EGFR mutations are among the most actionable genetic alterations in lung adenocarcinoma, present in 40-50% of Asian patients and 15-20% of Western patients. Patients with EGFR-mutated tumors respond dramatically to tyrosine kinase inhibitors, achieving progression-free survival measured in years rather than months. Identifying mutation status before surgery allows oncologists to plan optimal adjuvant therapy and counsel patients about long-term prognosis.
Current standard practice requires pathological tissue from biopsy or surgical resection to confirm EGFR mutation status, but these approaches have significant limitations. Biopsy samples only a fraction of the tumor and may miss mutations present in other regions, while liquid biopsy (circulating tumor DNA) has limited sensitivity for early-stage nodules. A non-invasive imaging-based approach that predicts EGFR status preoperatively would allow earlier treatment planning without waiting for pathology results.
CT imaging features have been associated with EGFR mutation status, with ground-glass opacity components particularly common in EGFR-positive tumors. The consolidation-to-tumor ratio (CTR), which measures the proportion of solid versus ground-glass components in a nodule, has emerged as a potentially powerful imaging biomarker. Lower CTR values reflect more ground-glass opacity, which correlates with less aggressive, often EGFR-mutated tumors.
This study focused specifically on small nodules measuring 3 cm or less, a clinically important subset where non-invasive prediction is most valuable. These small nodules often present diagnostic uncertainty, and their size limits the quality of biopsy samples obtainable. A validated CT-based model for this size category could directly improve clinical decision-making for the most common early-stage presentations.
The study enrolled 492 patients with pathologically confirmed lung adenocarcinoma who underwent preoperative CT imaging between 2017 and 2022. All patients had nodules 3 cm or smaller confirmed on thin-slice CT (1 mm slice thickness). The dataset was split 80/20 into training (394 patients) and testing (98 patients) sets, with the training set used for all model development and feature selection.
A total of 1,130 radiomic features were initially extracted from three-dimensional nodule volumes using PyRadiomics, covering shape, first-order statistics, and texture categories. Features were derived from original CT images as well as filtered image types. The consolidation-to-tumor ratio was manually measured by two radiologists blinded to EGFR status, with a third resolving disagreements. Clinical features included age, sex, smoking history, and nodule location.
Seven machine learning algorithms were systematically compared: logistic regression, random forest, support vector machine, gradient boosting, XGBoost, LightGBM, and k-nearest neighbors. All algorithms were trained using identical feature sets and evaluated on the same held-out test set. Random forest was selected as the final algorithm based on superior AUC performance and calibration. Recursive Feature Elimination (RFE) reduced the 1,130 radiomic features to 12 most informative variables.
Three models were constructed and compared: a clinical-only model using demographics and CTR, a radiomics-only model using the 12 selected features, and a combined model integrating all three feature types. Model performance was assessed using AUC, sensitivity, specificity, accuracy, and decision curve analysis measuring net clinical benefit across a range of threshold probabilities. SHAP (SHapley Additive exPlanations) with TreeExplainer was used to quantify the contribution of each feature to individual predictions.
The combined model integrating radiomics, CTR, and clinical features achieved an AUC of 0.76 on the held-out test set, outperforming both the clinical-only model (AUC=0.60) and the radiomics-only model (AUC=0.69). The 16-point improvement in AUC over the clinical model and 7-point improvement over radiomics alone demonstrate that each feature category contributes complementary and non-redundant predictive information. All three models were well-calibrated based on Hosmer-Lemeshow testing.
EGFR-mutated patients in this cohort showed distinct clinical and imaging characteristics compared to wild-type patients. Female sex was significantly more prevalent among EGFR-positive cases (73.68% vs 51.22%), consistent with well-established epidemiological associations. CTR was significantly lower in EGFR-positive patients (mean 0.73 vs 0.60 for wild-type), confirming that more ground-glass opacity component correlates with EGFR mutation status.
Decision curve analysis demonstrated that the combined model provided the highest net clinical benefit across threshold probabilities from 0.20 to 0.60. This range encompasses the most clinically relevant decision thresholds for treatment planning. The clinical model showed essentially no benefit over treating all or no patients in this range, while the radiomics model showed intermediate benefit, confirming the superiority of feature integration.
Among the 12 selected radiomic features, texture features from wavelet-transformed images showed the highest individual importance scores. These features capture multi-scale intensity patterns within the nodule that are not visible to the naked eye but correlate with underlying tumor biology. The combination of these abstract texture measurements with the clinically interpretable CTR value creates a model that is both accurate and partially explainable.
SHAP analysis identified CTR as the single most influential feature in the combined model, with the highest mean absolute SHAP value of 0.03 across all test patients. Higher CTR values (more solid component) consistently pushed predictions toward EGFR wild-type, while lower CTR values (more ground-glass opacity) shifted predictions toward EGFR mutation. This directional consistency confirms that CTR captures a biologically meaningful signal.
Female sex was the second most important feature according to SHAP analysis, consistent with epidemiological data showing that women have significantly higher rates of EGFR mutation in lung adenocarcinoma regardless of smoking history. This association appears to reflect underlying biological differences in the molecular carcinogenesis pathway rather than confounding by tobacco exposure, as it holds even in never-smokers of both sexes.
