Predicting Targeted Therapy Resistance in Non-Small Cell Lung Cancer Using Multimodal Machine Learning

Background Resistance to tyrosine kinase inhibitors remains a major clinical challenge in the treatment of non-small cell lung cancer (NSCLC) with activating epidermal growth factor receptor (EGFR) mutations. Despite the efficacy of third-generation EGFR inhibitors, no standard tool currently exists to predict resistance using routinely available clinical data. In this study, we aim to develop a multimodal machine learning model to predict resistance in NSCLC patients using readily accessible clinical information. Methods We conducted a multi-institutional retrospective study to develop and evaluate a multimodal machine learning model for predicting therapy resistance in late-stage NSCLC patients with EGFR mutations. The study included 42 patients treated with EGFR-targeted therapy from Dartmouth-Hitchcock Medical Center and Ochsner Health System, using data including histology whole-slide images, next-generation sequencing results, and demographic and clinical variables. The modeling framework fused image and non-image data through a three-stage training process and was evaluated using 5-fold nested cross-validation. Model performance was assessed using the concordance index (C-index), Kaplan-Meier survival curves, and log-rank tests. Interpretability analyses were conducted using attention maps, feature importance coefficients, and cellular composition comparisons. Results The multimodal model achieved a mean C-index of 0.82 across cross-validation folds, outperforming image-only and non-image models (C-index 0.75 and 0.77, respectively). Stratified analyses across institutions confirmed consistent performance gains with the multimodal approach. Kaplan-Meier analysis revealed that the multimodal model significantly stratified patients into distinct hazard groups (log-rank P=0.04), which unimodal models failed to achieve. Key predictors included RB1 mutation and Hispanic ethnicity. Attention maps highlighted histologic regions with deformed nuclei, and cellular analysis revealed reduced inflammatory cell presence in high-risk patients. Conclusions This study presents a robust multimodal machine learning model for predicting therapy resistance in EGFR-mutant NSCLC, leveraging routinely collected clinical data without manual feature engineering. The model demonstrated superior performance over unimodal models and effective hazard stratification, suggesting utility for personalized treatment decisions. These findings underscore the potential of multimodal artificial intelligence (AI) tools to advance precision oncology, particularly in resource-limited settings. Further validation in larger, diverse cohorts is warranted.