K-means clustering for unsupervised participant grouping from fNIRS brain signal in working memory task

Abstract

Functional near-infrared spectroscopy (fNIRS) images the changes in oxygenated and deoxygenated hemoglobin blood concentration on the cortical surface of the brain. This technology has been widely used to assess human perceptual and cognitive functions, with many studies showing a positive correlation between working memory (WM) task difficulty and the hemodynamic response measured in frontal regions of the brain. In this work, an unsupervised machine learning (ML) method, k-means clustering, was used to study the relationship between WM task difficulty, user performance, and hemodynamic brain response. We used an fNIRS data set derived from 25 healthy participants performing a WM task with four difficulty levels, with data collected using a portable fNIRS device. For each participant, the raw fNIRS time-series data were processed and block-averaged to find hemodynamic responses in the four WM task conditions. A k-means clustering ML algorithm was used to identify clusters of participants based on hemodynamic responses across the four conditions. The elbow method suggested three optimal clusters (groups). We studied the task-induced hemodynamic response variation in the three groups as well as task performance (accuracy and reaction time). Clusters were related to WM task performance. At increasing task difficulty levels (1-back, 2-back), cluster membership predicted changes in task performance. Specifically, higher hemodynamic responses in the 2-back condition predicted poorer performance outcomes in the 1-back and 2-back conditions. At higher task difficulty levels, the hemodynamic response signals greater metabolic costs associated with diminishing performance, suggesting future directions for closed-loop systems for monitoring and predicting performance outcomes.