Using Machine Learning to Predict the Prognosis in Endometrial Cancer Patients Undergoing Fertility-Sparing Treatment

Cancer Med 2025 AI 5 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
Preserving Fertility in Young Endometrial Cancer Patients

Endometrial cancer typically affects women after menopause, but an estimated 4-14% of cases occur in women of reproductive age who have not yet had children. Standard treatment - surgical removal of the uterus - would permanently end any chance of pregnancy. For carefully selected younger patients, an alternative called fertility-sparing treatment (FST) uses hormone therapy (progestins) to shrink the tumor without removing the uterus.

FST can work well, but it carries real risks: about 30-40% of patients relapse (the cancer returns) after initially responding to treatment. Predicting who will succeed with FST and who will relapse is crucial for counseling patients and deciding whether the approach is appropriate. Currently, there is no reliable test to make this distinction before starting treatment.

This study from China used machine learning applied to both MRI-derived radiomic features and standard clinical data to build a prediction model for FST outcomes in 102 endometrial cancer patients - one of the larger datasets for this rare clinical scenario.

TL;DR: Young endometrial cancer patients may preserve fertility using hormone therapy instead of surgery, but 30-40% relapse. This study built an AI model to predict who will respond well versus relapse.
Pages 2-4
Building the Model: Radiomics Plus Clinical Factors

102 patients who underwent FST were included, split into a training set (71 patients) and a test set (31 patients). All patients had pretreatment MRI scans. From these scans, researchers manually outlined the tumor region and extracted radiomic features - hundreds of mathematical measurements describing the tumor's shape, texture, and intensity patterns.

Using a technique called LASSO regression (which automatically selects the most informative features while discarding redundant ones), 13 radiomic features were selected from an initial pool of over 800. Six machine learning models were built using these features and compared: logistic regression, SVM, random forest, XGBoost, LightGBM, and a neural network. Logistic regression performed best on the test set.

Separately, five clinical predictors were identified through statistical analysis: HE4 blood marker level, presence of myometrial invasion, polycystic ovary syndrome (PCOS) diagnosis, insulin resistance, and maximum tumor diameter. A combined model merging radiomic and clinical scores was also constructed and showed the strongest overall performance.

TL;DR: Radiomic features from MRI were combined with 5 clinical variables. LASSO selected 13 features from 800+ candidates. Logistic regression was the best single model; combining radiomic and clinical scores performed best overall.
Pages 4-6
Impressive Accuracy, Especially When Combined

The combined radiomics-plus-clinical model achieved an AUC of 0.941 in the training set and 0.907 in the independent test set - a very strong performance that held up well in unseen patients. This means the model correctly ranked approximately 90% of patient pairs (one relapse, one no-relapse) in the correct order by risk.

For comparison, the radiomics-only model achieved AUC 0.872 (train) / 0.853 (test), while the clinical-only model reached 0.888 (train) / 0.862 (test). The combined model was better than either alone, confirming that MRI texture information and blood/clinical data capture different aspects of the biology that both contribute to predicting outcomes.

Decision curve analysis confirmed that the combined model provided net clinical benefit across a range of decision thresholds - meaning it would actually help doctors make better treatment choices compared to treating all patients the same way. Calibration was also good, meaning predicted probabilities matched actual relapse rates closely.

TL;DR: The combined model hit AUC 0.941 (training) and 0.907 (test). Radiomics and clinical data each contributed unique information; together they outperformed either alone. Decision curve analysis confirmed real clinical utility.
Pages 6-7
Five Clinical Red Flags for Relapse Risk

The five clinical predictors identified for the model are worth understanding individually. HE4 is a protein biomarker measured in blood - elevated levels are associated with more aggressive endometrial cancer. Myometrial invasion (tumor growing into the uterine muscle) even in its early form is a sign of more invasive biology. Both factors suggest the tumor may not respond as well to hormone-only therapy.

PCOS (polycystic ovary syndrome) and insulin resistance are metabolic conditions that create a hormonal environment favoring endometrial cancer growth through chronically elevated estrogen and insulin. Their presence may indicate that the same underlying metabolic driver keeping the cancer alive could interfere with progestin-based treatment.

Maximum tumor diameter on imaging is a straightforward measure of disease burden. Larger tumors tend to be harder to eliminate with hormonal therapy alone. Together, these five factors create a practical clinical profile: metabolically driven, larger, invasive tumors in patients with elevated HE4 should be considered higher risk for FST failure.

TL;DR: Five factors predicted FST failure: elevated HE4, myometrial invasion, PCOS, insulin resistance, and larger tumor size. Together they paint a picture of metabolically aggressive tumors that resist hormone therapy.
Pages 7-8
Limitations and What Comes Next

The study has important limitations. With only 102 patients, this is a relatively small dataset for machine learning - the models may be overfit to this specific patient population. The retrospective design means data was collected from existing medical records rather than a prospectively planned study, which can introduce selection biases.

All patients came from a single center in China. Chinese patient demographics, local MRI scanner protocols, and FST selection criteria may differ from those used in Europe or North America, limiting how broadly these results apply. External validation at different hospitals with different patient populations is essential before this tool could be recommended for clinical use.

Despite these limitations, the study represents a meaningful step. The combination of radiomic and clinical features achieved near-0.91 AUC in independent test patients - a level that, if replicated externally, would be genuinely useful for clinical counseling. Prospective multicenter studies should now be designed to validate and refine this approach.

TL;DR: Key limitations: 102 patients from a single Chinese center, retrospective design. Strong results are promising but need validation at independent hospitals before clinical adoption.
Citation: Open Access, 2025. Available at: PMC12633650.