Predicting targeted therapy resistance in non-small cell lung cancer using multimodal machine learning

Area of Science:

• Oncology
• Artificial Intelligence
• Genomics

Background:

• Resistance to tyrosine kinase inhibitors (TKIs) is a significant clinical challenge in treating non-small cell lung cancer (NSCLC) with activating epidermal growth factor receptor (EGFR) mutations.
• Existing predictive tools for TKI resistance often rely on specialized data, lacking integration with routinely available clinical information.
• The development of predictive models using accessible clinical data is crucial for optimizing treatment strategies in EGFR-mutant NSCLC.

Purpose of the Study:

• To develop and evaluate a multimodal machine learning (ML) model for predicting therapy resistance in NSCLC patients with EGFR mutations.
• To utilize readily accessible clinical information, including histology images and next-generation sequencing (NGS) data, for resistance prediction.
• To assess the model's performance and interpretability for potential clinical application in personalized oncology.

Main Methods:

• A multi-institutional retrospective study involving 42 late-stage NSCLC patients with EGFR mutations treated with EGFR-targeted therapy.
• Data integration of histology whole-slide images, NGS results, and demographic/clinical variables into a multimodal ML framework.
• Model evaluation using 5-fold nested cross-validation, concordance index (C-index), Kaplan-Meier survival analysis, and interpretability techniques (attention maps, feature importance).

Main Results:

• The multimodal ML model achieved a mean C-index of 0.82, outperforming image-only (C-index=0.75) and non-image (C-index=0.77) models.
• The model significantly stratified patients into distinct hazard groups (log-rank P=0.04), demonstrating superior predictive capability.
• Key predictors identified include RB1 mutation and Hispanic ethnicity; attention maps highlighted specific histologic features associated with resistance.

Conclusions:

• A robust multimodal ML model can effectively predict therapy resistance in EGFR-mutant NSCLC using routinely collected clinical data.
• The model's superior performance and hazard stratification capabilities suggest its utility for personalized treatment decisions in precision oncology.
• Multimodal AI holds promise for advancing cancer treatment, especially in resource-limited settings, warranting further validation in diverse cohorts.