Decision Tree Model for Predicting the Overall Survival of Endometrial Cancer Patients with Adenomyosis

Adenomyosis is a benign gynecological condition characterized by the presence of endometrial glands and stroma within the myometrium, often resulting in an enlarged uterus and symptoms such as dysmenorrhea, menorrhagia, and chronic pelvic pain. 1 , 2 Although traditionally regarded as a non-neoplastic disorder, adenomyosis and endometrial cancer share several pathogenic features, including estrogen-dependent proliferative signaling, chronic inflammatory microenvironmental alterations, and overlapping molecular events such as PTEN loss and KRAS mutations. 3–5 These shared pathways have prompted investigation into the clinical significance of adenomyosis coexisting with endometrial malignancy. EC is the most common malignancy of the female reproductive tract in developed countries. It typically presents with abnormal uterine bleeding, and its prognosis largely depends on early detection and histopathological characteristics. 6 , 7 The coexistence of adenomyosis and EC in surgical specimens has raised questions about a potential pathophysiological link and the impact of adenomyosis on tumor behavior, staging accuracy, and long-term prognosis. 8 , 9 Prior studies have reported that adenomyosis is associated with favorable clinicopathological features, including lower tumor grade, reduced lymphovascular space invasion (LVSI), and earlier stage at diagnosis. 10 , 11 However, whether adenomyosis confers independent prognostic significance after adjustment for established risk factors remains uncertain, as several large cohort studies and meta-analyses have yielded inconsistent conclusions. 10–12 Decision tree models have emerged as interpretable tools in oncology, offering visual frameworks for clinical decision support that reveal complex variable interactions less apparent in multiplicative hazard models. 13–15 However, when applied as binary classifiers predicting vital status, these models do not incorporate censoring or the timing of events, and thus complement rather than replace time-to-event analyses such as Cox proportional hazards regression. In the context of EC, decision tree approaches remain underutilized for survival prediction, particularly in the subpopulation of patients with coexisting adenomyosis. This study had two distinct objectives. The first was to evaluate whether coexisting adenomyosis is associated with clinicopathological features and survival outcomes in a retrospective EC cohort. The second was to develop an exploratory, interpretable decision tree model for overall survival prediction, incorporating adenomyosis among other clinically established candidate variables, and to identify the most informative predictors within this framework. While prior investigations from our group and others have characterized the baseline clinicopathological profiles of patients with coexisting adenomyosis, 1 , 8 they did not evaluate long-term survival outcomes using machine learning paradigms or address non-proportional hazards. The current study represents a legitimate extension of this prior research, providing a novel and distinct contribution by integrating standard Cox modeling, Restricted Mean Survival Time (RMST) diagnostics, and machine learning decision tree and time-to-event survival tree architectures to map the prognostic landscape of these patients.

Of 400 patients, 69 (17.3%) had adenomyosis and 331 (82.7%) did not. Age and menopausal status were similar between groups (mean age 59.16 ± 7.81 vs 58.51 ± 9.96 years, p =0.609; postmenopausal 76.8% vs 74.6%, p =0.703). Parity and mode of delivery did not differ ( p =0.054 and p =0.288, respectively). Gravidity distributions showed a small but statistically significant difference ( p =0.031). Median preoperative CA-125 was higher in the adenomyosis group (16.70 [8.90–28.00] vs 14.20 [9.50–22.40] U/mL, p<0.001), although both medians remained below the clinical threshold of 35 U/mL. Surgical approach was comparable (p=0.574). The median follow-up time was 98 months (IQR 72–134). Patients with adenomyosis were more often early stage (IA1–IIB) (97.1% vs 88.5%); the odds of advanced stage were lower (OR 0.23; 95% CI 0.05–0.98; p=0.047). LVSI was less frequent with adenomyosis (10.1% vs 26.3%; OR 0.31; 95% CI 0.13–0.71; p=0.006). Myometrial invasion 50% showed a non-significant trend toward lower frequency (p=0.053). Histology, histologic and nuclear grades, and other anatomic involvement parameters were not significantly different (all p>0.05). Crude recurrence/metastasis rates were similar (10.1% vs 5.7%; p=0.183). During follow-up, 78 deaths and 52 recurrences occurred. In multivariable Cox models, coexisting adenomyosis was not an independent predictor of OS (HR 0.62, 95% CI 0.32–1.20, p=0.153) or DFS ( Figures 1 and 2 ). To resolve the crossing of survival curves beyond 125 months, RMST analysis up to 150 months confirmed no significant long-term survival differences (OS difference: +2.2 months, p=0.553). Landmark analysis revealed that adenomyosis showed a strong trend toward improved survival in the first 120 months (HR 0.42, p=0.052) that reversed in the late follow-up period (HR 1.48, p=0.461). Sensitivity analysis using LASSO penalized regression confirmed coefficient stability, with adenomyosis consistently excluded from the final survival predictors. The exploratory decision tree ( Figure 3 ) and the censored Survival Tree both identified older age, LVSI, and deep myometrial invasion as the primary determinants of survival, while adenomyosis was not selected in either model. The decision tree demonstrated good calibration (Brier score: 0.18; Hosmer–Lemeshow p=0.918). Figure 1 Kaplan–Meier curves for overall survival (OS) by adenomyosis status. Kaplan–Meier estimates of overall survival probability for patients with adenomyosis present (adenomyosis = 1, solid line) versus adenomyosis absent (adenomyosis = 0, dashed line). The log-rank p-value is displayed. Number-at-risk tables are provided below the x-axis. Censoring events are indicated by tick marks. The x-axis shows time in months. Overall Survival (OS) by Adenomyosis Status\nKaplan–Meier survival plot with two step curves and shaded confidence bands. The x-axis is labeled Follow-up Time (months), ranging from 0 to 180 with ticks at 0, 20, 40, 60, 80, 100, 120, 140, 160, 180. The y-axis is labeled Overall Survival Probability, ranging from 0.0 to 1.0 with ticks at 0.0, 0.2, 0.4, 0.6, 0.8, 1.0. Legend lists Adenomyosis Absent (N equals 331) as a dashed curve and Adenomyosis Present (N equals 69) as a solid curve. Text box states Log-rank p-value equals 0.18. An arrow annotation reads Crossover after 125m, pointing near 125 months where the two curves meet. Tick marks on curves indicate censoring. Number at risk table: at 0 months, 331 and 69; 20 months, 325 and 68; 40 months, 320 and 68; 60 months, 316 and 67; 80 months, 288 and 63; 100 months, 222 and 56; 120 months, 168 and 44; 140 months, 104 and 34; 160 months, 47 and 21; 180 months, 0 and 0. Kaplan–Meier plot of overall survival by adenomyosis status, with curves crossing after 125 months. Figure 2 Kaplan–Meier curves for disease-free survival by adenomyosis status. Kaplan–Meier estimates of disease-free survival probability for adenomyosis present (adenomyosis = 1, solid line) versus adenomyosis absent (adenomyosis = 0, dashed line). The log-rank p-value is displayed. Number-at-risk tables are provided below the x-axis. The x-axis shows time in months. Disease-Free Survival (DFS) by Adenomyosis Status is depicted in a Kaplan–Meier graph with two groups and an at-risk table. The x-axis shows Follow-up Time (months) from 0 to 180, while the y-axis indicates Disease-Free Survival Probability from 0.0 to 1.0. The Log-rank p-value is 0.25. The Adenomyosis Absent group (N=331) is represented by a dashed line and the Adenomyosis Present group (N=69) by a solid line. Both start at 1.0 at 0 months, with the solid line staying near 1.0 until about 120 months before declining. The