Explainable artificial intelligence model to predict brain states from fNIRS signals.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep Learning (DL) methods for classifying functional Near-Infrared Spectroscopy (fNIRS) signals often lack interpretability.
• Understanding which signal features drive classification is crucial for advancing brain-computer interfaces and neuroscience research.

Purpose of the Study:

• To develop and evaluate an explainable AI (xAI) system for fNIRS signal classification.
• To enable the decomposition of DL model outputs onto input variables for fNIRS signals.
• To identify specific brain regions and hemodynamic responses contributing to motor task and motor imagery classification.

Main Methods:

• Proposed an xAI-fNIRS system comprising a classification module and an explanation module.
• The classification module utilized two sliding window-based classifiers: 1-D Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM).
• The explanation module employed SHAP (SHapley Additive exPlanations) to interpret the CNN model's output.

Main Results:

• Achieved over 96% classification accuracy for both Motor Task (MT) and Motor Imagery (MI) datasets using CNN and LSTM models.
• The explanation module successfully identified critical channels and their corresponding oxy- and deoxy-hemoglobin levels contributing to classification.
• Demonstrated the system's ability to differentiate between overt and covert motor imagery based on fNIRS signals.

Conclusions:

• The developed xAI-fNIRS system provides high classification accuracy for distinguishing brain states from fNIRS data.
• The system offers valuable insights into the signal features and brain regions underlying motor control and imagery.
• This approach enhances the interpretability of DL models in neuroimaging, paving the way for more sophisticated BCI applications.