A Multi-Modal Deep Learning Framework for Early Classification of Neurodegenerative Disorders

Alzheimer Disease and Parkinson Disease are neurodegenerative conditions of great importance to the classification as they have a specific clinical presentation as well as they are based on various biomedical modalities. The heterogeneous nature of the disease is not represented in the traditional unimodal AI systems and thus have restricted reliability in the real-world generalization processes. NeuroFusion suggests a combined multi-modal deep learning structure to early-classes to meet this clinical need. The objective of the approach is to replicate multi-input clinical thinking, by exploiting correlating sources of information rather than the classical methodology of handling inputs separately. It uses MRI scans with EfficientNet to analyze spatial information about the brain region, voice samples developed with LSTM to encode time speech patterns and structured biomarkers that are categorized using XGBoost. Subsequently, there is a fusion layer fusion coupled with modality-specific predictions to receive a single diagnosis. The Experimental evaluation demonstrates that NeuroFusion achieves a classification accuracy of 92.3% on average, the best result in comparison with current unimodal and early- fusion baselines by far, and a new standard on multimodal neurodegenerative disease prediction.