Robust Brain State Decoding using Bidirectional Long Short Term Memory Networks in functional MRI.

Area of Science:

• Neuroscience
• Machine Learning
• Cognitive Science

Background:

• Brain state decoding from fMRI data is crucial for understanding cognitive processes.
• Analyzing temporal dynamics in fMRI is essential due to the inherent latency in BOLD responses.
• Conventional Long-Short Term Memory (LSTM) models struggle to capture future temporal context.

Purpose of the Study:

• To develop a robust method for brain state decoding using fMRI data.
• To address the limitation of conventional LSTMs in encoding future temporal information.
• To improve the accuracy of predicting brain states by incorporating bidirectional temporal context.

Main Methods:

• Utilized a bidirectional LSTM (bi-LSTM) architecture to process fMRI data.
• Fed fMRI sequences in both normal and reverse time-order to separate LSTM networks.
• Collated hidden activations from both forward and reverse directions to build the model's memory.

Main Results:

• The bidirectional LSTM model demonstrated robust brain state decoding.
• The bi-LSTM approach successfully encapsulated information from both past and future fMRI instances.
• Achieved an 18% improvement in accuracy compared to unidirectional LSTM models in predicting brain states.

Conclusions:

• Bidirectional LSTM effectively models fMRI BOLD response dynamics without delay adjustment.
• Incorporating future temporal context via bi-LSTM significantly enhances brain state decoding accuracy.
• This approach offers a promising advancement for analyzing dynamic brain activity in neuroimaging studies.