Abstract
The application of deep learning technologies in constructing infectious disease prediction models has significantly enhanced public health strategies; however, the imperative for medical data privacy often prevents institutions from sharing diverse datasets, leading to data silos and diminished predictive accuracy. To address these challenges, we propose a multi-layered privacy-preserving framework that balances security and computational performance. First, we introduce a Random Transmission Hybrid Homomorphic algorithm that integrates CKKS fully homomorphic encryption with Paillier semi-homomorphic mechanisms, optimized by a random transmission sequence. Experimental evaluations demonstrate that this hybrid approach achieves a 25% improvement in computational and communication efficiency compared to conventional homomorphic encryption methods by reducing ciphertext overhead and skipping redundant update cycles. Second, we developed the Data Selection-Distributed Selection Stochastic Gradient Descent (DS-DSSGD) algorithm to optimize the trade-off between training speed and predictive accuracy. By filtering insignificant gradient updates and focusing on high-contribution features, the DS-DSSGD algorithm ensures high model precision even under the increased computational demands of privacy-preserving technologies. Finally, these innovations are integrated into the XDP Privacy Data Sharing Platform, providing a secure environment for end-to-end data lifecycle management. Collectively, our results indicate that the proposed framework not only safeguards sensitive health information but also maintains the high-precision forecasting capabilities essential for effective epidemic response. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1038/s41598-026-38906-9.