Abstract
OBJECTIVE: Accurate cancer risk prediction is hindered by complex, multi-layered immune interactions, and traditional tissue biopsies are invasive and lack scalability for large-scale or repeated assessments. Peripheral blood offers a minimally invasive and accessible alternative for immune profiling. This study aims to develop CAMFormer, a deep learning framework that integrates multimodal peripheral blood-derived immune features for precise, non-invasive early cancer risk prediction. METHODS: CAMFormer combines mRNA expression, immune cell frequencies, and TCR diversity index, leveraging a cross-attention-based multimodal Transformer to capture cross-scale immune interactions. RESULTS: In five-fold cross-validation, CAMFormer achieved an AUC of 0.92 and an F1-score of 0.85 on the validation set, outperforming unimodal and baseline methods. CONCLUSION: These results highlight the potential benefits of integrating multimodal immune features with cross-attention mechanisms for early cancer detection and for guiding future personalized immunotherapy studies.