Discussion

In this population-based study, primary cardiac malignancies, survival patterns were largely associated with age, tumor biology, and disease stage. Surgery and chemotherapy were each associated with significantly higher overall and cause-specific survival, whereas radiotherapy demonstrated limited associations, potentially reflecting its use in selected or palliative cases. Multimodal treatment approaches were associated with lower mortality from other causes compared with surgery alone, but no significant association was observed in primary cardiac malignancy-specific mortality. Machine learning methods, including random survival forests and competing-risk models, provided additional insights into the relative importance of prognostic factors and revealed nonlinear treatment-outcome relationships, and improved individualized mortality prediction for this rare, aggressive cancer.

The early steep decline in survival, even among patients receiving active treatment, reflects the highly aggressive nature of cardiac sarcomas and the challenges in early detection. These tumors often present at advanced stages due to nonspecific symptoms and delayed diagnosis, resulting in rapid disease progression and poor outcomes [17]. Consistent with previous studies, surgical resection was linked to higher survival probabilities, particularly among patients with localized or regionally advanced disease [27, 34, 35]. When technically feasible, complete surgical resection remains the cornerstone of therapy, offering the best opportunity for long-term survival and potential cure [11, 34]. Achieving negative surgical margins is crucial for improving outcomes, underscoring the need for early diagnosis and referral to specialized centers with expertise in complex cardiac tumor resections [36, 37].

Chemotherapy was associated with lower mortality during the early follow-up period, supporting prior studies demonstrating that systemic therapy may improve early disease control and extend survival as part of a multimodality treatment approach [16, 38]. However, this early association diminished over time, likely reflecting the development of tumor resistance, cumulative treatment toxicity, and selection bias favoring healthier patients with greater treatment tolerance [16]. Similar temporal patterns have been reported in prior studies, where the initial efficacy of systemic therapy was followed by relapse or progression due to limited long-term disease control in cardiac sarcomas [37, 39, 40].

Radiotherapy, in contrast, was not significantly associated with primary cardiac malignancy-specific mortality but was linked to modestly lower mortality from other causes, which may reflect supportive-care benefits, improved local symptom control, or advances in modern delivery techniques, such as conformal and intensity-modulated radiotherapy [9, 41]. Although radiotherapy is often used to manage local disease and palliate tumor-related symptoms, durable remission is uncommon, likely due to anatomical constraints, proximity to critical cardiac structures, and limitations in delivering curative doses without cardiotoxicity [42, 43]. These findings are consistent with prior studies indicating that while radiotherapy may improve short-term local control and quality of life, its contribution to long-term survival remains limited in primary cardiac sarcomas [27, 34, 38, 40].

The combination of surgery and chemotherapy was associated with higher overall survival from both rare malignancy and other-cause mortality. Relative to surgery alone, all multimodal combinations were significantly associated with lower mortality from other-cause but showed no significant difference in primary cardiac malignancy-specific mortality. Overall, multimodal strategies may enhance overall survival even when their direct impact on cancer-specific outcomes is limited. The combination therapy is likely to enhance both immediate and sustained survival by targeting residual microscopic disease and reducing recurrence risk [44]. Moreover, potential benefits of radiotherapy are related to improved supportive care or selection of healthier patients for multimodal therapy [34, 40]. Insignificantly association with primary cardiac malignancy-specific mortality may reflect the limited efficacy of radiotherapy in achieving durable tumor control within the cardiac environment due to anatomic and dose constraints [38]. In the cohort of older patients with aggressive disease, the association of multimodal therapy with survival appeared attenuated, suggesting limited potential for substantial long-term improvements in overall survival [40].

Age and disease stage were the most important predictors of survival. Older patients were linked to higher mortality, likely due to comorbidity burden, limited physiological reserve, and lower tolerance for multimodal therapy [34]. Advanced-stage disease was associated with lower survival, highlighting the challenge of delayed detection, while angiosarcoma histology corresponded to less favorable survival, consistent with their clinical behavior reported in previous studies [40, 45].

Importantly, these therapeutic associations were less pronounced among patients aged 65 years and older, a population often characterized by reduced physiological reserve, treatment intolerance, and competing-risk health risks [34, 38]. The attenuation of treatment efficacy in older adults underscores the importance of age-tailored therapeutic strategies, including the use of comprehensive geriatric assessment, individualized dose optimization, and multidisciplinary cardio-oncology collaboration to balance efficacy, safety, and quality of life in this vulnerable population [38, 46]. The RF analysis revealed age, angiosarcoma histology, and cancer stage as the strongest predictors of competing mortality, followed by surgery and chemotherapy, aligning closely with conventional multivariable regression findings [31]. This internal concordance enhances model credibility and underscores the biological plausibility of results [47]. The ML regression tree revealed notable treatment-effect heterogeneity, indicating that younger patients with regional or distant disease were most likely to benefit from surgical therapy, which is aligned with previous oncologic research on sarcoma [32, 48].

The ML survival model showed acceptable discriminative ability and calibration, with a lower mean absolute prediction error, indicating stable performance and internal validity. These findings align with prior work suggesting that ML-based survival methods can accommodate complex, non-linear interactions and high-dimensional data more effectively than traditional regression models [31, 47, 49]. Clinically, these models can support precision oncology by generating individualized survival estimates and guiding shared decision-making for rare malignancies.

This study has several clinical implications. First, early recognition and referral to specialized centers remain crucial, as timely surgical resection was associated with higher survival [34, 35]. Second, chemotherapy was linked to lower early mortality though its long-term associations were less pronounced. Third, ML models may provide complementary information for prognostic assessment, risk stratification, and treatment planning in rare cancer settings [19, 50]. Combining ML with causal inference approaches could help disentangle treatment effects from confounding and better estimate individualized therapeutic benefit. Future studies should focus on external validity and the assessment of model generalizability across multi-institutional or international datasets.

Several limitations warrant acknowledgment. Due to the observational nature of the data, causal associations cannot be inferred. The retrospective design also introduces potential for residual confounding and selection bias. Observed survival differences may be partially influenced by immortal time bias, particularly for chemotherapy and radiotherapy, as patients must survive long enough after diagnosis to receive these treatments. Important clinical variables, including performance status, margin status, and detailed chemotherapy regimens, were unavailable, and comorbidity information was either not recorded or not included in this study. Additionally, sample size constraints limited precision in subgroup analyses, particularly for multimodal treatment combinations and older age groups. Treatment-era effects (early vs late SEER years) were not formally evaluated due to the rarity of primary cardiac malignancies and limited subgroup sizes. Therefore, evolving treatment practices over time, including surgical techniques, chemotherapy regimens, radiotherapy technology, and diagnostic imaging, may have influenced the observed associations. Because of limited treatment detail in SEER, treatments were used in the models as binary indicators of receipt versus non-receipt. This approach does not account for heterogeneity in timing, intensity, or surgical margin status and may lead to residual confounding, misclassification, and attenuation of treatment-related associations. The association of treatment modalities with other-cause mortality may reflect selection bias rather than treatment efficacy. 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. Moreover, any generalization of these predictive findings requires external validation in independent datasets. Nevertheless, this analysis provides a comprehensive evaluation of treatment outcomes and predictive modeling for one of the rarest and most lethal malignancies.