Results

Sample characteristics

A total of 416 primary cardiac malignancy patients were included in this study. The median age at diagnosis was 48 years (IQR: 35–62), increasing to 55 years (IQR: 38–67) for those who died from other causes. The cohort comprised 51% females (n = 212) and 49% males (n = 204). Racially, 78% were White, 12% Black, and 10% other racial groups. At diagnosis, 29% were localized, 31% regional, and 40% distant-stage cancer. Angiosarcoma accounted for 45% of cases. Treatment modalities included radiotherapy (22%), chemotherapy (51%), and surgery (66%), with surgical status missing for two patients. Outcomes showed 68 censored, 247 died from primary cardiac malignancy, and 101 died from other causes (Table 1).

Predictors | Overall(n = 416) | Censored(n = 68) | Mortality | p value
Primary cardiac malignancies(n = 247) | Other causes(n = 101)
Age in years† | 48 (35, 62) | 44 (32, 56) | 48 (35, 60) | 55 (38, 67) | 0.016
Sex |  |  |  |  | 0.15
Female | 212 (51%) | 37 (54%) | 132 (53%) | 43 (43%) | 
Male | 204 (49%) | 31 (46%) | 115 (47%) | 58 (57%) | 
Race/ethnicity |  |  |  |  | 0.77
Others | 43 (10%) | 7 (10%) | 25 (10%) | 11 (11%) | 
Black | 50 (12%) | 6 (9%) | 34 (14%) | 10 (10%) | 
White | 323 (78%) | 55 (81%) | 188 (76%) | 80 (79%) | 
Cancer stage |  |  |  |  | 0.039
Localized | 106 (29%) | 19 (41%) | 66 (29%) | 21 (22%) | 
Regional | 114 (31%) | 17 (37%) | 64 (28%) | 33 (35%) | 
Distant | 149 (40%) | 10 (22%) | 98 (43%) | 41 (43%) | 
(Missing) | 47 | 22 | 19 | 6 | 
Angiosarcoma [yes] | 188 (45%) | 28 (41%) | 120 (49%) | 40 (40%) | 0.24
Radiotherapy [yes] | 92 (22%) | 16 (24%) | 61 (25%) | 15 (15%) | 0.13
Chemotherapy [yes] | 212 (51%) | 41 (60%) | 131 (53%) | 40 (40%) | 0.018
Surgery [yes] | 274 (66%) | 58 (85%) | 152 (62%) | 64 (64%) | 0.001
(Missing) | 2 | 0 | 1 | 1 | 

Survival analysis

Figure 2 presents univariate Kaplan–Meier overall survival analyses stratified by treatment modalities: (A) radiotherapy, (B) surgery, and (C) chemotherapy. Across all modalities, survival probability declines sharply during the early follow-up period, reflecting high short-term mortality among patients with cardiac cancer. Patients who underwent surgery showed the highest and consistent survival benefit over the follow-up period (Fig. 2B). Radiotherapy and chemotherapy were also associated with higher short-term survival compared to no treatment, although these differences diminished over time (Fig. 2A and C). These findings indicate differences in survival across treatment groups; however, further multivariable analyses are needed to account for potential confounding factors.

Fig. 2: Kaplan–Meier survival curves for all-cause mortality stratified by treatment modalities: Panel A for radiotherapy, Panel B for surgery, and Panel C for chemotherapy

Competing-risk analyses demonstrate differences in the cumulative incidence of primary cardiac malignancy-specific mortality versus other-cause mortality across treatment groups (Fig. 3). The solid lines represent primary cardiac malignancy-specific mortality, while the dashed lines represent mortality from other causes. Patients who underwent surgical treatment showed lower cumulative incidence of primary cardiac malignancy-specific mortality compared with those who did not receive surgery. However, mortality from other causes did not differ significantly between the two groups over the follow-up period. Chemotherapy was associated with a reduction in primary cardiac malignancy-specific mortality during the first three years of follow-up. After this period, the association with cancer-specific mortality was diminished; however, chemotherapy remained significantly associated with lower mortality from other causes throughout the study period. Radiotherapy showed no substantial difference in primary cardiac malignancy-specific mortality during the initial years of follow-up. Over time, the cumulative incidence of primary cardiac malignancy-related death declined among patients who did not receive radiotherapy, while radiotherapy was associated with lower mortality from other causes across the full follow-up duration.

