Concurrent Endometrial Cancer in Women with Atypical Endometrial Hyperplasia: What Is the Predictive Value of Patient Characteristics?

Simple Summary The rate of concurrent endometrial cancer (EC) in women with atypical endometrial hyperplasia (AEH) is not negligible. Furthermore, among women with EC, about 12% may have a high-risk disease requiring lymph node status assessment. Given that endometrial sampling cannot exclude EC in women with AEH, knowing variables that increase the risk of malignancy can be helpful in clinical practice. Some patient characteristics were associated with this occurrence, representing possible risk factors on which to adjust treatment planning. No prediction models with internal validation showed the impact of patient characteristics in predicting EC after a preoperative diagnosis of AEH. The present study, using regressions and artificial neural networks, found recurrent patient characteristics in women with EC. However, they likely do not contain good/optimal discriminating information. Future predictive models should include other individual factors (e.g., genotypic variables) to move toward more personalized medicine. Abstract Background: The rate of concurrent endometrial cancer (EC) in atypical endometrial hyperplasia (AEH) can be as high as 40%. Some patient characteristics showed associations with this occurrence. However, their real predictive power with related validation has yet to be discovered. The present study aimed to assess the performance of various models based on patient characteristics in predicting EC in women with AEH. Methods: This is a retrospective multi-institutional study including women with AEH undergoing definitive surgery. The women were divided according to the final histology (EC vs. no-EC). The available cases were divided into a training and validation set. Using k-fold cross-validation, we built many predictive models, including regressions and artificial neural networks (ANN). Results: A total of 193/629 women (30.7%) showed EC at hysterectomy. A total of 26/193 (13.4%) women showed high-risk EC. Regression and ANN models showed a prediction performance with a mean area under the curve of 0.65 and 0.75 on the validation set, respectively. Among the best prediction models, the most recurrent patient characteristics were age, body mass index, Lynch syndrome, diabetes, and previous breast cancer. None of these independent variables showed associations with high-risk diseases in women with EC. Conclusions: Patient characteristics did not show satisfactory performance in predicting EC in AEH. Risk stratification in AEH based mainly on patient characteristics may be clinically unsuitable.


Introduction
Atypical endometrial hyperplasia (AEH) is a premalignant lesion that may reveal concurrent cancer on definitive histology in 40% of cases [1,2].Moreover, 10-15% of women with endometrial cancer (EC) may have a high-risk disease requiring lymph node status assessment [3][4][5].Current guidelines suggest nodal staging should be performed in intermediate-high-risk/high-risk diseases.The sentinel lymph node technique represents an acceptable alternative to systematic lymphadenectomy for lymph node staging in stages 1 and 2 [6].
It is known that these women often do not undergo oncological work-up, and definitive histology may reveal a stage of the disease that would have required more radical surgery [7].Moreover, it should be considered that conservative treatment may be an option in women of childbearing age [6].Therefore, an accurate diagnosis is of pivotal importance.
Some EC predictors may affect the rate of unexpected cancer in AEH [8][9][10][11][12].Matsuo et al. studied 211 women with EH and showed that older age, obesity, diabetes mellitus, and complex atypical hyperplasia were associated with concurrent EC in AEH [10].Given that endometrial sampling cannot exclude EC in women with AEH, knowing those variables that increase the risk of malignancy can be helpful in clinical practice.
Although some independent variables showed associations with cancer, there have been no attempts to build prediction models with related validation.In the latter studies, multivariate analyses were used to assess only the presence or absence of associations between independent variables and the presence of EC in women with a previous diagnosis of AEH [8][9][10][11][12].Testing independent variables in the training and validation set would allow us to test the real predictive power of patient characteristics to assess their clinical utility.Moreover, the study of variables in the sub-population of women with high-risk diseases would enable us to measure their impact in identifying patients needing more radical surgery.
The present study aimed to create and assess prediction models of EC in women with AEH using various regression and artificial intelligence (AI) methods.

