Deep multimodal predictome for studying mental disorders.

Area of Science:

• Neuroscience
• Genetics
• Artificial Intelligence

Background:

• Neuropsychiatric disorders present diagnostic challenges due to population heterogeneity.
• Existing multimodal studies often use basic neural networks, assuming equal data modality contributions.
• Understanding the neural and biological underpinnings of mental disorders requires advanced analytical approaches.

Purpose of the Study:

• To enhance the classification accuracy of neuropsychiatric disorders using a multimodal approach.
• To develop an adaptive fusion mechanism that weighs the influence of different data modalities.
• To identify neural and biological mechanisms associated with mental disorders through model introspection.

Main Methods:

• A multimodal classification framework combining structural/functional neuroimaging and genomic data.
• Development of advanced neural networks for feature learning, including a multilayer perceptron, autoencoder, and bi-directional LSTM with attention.
• Implementation of a linear attention module for adaptive fusion of modality-specific features.

Main Results:

• Achieved 92% accuracy in schizophrenia prediction, surpassing state-of-the-art models.
• The model effectively integrated diverse data types (neuroimaging, genomics) for improved classification.
• Post hoc analyses identified significant neural features, genes, and biological pathways linked to schizophrenia.

Conclusions:

• The proposed adaptive multimodal model significantly improves the prediction of mental disorders like schizophrenia.
• Interpretable features identified by the model offer insights into the etiological mechanisms of these conditions.
• This approach provides a powerful tool for advancing our understanding of complex brain disorders.