Predicting primary resistance to third-generation EGFR-TKIs in lung adenocarcinoma using a multisource cross-modal...

Area of Science:

• Oncology
• Artificial Intelligence
• Medical Imaging

Background:

• Lung adenocarcinoma patients often develop primary resistance to third-generation EGFR-TKIs.
• Predicting this resistance is crucial for effective treatment selection.
• Current prediction methods have limitations.

Purpose of the Study:

• To evaluate a multisource cross-modal Transformer (MC-Trans) model for predicting primary resistance to third-generation EGFR-TKIs.
• To assess the model's performance against unimodal models and human experts.

Main Methods:

• Retrospective analysis of clinical and CT imaging data from 222 lung adenocarcinoma patients.
• Development of the MC-Trans model using Table Transformer and Swin Transformer.
• Validation on training/validation and two external test cohorts.

Main Results:

• MC-Trans achieved an ROC-AUC of 0.89 in predicting primary resistance, outperforming unimodal models (tabular: 0.78, CT: 0.63).
• External validation showed MC-Trans performance comparable to human expert panels.
• The model also predicted disease progression risk in patients without primary resistance.

Conclusions:

• MC-Trans demonstrates high utility in predicting primary resistance to third-generation EGFR-TKIs in lung adenocarcinoma.
• This AI tool can aid clinicians in optimizing therapeutic strategies.
• MC-Trans offers potential for predicting disease progression beyond primary resistance.