Atypical endometrial hyperplasia (AEH) is a precancerous condition of the uterine lining. Women diagnosed with AEH are typically referred for surgical removal of the uterus (hysterectomy), but before surgery, many do not know whether they already harbor concurrent invasive cancer. Studies suggest that up to 40% of AEH cases contain concurrent endometrial cancer when the uterus is examined after surgery.
This high rate of concurrent cancer creates a clinical dilemma. If surgeons knew in advance which AEH patients already have cancer - particularly aggressive, high-risk cancer - they could plan more extensive operations (including lymph node assessment) from the start, rather than discovering cancer at pathology after a less complete initial surgery.
Conversely, if surgeons knew that a patient's AEH is very unlikely to harbor invasive cancer, they could counsel younger women about fertility-preserving options or plan a less extensive procedure for older women with multiple comorbidities. Accurate preoperative prediction of concurrent cancer is therefore clinically valuable from multiple directions.
This retrospective observational study collected data from 13 Italian gynecology centers, enrolling 629 eligible women with a preoperative AEH diagnosis who underwent hysterectomy between 2015 and 2020. The multi-center design is a strength: it captures patient diversity from multiple clinical contexts, making findings more generalizable than a single-institution study.
Patient characteristics collected included age, BMI, menopausal status, parity, diabetes, hypertension, Lynch syndrome, prior breast cancer, prior tamoxifen therapy (a breast cancer drug known to increase endometrial cancer risk), abnormal uterine bleeding, and the method of endometrial sampling performed. These are all variables routinely available in any gynecology clinic.
Multiple types of AI prediction models were built and compared: linear regression, quadratic regression, artificial neural networks (ANNs), support vector machines (SVMs), and random forests. All models were evaluated using 10-fold cross-validation - dividing the dataset into 10 parts, training on 9 and testing on 1 repeatedly - to provide reliable estimates of performance on unseen data.
Of 629 women with AEH, 193 (30.7%) had concurrent endometrial cancer at hysterectomy - confirming that concurrent cancer is a common finding, not a rare exception. Among those 193 cancer cases, 26 (13.4%) had high-risk disease that would have required more extensive surgery had it been known preoperatively, including lymph node assessment.
In the univariate analysis comparing women with and without concurrent cancer, several patient characteristics showed statistically significant associations: older age, higher BMI, postmenopausal status, Lynch syndrome, and prior breast cancer were all associated with a higher probability of concurrent cancer. These are consistent with known endometrial cancer risk factors.
Interestingly, some expected risk factors - including hypertension, diabetes, and abnormal uterine bleeding - did not show statistically significant differences between the cancer and non-cancer groups in this cohort. This highlights that even the most established risk factors may not reliably discriminate within a population already selected for having AEH.
Despite testing many variable combinations and multiple AI methods, prediction performance plateaued at a mean AUC of approximately 0.65 for regression models and 0.75 for artificial neural networks on the validation sets. An AUC of 0.75 represents moderate - not strong - discriminative ability: meaningfully better than random guessing, but far from the 0.90+ values considered clinically reliable for a decision-support tool.
Adding more patient characteristics to the models beyond the most informative variables did not improve test-set AUC, and in some cases caused a decline - a classic pattern of overfitting, where models learn noise in the training data rather than generalizable patterns. The training-set AUC was consistently higher than the test-set AUC, confirming this phenomenon.
The most recurrent patient characteristics across the best-performing models were age, BMI, Lynch syndrome, diabetes, and prior breast cancer. These five variables appeared repeatedly across different model types and variable combinations, suggesting they carry the most genuinely informative signal in the dataset - even if that signal is not strong enough for high-accuracy individual prediction.
The study's central conclusion is that patient demographic and clinical characteristics alone do not contain enough information to reliably predict concurrent endometrial cancer in AEH. While associations exist and the AI models found them, the prediction accuracy is too limited to safely guide major surgical decisions for individual patients.
This finding points to what is missing: deeper biological information about the tumor itself. Molecular and genetic variables - such as POLE mutation status, microsatellite instability, specific somatic mutations, or imaging-based tumor features - likely contain predictive information that patient-level demographic data cannot capture. Future models incorporating these richer inputs may achieve clinically actionable accuracy.
The result also has a direct practical implication: clinicians should not rely primarily on patient characteristics when counseling AEH patients about concurrent cancer risk. The current surgical standard of recommending hysterectomy for most AEH patients remains justified precisely because the risk of concurrent cancer cannot be reliably excluded by clinical risk profiling alone.
While acknowledging the current limitations, this study provides an important benchmarking contribution: it rigorously quantifies how much predictive information patient characteristics actually contain for this clinical question. Future researchers can use these AUC values as a baseline to determine whether new variables (imaging, molecular) add meaningful incremental prediction beyond demographic data.
The study suggests a promising path forward: combining patient characteristics with preoperative molecular testing of AEH specimens (e.g., testing for POLE mutations or mismatch repair deficiency, which are now part of standard endometrial cancer molecular classification) could provide the additional biological information needed to improve prediction accuracy substantially.
For the subset of AEH patients with Lynch syndrome - who showed strong association with concurrent cancer - the findings reinforce the importance of identifying Lynch syndrome carriers before surgery. Women with Lynch syndrome warrant particularly careful preoperative evaluation and should be counseled about their substantially elevated cancer risk when planning surgical management of AEH.