-based risk stratification system that classified patients into distinct prognostic groups. The model successfully differentiated survival outcomes in both the development and internal validation cohorts. Importantly, unlike established classification systems such as the INRG system, which primarily categorizes patients into broad risk groups for treatment allocation, the present model was designed to estimate individualized survival probabilities using routinely available clinical variables. Therefore, the proposed model should be considered a complementary prognostic tool rather than a replacement for existing biologically driven risk-classification systems. If validated in independent cohorts, such individualized risk estimates may provide additional information for prognostic assessment, follow-up planning, and patient counseling.

Several limitations should be acknowledged. First, this was a retrospective registry-based study and is therefore subject to potential selection bias, residual confounding, and misclassification of recorded variables. Second, although the SEER database provides a large population-based cohort, several important biological and molecular prognostic factors were unavailable, including MYCN amplification status, chromosome 11q aberration, DNA ploidy, histopathological classification, and urinary catecholamine metabolites. The absence of these variables may have limited model performance and clinical applicability. Third, complete-case analysis was performed, which may have introduced additional selection bias. Fourth, model development and validation were conducted using a random split of the same SEER dataset; therefore, the present study represents internal validation rather than true external validation. Although calibration and discrimination analyses demonstrated acceptable predictive performance, the possibility of model optimism and overfitting cannot be completely excluded. Future studies incorporating molecular, genetic, and clinical variables and validating the model in independent multicenter cohorts are needed before routine clinical application can be recommended.

Conclusions

In summary, we developed and internally validated six machine-learning models for predicting OS in pediatric NB using a population-based SEER cohort. Among these models, XGBoost demonstrated the best predictive performance and was subsequently used to establish a risk stratification system. The proposed model may facilitate individualized prognostic assessment based on routinely available clinical variables. However, given the retrospective study design, the absence of important biological predictors, and the lack of external validation, further evaluation in independent cohorts is required before clinical implementation.

Acknowledgments

We thank the SEER database for its open data access.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tp.amegroups.com/article/view/10.21037/tp-2026-0337/rc

