Using Machine Learning to Decode fNIRS Data

Project Supervision

• Professor Benjamin BECKER, The University of Hong Kong
• Miss. Michelle Tsang (Post-doctoral researcher), The University of Hong Kong

Project Link

Read the full article on Medium

Overview

This research investigates cognitive reappraisal through neurofeedback-based training using functional near-infrared spectroscopy (fNIRS). The study employs various analytical approaches including correlation matrices, repeated-measures ANOVA, and machine learning methods like Support Vector Machines (SVM).

Key Hypotheses

1. Brain activity in the lateral prefrontal cortex (lPFC) will increase across training runs during reappraisal trials
2. Cognitive reappraisal enhances connectivity in the lPFC ROI more strongly for the experimental neurofeedback group
3. SVM classifier can distinguish between regulation and view conditions
4. SVM classifier accuracy significantly differs from random guessing

Methods

fNIRS Data Preprocessing Pipeline

Significant Signal Processing Steps & Results

5. Wavelet-Based Motion Correction

Before vs After:

• Successfully removed sudden spikes (IQR > 1.5)
• Smoother signal trajectory while preserving underlying neural responses

6. Bandpass Filtering

Filtering Results:

• High-pass filter (0.01 Hz): Removed slow drifts
• Low-pass filter (0.4 Hz): Eliminated high-frequency noise
• Removed physiological noise:
  • Heart rate (~1-1.2 Hz)
  • Respiration (~0.3-0.6 Hz)
  • Blood pressure/Mayer waves (~0.1 Hz)

8. CBSI Motion Correction

Improvements:

• Enhanced signal quality through correlation-based signal improvement
• Stronger mirroring effect between HbO and HbR signals
• More distinct hemodynamic response patterns

Region of Interest (ROI)

Results

Connectivity Analysis

• Technique: Pearson correlation matrices between channel pairs
• Finding: Stronger S4/-D2-D4-D6 cluster correlation in neurofeedback group (r > 0.7)

Time Series Analysis

• Technique: Averaged HbO response across trials and subjects
• Finding: Peak activation difference between groups during 6-12s window (5 x 10⁻⁸ mol/L)

Machine Learning Classification

• Technique: Linear SVC with grid search CV (optimal C = 0.1954)
• Finding: 77.78% accuracy, significantly above chance (χ² = 4.43, p = 0.0353)

Training Effects

• Technique: Repeated measures ANOVA with post-hoc pairwise t-tests
• Finding: Sessions 2-3 improvement in neurofeedback group (p = 0.013, F = 2.0655)

Key Findings

• Significant increase in brain activity between neurofeedback training sessions 2 and 3
• Enhanced functional connectivity during reappraisal tasks in the neurofeedback group
• SVM classifier achieved 77.78% accuracy in distinguishing between conditions
• Statistical significance confirmed through chi-square test (χ² = 4.43, p = 0.0353)

Code Notebooks

• Time Series Analysis
• Training Effect Analysis
• SVM Classification

Future Directions

• Investigation of additional statistical predictors
• Development of more sophisticated temporal feature analysis
• Enhancement of signal-to-noise ratio in machine learning models

References

1. Berboth, S., & Morawetz, C. (2021). Amygdala-prefrontal connectivity during emotion regulation: A meta-analysis of psychophysiological interactions. Neuropsychologia, 153, 107767.
2. Bookheimer, S. Y., et al. (2011). Decoding Continuous Variables from Neuroimaging Data: Basic and Clinical Applications. Frontiers in Neuroscience, 5(Pt 1).
3. Buhle, J. T., et al. (2014). Cognitive reappraisal of emotion: a meta-analysis of human neuroimaging studies. Cerebral Cortex, 24(11), 2981–2990.
4. Full reference list available in the paper.