One of the most contested decisions in endometrial cancer surgery is whether to perform lymph node dissection. Lymph nodes are removed to assess whether cancer has spread beyond the uterus, but the procedure carries significant risks including lymphedema, nerve injury, and increased surgical time. For many patients whose cancer has not spread, the operation provides no survival benefit while adding harm.
Existing tools for predicting lymph node metastasis (LNM) rely on standard clinicopathological factors such as tumor grade and depth of myometrial invasion, but these alone lack the precision needed to reliably identify the roughly 15-20% of patients with node-positive disease. Most guidelines still recommend lymphadenectomy for intermediate or high-risk patients, resulting in many unnecessary procedures.
The ENDORISK model was developed to address this gap by combining multiple biomarkers and imaging findings into a probabilistic framework that better reflects the complexity of LNM risk. The goal is to identify patients with low enough risk that lymph node dissection can be safely omitted, while still capturing those at high risk who most need the procedure.
ENDORISK is built on a Bayesian network - a probabilistic graphical model that represents relationships between variables as a directed acyclic graph. Unlike standard logistic regression, a Bayesian network can encode conditional dependencies between predictors and naturally handles missing data, which is common in clinical settings.
The model integrates 9 predictors: tumor grade, estrogen receptor (ER) status, progesterone receptor (PR) status, L1CAM expression, p53 mutation status, serum Ca-125, platelet count (thrombocytes), imaging evidence of lymphadenopathy, and cervical cytology results. These were selected based on prior literature and were available from standard preoperative workup in most European gynecology centers.
Model development used data from 763 patients across 10 European ENITEC centers. Validation was performed in two independent external cohorts: the MoMaTEC cohort (446 patients) and the PIPENDO cohort (384 patients). This multi-cohort external validation is a rigorous standard that tests whether a model generalizes beyond its training population.
The ENDORISK model achieved an AUC of 0.82 for LNM prediction, indicating strong discriminatory ability between node-positive and node-negative patients. This performance was consistent across both external validation cohorts, demonstrating robust generalizability.
Critically, the model classified 55.8% of patients as having less than 5% LNM risk - a threshold at which the potential harms of lymphadenectomy are generally considered to outweigh the benefits. Within this low-risk group, only 1.6% of patients actually had lymph node metastasis (the false-negative rate), meaning that adopting this threshold would result in fewer than 2 patients per 100 having undetected nodal disease.
At the other end of the risk spectrum, the model identified a high-risk group in which lymphadenectomy is clearly warranted. This bimodal risk stratification - confidently identifying both low-risk patients who can skip the procedure and high-risk patients who need it - is the clinical objective a prognostic model for surgical decision-making must achieve.
If ENDORISK were adopted in clinical practice, approximately half of endometrial cancer patients could potentially avoid lymph node dissection based on the low-risk classification. This represents a substantial reduction in surgical morbidity across the endometrial cancer population, which numbers in the hundreds of thousands of new cases per year globally.
The model's strength is that it combines information routinely collected during preoperative workup - tissue biomarkers available from biopsy, standard blood tests, and preoperative imaging - without requiring additional specialized tests. This makes ENDORISK potentially deployable at any center that performs endometrial cancer surgery without major changes to the diagnostic workflow.
The incorporation of imaging-based lymphadenopathy and cervical cytology alongside molecular markers reflects the model's ability to integrate heterogeneous data types. Bayesian networks are particularly well-suited for this because they can formally quantify how each piece of information updates the probability of LNM, making the reasoning transparent to the clinician rather than operating as a black box.
Current clinical guidelines use systems such as the ESMO-ESGO-ESTRO risk stratification, which divides patients into low, intermediate, high-intermediate, and high-risk categories based primarily on tumor grade and myometrial invasion. While useful, these systems were designed to guide adjuvant therapy decisions rather than lymphadenectomy and do not explicitly model LNM probability.
ENDORISK adds meaningful value by directly targeting the LNM question with a probabilistic output rather than categorical risk groups. The probabilistic output allows the clinician and patient to discuss the specific estimated probability - for example, a 3% or 12% chance of nodal involvement - rather than a category label whose thresholds are inherently arbitrary.
The inclusion of molecular markers such as L1CAM and p53 status is particularly notable. These markers have independently been associated with LNM and aggressive disease behavior in endometrial cancer, and their formal integration into a validated multivariate model is a step beyond the traditional histopathological approach to risk assessment.
The strong external validation performance across two independent European cohorts provides substantial evidence that ENDORISK is a generalizable and reliable tool. However, prospective implementation studies are needed to evaluate whether using the model to guide intraoperative decisions actually improves patient outcomes compared to current practice.
A web-based or embedded calculator version of ENDORISK would facilitate point-of-care use, allowing surgeons to enter preoperative values and receive an instantaneous LNM probability estimate during surgical planning. Several online clinical prediction tools have successfully translated validated statistical models into usable formats for practicing clinicians.
This work also establishes a template for developing Bayesian network-based decision support tools in gynecological oncology. The integration of molecular, serological, and imaging information in a probabilistic framework represents a more sophisticated approach to surgical risk stratification than current practice and merits broader adoption across cancer types where similar lymphadenectomy trade-offs exist.