Abstract 2299: Risk assessment and development of a machine learning-based prediction model for survival in patients with medulloblastoma

Background: While clinical risk factors have long guided medulloblastoma (MB) treatment, the significance of genetic events in therapeutic strategies has not been fully elucidated. We aimed to explore whether revised molecular risk categories could inform management by analyzing long-term follow-up profiles, and concurrently to establish and validate an explainable prediction model for MB based on machine learning (ML) approaches, and to assess its prognostic implications in patients. Methods: A multicenter cohort of 729 patients from 2001 to 2023 was enrolled as a discovery cohort. Multivariate survival analyses assessed the associations between radiotherapy dose and survival by different revised molecular risk stratifications which integrated clinical and genetic events, followed by utilizing this cohort for training and internal validation of the model. The area under the receiver-operating-characteristic curve (AUROC) and decision curve analysis (DCA) were used to compare the predictive performance. The SHapley Additive exPlanation method was used to explain the final model. Results: Fifteen-year OS and DFS rates were 64.1% (95% CI, 57.9 to 71.0) and 59.1% (95% CI, 53.2 to 65.5) for clinical average-risk group (n = 554), and 55.9% (95% CI, 44.4 to 70.2) and 60.6% (95% CI, 51.9 to 70.7) for clinical high-risk group (n = 175), respectively. Within this revised high-risk stratification of SHH-MB, higher PF dose (≥ 55.8 Gy) combined with reduced CSI dose (23.4 ∼ < 30.6 Gy) was associated with better prognoses. In Gr.3-MB, MYC amplification (activation), as a high-risk marker of dissemination, indicated that CSI dose should be escalated to high intensity (36.0 Gy, p = 0.03) for survival benefits. In the revised model of Gr.4-MB, CDK6 activation well stratified the patients, with 10-year OS of 71.0% (95% CI, 54.0 to 93.5) for average-risk and 52.1% (95% CI, 39.5 to 68.8) for high-risk group. Furthermore, Clinical and molecular information from discovery and validation cohort was involved into model training. The Random survival forest method demonstrated the most optimal predictive efficacy among the six ML models, as evidenced by the training set AUROC of 0.73 (95% CI 0.68-0.78) for 5-year survival rates, 0.76 (0.70-0.82) for 10-year survival rates, and 0.86 (0.81-0.92) for 15-year survival rates in all patients. The final model could accurately predict survival rates in both internal and external validations, and has been deployed through a free, publicly available online software interface. Conclusion: The revised molecular risk stratification was suggested to potentially guide risk-adapted radiotherapy. Except that 36.0 Gy CSI dose is warranted for Gr.3-MB with MYC amplification, CSI reduction is recommended. Furthermore, this study, based on clinical and molecular information, constructs an available survival prediction tools for MB patients. Citation Format: Yu Su, Kaiwen Deng, Tao Jiang, Xiaoguang Qiu, Jiao Zhang, Hailong Liu, Yu Tian. Risk assessment and development of a machine learning-based prediction model for survival in patients with medulloblastoma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2299.