Accurate disease prediction is essential for improving patient outcomes. Privacy regulations like GDPR and HIPAA limit data sharing, hindering the development of robust predictive models across institutions. FL and multi-modal fusion frameworks counter these problems but are restricted in scalability, inter-client communication, and heterogeneity of data modalities. Techniques which provide privacy on data have an issue whereby they cause a reduction in performance or are computationally costly. This paper presents Mamba-Fusion for Disease prediction, a privacy-preserving framework for multi-modal data. It uses a hierarchical FL architecture to minimize the communication costs and improve the architecture's scalability solution and a Mixture of Experts (MoE) with LSTM based layers for dynamic temporal integration. The latest techniques like, differential privacy, secure aggregation protect both the data and its accuracy of the data as well. Experimental results on multi-modal clinical measurements, ECG, EEG, clinical notes, and demographic data support the applied framework. We have then used Mamba-Fusion to achieve 92:4% accuracy, 0:91 F-Score, and 0:96 AUC-ROC by keeping the privacy leakage at 0:02 and communication costs to 12:5 MB, which make it superior to conventional FL techniques. These results affirm Mamba-Fusion as an applications that are secure enough to support collaborative healthcare analytics on a large scale.