CAMAF: Context-Aware Multimodal Alignment Framework for explainable lung disease risk stratification

Abstract

Lung disease remains a leading cause of morbidity and mortality worldwide, necessitating early and accurate risk stratification to improve patient outcomes. Existing approaches rely on either clinical data or imaging modalities independently, limiting their ability to capture the intricate interactions between structured and unstructured information. To overcome this limitation, we propose the Context-Aware Multimodal Alignment Framework (CAMAF), a novel AI-driven approach that dynamically integrates structured clinical data from the MIMIC-IV dataset and unstructured imaging data from the CheXpert dataset for context-aware and interpretable lung disease risk stratification. Unlike conventional multimodal fusion techniques, CAMAF introduces a Context-Gated Attention (CGA) mechanism, which adaptively aligns modality contributions based on patient-specific contexts. Additionally, CAMAF employs TabTransformer and Vision Transformer (ViT) for high-fidelity feature extraction and leverages a Transformer Encoder for enhanced multimodal fusion, ensuring superior predictive performance. Experimental results demonstrate the efficacy of CAMAF, achieving 92.78% accuracy and 0.9457 AUC-ROC, outperforming state-of-the-art machine learning baselines and traditional clinical risk scores (e.g., CURB-65, BODE Index). Robustness analysis highlights its ability to maintain performance under noisy and missing data conditions, while SHAP and Grad-CAM provide interpretable explanations for clinical decision-making. Despite its advantages, CAMAF faces challenges in handling rare pathologies, which can be addressed through data augmentation and transfer learning. This work establishes CAMAF as a novel, clinically viable AI framework, bridging the gap between multimodal data integration, interpretability, and real-world clinical applicability in lung disease risk stratification.