Multi-modal federated learning with differential privacy for privacy-preserving healthcare AI

The growing adoption of artificial intelligence in healthcare highlights the need for models that can leverage heterogeneous patient data while preserving strict privacy requirements. This paper proposes a novel multi-modal federated learning framework with differential privacy for decentralized healthcare AI. The model integrates electronic health records and ECG time-series using modality-specific encoders and a shared latent fusion network, enabling comprehensive representation learning without centralizing sensitive data. Differential privacy is incorporated into local updates to provide formal guarantees against information leakage in federated aggregation. Extensive experiments on real-world healthcare datasets show that the proposed method achieves $$94.12%$$ accuracy, $$93.64%$$ precision, $$93.21%$$ recall, $$93.42%$$ F1-score, and $$95.03%$$ AUC, outperforming centralized, single-modality, and non-private baselines. The framework also converges $$32.4%$$ faster than single-modality federated learning, reaching $$90%$$ accuracy in 35 rounds. An ablation study confirms the contribution of multi-modal fusion and class balancing, while client variance analysis shows the lowest performance deviation (± 1.2%) under heterogeneous distributions. These results indicate that combining federated optimization, differential privacy, and multi-modal learning provides an effective framework for privacy-preserving clinical AI, with potential for deployment in distributed healthcare settings.