Study Design and Setting
This retrospective observational study included women with a preoperative diagnosis of AEH and then undergoing definitive surgery within 40 days between January 2015 and December 2020.The histological reference standard was represented by hysterectomy.Women with incomplete data, previous events of AEH managed with medical treatment, and AEH diagnosis on endometrial structural lesions (e.g., polyps) were excluded.
Based on Italian law, the Ethics Committee (Comitato Etico Regionale Marche) took note of the study (Prot.338/2021) [13].According to Italian law, patient consent was not mandatory in a retrospective study [13].
Patient characteristics taken into account were age, parity, menopausal status, body mass index (BMI), presence of hypertension, or diabetes, Lynch syndrome, hormonal therapy, smoking, previous tamoxifen therapy (previous use of Tamoxifen for 5 years), abnormal uterine bleeding, previous breast cancer, and endometrial sampling method (D&C: dilation and curettage, HSC-bio: hysteroscopically guided biopsy, and HSC-res: hysteroscopically endometrial resection).
Descriptive patient characteristics also included the type of surgery (laparotomic, laparoscopic, vaginal).According to the classification system for EC, women were classified as low-risk, intermediate-risk, high-intermediate-risk, and high-risk [6].

Sample Size Calculation
Based on previous data for prediction models, we considered a minimum of 10 events (EC) for each predictor variable [16].Since we included 16 independent variables, the minimum required sample size should consist of 160 cases of EC.

Statistical Methods
The final histology (dependent variable) divided the women into patients with or without cancer.The Kolmogorov-Smirnov test was used to test continuous variable distribution.The univariate analysis was used to compare all studied independent variables.The non-parametric Mann-Whitney U-test for two independent samples was used for not normally distributed values.Comparisons between categorical variables were performed using the Chi-squared test.
The performance of endometrial cancer prediction methods was assessed according to Table S1 scheme.The cancer prediction models were trained and tested on the definitive histological reference standard (hysterectomy).All statistical analyses were performed using MATLAB Software (MATLAB R2022a, The MathWorks, Inc., Natick, MA, USA).p < 0.05 was considered to indicate statistical significance.
Apart from the age and BMI variables, all the others were considered categorical.They were assigned the value 0 (absent) or 1 (present) and passed through the "categorical" function of Matlab.The regression models were implemented using the Matlab functions "fitlm", "fitglm", and "mnrfit".We specify that Matlab's "fitlm" function can handle categorical variables when the options "bisquare" or "logistic" or "fitglm" function was used with the "Distribution: binomial" option.
Regarding regressions, we also used the parameter "linear" for linear regression ("model contains an intercept and linear term for each predictor"), or "interactions" that is a model containing an intercept, a linear term for each predictor, and all products of pairs of distinct predictors (no squared terms) or "pure quadratic" that is a model containing an intercept term and linear and squared terms for each predictor, or "quadratic" for quadratic regression with interactions.We used 10-fold cross-validation in which the train set contained 80% of the cases and the test set 20%.This was implemented using the "cvpartition" function with the "kfold" option.In particular, we wanted to assess which variables had the most discriminating power to reduce the dimensionality of the model.
Other predictors were constructed using artificial neural network (ANN) algorithms using the "patternnet" Matlab function.After running several tests with different nodes, we built predictive models with many nodes proportional to the number of variables in the model.In this way, the complexity was proportional to the number of variables.We found the configurations that provided the best performance.Also, for ANN, we have implemented 10-fold cross-validation.We also used the SVM (support vector machine) and RF (random forest) AI algorithms.
Regarding the prediction models obtained on the same variables, we showed the five combinations of variables associated with the best AUC.Finally, recurrent patient characteristics were considered those variables present in at least four of five prediction models.