Fig. 3: Competing-risk analysis of mortality by treatment modality using cumulative incidence functions. Solid lines represent primary cardiac malignancy-specific mortality, while dashed lines represent mortality from other causes

Table 2 summarizes results of multivariable competing-risk models for primary cardiac malignancy-specific mortality and other-cause mortality. Age was positively associated with cancer-specific mortality (AHR: 1.01, 95% CI: 1.01–1.02) and other-cause mortality (AHR: 1.026, 95% CI: 1.01–1.04). Sex was not statistically significant. Disease stage at diagnosis showed the strongest association: compared to localized stage, regional was associated with a more than twofold higher hazard of other-cause mortality (AHR: 2.36; 95% CI: 1.34–4.15), while distant-stage was associated with a 69% higher hazard of primary cardiac malignancy-specific mortality (AHR: 1.69; 95% CI: 1.18–2.43) and approximately threefold higher for other-cause mortality (AHR: 3.01; 95% CI: 1.68–5.41). Angiosarcoma histology was associated with a 35% higher hazard of primary cardiac malignancy-specific mortality (AHR: 1.35; 95% CI: 1.01–1.82).

Predictors | AHR (95% CI)
Primary cardiac malignancy-specific mortality | Other causes of mortality
Age in years | 1.01 [1.01–1.02] | 1.03 [1.01–1.04]
Sex (ref. = Female)
Male | 0.84 [0.64–1.10] | 1.46 [0.96–2.21]
Race/ethnicity (ref. = Other)
Black | 1.10 [0.64–1.90] | 0.80 [0.33–1.95]
White | 0.97 [0.62–1.51] | 0.79 [0.41–1.51]
Cancer stages (ref. = Localized)
Regional | 1.24 [0.87–1.77] | 2.36 [1.34–4.15]
Distant | 1.69 [1.18–2.43] | 3.01 [1.68–5.41]
Angiosarcoma | 1.35 [1.01–1.82] | 0.71 [0.44–1.12]
Radiotherapy | 0.95 [0.70–1.30] | 0.57 [0.32–1.01]
Surgery | 0.57 [0.42–0.79] | 0.64 [0.39–1.06]
Chemotherapy | 0.63 [0.47–0.85] | 0.51 [0.32–0.81]

Among treatment modalities, surgery and chemotherapy were associated with lower primary cardiac malignancy-specific mortality. Surgery corresponded to a 42% lower hazard of primary cardiac malignancy-specific mortality (AHR: 0.58; 95% CI: 0.42–0.79). Chemotherapy was similarly related to lower hazards of both primary cardiac malignancy-specific mortality (37% reduction, AHR: 0.63, 95% CI: 0.47–0.85) and other-cause mortality (49% reduction, AHR: 0.51, 95% CI: 0.32–0.81). Radiotherapy was not statistically significantly related to primary cardiac malignancy-specific mortality but demonstrated a borderline association with lower other-cause mortality (AHR: 0.57; 95% CI: 0.32–1.01) (Table 2). It is noted here and throughout that the association of treatment modalities with other-cause mortality may reflect selection bias rather than treatment efficacy.

Table 3 presents multivariable competing-risk analyses examining the associations of multimodal treatments with primary cardiac malignancy-specific and other-cause mortality. The reference group included patients who received no therapeutic intervention. The combination of chemotherapy and surgery showed significantly higher survival from both primary cardiac malignancy-specific mortality (AHR: 0.72; 95% CI: 0.54–0.95) and other-cause mortality (AHR: 0.58; 95% CI: 0.36–0.93). Similarly, the combination of radiotherapy and surgery was associated with a 56% lower hazard of other causes of mortality (AHR = 0.44; 95% CI: 0.22–0.87), while no significant association was observed for primary cardiac malignancy-specific mortality. Both radiotherapy plus surgery and triple therapy (radiotherapy, surgery, and chemotherapy) showed lower but insignificant hazards for both mortality outcomes.

