WCLC 2025 - Posters & ePosters-PT1.03.05 An Ai-Based Predictive Tool for Druggable Mutations in Lung Cancer Using Nationwide Comprehensive Genomic Profiling Data

PT1.03.05 An Ai-Based Predictive Tool for Druggable Mutations in Lung Cancer Using Nationwide Comprehensive Genomic Profiling Data

Back to course

Pdf Summary

This study presents the development of an AI-based predictive tool designed to estimate the probability of detecting druggable mutations in lung cancer patients using clinical data available prior to comprehensive genomic profiling (CGP). Analyzing a nationwide dataset from Japan’s C-CAT database, covering 99.6% of CGP cases, the researchers retrospectively examined clinical data from 3,470 lung cancer patients (June 2019 – Nov 2023) to train an eXtreme Gradient Boosting (XGBoost) model. The model aimed to predict whether a patient’s tumor harbors mutations targetable by available therapies.<br /><br />Key predictive clinical features identified by SHapley Additive exPlanations (SHAP) included female sex, nonsmoking status, adenocarcinoma histology, and presence of distant metastases. These factors positively correlated with the presence of druggable mutations. The model demonstrated strong predictive performance with an area under the receiver operating characteristic curve (AUROC) of 0.85 in validation. Subsequently, the model was deployed as a web application to provide clinicians with an intuitive tool estimating the likelihood of identifying actionable mutations prior to CGP testing.<br /><br />The tool’s performance was independently tested on a separate cohort of 1,307 lung cancer patients (Dec 2023 – Nov 2024), achieving an AUROC of 0.77 and a Brier score of 0.19, indicating good calibration and predictive accuracy. Predominant mutations identified included EGFR (50.8%), ERBB2 (13.3%), KRAS G12C, MET exon 14 skipping, and other clinically actionable variants.<br /><br />By enabling more efficient selection of patients likely to benefit from CGP, this AI model may reduce unnecessary testing burden, optimize resource allocation, and facilitate broader access to precision oncology treatments. Approved by the University of Tokyo ethics committee, this research underscores the value of explainable AI in guiding personalized cancer medicine and supports clinical decision-making through an accessible web-based predictive platform.

Asset Subtitle

Hiroaki Ikushima

Meta Tag

Speaker Hiroaki Ikushima

Topic Tumor Biology – Translational Biology

Keywords

AI-based predictive tool

lung cancer

druggable mutations

clinical data

comprehensive genomic profiling

XGBoost model

SHAP feature importance

AUROC

precision oncology

web application