Exploring fNIRS-Based Brain State Recognition and Visualization through the use of Explainable Convolutional Neural...

Area of Science:

• Neuroimaging
• Machine Learning
• Biomedical Engineering

Background:

• Functional near-infrared spectroscopy (fNIRS) is a growing neuroimaging technique increasingly used in clinical research.
• Current machine learning applications in fNIRS lack interpretability and struggle with limited clinical data sample sizes.
• There is a need for interpretable models that can effectively analyze fNIRS data with minimal preprocessing.

Purpose of the Study:

• To develop an interpretable machine learning model for fNIRS data analysis.
• To achieve accurate classification with minimal human manipulation, channel selection, or feature extraction.
• To visualize biomarkers and identify important brain regions using class-specific gradient information.

Main Methods:

• Utilized all available fNIRS channels for analysis.
• Developed a novel interpretable machine learning model.
• Employed class-specific gradient information for biomarker visualization and region localization.

Main Results:

• The developed interpretable model achieved 6% higher accuracy than conventional Support Vector Machine (SVM) methods in within-subject classification.
• The model demonstrated physiological relevance by focusing on left-brain signals for right-hand tasks and right-brain signals for left-hand tasks.
• Class-specific gradients successfully visualized important brain regions and biomarkers.

Conclusions:

• The interpretable fNIRS-based machine learning model offers enhanced classification accuracy and interpretability.
• The model's ability to identify relevant brain regions supports its physiological validity.
• This approach holds potential for clinical applications in diagnosing neurological conditions and predicting treatment outcomes.