A causal multimodal framework for privacy-preserving early-stage cancer detection and adaptive testing.

Early detection of cancer at stage I is critical for improving survival rates, yet existing diagnostic tools often face trade-offs between sensitivity, specificity, and clinical scalability. While liquid biopsies, radiomics, and breathomics independently offer promise, their isolated use struggles with robustness, leading to false positives or missed early lesions. To overcome these challenges, this research proposes CausaLMED, a causal multimodal framework that integrates cfDNA fragmentomics, exhaled breathomics, imaging radiomics, and digital pathology embeddings through a causal graph-based fusion mechanism. Unlike conventional ensemble models, CausaLMED explicitly disentangles causal dependencies across modalities, thereby reducing bias from confounders such as lifestyle factors, imaging vendor variability, and population heterogeneity. The framework incorporates an uncertainty-aware adaptive testing policy, which dynamically selects the next diagnostic modality using a partially observable Markov decision process, ensuring cost-effectiveness while minimizing patient burden. Federated learning with differential privacy safeguards institutional data sharing, enabling large-scale, secure model training. Experimental validation on retrospective multimodal datasets demonstrates that CausaLMED achieves a 96.7% accuracy, 94.2% sensitivity for stage I cancers, and maintains 99.1% specificity, significantly outperforming single-modality baselines by over 8%. Moreover, the adaptive testing policy reduces unnecessary imaging referrals by 23%, highlighting both efficiency and clinical practicality. By unifying causal learning, adaptive diagnostics, and privacy-preserving collaboration, CausaLMED presents a transformative paradigm for clinically viable early-stage cancer detection.