CT-based deep learning models show high accuracy predicting EGFR mutation status in lung adenocarcinomaMachine learning models help predict lung cancer mutations from scans

This meta-analysis evaluated the diagnostic performance of machine learning-based radiomics models for predicting epidermal growth factor receptor (EGFR) mutation status in a population of 6,628 Chinese patients with lung adenocarcinoma. The study aimed to determine how different computational approaches and imaging modalities influence the accuracy of identifying these critical mutations.

The analysis compared several variables including model architecture (deep learning versus conventional machine learning), imaging modality (Computed Tomography [CT] versus Positron Emission Tomography-Computed Tometry [PET-CT]), and validation methods (internal versus external cohorts). The primary outcome was the diagnostic performance of these radiomics models in predicting EGFR mutation status.

The pooled results for the radiomics models showed a sensitivity of 71% (95% confidence interval: 68-74) and a specificity of 81% (95% confidence interval: 78-84). The summary area under the curve (AUC) for these models was 0.85 (95% confidence interval: 0.82-0.88).

Comparative analyses revealed significant differences based on methodology and technology. Deep learning models demonstrated a significantly higher AUC of 0.871 compared to 0.798 for conventional machine learning models (P = 0.012). Furthermore, CT-based models yielded higher accuracy with an AUC of 0.879 compared to 0.828 for PET-CT-based models (P = 0.038). The study also found that models validated on independent external cohorts performed significantly better than those using internal validation, showing an AUC of 0.922 versus 0.841 (P = 0.006).

Safety and tolerability data were not reported as the study focused on diagnostic modeling rather than clinical intervention. The findings suggest that CT-based deep learning models, particularly those validated on external cohorts, offer a promising noninvasive pathway for predicting EGFR mutations.

Methodological limitations included the retrospective nature of the studies included in the meta-analysis and substantial heterogeneity among those studies. These factors may impact the generalizability of the results to broader populations outside of the specific Chinese cohort studied.

Clinical implications suggest that machine learning-based radiomics, specifically CT-based deep learning models with external validation, show promise for noninvasive prediction of EGFR mutations in lung adenocarcinoma patients. However, clinical translation requires large-scale, prospective, multicenter trials with standardized workflows to confirm these findings in a real-world setting.

Several questions remain regarding the integration of these models into standard diagnostic pipelines and whether they can consistently outperform traditional biopsy methods across diverse ethnicities and imaging hardware. Further research is needed to establish standardized radiomics features for consistent cross-platform application.