388P - Adenox: A machine learning model to predict molecular targets in non-small cell lung cancer

Background

Targeted therapies significantly improve non-small cell lung cancer (NSCLC) survival, and identifying molecular targets is crucial for determining patient eligibility for treatment. However, the lengthy process of obtaining molecular test results delays treatment initiation. This study develops ‘Adenox’, a clinically practical machine learning model, to predict molecular targets using clinical features before test results are available.

Methods

This retrospective study included 281 NSCLC patients, comprising 80 (28.5%) with EGFR mutation, 107 (38.1%) with PD-L1 positivity, 17 (6%) with ALK mutation, 13 (4.6%) with KRAS mutation, 3 (1.1%) with ROS1 mutation, and 85 (30.2%) without molecular targets. Clinical features were categorized into four groups: demographics, imaging features, laboratory, and pathology results. A predictive model was developed using multivariable Logistic Regression and Random Forest. The model’s performance was assessed with receiver operating characteristic (ROC) curve analysis.

Results

Associated factors with mutations include: EGFR with gender, smoking, alcohol consumption, asthma, spiculation, tumor size, histology, and brain/bone metastases; ALK with histology and pleural metastases; ROS1 with family history; and PD-L1 positivity with smoking, histology, and brain metastases (p < 0.001). ROC analysis demonstrated predictive performance for EGFR (AUC=0.887), PD-L1 (AUC=0.679), ALK (AUC=0.798), KRAS (AUC=0.674), and ROS1 (AUC=0.911).

Evaluation of sensitivity, specificity, AUC, accuracy, PPV and NPV in predicting EGFR, ALK, KRAS, R0S1 mutations, and PD-L1 expression status