Ewing's sarcoma (ES) is the second most common malignant bone tumor in children, accounting for approximately 3% of all pediatric malignancies. It presents as a highly anaplastic, small round blue cell neoplasm arising primarily from the intramedullary portion of bone, though extraskeletal variants also occur. The disease disproportionately affects young patients, with 80% of cases diagnosed in individuals under 20 years old, and shows a male predominance with a 1.5:1 sex ratio. Its hallmark molecular feature is the EWS-FLI1 fusion protein, a product of the t(11;22) chromosomal translocation, which drives transcriptional dysregulation and promotes tumor progression.
The prognostic weight of lymph node involvement: Lymph node metastasis (LNM) in ES is an underappreciated but clinically significant event. Historically, ES has been described as spreading primarily hematogenously to the lungs and bones, with lymph node involvement considered relatively rare. However, when LNM does occur, the consequences are severe: the 5-year overall survival rate drops from 60.3% in node-negative patients to 45.9% in those with confirmed lymph node involvement. This 14.4 percentage-point difference in survival is clinically meaningful, yet there has been no validated tool for prospectively identifying patients at greatest risk of LNM before or at the time of diagnosis.
The gap this study fills: Predictive models for lymph node involvement have been developed in thyroid papillary carcinoma, bladder urothelial carcinoma, and early esophageal squamous cell carcinoma, but no equivalent model existed for ES prior to this work. The authors leveraged the Surveillance, Epidemiology, and End Results (SEER) database combined with a four-institution Chinese external cohort to build and externally validate six machine learning models, ultimately selecting the best-performing algorithm and deploying it as a publicly accessible web-based clinical calculator.
The study is grounded in the practical reality that LNM assessment in ES is often incomplete or overlooked in clinical staging, and that more careful lymph node evaluation, informed by a risk prediction tool, could improve treatment planning, guide diagnostic workup intensity, and support enrollment in clinical trials targeting high-risk patients.
The training cohort was drawn from SEER, a population-based cancer registry maintained by the National Cancer Institute that covers approximately 30% of the US population. Patients diagnosed with ES from 2010 to 2016 were extracted using SEER*Stat (version 8.3.5) software, yielding 923 eligible patients. Inclusion required pathologically confirmed primary ES with ICD-O-3/WHO 2008 morphology code 60, and complete clinicopathological and survival data. Patients with other concurrent primary tumors or missing key variables were excluded.
External validation cohort: The validation set comprised 51 ES patients recruited retrospectively from four independent Chinese institutions: the Second Affiliated Hospital of Jilin University, the Second Affiliated Hospital of Dalian Medical University, Liuzhou People's Hospital, and Xianyang Central Hospital. These patients were diagnosed between 2010 and 2018. Two investigators independently extracted and processed validation data, with a third investigator adjudicating any discrepancies to ensure consistency.
Variables collected: Fourteen clinicopathological variables were systematically recorded for both cohorts: race, age, sex, primary site (axial bone, limb bone, or other), laterality, T stage, M stage, surgical treatment (yes or no), radiation therapy, chemotherapy status, bone metastases, lung metastases, lymph node metastasis status, and survival time. LNM status was categorized as confirmed present or unable to evaluate versus confirmed absent, following SEER coding conventions.
Statistical framework: Continuous variables were reported as mean and standard deviation (SD) for normally distributed data. Categorical variables were described as counts and proportions. Univariate logistic regression was run across all 14 variables, and those reaching significance at P less than 0.05 entered multivariate logistic regression to identify independent predictors. The five variables identified by univariate analysis served as inputs to all six machine learning models. All statistical analyses used R version 4.0.5; ML algorithms and the web application were implemented in Python.
