An online explainable ensemble machine learning model for predicting epidermal growth factor receptor mutation status in lung adenocarcinoma.

Background: Non-invasive determination of epidermal growth factor receptor (EGFR) mutation status is essential for selecting lung adenocarcinoma patients suitable for EGFR-tyrosine kinase inhibitors (EGFR-TKIs). This study aimed to develop and validate an online ensemble machine learning (EML) model that combines multiple machine learning (ML) models to predict the EGFR mutation status in lung adenocarcinoma.

Methods: A total of 823 lung adenocarcinoma patients with known EGFR mutation status from three medical centers were divided into a training cohort (n=556) and a validation cohort (n=267) (ChiCTR2400083082 in the WHO International Clinical Trials Registry). Five ML models incorporating clinical and radiological characteristics-random forest (RF), logistic regression (LR), support vector machine (SVM), light gradient boosting machine (LightGBM), and extreme gradient boosting (XGBoost)-along with a CT-based deep learning (DL) model were constructed to predict EGFR mutation status. Subsequently, an EML model was created by combining these models. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), and the SHapley Additive exPlanation (SHAP) method was used to explain the EML model.

Results: In the training cohort, the AUCs for the RF, LR, SVM, LightGBM, XGBoost, DL, and EML were 0.851, 0.790, 0.810, 0.835, 0.853, 0.884, and 0.928, respectively. In the validation cohort, the AUCs for the RF, LR, SVM, LightGBM, XGBoost, DL, and EML were 0.753, 0.744, 0.732, 0.749, 0.751, 0.754, and 0.813, respectively. The Delong test indicated that the AUC of the EML model showed outstanding performance compared to the single models in both the training and validation cohorts. Decision curve analysis indicated that the EML model provided a clinically useful net benefit, and calibration curves showed good agreement. SHAP analysis identified predictive characteristics ranked by their contribution to the EML model: DL score, long-axis diameter, smoking history, pleural retraction, texture, vascular convergence, sex, air bronchogram, and bubblelike lucency. These characteristics were further used to develop an online web tool.

Conclusions: The EML model could serve as a non-invasive and accurate method for predicting EGFR mutation status in lung adenocarcinoma.