Abstract
Uterine sarcomas are rare and aggressive tumors with heterogeneous outcomes, making accurate risk stratification crucial for personalized treatment. This study introduced a novel semi-supervised clustering approach, survival-guided adaptive Kmeans (SGAKmeans), for enhanced mortality risk stratification in early-stage uterine sarcoma patients. SGAKmeans uniquely integrated clinical characteristics and survival information, leveraging domain knowledge and soft pairwise constraints to adaptively adjust distance calculations. Using data from 1,836 uterine sarcoma patients in localized or regional stages in the SEER database, SGAKmeans identified three distinct risk groups with significantly different survival outcomes: high-risk (n = 293, 90.1% mortality, median survival 17 months), medium-risk (n = 767, 59.6% mortality, median survival 66 months), and low-risk (n = 776, 21.3% mortality). The method outperformed eight traditional clustering approaches in risk stratification performance and demonstrated robustness across various data distribution scenarios. Notably, the stratified groups showed differential responses to radiotherapy: high-risk patients benefited significantly (hazard ratio: 0.695, 95% CI 0.541-0.894), medium-risk patients showed no significant difference (0.910, 95% CI 0.747-1.110), while low-risk patients exhibited worse outcomes with radiotherapy (1.826, 95% CI 1.278-2.607). These findings highlighted the potential of SGAKmeans for more nuanced risk stratification and personalized treatment decisions in early-stage uterine sarcoma management, promoting precision medicine.