Peer Review File: Available at https://tp.amegroups.com/article/view/10.21037/tp-2026-0337/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tp.amegroups.com/article/view/10.21037/tp-2026-0337/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Maris JM. Recent advances in neuroblastoma. N Engl J Med 2010;362:2202-11. [Crossref] [PubMed]
2. Mlakar V, Jurkovic Mlakar S, Lopez G, et al. 11q deletion in neuroblastoma: a review of biological and clinical implications. Mol Cancer 2017;16:114. [Crossref] [PubMed]
3. Tonini GP, Capasso M. Genetic predisposition and chromosome instability in neuroblastoma. Cancer Metastasis Rev 2020;39:275-85. [Crossref] [PubMed]
4. Gatta G, Botta L, Rossi S, et al. Childhood cancer survival in Europe 1999-2007: results of EUROCARE-5--a population-based study. Lancet Oncol 2014;15:35-47. [Crossref] [PubMed]
5. Lu W, Wang Q, Xu C, et al. SUMOylation is essential for Sirt2 tumor-suppressor function in neuroblastoma. Neoplasia 2021;23:129-39. [Crossref] [PubMed]
6. Croteau N, Nuchtern J, LaQuaglia MP. Management of Neuroblastoma in Pediatric Patients. Surg Oncol Clin N Am 2021;30:291-304. [Crossref] [PubMed]
7. Chung C, Boterberg T, Lucas J, et al. Neuroblastoma. Pediatr Blood Cancer 2021;68:e28473. [Crossref] [PubMed]
8. Park JR, Kreissman SG, London WB, et al. Effect of Tandem Autologous Stem Cell Transplant vs Single Transplant on Event-Free Survival in Patients With High-Risk Neuroblastoma: A Randomized Clinical Trial. JAMA 2019;322:746-55. [Crossref] [PubMed]
9. Irwin MS, Park JR. Neuroblastoma: paradigm for precision medicine. Pediatr Clin North Am 2015;62:225-56. [Crossref] [PubMed]
10. Brodeur GM, Pritchard J, Berthold F, et al. Revisions of the international criteria for neuroblastoma diagnosis, staging, and response to treatment. J Clin Oncol 1993;11:1466-77. [Crossref] [PubMed]
11. Monclair T, Brodeur GM, Ambros PF, et al. The International Neuroblastoma Risk Group (INRG) staging system: an INRG Task Force report. J Clin Oncol 2009;27:298-303. [Crossref] [PubMed]
12. Zhuo X, Xia L, Tang W, et al. A practical nomogram and risk stratification system for predicting survival outcomes in neuroblastoma patients: a SEER population-based study. J Cancer Res Clin Oncol 2023;149:12285-96. [Crossref] [PubMed]
13. Amano H, Uchida H, Harada K, et al. Scoring system for diagnosis and pretreatment risk assessment of neuroblastoma using urinary biomarker combinations. Cancer Sci 2024;115:1634-45. [Crossref] [PubMed]
14. Tran KA, Kondrashova O, Bradley A, et al. Deep learning in cancer diagnosis, prognosis and treatment selection. Genome Med 2021;13:152. [Crossref] [PubMed]
15. Nguyen TT, Ho CT, Bui HTT, et al. Multidimensional Machine Learning for Assessing Parameters Associated With COVID-19 in Vietnam: Validation Study. JMIR Form Res 2023;7:e42895. [Crossref] [PubMed]
16. Sajda P. Machine learning for detection and diagnosis of disease. Annu Rev Biomed Eng 2006;8:537-65. [Crossref] [PubMed]
17. Senders JT, Staples P, Mehrtash A, et al. An Online Calculator for the Prediction of Survival in Glioblastoma Patients Using Classical Statistics and Machine Learning. Neurosurgery 2020;86:E184-92. [Crossref] [PubMed]
18. Xu Q, Lu X. Development and validation of an XGBoost model to predict 5-year survival in elderly patients with intrahepatic cholangiocarcinoma after surgery: a SEER-based study. J Gastrointest Oncol 2022;13:3290-9. [Crossref] [PubMed]
19. Li C, Liu M, Zhang Y, et al. Novel models by machine learning to predict prognosis of breast cancer brain metastases. J Transl Med 2023;21:404. [Crossref] [PubMed]
20. Zhong X, Lin Y, Zhang W, et al. Predicting diagnosis and survival of bone metastasis in breast cancer using machine learning. Sci Rep 2023;13:18301. [Crossref] [PubMed]
21. Kinoshita F, Takenaka T, Yamashita T, et al. Development of artificial intelligence prognostic model for surgically resected non-small cell lung cancer. Sci Rep 2023;13:15683. [Crossref] [PubMed]
22. Takefuji Y. Beyond XGBoost and SHAP: Unveiling true feature importance. J Hazard Mater 2025;488:137382. [Crossref] [PubMed]
23. Zuo W, Yang X. A machine learning model predicts stroke associated with blood cadmium level. Sci Rep 2024;14:14739. [Crossref] [PubMed]
24. Sayeed MA, Rahman A, Rahman A, et al. On the interpretability of the SVM model for predicting infant mortality in Bangladesh. J Health Popul Nutr 2024;43:170. [Crossref] [PubMed]
25. Fan R, Yu N, Li G, et al. Machine-learning model comprising five clinical indices and liver stiffness measurement can accurately identify MASLD-related liver fibrosis. Liver Int 2024;44:749-59. [Crossref] [PubMed]
26. Shrivastava T, Singh V, Agrawal A. Autism spectrum disorder detection with kNN imputer and machine learning classifiers via questionnaire mode of screening. Health Inf Sci Syst 2024;12:18. [Crossref] [PubMed]
27. Staffa SJ, Zurakowski D. Statistical Development and Validation of Clinical Prediction Models. Anesthesiology 2021;135:396-405. [Crossref] [PubMed]
28. Dong B, Zhang H, Duan Y, et al. Development of a machine learning-based model to predict prognosis of alpha-fetoprotein-positive hepatocellular carcinoma. J Transl Med 2024;22:455. [Crossref] [PubMed]
29. Liu S, Yin W, Lin Y, et al. Metastasis pattern and prognosis in children with neuroblastoma. World J Surg Oncol 2023;21:130. [Crossref] [PubMed]
30. Oldridge DA, Truong B, Russ D, et al. Differences in Genomic Profiles and Outcomes Between Thoracic and Adrenal Neuroblastoma. J Natl Cancer Inst 2019;111:1192-201. [Crossref] [PubMed]
31. Vo KT, Matthay KK, Neuhaus J, et al. Clinical, biologic, and prognostic differences on the basis of primary tumor site in neuroblastoma: a report from the international neuroblastoma risk group project. J Clin Oncol 2014;32:3169-76. [Crossref] [PubMed]
32. Morgenstern DA, Bagatell R, Cohn SL, et al. The challenge of defining “ultra-high-risk” neuroblastoma. Pediatr Blood Cancer 2019;66:e27556. [Crossref] [PubMed]
33. Sokol E, Desai AV, Applebaum MA, et al. Age, Diagnostic Category, Tumor Grade, and Mitosis-Karyorrhexis Index Are Independently Prognostic in Neuroblastoma: An INRG Project. J Clin Oncol 2020;38:1906-18. [Crossref] [PubMed]
34. Wang JX, Cao ZY, Wang CX, et al. Prognostic impact of tumor size on patients with neuroblastoma in a SEER-based study. Cancer Med 2022;11:2779-89. [Crossref] [PubMed]