ESTRO meets Asia 2024 - Abstract Book

S159

Interdisciplinary – Head & neck

ESTRO meets Asia 2024

21

Digital Poster

metachronous metastatic sites and risk factors for NPC via an interpretable machine learning system

Liuling Wang 1 , Zirong Li 2 , Linghui Yan 1 , Jianming Ding 1 , Yuhao Lin 1 , Xiaoting Lin 1 , Mengting Xu 1 , Qichao Zhou 2 , Xiao Liu 2 , Chuanben Chen 1 , Zhaodong Fei 1 1 Clinical Oncology School of Fujian Medical University, Fujian Cancer Hospital, Fuzhou, China. 2 Department of Research Algorithms, Manteia Technologies Co.,Ltd, Xiamen, China

Purpose/Objective:

Distant metastasis remains a knotty issue for treatment failure in nasopharyngeal carcinoma (NPC) patients, it is a focus on how to accurately predict metastasis. Machine learning (ML) addresses complex issues effectively and Shapley Additive exPlanations (SHAP) analysis overcomes interpretability limitations of ML, so we aim to predict metachronous metastatic sites and corresponding risk factors after treatment of non-metastatic NPC via an interpretable ML system.

Material/Methods:

We retrospectively included 4,305 observations and predict different sites of distant metastasis in 3-year span via an interpretable ML system, including logistic regression (LR), Decision tree (DT), Random Forest (RF), Extreme Gradient Boosting (XGB) and a stacked model. Performance and relevant covariates were analyzed.

Results:

The predictive power of all models decreased generally over time. Models worked better in predicting the onset and bone metastasis, while were less effective for lung and liver metastasis. For the main relevant covariates, it was pre-EB, post-EB and N stage for onset and bone metastasis, while less ideal in lung and liver metastasis. The risk contribution ratio of relevant factors for different metastatic sites varied and SHAP enhanced consistency and interpretability.