Abstract C003: Interpretable machine learning for predicting second malignant neoplasms in childhood cancer survivors
Mohamed Osman, Sara E. Abouelnaga, Omar Ibrahim, Ahmed ElhadidyAbstract
Background:
Advances in cancer treatment have markedly improved survival for children with cancer. As more survivors live into adulthood, second malignant neoplasms (SMNs) have become one of the most serious late effects and the second leading cause of death in five-year survivors, after recurrence of the primary tumor. In the Childhood Cancer Survivor Study, the 30-year cumulative incidence of SMNs reached 20.5%. Identifying who is at greatest risk could help personalize long-term surveillance. We developed interpretable machine learning models to predict the occurrence of SMNs from routinely collected clinical data.
Methods:
Using the Surveillance, Epidemiology, and End Results (SEER) program, we identified patients diagnosed with a primary cancer at age ≤18 years who survived more than five years. Patients were grouped into hematologic and non-hematologic malignancies. A light gradient boosting model (LightGBM) was trained with 10-fold cross-validation and Bayesian hyperparameter optimization to predict the occurrence of SMNs. SHapley Additive exPlanations (SHAP) was used to interpret models' predictions.
Results:
We identified 58,655 childhood cancer survivors diagnosed between 1975 and 2012; of these, 24,245 (41.33%) had hematologic malignancies. Among hematologic cancer survivors, 800 (3.30%) developed an SMN; among non-hematologic survivors, 1,239 (3.60%) did. Survivors who developed an SMN after a hematologic cancer were more likely to have had Hodgkin lymphoma (47.75% of cases) and to have received radiation (47.25% vs. 23.29%). Among non-hematologic cancers, the most common tumor sites were brain (23.79%), thyroid (8.35%), and bone (7.68%). For predicting SMN occurrence, LightGBM models reached mean training/testing AUCs (area under the curve) of 75.84%/71.18% (hematologic) and 73.93%/64.48% (non-hematologic) under 10-fold cross-validation. Despite the rarity of SMNs, the hematologic occurrence model retained class balance, correctly identifying 54.70% of survivors who developed an SMN and 77.30% of those who did not. Time-to-event models reached testing AUCs of 67.96% and 69.17% for predicting 10- and 15-year risk of SMNs after hematologic malignancies, and 78.52% and 80.43% after non-hematologic malignancies. SHAP analysis linked Hodgkin lymphoma, older age at diagnosis, prior radiation, and female sex to increased SMN risk after hematologic cancers. For non-hematologic cancers, SHAP identified older age, metastasis, and tumor size as the leading contributors. The final models were deployed in a visual web application that returns an individual's SMN probability.
Conclusions:
Interpretable machine learning can stratify SMN risk in childhood cancer survivors from readily available clinical data while revealing the factors behind each prediction, supporting personalized surveillance and earlier detection of subsequent malignancies.
Citation Format:
Mohamed Osman, Sara E. Abouelnaga, Omar Ibrahim, Ahmed Elhadidy. Interpretable machine learning for predicting second malignant neoplasms in childhood cancer survivors [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Bridging Discovery and Clinical Impact in Pediatric Cancer; 2026 Sep 22-25; Philadelphia, PA. Philadelphia (PA): AACR; Cancer Res 2026;86(18_Suppl_1):Abstract nr C003.