Main Session
Sep 29
PQA 07 - Head and Neck Cancer, Lung Cancer/Thoracic Malignancies, and Nursing and Supportive Care

3569 - Development and Validation of a Machine Learning-Based Survival Prediction Model for Patients with EGFR/ALK/ROS1-Positive NSCLC and Brain Metastases Following Initial Intracranial Radiotherapy

03:45pm - 05:00pm ET
Poster Hall - Exhibit Hall A
Screen: 24
POSTER

Presenter(s)

Erha Munai, - School of Medicine, Chongqing University, Chongqing, China, Chongqing,

E. Munai1, W. Zhou2, Z. Tang3, L. Wang2, and Y. Wu2; 1Department of Radiation Oncology, Chongqing University Cancer Hospital, Chongqing, China, Chongqing, China, 2Department of Radiation Oncology, Chongqing University Cancer Hospital, Chongqing, China, 3Radiation Oncology Center, Chongqing University Cancer Hospital, Chongqing, China

Purpose/Objective(s): Prognosis among patients with driver-gene-positive non-small cell lung cancer (NSCLC) and brain metastases (BMs) exhibits significant heterogeneity. Current prognostic scoring systems fail to integrate key molecular characteristics and therapeutic interventions adequately. This study aimed to develop and validate a machine learning (ML) model to provide highly accurate, individualized survival predictions for this population following their first course of intracranial radiotherapy (RT).

Materials/Methods: This retrospective study enrolled 370 patients with EGFR (n=304), ALK (n=40), or ROS1 (n=26) altered NSCLC and BMs who received initial intracranial RT. The cohort was randomly assigned to a training and a validation set in a 7:3 ratio. Five ML algorithms, Gradient Boosting Machine (GBM), Random Survival Forest (RSF), Lasso-Cox, CoxBoost, and Survival Support Vector Machine (Survival-SVM), were utilized to construct the survival prediction models. Predictive performance was comprehensively evaluated using time-dependent area under the receiver operating characteristic curve (AUC), calibration curves, decision curve analysis (DCA), and the concordance index (C-index). SHapley Additive exPlanations (SHAP) were incorporated to elucidate model interpretability and the prognostic contribution of individual features.

Results: Among the five algorithms, the GBM model demonstrated the best predictive performance, achieving AUC values of 0.869 and 0.857 on the training and validation sets, respectively. Furthermore, this model exhibited high calibration consistency and yielded the highest net clinical benefit in DCA. SHAP feature importance analysis revealed that TKI (tyrosine kinase inhibitor) generation (with 3rd-generation TKIs offering the greatest benefit), RT modality (favoring stereotactic radiotherapy [SRT]), number of BMs (=4 indicating better prognosis), KPS (=80 associated with improved survival), and the absence of baseline neurological symptoms were the top five core prognostic determinants. The risk scoring system generated by this model enabled precise risk stratification, effectively classifying patients into subgroups with significantly distinct survival outcomes (P < 0.0001).

Conclusion: For patients with EGFR/ALK/ROS1-positive NSCLC and BMs, the proposed GBM-based ML model provides highly accurate, individualized prognostic assessments following initial intracranial RT. By deeply integrating TKI generation, RT parameters, and clinical baseline characteristics, this predictive tool can effectively assist clinicians in optimizing personalized treatment decisions and formulating precise follow-up management strategies.