Study Population
The study period included 730 consecutive women with AEH undergoing major surgery.After excluding 101 cases, 629 eligible women were analyzed (Figure 1).Patient characteristics are reported in Table 1.The surgical approach is reported in Table S2.Univariate analysis results comparing women with or without cancer on hysterectomy are reported in Table 2.
for quadratic regression with interactions.We used 10-fold cross-validation in which the train set contained 80% of the cases and the test set 20%.This was implemented using the "cvpartition" function with the "kfold" option.In particular, we wanted to assess which variables had the most discriminating power to reduce the dimensionality of the model.
Other predictors were constructed using artificial neural network (ANN) algorithms using the "patternnet" Matlab function.After running several tests with different nodes, we built predictive models with many nodes proportional to the number of variables in the model.In this way, the complexity was proportional to the number of variables.We found the configurations that provided the best performance.Also, for ANN, we have implemented 10-fold cross-validation.We also used the SVM (support vector machine) and RF (random forest) AI algorithms.
Regarding the prediction models obtained on the same variables, we showed the five combinations of variables associated with the best AUC.Finally, recurrent patient characteristics were considered those variables present in at least four of five prediction models.

Study Population
The study period included 730 consecutive women with AEH undergoing major surgery.After excluding 101 cases, 629 eligible women were analyzed (Figure 1).Patient characteristics are reported in Table 1.The surgical approach is reported in Table S2.Univariate analysis results comparing women with or without cancer on hysterectomy are reported in Table 2.
From here on, we built models to predict whether AEH patients also had concomitant EC.Although Table 2 showed variables not significantly related to cancer, we initially decided not to discard any variables since the models can capture correlations between unrelated variables to cancer.From here on, we built models to predict whether AEH patients also had concomitant EC.Although Table 2 showed variables not significantly related to cancer, we initially decided not to discard any variables since the models can capture correlations between unrelated variables to cancer.

Regression-Based Predictors
To test the real predictive power of different kinds of regressions (linear, interactions, quadratic, and purequadratic), we trained the same models across multiple data sets (training) and tested them on different sets (test).In particular, we wanted to assess which variables had the most discriminating power to reduce the dimensionality of the model.We then performed the linear regression with all numbers and combinations of variables.Regarding the regressions performed on the same variables, we showed the five combinations of variables associated with the best AUC (Table S3).It is noted that the AUC calculated in the train set increased as the number of variables increased.This is perfectly normal and does not necessarily imply a good performance in prediction.The AUC calculated on the test sets increased until it reaches the maximum of about 0.65 in correspondence with a 5-variable model, among which the most common were age, BMI, Lynch syndrome, abnormal uterine bleeding, and previous breast cancer.Linear regression models that used more than these five variables did not improve performance or even progressively deteriorated it despite increasing the AUC of the training set.
We also tested whether a quadratic regression improved performance thanks to capturing nonlinear interactions between variables (Table 3).This model achieved the maximum AUC (0.65) in the test set using a combination of four variables.The most recurrent patient characteristics were age, BMI, and Lynch syndrome.However, the best linear and quadratic regressions yielded a similar performance with a sensitivity and a specificity of around 61%.
We did not report the results of the interactions and pure quadratic regressions because they returned slightly lower performance than the previous tests.

