O2073: Interpretable Machine Learning for Predicting Recurrence in Pediatric Extremity Sarcomas: A Proof-of-Concept Using National Cancer Registry Data

Background: Pediatric extremity soft tissue sarcoma (eSTS) carries a risk of recurrence that necessitates long-term management and surveillance planning. While machine learning (ML) has demonstrated promise in risk stratification, the clinical adoption of ML is limited due to the opaque nature of the way in which the decision-making processes in a given model are made. Explainable AI (XAI) offers promising capabilities of bridging the gap between ML and clinical implementation by increasing the transparency of how individual features contribute to predictions. We present a proof-of-concept study demonstrating the use of interpretable ML using XAI to estimate recurrence risk in pediatric extremity sarcomas using national registry data.

Methods: This was a retrospective cohort study using the SEER database (2000–2021). We identified 890 pediatric patients ( < 20 years old) with eSTS (ICD-O C49.1, C49.2) and valid tumor size and treatment records. Recurrence was inferred using cause-specific mortality as a surrogate. A logistic regression model with class weighting was trained using features including tumor size, grade, surgical modality, chemotherapy, and radiation. The dataset was split into a 70% training and 30% stratified test set. To enhance model transparency, we applied Local Interpretable Model-Agnostic Explanations (LIME), an explainable AI (XAI) technique, to generate individualized visualizations of prediction logic. For representative patients, LIME identified the clinical features most responsible for the predicted recurrence risk and indicated the direction and magnitude of each feature’s influence on the model’s output.

Results: The model achieved an accuracy of 70% on the test set (n=178), with a sensitivity of 74%, specificity of 69%, and an area under the receiver operating characteristic curve (AUC) of 0.792, indicating good discriminative performance. Figure 1 displays side-by-side LIME plots comparing a representative low-risk and high-risk patient from the test set. In the low-risk patient with a 30 mm moderately differentiated tumor, features associated with decreased risk of recurrence included limb-sparing resection without the need for chemotherapy or radiation (predicted risk of recurrence: 5%). In the high-risk patient with a 71 mm undifferentiated tumor, tumor grade and chemotherapy were associated with increased odds of recurrence whereas lack of radiation treatment was associated with decreased odds (predicted risk of recurrence: 78%). These plots demonstrate how different clinical feature combinations influence recurrence predictions in ML at the individual level.

Conclusion: This study demonstrates that interpretable machine learning models can deliver accurate recurrence risk predictions in pediatric eSTS while maintaining transparency. LIME-based visualizations show how clinical features contribute to individual predictions. The top predictors in our example patients were chemotherapy and tumor grade - likely reflecting treatment intensity associated with higher underlying risk. As machine learning becomes more integrated into oncology, XAI can bridge computational tools with actionable patient care.