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

The aim of this study was to investigate the utility of a multisource cross-modal Transformer (MC-Trans) model in predicting primary resistance to third-generation epidermal growth factor receptor tyrosine kinase inhibitors (3rd-EGFR-TKIs) in patients with lung adenocarcinoma. A retrospective analysis of clinical and imaging data from 222 lung adenocarcinoma patients treated with 3rd-EGFR-TKIs was conducted. Patients were allocated to a training/validation cohort (n = 136) and two external test cohorts (n = 34 and n = 52). The Table Transformer and a Swin Transformer-based model were employed to extract features from tabular and CT imaging data, respectively, to construct the MC-Trans model. The results demonstrated that MC-Trans exhibited excellent performance in predicting primary resistance, with an ROC-AUC of 0.89, which was significantly superior to those of the unimodal models (tabular model: 0.78; CT model: 0.63). Furthermore, predictions on external Test Cohort 2 revealed that the predictive performance of MC-Trans was comparable to that of a human expert panel. The study also revealed that MC-Trans could predict the risk of disease progression even among patients without primary resistance. In conclusion, MC-Trans may serve as a valuable tool to assist in determining the therapeutic response of lung adenocarcinoma patients to 3rd-EGFR-TKIs. Keywords: Artificial intelligence; Lung Cancer; Multimodal models; Third-generation EGFR-TKIs; Primary Resistance.