A Bilayer Feature Fusion Framework for Pan-Cancer Survival Prediction Based on Multihead Attention and Adaptive...

In the field of precision medicine, pan-cancer survival prediction is crucial for individualized oncology diagnosis and treatment. Although multimodal data fusion techniques have significantly improved prediction accuracy, existing studies generally overlook the sensitivity of medical data and the need for privacy protection.

To address the aforementioned problem, this study aims to propose a bilayer feature fusion framework based on the multihead attention mechanism and adaptive differential privacy, which balances precise feature extraction and sensitive data protection.

Specifically, the multihead attention mechanism was integrated into bilayer feature extraction and fusion. Layer-wise relevance analysis was used to calculate the correlation between features and outcomes, and Laplacian noise was adaptively added based on the calculation results to achieve collaborative optimization of precise feature extraction and sensitive data protection. Additionally, the concordance index (C-index) and 5-fold cross-validation were used to compare the proposed method with state-of-the-art approaches.

The proposed model achieved superior performance in both pan-cancer and single-cancer survival prediction, validated via the C-index and 5-fold cross-validation. In pan-cancer scenarios, the trimodal combination of clinical data, messenger RNA data, and microRNA expression data achieved the highest C-index of 0.799, outperforming multiple existing multimodal survival prediction approaches. After adaptive Laplacian noise injection for privacy protection, the model's accuracy decreased by only 0.01-0.03 while satisfying ϵ-differential privacy. For single-cancer prediction, the proposed method achieved higher C-index values in 18 out of 20 cancer types compared with representative deep learning-based survival models, with 7 types showing significant improvements (C-index difference >0.1) and more stable performance distributions. Furthermore, pan-cancer-trained models generally outperformed single-cancer-trained counterparts in most cancer types, highlighting the value of shared predictive features across cancers.

This study provides a solution that balances prediction accuracy and privacy security for pan-cancer survival prediction, laying the foundation for the efficient use of medical data under privacy protection. Future work may further integrate pathological images and proteomics data to expand the model's applications in cancer subtype classification and biomarker discovery.