Unsupervised Dynamic Time Warping Clustering for Robust Functional Network Identification in fNIRS Motor Tasks.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Functional near-infrared spectroscopy (fNIRS) is crucial for non-invasive brain-computer interfaces (BCIs).
• Interpreting fNIRS signals is challenging due to hemodynamic response variability and temporal jitter.
• Standard linear methods struggle with non-linear temporal shifts in brain activity.

Purpose of the Study:

• To validate an unsupervised Dynamic Time Warping (DTW) clustering framework for robust motor network identification from fNIRS data.
• To accommodate non-linear temporal shifts in hemodynamic responses for improved functional connectivity analysis.
• To compare the DTW framework's performance against traditional Pearson correlation methods.

Main Methods:

• Utilized a public fNIRS dataset (N=30) involving right-hand, left-hand, and foot tapping tasks.
• Implemented a preprocessing pipeline including Wavelet Motion Correction and Common Average Referencing (CAR).
• Applied unsupervised DTW clustering via Z-score normalized DTW distance matrices and hierarchical clustering.

Main Results:

• The DTW framework achieved 53.17% network identification accuracy, significantly outperforming Pearson correlation (48.06%, p < 0.05).
• Successfully identified distinct, somatotopically appropriate motor networks for hand and foot tasks.
• Demonstrated superior performance in capturing functional networks despite temporal signal variations.

Conclusions:

• Unsupervised DTW clustering offers a robust, data-driven approach for analyzing fNIRS data in BCIs.
• This method overcomes limitations of conventional linear techniques in detecting functional connectivity with temporal jitter.
• The DTW framework shows significant potential for advancing next-generation asynchronous BCIs.