dashed line gradually declines to around 0.85-0.9 between 120 and 140 months. A crossover occurs after 125 months at approximately 0.85 probability. The at-risk table shows the number of participants at intervals: Adenomyosis Absent: 331, 331, 327, 323, 319, 283, 212, 145, 103, 49, 0; Adenomyosis Present: 69, 69, 69, 68, 68, 60, 49, 40, 29, 21, 0. Kaplan–Meier plot of disease-free survival by adenomyosis status with curves crossing after 125 months. Figure 3 Decision-tree model for overall survival prediction. Interpretable decision tree classifier trained on age, adenomyosis, FIGO-2023 stage (advanced vs early), LVSI, deep myometrial invasion (MI) ≥50%, histology (endometrioid vs non-endometrioid), grade 3, CA-125 > 35 U/mL, diabetes mellitus, and gravidity. The optimal model selected by 5-fold cross-validated grid search was depth-limited (max_depth = 3, min_samples_leaf = 15) with balanced class weighting. Test-set performance (internal validation): accuracy 0.81, balanced accuracy 0.72, AUC 0.76, Brier score 0.18 and the Hosmer-Lemeshow goodness-of-fit calibration test (p=0.918). Node annotations show Gini impurity, weighted sample counts (non-integer values reflect balanced class weighting), value distributions [Alive, Deceased], and predicted class. Adenomyosis was not selected as a splitting variable at any node. Decision tree model for overall survival prediction. The root node (N=400) uses age as the splitting predictor, with a Gini impurity of 0.31, classifying as alive. Node 1 (N=245) splits by LVSI, with a Gini impurity of 0.20, leading to Leaf Node 1 (N=212) predicting ′ALIVE (Low Risk)′ and Leaf Node 2 (N=33) predicting ′ALIVE (Moderate Risk)′. Node 2 (N=155) splits by deep myometrial invasion (MI), with a Gini impurity of 0.44, leading to Leaf Node 3 (N=110) predicting ′ALIVE (Moderate Risk)′ and Leaf Node 4 (N=45) predicting ′DECEASED (High Risk)′. Performance metrics include accuracy of 0.81, balanced accuracy of 0.72, AUC-ROC of 0.76, Brier score of 0.18 and Hosmer-Lemeshow goodness-of-fit test p=0.918, indicating high calibration. Decision tree for survival prediction based on age, LVSI and deep MI, with performance metrics. Abbreviations : AUC, area under the receiver operating characteristic curve; LVSI, lymphovascular space invasion; MI, myometrial invasion; FIGO-2023, International Federation of Gynecology and Obstetrics 2023 staging. Kaplan–Meier curves for overall survival (OS) by adenomyosis status. Kaplan–Meier estimates of overall survival probability for patients with adenomyosis present (adenomyosis = 1, solid line) versus adenomyosis absent (adenomyosis = 0, dashed line). The log-rank p-value is displayed. Number-at-risk tables are provided below the x-axis. Censoring events are indicated by tick marks. The x-axis shows time in months. Kaplan–Meier curves for disease-free survival by adenomyosis status. Kaplan–Meier estimates of disease-free survival probability for adenomyosis present (adenomyosis = 1, solid line) versus adenomyosis absent (adenomyosis = 0, dashed line). The log-rank p-value is displayed. Number-at-risk tables are provided below the x-axis. The x-axis shows time in months. Decision-tree model for overall survival prediction. Interpretable decision tree classifier trained on age, adenomyosis, FIGO-2023 stage (advanced vs early), LVSI, deep myometrial invasion (MI) ≥50%, histology (endometrioid vs non-endometrioid), grade 3, CA-125 > 35 U/mL, diabetes mellitus, and gravidity. The optimal model selected by 5-fold cross-validated grid search was depth-limited (max_depth = 3, min_samples_leaf = 15) with balanced class weighting. Test-set performance (internal validation): accuracy 0.81, balanced accuracy 0.72, AUC 0.76, Brier score 0.18 and the Hosmer-Lemeshow goodness-of-fit calibration test (p=0.918). Node annotations show Gini impurity, weighted sample counts (non-integer values reflect balanced class weighting), value distributions [Alive, Deceased], and predicted class. Adenomyosis was not selected as a splitting variable at any node.

