Abstract
Background/Objectives: Pediatric sarcomas are a biologically diverse group of mesenchymal tumors associated with morbidity due to recurrence, despite aggressive multimodal treatment. Reliable predictors of early recurrence remain limited. This exploratory study aimed to identify clinical features associated with first tumor recurrence using a machine learning approach tailored to low-event settings. Methods: We conducted a retrospective, single-center cohort study of 23 pediatric patients with histologically confirmed sarcoma. Forty-six baseline variables were extracted per patient, including clinical, histological, and comorbidity data. Tumor recurrence was the primary binary endpoint. A LASSO-regularized logistic regression model was developed using leave-one-out cross-validation (LOOCV) to identify the most informative predictors. Dimensionality reduction (PCA) and SHAP-value analyses were used to visualize patient clustering and interpret variable contributions. Results: The model identified a four-variable risk signature comprising histological grade, primary tumor width, arterial hypertension, and extremity localization. Each additional tumor grade or centimeter of width approximately doubled the odds of recurrence (OR 2.18 and 2.04, respectively). Hypertension and limb location were associated with a 1.7 and 1.9 odds ratio of recurrence, respectively. The model achieved a balanced accuracy of 0.61 ± 0.08 and AUROC of 0.47 ± 0.12, reflecting limited discriminative power. PCA mapping revealed distinct outlier patterns correlating with high-risk profiles. Conclusions: Even in a small cohort, classical prognostic markers, such as tumor grade and size, retained predictive relevance, while hypertension emerged as a novel, potentially modifiable cofactor or indicator for recurrence. Although model performance was modest, the findings are hypothesis-generating and warrant validation in larger prospective datasets.