LungGuard- A Multimodal Deep Learning System for Early Lung Cancer Detection via Fusion of CT Imaging, Clinical Biomarkers, and Demographic Data

Early detection of lung cancer significantly improves patient prognosis by enabling timely intervention. However, single‐modality methods (e.g., CT imaging alone) often suffer from false positives, variable sensitivity, or lack of context from patient history. This work proposes LungGuard, a multimodal deep learning framework that fuses three complementary data sources—low‐dose CT scans, key clinical biomarkers (e.g. serum tumor markers, smoking history, family history), and demographic features (age, sex, environmental exposures)—to detect lung cancer at early stages, especially in high-risk populations. LungGuard uses a 3D convolutional neural network backbone to process volumetric CT scans, a fully connected network for the clinical biomarkers and demographics, and an attention-based fusion module to combine features. The model is trained and validated on a dataset of ~2,500 patients drawn from multiple centres, with ground truth from histopathology and clinical follow-ups. Data augmentation (rotation, scaling, contrast variation), transfer learning from pre-trained models, and techniques such as focal loss are used to address class imbalance. Preliminary results show that LungGuard achieves AUC ≈ 0.96, sensitivity ≈ 92%, specificity ≈ 90% in early-stage (I & II) lung cancer detection, outperforming radiologists in a controlled observer study by ~5% in accuracy. Explainability methods (Grad-CAM, SHAP) are used to highlight imaging regions and biomarker contributions. The contributions of this work are: (1) demonstrating that fusion of imaging + clinical + demographic data improves early detection; (2) providing a validated model on multi-centre data with robust performance; (3) integrating interpretability to increase clinical trust. Future work aims at external validation in prospective cohorts, inclusion of molecular/omics data, and deployment in low-resource settings.