Mamba-fusion for privacy-preserving disease prediction

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Data Privacy

Background:

• Accurate disease prediction is crucial but hindered by privacy regulations (GDPR, HIPAA) limiting data sharing.
• Existing federated learning (FL) and multi-modal fusion methods face challenges in scalability, communication, and data heterogeneity.
• Privacy-preserving techniques often compromise model performance or increase computational burden.

Purpose of the Study:

• To introduce Mamba-Fusion, a novel privacy-preserving framework for multi-modal disease prediction.
• To address the limitations of current FL and fusion techniques in terms of scalability, communication efficiency, and data privacy.
• To enable secure, large-scale collaborative healthcare analytics.

Main Methods:

• A hierarchical federated learning (FL) architecture to reduce communication costs and enhance scalability.
• A Mixture of Experts (MoE) with LSTM-based layers for dynamic temporal integration of multi-modal data.
• Integration of advanced privacy techniques like differential privacy and secure aggregation.

Main Results:

• Mamba-Fusion demonstrated superior performance compared to conventional FL techniques.
• Achieved 92.4% accuracy, 0.91 F-Score, and 0.96 AUC-ROC on multi-modal clinical data (ECG, EEG, notes, demographics).
• Maintained low privacy leakage (0.02) and communication costs (12.5 MB).

Conclusions:

• Mamba-Fusion offers a scalable, privacy-preserving solution for multi-modal disease prediction.
• The framework effectively balances data protection with high predictive accuracy.
• Mamba-Fusion supports secure, large-scale collaborative healthcare analytics.