The interpretability provided by SHAP force plots for individual patients allows clinicians to understand why the model made each specific prediction. A patient with low CTR, female sex, and specific texture patterns would receive a high-probability EGFR-positive prediction with visible contributions from each variable. This transparency distinguishes the model from black-box approaches and supports the trust required for clinical adoption.
The 12 selected radiomic features included shape irregularity measures, first-order intensity statistics, and GLCM texture features from wavelet-filtered images. The inclusion of wavelet features reflects that patterns at different frequency scales within the nodule contribute independent information about mutation status. Shape features capturing nodule margin irregularity and spiculation also contributed significantly, consistent with known radiological associations between speculation and EGFR wild-type tumors.
Recursive Feature Elimination with cross-validation (RFECV) systematically reduced 1,130 radiomic features to 12 by iteratively removing the least important features and measuring the resulting change in model performance. This aggressive reduction from over a thousand to just 12 features serves two purposes: preventing overfitting by removing noise features, and creating a more interpretable and practically deployable model that requires fewer computational resources.
The comparison of seven algorithms on identical training data revealed that random forest consistently outperformed other approaches in AUC and calibration. Gradient boosting and XGBoost showed similar but marginally lower performance. Logistic regression underperformed significantly, suggesting that nonlinear relationships between features and EGFR status require tree-based approaches to capture adequately. K-nearest neighbors showed the worst generalization, likely due to the curse of dimensionality even after feature selection.
Feature importance from random forest (based on mean decrease in impurity) and SHAP values showed high concordance, with the same top features identified by both methods. This consistency across two different importance calculation approaches strengthens confidence that the identified features reflect genuine biological signal rather than algorithmic artifacts. When importance rankings differ between methods, it often indicates unstable or collinear features.
Inter-reader reliability for CTR measurement was quantified using intraclass correlation coefficients, confirming high agreement between the two primary radiologists. This reliability assessment is critical for a clinical biomarker: if CTR measurement varies substantially between readers, its clinical utility is limited regardless of its theoretical predictive value. The high ICC demonstrates that CTR is reproducibly measurable in practice, a requirement for real-world deployment.
The clinical value of preoperative EGFR status prediction lies in enabling early initiation of targeted therapy planning for patients who are EGFR mutation positive. Patients identified preoperatively as likely EGFR-positive could be discussed in multidisciplinary tumor boards for adjuvant osimertinib therapy planning, with the pathological confirmation serving to confirm rather than initiate treatment discussions. This could shorten the time from surgery to adjuvant therapy initiation.
For patients in whom bronchoscopic or CT-guided biopsy is being considered, a high model-predicted probability of EGFR mutation could influence the type of molecular testing ordered from the biopsy sample. Comprehensive next-generation sequencing panels are more expensive than targeted EGFR PCR testing, but a high prior probability of EGFR mutation makes targeted testing cost-effective. Conversely, a low probability might prompt broader genomic profiling.
The focus on nodules 3 cm or smaller is particularly relevant because this size category includes most screen-detected lung cancers and stage I adenocarcinomas. These are precisely the patients for whom detailed preoperative molecular characterization is most valuable, as curative surgery is feasible and adjuvant therapy decisions are being weighed. In advanced disease, liquid biopsy and rebiopsy are standard options, but early-stage disease lacks these reliable alternatives.
Integration of this model into existing CT reporting workflows would require only the addition of CTR measurement to standard radiology reports. The radiomic features would be computed automatically from the same DICOM images already acquired, and clinical variables are available from the electronic health record. The implementation burden on radiologists and clinicians would therefore be minimal relative to the potential benefit of preoperative molecular stratification.
This study demonstrates that a random forest model combining CT radiomics, consolidation-to-tumor ratio, and clinical features achieves an AUC of 0.76 for EGFR mutation prediction in small lung adenocarcinoma nodules, substantially outperforming any single feature category. The integration of interpretable features like CTR and sex with abstract radiomic texture information creates a model that is both accurate and explainable, addressing two key requirements for clinical translation.
The identification of CTR as the dominant predictive feature provides mechanistic insight into why the model works. Ground-glass opacity reflects lepidic growth pattern, which is histologically associated with less invasive, well-differentiated adenocarcinoma subtypes that are disproportionately EGFR-mutated. The model essentially quantifies and formalizes this existing radiological intuition while adding precision through radiomic texture analysis that captures subtle patterns beyond what radiologists routinely assess.
Limitations include the single-institution retrospective design, which limits generalizability to different scanners and patient populations, and the relatively modest AUC of 0.76 that reflects residual unpredictability in the relationship between imaging features and genotype. Approximately one in four predictions will be incorrect at any reasonable operating threshold, meaning the model supplements rather than replaces molecular testing. External prospective validation is essential before clinical implementation.
Future work should incorporate dynamic contrast enhancement CT data, PET metabolic features, and multi-center validation to improve both model performance and generalizability. Integration with liquid biopsy results in patients where both are available could create a multimodal prediction approach with higher accuracy than imaging alone. Ultimately, this type of interpretable preoperative model represents a step toward precision oncology where treatment decisions are informed by imaging at the point of initial diagnosis rather than waiting for pathology.