Endometrial cancer is one of the most common cancers of the female reproductive system, and its incidence has been rising - particularly in Taiwan, where it is the most common gynecological malignancy. While early-stage disease can often be cured with surgery, recurrence remains a significant clinical problem. Identifying which patients are most likely to experience recurrence is critical for tailoring follow-up care and adjuvant therapies.
Traditional approaches to predicting recurrence rely on clinical factors such as cancer stage, histological type, and treatment history. However, these factors alone do not capture the full complexity of the disease. Machine learning offers a promising alternative by analyzing large datasets with many variables simultaneously to identify patterns that predict recurrence.
This study used data from 599 patients at Chung Shan Medical University Hospital in Taiwan to compare several machine learning approaches for recurrence prediction. A key challenge was that the dataset was imbalanced - far more patients had not recurred than had, which can bias standard machine learning algorithms toward predicting no recurrence even when recurrence is present.
The study compared five machine learning methods across two categories. Supervised learning methods - which learn from labeled training examples - included the Back Propagation Neural Network (BPNN) and Support Vector Machine (SVM). Unsupervised learning methods - which identify patterns without labeled data - included Fuzzy C-Means (FCM), Ant K-means (AK), and the novel Artificial Immune System (AIS).
The AIS algorithm mimics how the biological immune system works, using concepts like antibodies, antigens, and clonal selection. It identifies data patterns without requiring pre-labeling, making it particularly useful in scenarios where disease categories are not sharply defined. The AIS approach includes a hormone concentration matrix that guides classification decisions in ambiguous cases.
A critical methodological contribution was the use of bootstrap sampling to address the class imbalance problem. The cancer recurrence group had only 38 patients versus 561 controls. Without correction, algorithms learn to mostly predict non-recurrence and appear highly accurate while missing nearly all true recurrences. Bootstrap sampling synthetically resampled the minority class to create a balanced dataset, dramatically improving the algorithms' ability to detect true recurrences.
The dataset included 20 clinical variables, ranging from age, histological type, and cancer stage to treatment details such as chemotherapy, radiation dose, and surgical margins. Not all of these variables are equally informative for predicting recurrence, and including unnecessary variables can actually reduce a model's accuracy and interpretability.
The researchers developed a novel feature selection method called SFS-AIS (Sequential Forward Selection combined with AIS). This method iteratively adds features that most improve classification performance, ultimately selecting the 11 most important from the original 20. This was compared against two established methods: the Relief algorithm and the Information Gain method.
The SFS-AIS method was evaluated alongside two established approaches - the Relief algorithm (which scores features based on their ability to distinguish between cases that are neighbors in the feature space) and Information Gain (which selects features based on how much they reduce uncertainty about the target variable). The SFS-AIS method provided the best overall feature combination for most classifiers.
Before applying bootstrap sampling, the clinical imbalance caused severe problems: models appeared accurate overall but had near-zero sensitivity (ability to detect true recurrences). For example, one model (BPNN) achieved 93.59% accuracy but identified essentially zero recurrences correctly, because it simply predicted non-recurrence for every patient. Accuracy figures without sensitivity are meaningless in this context.
After bootstrap sampling with 11 selected features, performance improved dramatically across all algorithms. The SVM achieved 97.51% accuracy, 95.02% sensitivity, and 99.29% specificity - the best performance among supervised learning methods. The AIS algorithm achieved 83.35% accuracy, 77.53% sensitivity, and 92.31% specificity - the best among unsupervised methods.
Across all three feature selection methods (SFS-AIS, Relief, and Information Gain), histological type and chemotherapy were consistently identified as the most important predictors of recurrence. Behavior code (the clinical behavior classification of the tumor) and radiotherapy details also emerged as significant predictors across multiple methods.
The consistent identification of histological type as a top recurrence predictor aligns with established clinical knowledge. Non-endometrioid subtypes (such as serous and clear cell carcinomas) carry substantially higher recurrence risks than the more common endometrioid type. Histological grade also reflects how abnormal the cancer cells look, with higher grades indicating more aggressive tumors.
Chemotherapy status being a top predictor reflects the reality that patients who receive chemotherapy typically have higher-risk disease to begin with, making it a proxy marker for disease severity rather than necessarily a cause of recurrence. The finding that behavior code - which captures the clinical aggressiveness classification of the tumor - also emerged as important supports the idea that intrinsic tumor biology is a key driver.
The AIS algorithm offers a notable advantage over supervised methods like SVM: it does not require pre-labeled training data, and it can adapt to unknown class structures. This means it could potentially be deployed in new clinical settings without requiring an existing labeled dataset, making it more flexible for real-world clinical use where labeled historical data may not always be available.
This study demonstrates that the artificial immune system algorithm - inspired by how the body's own immune system identifies and responds to threats - can effectively identify patterns predicting endometrial cancer recurrence. Combined with bootstrap sampling to address data imbalance, the AIS method achieves clinically meaningful accuracy without requiring labeled training examples.
The practical implication is that clinicians could use a combination of histology, chemotherapy history, behavior code, and radiotherapy details to stratify patients for more intensive follow-up or adjuvant treatment. The optimal 11-feature model offers a manageable set of variables that could be incorporated into a clinical decision support tool.
Future work should explore longitudinal studies that can capture how recurrence risk evolves over time with treatment. Additionally, validating these findings in external datasets from different institutions and healthcare systems would strengthen the generalizability of these predictions. The biomimetic AIS approach may have value in other cancer types where data imbalance and unsupervised classification challenges are common.