Endometrial cancer (EC) is the second most common gynecological malignancy worldwide, with incidence rising year over year. One of the most critical factors determining a patient's prognosis is whether cancer has spread to the lymph nodes - the small immune organs distributed throughout the body that cancer cells often use as an early route of dissemination.
Patients with pelvic or para-aortic lymph node metastasis (LNM) face significantly worse outcomes than those without nodal involvement. Knowing whether lymph nodes are affected before surgery allows surgeons and oncologists to plan more aggressive or more conservative treatment accordingly. The challenge is that many patients who appear to have early-stage disease on preoperative imaging actually have hidden lymph node spread when lymph nodes are examined under a microscope after surgery.
Current practice involves either sentinel lymph node biopsy (SLNB) - removing only the first few lymph nodes the tumor would drain into - or full systematic lymph node dissection (SLND). SLNB is less invasive but has limitations: technical difficulty, variable detection rates, and uncertainty about whether micro-metastases detected by advanced pathology require treatment. A reliable preoperative prediction model that identifies which patients are at high risk for LNM could guide more targeted use of these procedures.
Most existing models rely on pathological features only available after surgery (like lymphovascular space invasion), limiting their usefulness for preoperative planning. This study aimed to develop a model using only information available before the operation, including a novel addition: molecular tumor subtype.
The study enrolled 465 patients with clinically early-stage EC (no evidence of lymph node involvement on preoperative imaging) treated at Qilu Hospital of Shandong University. Tumors were classified using the TCGA molecular framework into four subtypes: POLE-ultramutated, MSI-H (microsatellite instability-high), CN-L (copy-number low, p53 wild type), and CN-H (copy-number high, p53 abnormal). Because POLE cases were rare, all non-CN-H subtypes were grouped together as "non-CN-H" for modeling purposes, based on evidence that p53 abnormality is associated with worse prognosis and higher LNM risk.
Thirteen preoperative candidate variables were evaluated: age, menopausal status, CA125 (a blood tumor marker), neutrophil-to-lymphocyte ratio (NLR) (a blood count ratio that reflects systemic inflammation and immune status), tumor grade, histological subtype (endometrioid vs. non-endometrioid), myometrial invasion depth on MRI, cervical involvement, and molecular typing. LASSO regression - a statistical variable selection method - was applied to identify the most predictive subset from these 13 candidates.
LASSO selected five key predictors: molecular typing (CN-H vs. non-CN-H), histological subtype, depth of myometrial invasion, NLR, and serum CA125. Notably, lymphovascular space invasion (LVSI) was intentionally excluded from the analysis because it is only definitively determined after surgery - the model must work from preoperative data alone.
Five machine learning algorithms were trained: Logistic Regression (LR), Random Forest (RF), Gradient Boosting Machine (GBM), Decision Tree, and Support Vector Machine (SVM). Logistic Regression was ultimately selected despite comparable performance to Random Forest, because of its interpretability, simplicity, and ease of integration into the clinical nomogram and web tool that were built to deploy the model.
Among the 465 patients, 9.7% had lymph node metastasis (45 patients). Several preoperative factors were significantly associated with LNM: elevated CA125 (60% of LNM patients vs. 23% of non-LNM patients had CA125 above 35 U/mL), deep myometrial invasion (57.8% vs. 18.8% with invasion greater than 50%), non-endometrioid histological subtype (37.8% vs. 8.1%), and elevated NLR (42.9% vs. 23.2% with NLR above 2.838).
The CN-H (p53 abnormal) molecular subtype was significantly more common in the LNM group: 66.7% of LNM patients had CN-H tumors, compared to 49% in the non-LNM group (p = 0.025). This confirms that molecular subtyping, beyond traditional clinical variables, adds independent predictive value for lymph node spread - and provides a biological rationale: p53-abnormal tumors are more genomically unstable and biologically aggressive.
Interestingly, age, tumor grade, cervical involvement, and menopausal status were not significantly different between the LNM and non-LNM groups in this cohort - highlighting that traditional clinical intuition about risk factors does not always translate to statistical prediction power, and that molecular information adds important signal beyond morphology alone.
The final logistic regression model achieved an AUC of 0.843 in the training cohort and 0.809 in the independent testing cohort - both indicating good discriminative ability. Calibration curves showed that predicted probabilities matched observed LNM frequencies closely, meaning the model provides accurate probability estimates, not just rankings. Decision curve analysis confirmed that using the model to guide clinical decisions would provide net benefit across a range of risk thresholds.
The model was used to generate a nomogram - a graphical tool that allows clinicians to calculate a patient's individualized LNM risk by adding up points corresponding to each predictor's value. A total score above 151.934 points identifies patients as high-risk for LNM, while those below are classified as low-risk. This threshold was validated to significantly distinguish the two groups (p less than 0.001).
The researchers deployed the nomogram as a free online web calculator (built using the Shiny R framework), allowing clinicians worldwide to enter a patient's preoperative data and immediately receive an individualized LNM risk estimate. This kind of clinician-facing tool is what transforms a research model into a practical clinical resource - reducing the barrier between AI research and bedside application.
Sensitivity analysis confirmed that the model performed similarly when restricted to patients who underwent systematic lymph node dissection (AUC 0.710) compared to the full cohort (AUC 0.809), with no statistically significant difference (p = 0.502). This demonstrates robustness to the type of lymph node staging procedure used.
The practical value of this model lies in its ability to stratify patients before surgery into those who likely need comprehensive lymph node evaluation and those who can be safely managed with less invasive staging. For high-risk patients (above the model threshold), systematic lymph node dissection or at minimum thorough SLNB with backup dissection is recommended. For low-risk patients, the model provides evidence-based reassurance that extensive nodal surgery may not be required.
This kind of risk stratification has real clinical impact: systematic lymph node dissection is associated with complications including lymphedema (fluid accumulation from disrupted lymphatic drainage) and longer operative time. Avoiding unnecessary dissections in low-risk patients improves quality of life without compromising cancer control, while ensuring high-risk patients receive the thorough staging they need.
A key advance of this model over previous approaches is the inclusion of molecular typing as a predictor. The 2023 FIGO staging system and current international guidelines (NCCN, ESMO) now incorporate molecular classification into EC management algorithms. A preoperative LNM model that also uses molecular typing is therefore aligned with the direction of clinical practice and can be applied within existing guidelines frameworks.
This study presents the first validated machine learning model to integrate molecular tumor subtyping with routinely available clinical variables for preoperative lymph node metastasis prediction in early-stage endometrial cancer. It demonstrates that combining molecular information (CN-H vs. non-CN-H) with five accessible clinical features produces an accurate and clinically useful prediction tool.
The deployment of the model as a free online nomogram tool represents an important step toward translating research findings into clinical practice. Rather than requiring clinicians to perform complex calculations, the web tool provides immediate, individualized risk estimates at the point of care - the kind of implementation that makes AI models actionable in real hospital settings.
Limitations of the study include its retrospective design and single-institution derivation, the relatively small testing cohort (n = 102), and the need for prospective multicenter validation across diverse patient populations, healthcare settings, and molecular testing platforms. Future work should also explore whether incorporating additional molecular markers beyond CN-H classification could further improve model accuracy.