Endometrial cancer is the most common gynecologic malignancy in developed countries, and its incidence is rising. While approximately 75% of cases are diagnosed at an early stage (FIGO stages I or II) - where the five-year survival rate ranges from 74% to 91% - a meaningful subset of these patients will still experience recurrence after treatment.
When endometrial cancer recurs or spreads to distant sites, complete recovery becomes very difficult and treatment shifts toward palliative care. Identifying which early-stage patients are most at risk of recurrence would allow doctors to tailor follow-up intensity and additional treatments more precisely.
Traditionally, the FIGO staging system - based on how far cancer has physically spread at diagnosis - has been the main tool for estimating outcomes. However, other factors such as patient age, tumor grade, cancer subtype, and treatment choices also play important roles that a simple staging number cannot capture.
This pilot study explored whether machine learning - a form of artificial intelligence that finds patterns in complex data - could better predict which early-stage endometrial cancer patients would experience recurrence by combining multiple clinical variables simultaneously.
The study enrolled 75 patients with early-stage endometrial cancer (FIGO stage I or II) who underwent surgery at a single Japanese institution between 2013 and 2019. Of these, 63 had no recurrence within five years and 12 experienced recurrence.
Each patient was represented by 16 clinical features: age, body weight index, gravidity, parity, FIGO stage, tumor grade, hypertension, diabetes, three tumor marker measurements (CA125, CA19-9, CEA), the type of hysterectomy performed, whether lymph node dissections were performed, and the number of chemotherapy cycles received.
Five different machine learning classifiers were tested: support vector machine (SVM), random forest (RF), decision tree (DT), logistic regression (LR), and boosted tree. Each of these algorithms uses a different mathematical strategy to find patterns linking the 16 input features to whether a patient will relapse.
The dataset was split randomly - 80% for training the models and 20% for testing them on unseen patients. Performance was measured using accuracy (proportion of correct predictions) and AUC (area under the receiver operating characteristic curve), which captures how well a model distinguishes recurrence from non-recurrence cases across different decision thresholds.
The random forest (RF) classifier achieved the highest accuracy at 80%, followed by support vector machine at 77%, logistic regression at 73%, boosted trees at 70%, and decision tree at 66%. However, when measured by AUC - which is a more reliable metric for imbalanced datasets - logistic regression performed best at 0.53, suggesting the models were not dramatically better than chance at ranking patients by recurrence risk.
All AUC values were modest, ranging from 0.47 to 0.53 across models. For comparison, a completely random guess would produce an AUC of 0.50. These numbers indicate the models captured some signal but were far from clinically useful on this small dataset.
The tree-based models (RF, DT, boosted tree) allowed researchers to examine which features were most important to their predictions. Across all three models, age, FIGO stage, and tumor markers (CEA and CA125) consistently ranked as the most informative predictors of recurrence. The number of chemotherapy courses was also considered important.
Looking at the raw patient data, the recurrence group had a median age of 69 years versus 57 in the non-recurrence group - a statistically significant difference. Carcinosarcoma (a more aggressive cancer subtype) was also more frequent in the recurrence group, while standard endometrioid carcinoma was more common in the non-recurrence group.
The authors are candid that the main limitation of this study is the small sample size. With only 75 patients and 12 recurrence events, the dataset is too small for machine learning algorithms to reliably identify complex non-linear patterns. Deep learning and advanced AI techniques generally require thousands or tens of thousands of patients to work effectively.
Traditional approaches like nomograms - graphical prediction tools based on simple statistical models - have been used for endometrial cancer prognosis before. These tools work well for capturing linear relationships between a few variables, but they cannot model the complex, non-linear interactions between many factors that are present in real patient data.
Machine learning's theoretical advantage is precisely its ability to model these non-linear relationships and interactions without requiring the researcher to manually specify which factors matter. A model might learn, for example, that high age combined with high CA125 and carcinosarcoma histology is particularly dangerous - a three-way interaction that a traditional statistical model might miss.
The authors note that for AI to be truly useful in this setting, multi-institutional databases with thousands of patients would be needed. This could be achieved through national registries or collaborative research networks that pool data across hospitals and countries.
Accurately predicting which patients are at high risk of recurrence has several direct clinical benefits. It can guide decisions about adjuvant treatment - the additional therapies (chemotherapy, radiation) given after surgery. High-risk patients might benefit from more intensive post-surgical treatment while low-risk patients could be spared unnecessary side effects.
Recurrence prediction also informs follow-up scheduling. Current practice typically involves regular check-ups with imaging and blood tests for all patients, but a risk-stratified approach could mean more frequent, intensive monitoring for high-risk patients and less frequent visits for low-risk patients.
For patients who are found to be at low recurrence risk, an AI prediction tool could provide meaningful reassurance, reducing anxiety and potentially improving quality of life during the years of post-treatment surveillance.
The study demonstrates proof-of-concept that clinical variables routinely collected during standard care - without requiring expensive genomic testing or specialized imaging - can feed directly into machine learning models. This makes the approach potentially applicable even in resource-limited healthcare settings.
This study confirms it is feasible to apply machine learning to predict endometrial cancer recurrence from routine clinical data. The fact that even a 75-patient dataset produced AUC values above chance demonstrates that the approach has a real signal worth pursuing.
The most important next step is larger datasets. The authors estimate that datasets of over ten thousand patients would be needed for AI models to reach clinically meaningful predictive performance. Multi-institutional collaborations or national cancer registry data would be the most practical path to achieving this.
Future studies should also explore incorporating additional data types, such as molecular or genomic markers, imaging features, and pathology findings. Modern AI architectures can integrate these diverse data streams simultaneously, potentially unlocking prediction accuracy not achievable from clinical variables alone.
This pilot lays the conceptual and methodological groundwork for larger studies and illustrates both the promise and current limitations of machine learning for endometrial cancer recurrence prediction. It represents an important early step toward AI-driven personalized follow-up care.