Risk factors for misclassification in predicting EGFR mutation status using PET/CT imaging in non-small cell lung...

This study aims to develop 10 machine learning models based on positron emission tomography/computed tomography (PET/CT) radiomic features to predict epidermal growth factor receptor (EGFR) mutations in non-small cell lung cancer (NSCLC) patients and to identify risk factors contributing to model misclassification.

This study included 277 NSCLC patients from Zhongnan Hospital, Wuhan University, who underwent pretreatment ¹⁸F-FDG PET/CT and EGFR mutation testing. A PET/CT signature (PCS)-nomogram was developed by comparing 10 machine learning algorithms for EGFR prediction. Leave-one-out cross-validation generated model-specific EGFR mutation probabilities for individual patients, and performance disparities were analyzed across clinical subgroups. Model performance was assessed using the receiver operating characteristic curve, Youden's index, decision curve analysis, and DeLong's test.

The PCS-nomogram model, constructed using the partial least squares generalized linear models (plsRglm) algorithm, achieved optimal performance in predicting EGFR mutations in NSCLC patients (training cohort: area under the curve [AUC] = 0.80; validation cohort: AUC = 0.82). Smoking history caused statistically significant performance deterioration in seven of 10 machine learning models (|ΔYouden's index| ≥ 0.1). The PCS model demonstrated higher predictive performance in never-smokers than in smokers (AUC = 0.90 vs. 0.64; p < 0.05).

A plsRglm-based PCS-nomogram model was proposed for the noninvasive prediction of EGFR mutations in NSCLC patients. Compared with smokers, radiomics-based EGFR mutation prediction demonstrated superior performance in never-smokers.