Of the 974 total ES patients in the combined cohort, 13.86% (135 of 974) had confirmed or unevaluable lymph node metastasis, with 128 of these from the SEER training set and 7 from the external validation cohort. The median age across both sets was 17 years (IQR 12-27), consistent with the known pediatric and young-adult predominance of ES. Male patients accounted for 57.4% of the combined cohort. In the training set, the racial distribution was 81.8% White, 8.1% other (primarily Asian), and 4.1% Black, while the validation set was entirely composed of patients coded as "other" (100% Asian), reflecting the Chinese institution sources.
Comparing sets: Race (P less than 0.001) and radiation history (P = 0.001) were the only two baseline variables that differed significantly between training and validation cohorts. Radiation therapy was more common in the validation set (43.1% vs. 21.7%), likely reflecting institutional practice patterns. All other variables, including age, sex, primary site, T stage, M stage, surgery rate, chemotherapy use, bone metastases, and lung metastases, were statistically comparable between the two cohorts, supporting the validity of using the SEER set for model training and the Chinese multicenter set for external validation.
Characteristics of LNM-positive patients: When comparing patients with and without LNM across the full cohort, several variables showed significant differences. T stage distribution differed markedly (P less than 0.001): among LNM-positive patients, only 18.5% had T1 disease versus 38.1% in LNM-negative patients, while the TX (unclassified) category was 28.1% in LNM-positive versus 14.7% in LNM-negative patients. M stage was strongly associated: 58.5% of LNM-positive patients had M1 disease versus 27.9% of LNM-negative patients (P less than 0.001). Lung metastases were present in 41.5% of LNM-positive patients compared to 15.1% of LNM-negative patients (P less than 0.001). Surgery rates were lower in LNM-positive patients (46.7% vs. 60.1%, P = 0.005). Median survival was substantially worse in LNM-positive patients: 22.27 months versus 31.99 months (P less than 0.001).
Univariate logistic regression identified five variables with P less than 0.05: race, T stage, M stage, surgery, and lung metastases. These five variables were carried forward into multivariate logistic regression to determine which remained independently predictive after mutual adjustment. Two variables that appeared significant in univariate analysis, specifically surgery, did not retain significance in the multivariate model (OR = 1.127, 95% CI 0.738-1.721, P = 0.581), suggesting its univariate association with LNM was confounded by other disease severity markers rather than representing a true independent predictor.
Four independent predictors: The multivariate analysis confirmed four independent predictors of LNM in ES. Race (Black vs. White) carried an OR of 2.270 (95% CI 1.020-5.052, P = 0.045), indicating approximately 2.3-fold higher odds of LNM in Black patients compared to White patients. T stage showed a graded relationship: T2 versus T1 carried OR = 1.733 (95% CI 1.044-2.876, P = 0.033), while TX versus T1 was even stronger at OR = 2.712 (95% CI 1.511-4.870, P = 0.001), suggesting that tumors too advanced or poorly characterized to be staged carry the highest risk. M1 disease conferred OR = 2.038 (95% CI 1.157-3.591, P = 0.014), and the presence of lung metastases carried OR = 1.877 (95% CI 1.067-3.301, P = 0.029).
Clinical interpretation of the predictors: The finding that M stage was the strongest individual predictor aligns with the known biology of ES: 15-30% of patients present with distant metastases at diagnosis, and once distant spread is established, lymphatic involvement is substantially more likely. The T stage relationship is consistent with observations in other bone sarcoma subtypes: larger tumors (T2, T3) and tumors that have extended to periosteum or surrounding soft tissue can access regional lymphatics, which are absent in normal bone but present in periosteal and soft tissue compartments. The lung-LNM correlation likely reflects a bidirectional relationship where systemic hematogenous spread and lymphatic spread co-occur as markers of broadly aggressive tumor biology.
Race as a predictor: The finding that Black race carried 2.27-fold higher odds of LNM is consistent with documented racial disparities in ES outcomes. Prior data show that Black ES patients have lower 10-year survival rates and higher rates of metastatic disease at diagnosis compared to White patients. Hispanic patients more frequently present with tumors larger than 10 cm, and tumor size is itself associated with lymphatic spread, creating interlinked biological and social factors that influence LNM risk.
