104 - Multimodal Artificial Intelligence Model for Predicting Treatment Efficacy and Optimizing Upfront Cranial Radiotherapy Selection in Third-Generation EGFR-TKI-Treated NSCLC with Brain Metastases
Presenter(s)
Y. Pang1, X. Zhu2, L. Peng3, K. Pan4, S. Yu1, B. Xia4, L. Zhang5, W. Yu6, J. Gong7, J. Ni1, and Z. Zhu8; 1Fudan University Shanghai Cancer Center, Shanghai, China, 2Department of Radiation Oncology, Shanghai Chest Hospital, Shanghai Jiaotong University School of Medicine, Shanghai, China, 3Department of Oncology, Tongji Medical College, Tongji Hospital, Huazhong University of Science and Technology, Wuhan, China, 4Hangzhou Cancer Hospital, Hangzhou, China, 5Tongji Hospital,Tongji Medical University,Huazhong University of Science and Technology, Wuhan, China, 6Department of Radiation Oncology, Shanghai Chest Hospital, School of Medicine, Shanghai Jiao Tong University, Shanghai, China, 7Department of Oncology, Shanghai Medical College, Fudan University, Shanghai, China, 8Department of Radiation Oncology, Fudan University Shanghai Cancer Center, Shanghai, China
Purpose/Objective(s): Artificial intelligence model could predict treatment efficacy and identify candidates for upfront cranial radiotherapy in EGFR-mutant non-small cell lung cancer with brain metastases treated with third-generation EGFR tyrosine kinase inhibitors
Materials/Methods:
From a multicenter observational study, three cohorts of first-line third-generation EGFR-TKI-treated NSCLC patients with BMs were analyzed. A multimodal AI model, integrating baseline brain MRI imaging and clinical data, was developed and validated to predict intracranial response to first-line EGFR-TKI monotherapy using the construction (n=138) and validation cohorts (n=60). Patients were stratified into high-risk and low-risk groups based on predicted treatment efficacy. The model’s clinical utility in identifying ucRT candidates was assessed in an exploratory cohort (n=444), including patients with (n=118) and without ucRT (n=326). Notably, ucRT was defined as cranial radiotherapy performed between disease diagnosis to the initial progressive disease following EGFR-TKI treatment. Proteomic analysis of resected BM samples elucidated the model’s biological mechanisms.
Results:
The AI model demonstrated strong predictive accuracy, with AUCs of 0.80 (construction cohort) and 0.83 (validation cohort). Among those receiving EGFR-TKI monotherapy, those in the high-risk group exhibited significantly shorter iPFS (p=0.0014) and PFS (p=0.0018) compared to those in the low-risk group. In the exploratory cohort, ucRT was associated with longer median iPFS (38.0 vs. 20.6 months; p=0.0005) and PFS (18.6 vs. 11.7 months; p=0.0002) in high-risk patients (n=238), while no significant differences were observed in low-risk patients (n=206). Consistent results were found after propensity score matching, time-dependent Cox analyses and landmark analyses. Proteomic analysis revealed significant activation of the mitochondrial oxidative phosphorylation pathway in high-risk BMs.
Conclusion:
The multimodal AI model accurately predicts intracranial response in first-line third-generation EGFR-TKI-treated NSCLC patients with BMs and may aid in identifying ucRT candidates, warranting further validation. (NCT06604689)