Optimized Lung Cancer Prediction using Adaptive Federated Multimodal Transformer Framework with Self-Supervised Feature Evolution

Early detection of lung cancer requires predictive systems capable of analysing complex and heterogeneous medical data with high reliability. Conventional machine learning approaches often depend on centralized datasets and static feature selection, limiting their adaptability in distributed clinical environments. Objective: This study proposes a comprehensive multimodal framework that integrates structured clinical records, radiological imaging features, genomic mutation profiles and environmental exposure variables into a unified predictive architecture. The methodology combines self-supervised representation learning, transformer-based crossmodal fusion, federated collaborative training and dynamic feature evolution to enhance robustness and scalability. Methods: Self-supervised learning strengthens intrinsic feature representations before supervised classification, improving generalization across diverse patient groups. A transformer-based fusion mechanism captures interdependencies among modalities through cross-attention operations, enabling patient-specific feature weighting. Federated collaborative learning allows multiple institutions to train a shared global model without exchanging raw medical data, preserving privacy while improving cross-hospital generalization. In addition, a dynamic feature evolution strategy monitors distributional shifts and adjusts feature importance to maintain long-term predictive stability. Results: Experimental evaluation demonstrates superior performance compared to earlier hybrid ensemble models, achieving improvements in accuracy, sensitivity and calibration reliability. The federated configuration maintains stable performance across distributed datasets while reducing institutional bias. The proposed framework offers a scalable and adaptive solution for realworld lung cancer screening and clinical decision support.