Gynecological endometrioid adenocarcinoma with squamous differentiation (GE-ASqD) is a specific histological subtype of endometrial and ovarian cancer. Endometrioid adenocarcinoma is the most common type of endometrial cancer (accounting for 80% of all endometrial cancer cases). In some of these tumors, a portion of the cells undergo squamous differentiation - they take on features resembling squamous cells (the flat, scale-like cells that line the skin and other surfaces) rather than remaining purely glandular. This mixed pattern is what defines GE-ASqD.
Squamous differentiation in endometrial cancer has drawn scientific attention because it may carry prognostic significance. Some studies suggest that the grade of the squamous component matters: low-grade squamous differentiation may have different implications than high-grade. One study found that squamous differentiation in low-to-moderate grade endometrioid endometrial cancer raised the recurrence probability by a factor of 5.6 times - a striking finding that warrants further investigation.
Despite this clinical interest, the actual 5-year survival rates for GE-ASqD patients are rarely reported in the literature, and no reliable predictive models exist for this specific subtype. The authors of this study aimed to fill that gap by building machine learning models to predict 5-year overall survival in a large cohort of GE-ASqD patients drawn from a national cancer database.
The study used data from the Surveillance, Epidemiology, and End Results (SEER) database, a large population-based cancer registry in the United States that tracks cancer diagnoses, treatments, and outcomes. Patients diagnosed with GE-ASqD between 2004 and 2015 were identified using specific cancer diagnosis codes. Only patients who survived beyond 5 years (or died within that period, providing a complete outcome) were included - a total of 907 patients with survival data qualified for machine learning model development. The data was split 70% for training and 30% for testing.
Nine clinical variables were used to build the models: age at diagnosis, race, sequence number (whether this was the patient's first cancer or they had previous primaries), marital status, cancer stage, surgery status, radiation status, chemotherapy status, and regional lymph nodes examined. Tumor TNM staging (T = tumor size/extent, N = lymph node involvement, M = distant metastasis) was also included.
Five machine learning algorithms were trained and compared: Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), Light Gradient Boosting Machine (LGBM), and Gradient Boosting Machine (GBM). The goal was to classify patients as likely 5-year survivors or not, and models were evaluated using AUC, accuracy, precision, recall, F1-score, and mean squared error.
Multivariate survival analysis identified the following as independent predictors of overall survival: age (younger patients aged 18-66 had 73% lower mortality risk than those 67-95), surgery (patients who received surgery had 72% lower mortality risk), lymph node examination (performing regional lymph node examination reduced mortality risk by 46%), tumor extent (T3 stage carried 3.45x higher mortality risk vs. T1), lymph node involvement (N1 carried 2.9x higher risk vs. N0), and distant metastasis (M1 carried 3.03x higher risk vs. M0).
Interestingly, while radiation was associated with higher mortality risk in multivariate analysis (HR 1.44), this likely reflects confounding - radiation is typically given to higher-risk patients, so the association reflects disease severity rather than radiation causing harm. Chemotherapy was independently associated with improved survival (HR 0.60) after adjusting for other factors.
In the training dataset, GBM (gradient boosting machine) achieved the best overall performance with an AUC of 0.856. Random Forest was second (AUC 0.841). However, on the independent test dataset - which measures how well models generalize to new patients - Logistic Regression performed best with an AUC of 0.779, followed by Gradient Boosting (0.767), Random Forest (0.753), Decision Tree (0.738), and LGBM (0.734). Calibration curves confirmed that all five models were well-calibrated, meaning predicted probabilities closely matched actual outcomes.
The finding that all five machine learning models performed reasonably well (test AUC 0.734-0.779) with relatively simple clinical variables suggests that demographic and treatment factors carry substantial predictive power for 5-year survival in GE-ASqD. The models do not require any genomic or molecular data - only routinely collected clinical information from cancer registries.
The fact that logistic regression - the simplest of the five algorithms - performed best on the test set despite being outperformed by more complex algorithms on the training set is a textbook demonstration of the bias-variance trade-off in machine learning. More complex models like random forest and gradient boosting can over-fit training data, learning noise specific to the training examples rather than generalizable patterns. For relatively small datasets with limited features, simpler models often generalize better.
The most important variable across nearly all models was age at diagnosis - younger patients consistently had dramatically better survival. N stage (lymph node involvement) was consistently the second most important feature, highlighting that lymph node metastasis is the single greatest disease-specific risk indicator. This aligns with established oncology knowledge and validates the models' clinical coherence.
This study provides the first machine learning-based survival prediction models specifically designed for GE-ASqD, a cancer subtype that has been understudied despite its distinct clinical features. The models achieve AUC values around 0.74-0.78 on independent test data - clinically useful but not yet at the level needed for autonomous decision-making.
The authors envision these models being used in personalized medicine to guide treatment planning: identifying high-risk patients who warrant more aggressive intervention, and low-risk patients who might avoid unnecessary treatment toxicity. In an era where targeted therapies and immunotherapy (such as checkpoint inhibitors combined with PARP inhibitors) are transforming gynecologic oncology outcomes, better survival stratification tools become increasingly important for directing the right treatments to the right patients.
Key limitations include the use of SEER data, which lacks some clinically relevant variables (such as molecular subtype, body mass index, and detailed treatment response data), and the binary outcome variable (5-year overall survival) which does not capture recurrence or quality-of-life outcomes. Future work should incorporate molecular markers and prospective data to build more comprehensive predictive models for this specific histological subtype.