Alasmari Sultan, AlGhamdi Rayed, Tejani Ghanshyam G, Kumar Sharma Sunil, Mousavirad Seyed Jalaleddin
Department of Information Systems, College of Computer and Information Sciences, Majmaah University, Majmaah, Saudi Arabia.
College of Technology and Business, Riyadh Elm University, Riyadh, Saudi Arabia.
Front Physiol. 2025 Apr 23;16:1563185. doi: 10.3389/fphys.2025.1563185. eCollection 2025.
Heart disease remains a leading cause of mortality globally, and early detection is critical for effective treatment and management. However, current diagnostic techniques often suffer from poor accuracy due to misintegration of heterogeneous health data, limiting their clinical usefulness.
To address this limitation, we propose a privacy-preserving framework based on multimodal data analysis and federated learning. Our approach integrates cardiac images, ECG signals, patient records, and nutrition data using an attention-based feature fusion model. To preserve patient data privacy and ensure scalability, we employ federated learning with locally trained Deep Neural Networks optimized using Stochastic Gradient Descent (SGD-DNN). The fused feature vectors are input into the SGD-DNN for cardiac disease classification.
The proposed framework demonstrates high accuracy in cardiac disease detection across multiple datasets: 97.76% on Database 1, 98.43% on Database 2, and 99.12% on Database 3. These results indicate the robustness and generalizability of the model.
Our framework enables early diagnosis and personalized lifestyle recommendations while maintaining strict data confidentiality. The combination of federated learning and multimodal feature fusion offers a scalable, privacy-centric solution for heart disease management, with strong potential for real-world clinical implementation.
心脏病仍然是全球主要的死亡原因，早期检测对于有效治疗和管理至关重要。然而，由于异构健康数据的错误整合，当前的诊断技术往往准确性较差，限制了它们的临床实用性。
为了解决这一局限性，我们提出了一种基于多模态数据分析和联邦学习的隐私保护框架。我们的方法使用基于注意力的特征融合模型整合心脏图像、心电图信号、患者记录和营养数据。为了保护患者数据隐私并确保可扩展性，我们采用联邦学习，结合使用随机梯度下降优化的局部训练深度神经网络（SGD-DNN）。融合后的特征向量被输入到SGD-DNN中进行心脏病分类。
所提出的框架在多个数据集上的心脏病检测中表现出高精度：在数据库1上为97.76%，在数据库2上为98.43%，在数据库3上为99.12%。这些结果表明了该模型的稳健性和通用性。
我们的框架能够实现早期诊断和个性化生活方式建议，同时保持严格的数据保密性。联邦学习和多模态特征融合的结合为心脏病管理提供了一种可扩展的、以隐私为中心的解决方案，具有在现实世界临床应用中的强大潜力。
