Endometrial carcinoma (EC) is the most common gynecological malignancy in OECD countries, with 417,000 new cases and over 97,000 deaths worldwide in 2020. Five-year survival ranges from 85% for stage I disease to only 25% for stage IV, and treatment decisions depend heavily on accurate risk stratification.
The standard statistical approach for survival analysis is Cox proportional hazards (CPH) regression, which models the relationship between clinical variables and time-to-event outcomes. While widely used, CPH assumes linear, additive relationships between variables - an assumption that may not hold when the effect of one factor depends on another.
Advanced machine learning models like deep learning and ensemble methods can capture non-linear interactions but suffer from a critical clinical barrier: they are black boxes whose predictions cannot be explained or traced back to specific patient characteristics. Clinicians are understandably reluctant to act on recommendations they cannot understand or audit. This study tested whether interpretable tree-based models could match CPH performance without sacrificing transparency.
The study used retrospective data from 842 EC patients treated at Helsinki University Hospital, one of the largest EC cohorts analyzed with machine learning methods. Patients were drawn from the ProMisE (Proactive Molecular Risk Classifier for Endometrial Cancer) validation cohorts, which include molecularly profiled patients - a critical advantage for testing biomarker-enhanced models.
Two feature sets were defined. Feature Set I (FSI) contained 7 standard clinical variables available at diagnosis: age, stage, grade, histological subtype, myometrial invasion, lymphovascular space invasion, and peritoneal cytology. Feature Set II (FSII) expanded FSI by adding 4 molecular biomarkers: L1CAM expression, estrogen receptor (ER) status, peritoneal washing cytology, and tumor size.
Two modeling approaches were compared: Cox proportional hazards (CPH), the clinical standard, and optimal survival tree (OST), a tree-based method that produces visual decision trees. Performance was evaluated using C-index (discrimination) and Integrated Brier Score (IBS) (calibration - how well predicted probabilities match actual outcomes).
Both CPH and OST models achieved strong prognostic performance. Across feature sets and models, C-index values ranged from 0.84 to 0.86. The C-index (concordance index) measures how often the model correctly ranks a patient who died before another patient as having higher predicted risk, with 1.0 being perfect and 0.5 being random.
For the clinical-only feature set (FSI), CPH was preferred over OST. CPH's linear assumptions align well with the relatively simple relationships between standard clinical variables and survival, and it generalizes better with fewer features. The OST model with 7 clinical features offered no advantage over the well-established CPH approach.
For the biomarker-supplemented feature set (FSII), the advantage shifted: OST was favored over CPH. The molecular biomarkers introduced non-linear interactions - for example, the effect of L1CAM may differ between patients with different ER status - that CPH cannot model but that tree-based splitting can capture naturally.
Examining the OST decision path structure revealed that L1CAM (L1 cell adhesion molecule) expression and estrogen receptor (ER) status were identified as important risk factors specifically within the p53-abnormal EC subgroup. p53-abnormal EC (equivalent to the POLE-wild-type, mismatch repair-proficient, p53-abnormal ProMisE class) is the highest-risk molecular subtype.
L1CAM is a cell surface glycoprotein associated with aggressive tumor behavior and poor prognosis in endometrial cancer. Its importance emerging from the OST decision tree structure - without being explicitly programmed - validates the biological relevance of the model's learned decision rules and supports existing evidence of L1CAM as an EC prognostic marker.
The tree's visual interpretability means a clinician can trace the exact path from a patient's specific molecular and clinical profile to the model's risk prediction, examining which features drove the classification at each branch. This is fundamentally different from a black-box model that produces a number without explanation.
Beyond discrimination (C-index), the study emphasized calibration - measured by the Integrated Brier Score (IBS) - as an essential and underappreciated component of clinical prediction models. A well-calibrated model is one where predicted 5-year survival probabilities of, say, 70% correspond to actual 70% survival rates. Poor calibration misleads clinical interpretation even when ranking is accurate.
The authors argue that many published ML models for cancer prognosis report only discrimination metrics while ignoring calibration, creating an incomplete picture of clinical utility. They call for systematic reporting of both metrics as a standard in clinical prediction model development.
The study's core message is that interpretability need not come at the cost of performance: well-designed tree models matched CPH accuracy while providing visual, explainable decision paths. This makes OST-type models candidates for clinical decision support tools that physicians can actually trust, review, and act on with confidence.