Methods

Study design and sample

This study used a retrospective cohort from the U.S. National Cancer Institute’s Surveillance, Epidemiology, and End Results (SEER-18) registries spanning 2000–2018. The SEER database provides comprehensive, population-based data on cancer incidence and survival, covering about 34.6% of the U.S. population [28]. From an initial dataset of roughly 6.9 million records, eligibility criteria identified cases with complete demographic, clinical, and treatment information. Patients with primary cardiac tumors were identified using the SEER Cause of Death (COD) to Site Recode variable (ICD-O-3/WHO 2008) under the category “Soft Tissue Including Heart.” Of 55,261 eligible cases, further restriction to the SEER Rare Tumors List (subsection 50.5), “Soft Tissue Sarcoma of the Heart,” yielded the target cohort of 422 patients. After excluding cases with missing or incomplete key variables, the final analytic sample included 416 patients with histologically confirmed primary cardiac malignancies (Fig. 1B). Institutional Review Board (IRB) approval was not required, as this study used de-identified SEER data.

Fig. 1: Study design and sample selection process. Panel A illustrates the competing-risk framework and Panel B depicts the process to derive the analytical sample

Primary outcomes

The primary outcomes of interest were overall survival (OS) and cause-specific mortality (CSM). Overall survival was defined as death from any cause during the follow-up period, whereas cause-specific mortality represented death directly attributable to cardiac sarcoma. Deaths from other causes were treated as competing events to account for their potential influence on outcome estimation. Although SEER derives cause-of-death information from death certificates using standardized coding procedures, misclassification may occur, particularly when distinguishing cancer-specific mortality in complex clinical scenarios. This competing-risk framework provided a comprehensive evaluation of patient survival and treatment effects while appropriately distinguishing disease-specific mortality from unrelated deaths in this rare and aggressive malignancy (Fig. 1A). Details of data and variables can be found from SEER website (https://seer.cancer.gov/).

Key predictors

The primary predictors in this study were surgery, radiotherapy, chemotherapy, and their various combinations. Each treatment modality was coded as a binary variable: “yes” if the patient received the respective treatment in any form and “no” if the patient did not receive any.

Patients were categorized as having received surgery if they underwent any form of tumor resection regardless of the extent of operation. Chemotherapy was classified as “yes” if systemic anticancer treatment, including neoadjuvant, adjuvant, or palliative, was administered. In the SEER data, chemotherapy is originally coded as “yes” versus “no/Unknown,” without details on agents, dose, cycles, or timing; accordingly, we classified chemotherapy as “yes” if recorded as administered and “no” otherwise. For radiotherapy, although multiple categories exist (e.g., beam radiation, radiation not otherwise specified, none/unknown, refused, recommended but unknown if administered), they do not consistently capture dose, fractionation, modality, or treatment intent, and some categories have sparse counts. To ensure analytic stability, and address sparse data, radiotherapy was also dichotomized as “yes” (any form of radiation recorded) versus “no” (none/unknown or refused). These definitions are consistent with prior SEER-based oncology studies of rare cancers [3, 29].

Combinations of these modalities, such as surgery plus chemotherapy, surgery plus radiotherapy, or all three defined as triple therapy, were included to assess potential synergistic or additive effects on survival. This binary and combinatorial classification facilitated comprehensive modeling of treatment-related predictors in predictive analyses.

Covariates

Explanatory variables were selected based on clinical relevance and prior evidence from cardiac cancer research. Age at diagnosis was treated as a continuous variable to capture its linear relationship with survival outcomes. Sex was categorized as male or female. Race was grouped into three categories (White, Black, and Other), with the “Other” category combining all remaining racial groups due to their small counts. Key clinical variables included cancer stage at presentation (localized, regional, and distant), reflecting disease progression and extent of metastasis. Tumor histology, particularly angiosarcoma (yes/no), was included as a distinct predictor given its aggressive biological behavior and poor prognosis compared with other cardiac tumor subtypes. These variables were included in both statistical and machine learning models to evaluate their independent and joint effects on survival outcomes.

Analytic approaches

Statistical analysis

Descriptive statistics summarized demographic, clinical, and treatment characteristics. Categorical variables were reported as frequencies and percentages, and continuous variables as medians with interquartile ranges (IQRs). Overall survival was estimated using the Kaplan–Meier method and compared across groups with log-rank tests. Within the competing-risk framework, primary cardiac malignancy-specific mortality was the primary outcome, and deaths from other causes were treated as competing events (Fig. 1A). Cumulative incidence functions (CIFs) estimated cause-specific and other-cause mortality, and Fine–Gray sub-distribution hazard models evaluated associations between treatment modalities and mortality outcomes, adjusting for demographic and clinical covariates. Results were expressed as adjusted hazard ratios (AHRs) with 95% confidence intervals (CIs), with p < 0.05 indicating statistical significance. Given the observational nature of SEER data, analyses were designed to estimate associations rather than causal effects, recognizing the potential influence of selection bias and unmeasured confounding.

Machine learning analysis

To enhance prediction accuracy and capture complex interactions, competing-risk random forest (RF) and random survival forest (RSF) models were applied [30–32]. These tree-based ensemble methods model nonlinear effects and estimate variable importance [31]. Survival regression trees were further developed to classify patients into distinct risk subgroups based on clinical and treatment characteristics, providing interpretable frameworks for individualized survival prediction [33].

Validation and performance evaluation

Data was randomly divided into training (70%) and testing (30%) sets. Models were developed using the training data and evaluated on the testing data. Performance was assessed through ROC curves and Area Under the Curve (AUC) for discrimination, calibration plots for agreement between predicted and observed outcomes, and Brier scores for overall prediction error. The ML may be subject to overfitting due to the limited sample size and the number of candidate predictors, and that the reported performance metrics should therefore be interpreted cautiously. All analyses were performed using R software (version 4.5.2; R Foundation for Statistical Computing, Vienna, Austria).