This retrospective cohort study was conducted at the Department of Obstetrics and Gynecology, Kartal Dr. Lütfi Kırdar City Hospital, a tertiary care academic institution in Istanbul, Turkey. Medical records of patients who underwent primary surgical management for histologically confirmed EC between January 1, 2008, and December 31, 2018, were reviewed. Inclusion criteria were histologically confirmed primary endometrial carcinoma, complete clinicopathological documentation, and adequate follow-up data. Exclusion criteria comprised incomplete medical records, inadequate follow-up (defined as fewer than six months of postoperative documentation), prior pelvic malignancy, and receipt of neoadjuvant therapy. Among 513 eligible cases, 113 were excluded; the final cohort comprised 400 patients. Demographic, clinical, and comprehensive clinicopathological data were extracted from hospital medical records and pathology reports. All cases were retrospectively restaged according to the 2023 International Federation of Gynecology and Obstetrics (FIGO) staging system, 17 using the operative and pathology reports, with particular attention to revised criteria for myometrial invasion assessment and substaging. All diagnoses were based on final histopathological examination of surgical specimens. When clinically indicated, advanced preoperative imaging modalities, including magnetic resonance imaging (MRI) and contrast-enhanced computed tomography (CT), were utilized to assess disease extent and inform surgical planning. Treatment modalities, including surgical staging, chemotherapy, radiotherapy, and brachytherapy, were individualized based on through tumor board recommendations, consistent with international guidelines. 7 , 16 Overall survival (OS) was defined as the interval between the date of histopathological diagnosis and death from any cause or the date of the last documented follow-up. Disease-free survival (DFS) was defined as the time from diagnosis to the first documented recurrence of disease or death, whichever occurred first. Patients who remained event-free were censored at their last recorded follow-up visit. Median follow-up time was estimated using the reverse Kaplan–Meier method. The presence or absence of adenomyosis was determined histopathologically from the final surgical specimen. Patients were stratified into two groups: adenomyosis-present and adenomyosis-absent groups for comparative analyses. The study protocol was reviewed and approved by the Institutional Review Board of the hospital (Approval No. 2025/010.99/17/17; Date: June 25, 2025). This study was conducted in compliance with the ethical principles of the Declaration of Helsinki. Given its retrospective nature, the ethics committee waived the requirement for informed consent. All collected medical records were thoroughly anonymized prior to analysis, and all personal identifiers (such as names, national identification numbers, and hospital registration numbers) were permanently omitted. Access to the study database was strictly limited to the primary investigators to preserve patient privacy and data security. All statistical analyses were conducted using JASP (version 0.19.3) and Python (version 3.10). Preoperative CA-125, body mass index, and parity had minor missingness (<5% each), which was handled via complete-case analysis and validated using multiple imputation. Survival outcomes (OS and DFS) were evaluated using Kaplan–Meier curves and multivariable Cox proportional hazards regression. Because Kaplan–Meier curves crossed after 125 months, violating the proportional hazards assumption for adenomyosis, we performed Restricted Mean Survival Time (RMST) analysis up to a 150-month horizon and a landmark analysis split at 120 months to capture time-varying effects. To prevent model overfitting, LASSO penalized regression was conducted as a stability check. Finally, to complement the exploratory binary decision tree model and respect censoring, a formal time-to-event Survival Tree was constructed using log-rank splitting. Calibration was verified using Brier scores and Hosmer–Lemeshow tests. A p-value of <0.05 was considered statistically significant.

In this retrospective cohort of 400 endometrial cancer patients, coexisting adenomyosis was associated with favorable clinicopathological features, including earlier stage, reduced lymphovascular space invasion, and shallower myometrial invasion. However, adenomyosis did not independently predict overall or disease-free survival after adjustment for established prognostic factors in multivariable Cox regression. The exploratory decision tree model identified age, LVSI, and deep myometrial invasion ≥50% as the primary determinants of survival, while adenomyosis was not selected as a discriminating variable at any node of the tree. The time-varying pattern observed in the Kaplan–Meier curves, with crossing survival functions beyond 125 months, warrants further investigation in larger cohorts. From a clinical perspective, routine pathological reporting of adenomyosis in endometrial cancer specimens is reasonable, as it contributes to comprehensive histopathological characterization. However, treatment decisions should continue to be guided by established risk factors such as age, stage, LVSI, histological type, grade, and CA-125 levels, and adenomyosis should not be used as a standalone prognostic marker or as a criterion for therapeutic de-escalation. The decision tree model, while offering an interpretable framework for risk stratification, demonstrated only modest discrimination and requires external validation in multicenter cohorts before clinical application. Future studies integrating molecular profiling with clinical data may further clarify the biological relationship between adenomyosis and endometrial cancer behavior.

