Endometrial cancer (EC) is the fourth most common cancer in American women, with an estimated 66,200 new cases and 13,030 deaths projected for 2023. While the current FIGO staging system provides a general prognosis framework, it has known limitations in predictive accuracy. Survival outcomes depend on a complex combination of factors: age, tumor size, histological type, depth of invasion, lymph node involvement, and treatment choices - all interacting with each other in non-linear ways.
Traditional survival analysis relies heavily on the Cox proportional hazards model, a statistical approach that estimates how each risk factor individually contributes to the probability of dying over time. While widely used and well-understood, the Cox model treats each variable as independently influencing survival. This assumption is often violated in reality: for instance, lymph node metastasis and tumor stage are highly correlated, and chemotherapy decisions depend on tumor grade. Ignoring these relationships reduces predictive accuracy.
Machine learning offers an alternative. Bayesian networks are graphical models that explicitly represent and visualize the relationships (dependencies) between variables. Rather than assuming independence, they model how factors influence each other, potentially capturing the complex, interacting nature of cancer prognosis more accurately than conventional statistical methods.
Researchers used the SEER database (Surveillance, Epidemiology, and End Results) - a large US cancer registry covering 34.6% of the American population - to extract records for 618 patients diagnosed with EC between 2010 and 2015 who had undergone radical resection. The dataset was split 75/25 into training (464 patients) and test (154 patients) sets. For external validation, 104 additional EC patients from a Chinese hospital were used.
Thirteen clinical variables were analyzed: age, tumor grade, histological type, tumor stage, radiotherapy timing and type, chemotherapy, lymph node resection, lymph node metastasis, tumor size, depth of invasion, distant metastasis, and primary tumor site. The primary outcome was 5-year survival status (alive vs. dead).
Two models were built and compared head-to-head. The Cox proportional hazards model used univariate screening followed by multivariate analysis to identify independent prognostic factors. The Bayesian network used the tree augmented naive (TAN) algorithm - a variant that relaxes the assumption that all variables are independent by allowing each variable to have additional dependencies modeled as a tree structure. Variable importance was ranked using the MMFV (Mean Multi-state Fussell-Vesely) metric, which measures each variable's contribution to predicting the survival outcome.
On the internal test set, the Bayesian network achieved 74.68% accuracy, AUC 0.787, and C-index 0.72. The Cox model achieved 68.83% accuracy, AUC 0.723, and C-index 0.71. All three metrics consistently favored the Bayesian network. The AUC - which measures the model's ability to distinguish survivors from non-survivors across different probability thresholds - was meaningfully higher for the Bayesian network.
On the external Chinese hospital cohort, the performance gap was even larger: Bayesian network AUC 0.849, C-index 0.81 vs. Cox model AUC 0.786, C-index 0.79. This improvement on an independent external cohort is particularly important - it confirms the Bayesian network's superiority is not just an artifact of overfitting to the training data but reflects genuine generalization to new patients from a different healthcare system and geographic setting.
The C-index (concordance index) deserves special mention as a metric designed specifically for survival analysis. Unlike AUC, which treats survival as a binary classification, the C-index also accounts for the timing of events - whether the model correctly ranks patients by their survival time. Both models achieved similar C-indices (~0.71-0.72 on internal validation), but the Bayesian network was substantially better on external validation (0.81 vs 0.79).
The Bayesian network's MMFV importance ranking revealed a clear hierarchy of prognostic factors: tumor size was the single most important predictor of 5-year survival, followed by lymph node metastasis, distant metastasis, chemotherapy, lymph node resection, tumor stage, depth of invasion, tumor grade, histological type, age, primary tumor site, radiotherapy timing, and radiotherapy itself (least important of those tested).
The importance of tumor size as the top predictor - ranked above even formal tumor stage - is clinically meaningful. While FIGO staging incorporates multiple factors, this analysis suggests that the raw physical size of the tumor carries the most information about likely survival. Tumors 4 cm or larger were associated with meaningfully lower survival probabilities compared to tumors under 4 cm across all other conditions.
Interestingly, the Bayesian network captured relationships that the Cox model missed. For example, Cox multivariate analysis did not identify lymph node metastasis as an independent risk factor (because stage III already captured this information), while the Bayesian network properly modeled the relationship between lymph node status and stage, ranking lymph node metastasis as the second most important factor. This shows the value of modeling variable interdependencies rather than treating each factor as if it acted alone.
The key conceptual advantage of Bayesian networks is their directed acyclic graph structure, where arrows between variables represent probabilistic dependencies. A clinician can visually inspect the network to see which factors directly influence others. For example, the network might show that histological type depends on age and tumor grade, and that chemotherapy decisions are linked to histological type - connections that Cox models treat as independent effects.
This visual interpretability makes Bayesian networks useful not just for prediction but for understanding disease biology and treatment relationships. The network revealed that histological type was correlated with age at diagnosis, tumor grade, and chemotherapy - a biologically plausible finding that aggressive subtypes (serous, clear cell) tend to occur in older patients and require chemotherapy more often.
The authors argue for a complementary approach: using Bayesian networks for superior predictive accuracy while retaining Cox models for clear interpretation of individual variable effects. This combined approach - where the Bayesian network ranks risk and the Cox model explains why - could support more nuanced shared decision-making between oncologists and patients about post-surgical treatment intensity.
This study demonstrates that a Bayesian network trained on large-scale population data can predict 5-year EC survival with meaningful accuracy and be validated externally on patients from a different country and healthcare context. The consistency of results across both the SEER-based internal validation and the Chinese hospital external validation strengthens confidence in the model's generalizability.
Key limitations include the relatively modest sample sizes (618 training + 104 validation), which limit the Bayesian network's complexity - more variables and their interactions could be modeled with larger datasets. Several known prognostic factors were not included due to data unavailability in SEER: menopausal status, vaginal bleeding patterns, and lymphovascular space invasion (LVSI) are all established EC prognostic markers that could improve model accuracy if incorporated.
The external validation accuracy of 68.27% (lower than the 74.68% on SEER data) suggests that institutional and population differences do affect model performance, and that local calibration of the model may be needed for specific clinical settings. Despite these limitations, this study represents a meaningful demonstration that Bayesian networks can outperform traditional statistical approaches for EC prognosis prediction and merit further development as clinical decision support tools.