Generalizability and Fairness in Deploying AI for Lung Cancer Biomarker Prediction

To the Editor We read with great interest the study by Rakaee et al evaluating open-source artificial intelligence (AI) models for EGFR prediction in lung adenocarcinoma. This work importantly highlights the promise of AI as a scalable complement to molecular testing, particularly in resource-limited settings.

However, its findings raise critical concerns that must be addressed before clinical deployment. The study demonstrated that even a well-performing, fine-tuned model like EAGLE exhibits significant performance degradation in Asian patients (area under the curve [AUC], 0.68) and in pleural samples (AUC, 0.66).

This variability underscores that ancestry-associated morphologic variation and specimen context can influence algorithm behavior. The decline in Asian patients persisted even after accounting for higher EGFR variant prevalence, suggesting inherent challenges in model generalization across ancestries.

This poses a substantial risk of perpetuating or exacerbating health care disparities if such tools are deployed without subgroup-specific validation. Therefore, we urge a paradigm shift in the validation framework for AI pathology models.

Moving forward, validation must be ancestry aware and tissue-context aware.