Our findings indicate that adenomyosis is associated with more favorable pathological characteristics, including earlier stage at diagnosis, lower frequency of LVSI, and lower myometrial invasion. However, adenomyosis did not emerge as an independent prognostic factor for either OS or DFS in multivariable analyses, adjusting for established risk factors. This distinction between favorable univariate associations and the absence of independent prognostic significance is critical and aligns with the interpretation that adenomyosis correlates with less aggressive tumor biology without itself conferring an independent survival advantage once established prognostic variables are accounted for. The favorable clinicopathological profile observed in our adenomyosis cohort is consistent with prior reports. A recent meta-analysis by An et al 11 encompassing multiple cohorts reported that coexisting adenomyosis was associated with lower rates of LVSI and reduced myometrial invasion depth. Ulger et al 10 similarly demonstrated in a population-based study that women with adenomyosis had reduced tumor aggressiveness in endometrioid-type EC. La Torre et al 5 recently extended these observations to include immunohistochemical and molecular correlates, reporting associations between adenomyosis and favorable molecular subtypes. These converging data suggest that the adenomyosis-associated microenvironment, characterized by chronic estrogen stimulation, localized inflammation, and altered stromal-epithelial interactions, may select for or coexist with biologically indolent tumor phenotypes rather than directly modifying tumor progression. Notably, our Kaplan–Meier analysis revealed a time-varying relationship between adenomyosis and survival. While patients with coexisting adenomyosis experienced an early survival trend, this advantage diminished over a 10-year period, as confirmed by our RMST and landmark analyses. This temporal heterogeneity suggests that while coexisting adenomyosis correlates with a less aggressive tumor biology at presentation, this protective effect does not translate into long-term overall or recurrence-free survival superiority. Furthermore, our machine learning decision tree and survival tree analyses both selected age, LVSI, and deep myometrial invasion as the driving prognostic indicators, while adenomyosis was omitted. This dual-modeling approach suggests that treatment decisions should remain guided by established high-risk factors rather than the presence of coexisting adenomyosis. The significant difference in median preoperative CA-125 between groups (16.70 vs 14.20 U/mL) merits consideration. Although statistically significant, both medians remained well below the conventional threshold of 35 U/mL. This finding likely reflects the known capacity of adenomyosis to cause mild CA-125 elevation through endometrial gland disruption and peritoneal irritation rather than tumor-related production, and its clinical relevance in the context of cancer prognosis is limited. Our findings align with the work of Li et al, 15 who developed a decision tree algorithm differentiating malignant from benign ovarian tumors with strong predictive performance, and Fu et al, 17 who applied a classification and regression tree model to improve diagnostic accuracy in endometrial carcinoma. While those studies focused on diagnostic stratification, our model targeted prognostic prediction and confirmed that age and markers of disease extent remain the most influential determinants. Both the Cox regression and the decision tree converge on the same conclusion: established prognostic factors drive survival outcomes, and adenomyosis does not add independent predictive information. The decision tree framework nonetheless offers an interpretable, visual complement to traditional regression by highlighting hierarchical decision rules that may be useful for clinical education and bedside risk communication, provided that external validation confirms its generalizability. This study has several limitations that merit consideration. Its retrospective design introduces potential selection and information biases. The exclusion of 113 patients (22%) may introduce additional selection bias, although the excluded cases were predominantly attributable to incomplete documentation rather than systematic clinical differences. The diagnosis of adenomyosis relied on histopathological examination of surgical specimens, which may underestimate prevalence compared with magnetic resonance imaging, the current gold standard for in vivo adenomyosis detection. 9 The retrospective application of FIGO-2023 staging to historical cases introduces the possibility of staging inconsistencies, although restaging was performed by a senior pathologist using standardized criteria. The events-per-variable ratio for the Cox models was approximately 8–10, which is borderline relative to the recommended minimum of 10 and may limit the stability of coefficient estimates. The decision tree classifier, trained on a binary vital-status outcome, does not incorporate censoring or event timing, and thus provides a limited representation of the survival experience. The model has not been externally validated, and its generalizability to other populations and practice settings remains unknown. Calibration was assessed by Brier score but a formal calibration plot was not generated. Future prospective, multicenter studies should address these limitations by employing larger sample sizes, survival tree methodologies, external validation, and integration of molecular profiling data to elucidate potential biological links between adenomyosis and tumor behavior.