Abstract
Our study developed a deep neural network model for adenosarcoma. The performance of this model was superior to that of the traditional Cox proportional hazard model. In addition, a personalized survival prediction system was developed based on our deep survival learning model, which provided more accurate prognostic information for adenosarcoma patients.
Introduction
Uterine adenosarcoma is a rare malignancy, accounting for only about 5% of uterine sarcomas. Due to its rarity, clinical guidelines and staging systems, such as those from FIGO, are often based on limited case reports rather than large-scale datasets.
The standard tool for predicting cancer survival has traditionally been the Cox Proportional Hazard (CPH) model. However, CPH assumes a linear relationship between variables and survival time—an assumption that rarely holds true in the complex, multidimensional reality of oncology. For adenosarcoma, where factors like "sarcomatous overgrowth" can drastically change a patient's outlook regardless of their stage, a more flexible approach is needed. This study applies deep learning to capture the nonlinear correlations inherent in patient data to provide more accurate, personalized prognostic information.
Methods and Materials
Dataset and Feature Selection
The study utilized the Surveillance, Epidemiology, and End Results (SEER) database, covering approximately 34.65% of the U.S. population. A total of 797 uterine adenosarcoma patients diagnosed between 1973 and 2014 were included. The data was randomly partitioned into a training set (65%), a validation set (18%), and an external testing set (18%).
Out of 133 initial variables, 15 were selected via correlation matrix analysis to serve as model inputs. These included age, tumor diameter, grade, lymph node status, and surgical type.
Model Architecture
The authors developed a Deep Survival Learning (DSL) model based on the Neural Multi-Task Logistic Regression (N-MTLR) architecture. Unlike standard neural networks that treat survival as a simple classification task, N-MTLR models the probability of survival across a sequence of discrete time intervals. This allows the model to handle "censored" data (patients who left the study or were still alive at the last follow-up) more effectively. The final architecture consisted of a four-layer neural network with 50 neurons per layer, optimized using the Adam optimizer with L2 regularization.
Results
Superior Predictive Performance
The DSL model showed superior predictive power across primary metrics compared to the traditional CPH model:
- Concordance Index (C-index): The DSL model achieved 0.774 in the testing set, compared to 0.726 for the CPH model.
- Integrated Brier Score (IBS): The DSL model scored 0.14, outperforming the CPH model's 0.17. A lower IBS indicates higher accuracy in probabilistic predictions.
Beyond Traditional Staging
The study revealed significant limitations in current staging systems. While traditional Stage I patients had significantly better survival, the differences between Stages II, III, and IV were often statistically unclear.
Instead of broad stages, the DSL model generates a continuous risk score (0–8). This allowed researchers to re-group patients into three distinct risk categories that showed much clearer survival differentiation than the traditional four-stage system. Furthermore, the model can generate an individual survival probability curve for a specific patient, offering granular prognosis rather than a generic 5-year survival rate.
Discussion and Limitations
While the DSL model represents a significant advancement, several limitations remain. The SEER database lacks specific molecular markers (like Ki-67 or p53) and detailed pathology (like the exact percentage of sarcomatous overgrowth), which are critical prognostic factors. Additionally, as a retrospective study of historical data, it may not fully account for recent shifts in surgical techniques or chemotherapy regimens.
Future work should focus on modeling Progression-Free Survival (PFS) and integrating the model into a clinical online interface. This would allow clinicians to input patient data and receive real-time, personalized survival distributions to guide treatment decisions.
Summary updated