Endometrial cancer (EC) is typically diagnosed through uterine sampling before surgery, allowing surgeons to plan a comprehensive operation that includes lymph node removal for staging. But in a significant subset of cases, EC is discovered incidentally - the cancer is found unexpectedly in a uterus removed for a different reason (such as fibroids, abnormal bleeding, or prolapse). These patients have already had their hysterectomy before cancer was suspected.
This creates a serious clinical dilemma. Lymph node metastasis (LNM) - cancer cells spreading to regional lymph nodes - is the most critical factor determining whether a patient with EC needs additional treatment. Current guidelines suggest re-operation to remove lymph nodes for staging when risk is uncertain. But lymph node surgery carries real risks: lymphedema (persistent leg swelling), nerve damage, and the general risks of a second major abdominal operation.
The clinical question is therefore: which incidentally found EC patients actually have lymph node spread and need re-staging surgery, and which can be safely observed? A reliable prediction model could spare low-risk patients an unnecessary second operation while ensuring high-risk patients receive appropriate staging and treatment. This study builds and validates exactly such a model using deep learning and data from nearly 4,500 patients.
Data came from two Fudan University affiliated hospitals in China. The main cohort was 3,920 patients who underwent hysterectomy and had their lymph node status confirmed, used for model training and validation (75/25 split). A separate prospective cohort of 572 patients was collected subsequently to test whether the model performed on newly collected, never-before-seen cases.
Forty-three clinical and pathological variables were collected for each patient, including lymphovascular space invasion (LVSI), depth of myometrial invasion, tumor grade, the MELF pattern (a specific invasive growth pattern linked to higher LNM risk), progesterone receptor (PR) status, CA125 blood level, molecular classification (POLE, MSI, p53-mutated, or non-specific subtype), and various hormonal laboratory values.
The Fully-Connected Network (FC Network) was built with an architecture of progressively smaller layers: input layer (96 dimensions after encoding) feeding into layers of 256, 128, 64, and 20 neurons, then an output layer classifying LNM as positive or negative. Dropout regularization (rate 0.5) was used to prevent overfitting. Missing data - a universal challenge with real clinical datasets - was handled using KNN imputation (k-nearest neighbors), which outperformed four competing methods including mean imputation and multiple imputation by chained equations.
The FC Network was compared against seven other machine learning models: logistic regression, support vector machine, decision tree, random forest, gradient boosting, extreme gradient boosting (XGBoost), and k-nearest neighbors. The FC Network achieved the highest sensitivity of 0.946 with AUC 0.877 on validation and AUC 0.840 on the prospective cohort.
Sensitivity is particularly important here because missing a true positive (failing to identify a patient with LNM) has serious consequences - that patient would not receive lymph node surgery and could be undertreated. The Random Forest model, by contrast, achieved sensitivity of only 0.055 - meaning it missed 94.5% of actual LNM cases, making it clinically unusable for this application despite potentially better specificity or AUC in other metrics.
A simplified clinical nomogram was also developed using only six variables selected by LASSO regression and clinical relevance: LVSI, myometrial invasion depth, tumor grade, MELF pattern, PR status, and CA125 level. Despite using far fewer variables, this nomogram achieved AUC 0.890 on validation and AUC 0.885 on the prospective cohort - slightly higher AUC than the FC Network but with lower sensitivity. The nomogram's advantage is that it can be computed manually from a printed chart, making it practical in settings without computing infrastructure.
The practical output of both models is a predicted probability of LNM for each patient. Rather than a simple binary classification, the study defined three risk tiers: low risk (below 5% probability), suitable for observation without re-staging; medium risk (5-25%), where lymph node dissection should be considered based on patient factors and preferences; and high risk (above 25%), where lymph node dissection is advised.
The most clinically significant of the six nomogram variables is LVSI (lymphovascular space invasion) - the presence of tumor cells within blood vessels or lymphatic channels of the uterus. LVSI is a direct indicator that cancer cells are already accessing the routes by which they would spread to lymph nodes. Its presence dramatically increases LNM probability regardless of other factors.
Other important predictors include depth of myometrial invasion (deeper invasion means higher stage and higher LNM risk), tumor grade (higher grade means more aggressive biology), and MELF pattern (a distinctive infiltrative growth pattern associated with lymphatic invasion). PR positivity was associated with lower LNM risk - consistent with the known biology that hormone receptor-positive EC tends to be lower grade and less aggressive. CA125 elevation signals more advanced disease and active tumor biology.
The choice of FC Network as the primary model - prioritizing sensitivity over other metrics - reflects a deliberate clinical value judgment. In the context of predicting lymph node metastasis, the two types of errors have very different consequences. A false positive (predicting LNM when there is none) leads to an unnecessary second surgery with real but manageable risks. A false negative (missing actual LNM) means a patient with cancer in their lymph nodes receives no further staging or treatment, potentially allowing disease progression.
Most machine learning models optimize for overall accuracy or AUC - metrics that treat false positives and false negatives equally. Random Forest, for example, achieved high overall accuracy but sensitivity near zero, meaning it classified nearly every case as LNM-negative. This high-accuracy but clinically useless model illustrates why choosing the right evaluation metric for the clinical context is as important as model design.
The authors recommend a two-stage approach: use the FC Network for initial screening (high sensitivity catches all true LNM cases) and the nomogram for clinical communication and shared decision-making with patients (simpler, interpretable). The nomogram is particularly useful for discussing the probability estimate with patients, who can see which of their six risk factors are present and understand intuitively why their risk is calculated as it is.
This study provides the largest dataset analysis (nearly 4,500 cases across retrospective and prospective cohorts) specifically focused on the under-studied population of incidentally found EC. By focusing on this specific clinical scenario rather than EC as a whole, the models are calibrated for exactly the patient population where re-staging decisions are most difficult and consequential.
The combination of a high-sensitivity deep learning model for triage and a simple nomogram for clinical communication represents a pragmatic two-tool approach. In centers with computational infrastructure, the FC Network provides maximum sensitivity. In settings where computing tools are unavailable or where physicians prefer transparent calculations, the six-variable nomogram provides comparable AUC with a formula that can be printed and used at the bedside.
Key limitations include the single-country origin (Chinese academic medical centers), the focus on post-hysterectomy findings rather than pre-operative biopsy-diagnosed cases, and the absence of prospective randomized evidence that using this model to guide re-staging decisions actually improves patient outcomes compared to current standard approaches. The next step would be a prospective intervention study where model predictions are used to guide re-staging decisions and patient outcomes are tracked over time.