Six machine learning algorithms were trained on the SEER cohort using the five variables identified by univariate logistic regression (race, T stage, M stage, surgery, and lung metastases) as inputs and LNM status as the outcome. The algorithms evaluated were: random forest (RF), naive Bayes classifier (NBC), decision tree (DT), XGBoost (XGB), gradient boosting machine (GBM), and logistic regression (LR). This selection covers a range of model families, from simple interpretable classifiers (DT, LR) to ensemble methods that aggregate many weak learners into stronger predictors (RF, XGB, GBM), and a probabilistic approach (NBC).
Internal validation performance: 10-fold cross-validation was applied within the SEER training set to estimate each model's generalizability. In 10-fold cross-validation, the training data are divided into 10 equal subsets; each fold is held out as a mini-test set in turn while the model trains on the remaining nine, and performance is averaged across all ten iterations. The average AUC values across the six models ranged from 0.705 to 0.764. The RF model achieved the highest average AUC of 0.764 (standard deviation = 0.034), outperforming XGB (the second-best ensemble method) and all other classifiers in cross-validation.
External validation performance: ROC curve analysis was applied to the 51-patient multicenter Chinese validation cohort, providing an independent test of real-world generalizability. AUC values in external validation ranged from 0.612 to 0.727 across the six algorithms. Again, the RF model performed best with an AUC of 0.727. The performance drop from internal cross-validation (AUC 0.764) to external validation (AUC 0.727) represents a reduction of approximately 0.037, which is modest and well within acceptable bounds for clinical prediction models, especially given the substantial demographic differences between the predominantly White SEER cohort and the entirely Asian validation cohort.
Variable importance across models: Across all six algorithms, T stage, M stage, and lung metastases consistently ranked as the top three predictors. Race and surgery ranked lower. In the RF model specifically, the importance ranking from highest to lowest was: M stage, T stage, lung metastases, surgery, race. This consistent pattern across all six model types strengthens confidence that these clinical variables represent genuine biological drivers of LNM risk rather than artifacts of any single algorithm's structure.
A core contribution of this study is the deployment of the best-performing random forest model as a publicly accessible web-based calculator built using Streamlit, a Python-based web application framework. The calculator accepts five clinical inputs corresponding to the variables in the RF model: race, T stage, M stage, presence of lung metastases, and surgical status. Clinicians enter these values for an individual ES patient and receive an estimated probability of lymph node metastasis.
Design rationale: The web calculator addresses a recognized gap in how AI models transition from research publications to bedside utility. Many published ML models remain locked in academic papers with no accessible implementation, making clinical adoption essentially impossible. By converting the RF model into a point-and-click interface requiring no programming expertise or specialized software, the authors explicitly designed this tool for integration into routine clinical workflows at the time of initial staging and workup.
Intended clinical workflow: At the time of ES diagnosis, clinicians assemble staging information including tumor size (driving T stage assignment), distant metastasis status (M stage), and imaging evidence of lung metastases, all of which are obtained through standard staging CT and PET scans. These values, combined with race and surgical treatment decision, can be entered into the calculator to generate a patient-specific LNM risk probability. High-risk outputs could prompt more thorough regional lymph node evaluation, including fine needle aspiration cytology (FNAC) of suspicious nodes, sentinel lymph node biopsy (SLNB) where feasible, or more intensive FDG-PET evaluation of the regional nodal basins.
The authors note several diagnostic approaches for evaluating suspected LNM in ES. FDG-PET scanning can identify involved nodes but has reduced reliability for small-volume nodal disease, with possible false positives from benign inflammatory processes. SLNB offers targeted sampling but loses accuracy in previously irradiated fields where lymphatic channels may be distorted. A novel ES-specific probe, CS2-N-E9R, targeting the E/F fusion protein, has shown early promise for sensitive and selective identification of LNM but remains investigational.