Therapeutic combination | AHR (95% CI)
Primary cardiac malignancy-specific mortality | Other causes of mortality
Radiation + surgery | 1.06 [0.75–1.48] | 0.44 [0.22–0.87]
Chemotherapy + surgery | 0.72 [0.54–0.95] | 0.58 [0.36–0.93]
Chemotherapy + radiation | 0.96 [0.68–1.35] | 0.59 [0.31–1.12]
Radiation + surgery + chemotherapy | 0.99 [0.66–1.47] | 0.48 [0.21–1.10]

Table 4 summarizes results of multivariable competing-risk models comparing multimodal treatment approaches to the reference group of ‘surgery only’. Across treatment groups, none of the combination therapies showed a statistically significant difference in primary cardiac malignancy-specific mortality relative to surgery alone. However, mortality from other causes was lower among patients receiving additional treatments: surgery plus chemotherapy (AHR: 0.54, 95% CI: 0.29–0.98), surgery plus radiation (AHR: 0.29, 95% CI: 0.08–0.96), and surgery combined with chemotherapy and radiation (AHR: 0.34, 95% CI: 0.14–0.83). All models were adjusted for age, sex, race, cancer stage, and angiosarcoma subtype.

Therapeutic combination | AHR (95% CI)
Primary cardiac malignancy-specific mortality | Other causes of mortality
Surgery + chemotherapy (ref. Surgery only) | 0.78 [0.51–1.19] | 0.54 [0.29–0.98]
Surgery + radiation (ref. Surgery only) | 1.32 [0.73–2.38] | 0.29 [0.08–0.96]
Surgery + chemotherapy + radiation (ref. Surgery only) | 1.01 [0.62–1.65] | 0.34 [0.14–0.83]

Machine learning modeling

Figure 4 shows the ranking of variables based on importance scores derived from the random forest (RF) model. Age had the highest importance score (importance score ≈ 8), indicating it contributed most to the prediction model mortality. Angiosarcoma and cancer stage followed closely (importance score ≈ 6,7), reflecting the substantial contributions to mortality prediction within the model. Chemotherapy and surgery showed moderate importance (importance score ≈ 2–6), highlighting their therapeutic relevance. In contrast, sex, race, and radiotherapy had low importance scores, indicating minimal contribution to mortality prediction in this analysis.

Fig. 4: Variable importance scores from the random forest (RF) model for competing-risk of mortality

Figure 5 presents the regression tree from the machine learning competing-risk model, illustrating how demographic, clinical, and treatment factors were used to stratify survival outcomes. Age was the primary split variable, indicating that it contributed most to survival stratification. Among patients aged ≤ 65 years, cancer stage was the key secondary determinant, with localized disease corresponding to most favorable survival outcomes. Surgical intervention was consistently associated with higher survival probabilities across stages, particularly among patients with localized and regional disease. Among patients aged > 65 years, chemotherapy emerged as the primary stratifying variable. In this subgroup, surgery showed limited differences in survival among patients who did not receive chemotherapy. Among older patients who received chemotherapy, surgery corresponded to slightly higher survival probabilities, although overall survival declined rapidly over time across subgroups.

Fig. 5: Regression tree from a machine learning competing-risk survival model

The machine learning regression tree model demonstrated solid predictive performance in the test dataset (Fig. 6). The ROC curve (Panel A) yielded an AUC of 0.72, indicating acceptable discriminative ability to distinguish patients with higher versus lower observed mortality risk. The calibration plot (Panel B) showed close alignment between observed and predicted probabilities, with a mean absolute error of 0.035 and a 90th percentile error of 0.105, confirming reasonable calibration. Prediction error analysis (Panel C) showed lower prediction error for the machine learning forest model (gray line) compared to the reference model (black line). Overall, the model demonstrated a balance between discrimination and calibration in the test data, with lower prediction error, indicating stable predictive performance within the sample.

Fig. 6: Model performance evaluation on the test dataset. A Receiver Operating Characteristic (ROC) curve with area under the curve (AUC) = 0.72 for the machine learning model. B Calibration plot comparing predicted versus observed outcomes. C Prediction error assessment. The machine learning model was compared with a Cox proportional hazards model, which served as the reference model