Artificial Intelligence-Based Predictors
To assess if better performance could be obtained, we have adopted algorithms ("patternnet" function) that use artificial intelligence to capture deeper non-linearities present in the data, providing greater efficiency than regression algorithms.We tested networks with a proportional number of nodes for each number of variables N (from 2 to 15).For example, from 2*N to 5*N nodes.The best results were obtained with 3*N nodes (Table 4) reaching an AUC in the test set of about 0.75, including ten variables.Sensitivity ranged from 66% to 76%, while specificity ranged from 59% to 67%.These values may not be considered satisfactory for a prediction model.The most recurrent patient characteristics among the best five predictive models were: age, body mass index, Lynch syndrome, diabetes, and previous breast cancer.The performance did not improve significantly with a higher number of nodes.Unfortunately, with a slight improvement compared to the regression models, these use many more variables (i.e., 10 to 12).In this last case, the AUC reached a value of 0.76 (CI 95%: 0.706-0.814),with a sensitivity and specificity of 68.1% and 68.9%, respectively.These results demonstrate a relationship between the indicated variables and the EC, but they do not reach significant values for a clinical impact.We also tested a feedforward, fully connected neural network for regression implemented using the "fitrnet" Matlab function, but the performance was slightly lower than the previous algorithm (patternnet).underestimation of EC than D&C (28% vs. 35%, respectively).In this regard, in women with AEH, endometrial sampling procedures should prefer a hysteroscopic assessment, especially in patients who are candidates for fertility-sparing treatment [6].The explanation of our results cannot be separated from further research implications.It is known that there are estrogen-related forms of EC in which BMI and other clinical variables play a pivotal role [23].On the other hand, there are non-estrogen-dependent ECs where the patient characteristics are less decisive [23].The Proactive Molecular Risk Classifier for Endometrial Cancer (ProMisE) groups stratified different phenotypes [24].The POLE-mt women were younger and had the lowest BMI, most frequently at FIGO stage 1.The p53-wt group includes women with high BMI.The p53-abn women have advanced age, with a high prevalence of adjuvant treatment and less frequently at FIGO stage 1.Finally, MMR-d women provided intermediate values for all clinical features between the four ProMise groups [24].This clinical characterization can suggest different pathogenetic mechanisms and explain how the same patient characteristics at different values may identify different sub-populations of women with EC.Likewise, this may explain why such independent variables may appear to have little discrimination in prediction models when genetic mutation assessment is not included.
Recent studies added information on the progression model of endometrial hyperplasia to cancer [23].It is well known that the evolution of cancer is driven by estrogenic stimulation.However, new evidence suggests a crucial role of some gene mutations (PTEN, PIK3CA, FGFR2, ARID1A, and MYC) in cancer progression from endometrial hyperplasia [23].In endometrial hyperplasia progressing to cancer, more mutations in oncogenic signaling pathways have been found as occurs in endometrial cancer [23].We know that areas of AEH may coexist in women with EC.Also, for AEH, a gene panel assessment could be added to standard histological evaluation to better discriminate those cases of unrecognized ECs and modify the management accordingly.
Blood chemical markers could represent further predictive factors to be investigated.Increased preoperative and postoperative D-Dimer levels predicted high-risk EC [25].Furthermore, the combination of HE4, d-dimer, fibrinogen, and CA199 showed good diagnostic accuracy in predicting EC in symptomatic women [26].The study of blood chemical markers and genotypic variables could increase the predictive power of the various models to have a clinical impact.
The limitations of the present study include (i) its retrospective nature; (ii) some independent variables of importance may be missing.Endometrial ultrasound characterization could have a predictive weight; (iii) the reproducibility of endometrial hyperplasia is poor among pathologists.However, the latest atypical and non-atypical classifications can mitigate this limitation [14,15].The study's strengths include the appropriate sample size of EC cases in relation to the number of variables studied.Endometrial sampling methods included 85% of women undergoing hysteroscopy, minimizing blind biopsies.It should be noted that our study had several referral cancer centers to make the results more generalizable.

Conclusions
The study of patient characteristics for predicting EC in women with AEH showed disappointing performance using regressions and ANNs.Although some variables were recurrent in EC, they likely do not contain good/optimal discriminating information.So, a risk stratification of AEH based on patient characteristics can be misleading in clinical practice.As the first predictive models in this area, they could be regarded as the starting point for future models that should include other individual factors (e.g., genotypic variables) to move towards more personalized medicine.

Supplementary Materials:
The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/cancers16010172/s1,Table S1: Scheme used to evaluate the performance of prediction models; Table S2: Surgical characteristics; Table S3: Results of the linear regression

Table 2 .
Univariate analysis comparing women with or without endometrial cancer at hysterectomy.

Table 3 .
Results of the quadratic regression performed on all number and variable combinations.

Table 4 .
Results of the Patternnet algorithm trained with the 10-fold cross-validation.