Underestimation of LNM prevalence in SEER: The authors explicitly acknowledge that the 13.86% LNM rate in the SEER database likely underestimates true prevalence. SEER records lymph node status as it was clinically evaluated and documented, but ES staging does not universally require systematic lymph node dissection or biopsy. Many patients classified as node-negative may simply have had inadequate nodal evaluation, meaning the model was trained on data where the outcome itself is subject to ascertainment bias. A more rigorous staging protocol across the training cohort could have yielded a higher true LNM rate and potentially shifted the model's decision boundaries.
Missing treatment detail: The SEER database does not capture specific chemotherapy drug regimens, dosing intensity, or radiotherapy dose. ES treatment typically involves multi-drug regimens such as VDC/IE (vincristine, doxorubicin, cyclophosphamide alternating with ifosfamide and etoposide), and response to induction chemotherapy strongly influences tumor biology and metastatic behavior. The absence of treatment response data means the model cannot account for how neoadjuvant therapy modifies the tumor microenvironment and potentially alters LNM risk at the time of surgical evaluation.
No radiomics or imaging features: The study relies exclusively on clinicopathological variables. The authors acknowledge that incorporating imaging-derived features, specifically radiomics extracted from CT or MRI of the primary tumor, could substantially improve predictive power. Radiomic features capturing tumor heterogeneity, shape irregularity, and peritumoral infiltration have been validated in other bone tumors as predictors of lymphatic and distant metastasis. Integration of such features would require more complex data pipelines but could meaningfully raise the model's discriminative ability above AUC 0.727.
Validation cohort homogeneity and size: The external validation set comprised only 51 patients, all from Chinese institutions, creating two limitations. First, the small size limits the statistical precision of the external AUC estimate. Second, the entirely Asian composition means the model's performance has not been tested in a racially diverse external cohort, even though race was identified as an independent predictor. Future multi-ethnic, prospective external validation across multiple healthcare systems is explicitly recommended by the authors before clinical deployment at scale.
The practical implication of this model, if validated prospectively, is a shift in how ES staging is approached at the time of initial diagnosis. Currently, regional lymph node evaluation in ES lacks a standardized protocol: some centers perform systematic biopsy of clinically suspicious nodes, others rely on FDG-PET, and many document lymph node status based on imaging alone without pathological confirmation. A risk stratification tool that outputs a quantitative probability of LNM for each individual patient could create a logical framework for deciding which patients warrant more intensive nodal evaluation, thereby reducing both under-staging in high-risk patients and unnecessary invasive procedures in low-risk patients.
Sentinel lymph node biopsy considerations: The authors advocate for SLNB as a targeted diagnostic intervention for patients identified as high-risk by the model, though they note an important practical caveat: SLNB accuracy is degraded in previously irradiated fields because distorted lymphatic channels may lead to spurious sentinel node identification that bypasses the true draining node. Given that radiation is commonly used in ES treatment, planning SLNB before rather than after radiotherapy, in patients flagged as high-risk, could maximize diagnostic yield.
Integration with molecular markers: The current model relies on anatomical staging variables that are already available at diagnosis. A logical next iteration would integrate molecular markers, particularly EWS-FLI1 fusion transcript levels, MMR pathway gene expression profiles (which have been implicated in ES invasion and migration), and circulating tumor DNA (ctDNA), into a multimodal prediction framework. ctDNA approaches are increasingly feasible in pediatric sarcomas and could provide dynamic risk updating across the treatment course rather than a single baseline prediction.
Multimodal radiomics integration: Incorporating CT or MRI radiomics from the primary tumor is the most technically accessible next step that could substantially improve the model. Texture features reflecting tumor heterogeneity and morphological features capturing periosteal extension, both of which are mechanistically linked to lymphatic access, are already being extracted in other bone tumor studies and could be retrospectively computed from existing imaging archives. A combined clinico-radiomic RF model, validated prospectively across multiple ethnic populations and healthcare systems, would represent the highest-value near-term development path for this line of research.