Multimodal Deep Learning Framework for Decoding Treatment Response in NSCLC: Biomarker Discovery for Immune Checkpoint Inhibitors

Non-small cell lung cancer (NSCLC) remains one of the leading causes of cancer-related mortality worldwide, with therapeutic progress often limited by late-stage diagnosis and the variable effectiveness of existing treatments. Immune checkpoint inhibitors (ICIs) have emerged as a promising therapeutic approach, but their clinical success depends critically on accurate biomarker-based patient stratification. Current strategies for integrating biomarkers, however, remain fragmented and lack the robustness needed for reliable prediction of treatment outcomes. To address this gap, we propose a novel multimodal deep learning framework that integrates CT images, PET/CT scans, and curated biomarker profiles to improve prediction of ICI treatment response in NSCLC patients. Our model employs a pre-trained Xception encoder for advanced image feature extraction, an Encoder Attention Network for semantic representation learning, and a DenseNet-inspired metadata processor for structured biomarker data. Multimodal features are fused through a Block Dense Attention Convolutional Module with Self-Attention Multi-Head (BDAC-SAMH) mechanism, enabling richer interactions across modalities. Experimental evaluation demonstrates that our framework achieves 94.3% accuracy, 92.1% sensitivity, and 93.5% specificity, significantly outperforming conventional CNN-based unimodal methods. Importantly, the proposed system improves prediction of key ICI biomarkers, including PD-L1, TMB, and MSI, by 15.6% compared to unimodal baselines while revealing novel biomarker interactions. This highlights its potential to guide personalized immunotherapy strategies in NSCLC.