K-Means Clustering Machine Learning Approach Reveals Groups of Homogeneous Individuals With Unique Brain Activation, Task, and Performance Dynamics Using fNIRS.

Wearable functional near-infrared spectroscopy (fNIRS) for measuring brain function, in terms of hemodynamic responses, is pervading our everyday life and holds the potential to reliably classify cognitive load in a naturalistic environment. However, human's brain hemodynamic response, behavior, and cognitive and task performance vary, even within and across homogeneous individuals (with same training and skill sets), which limits the reliability of any predictive model for human. In the context of high-stakes tasks, such as in military and first-responder operations, the real-time monitoring of cognitive functions and relating it to the ongoing task, performance outcomes, and behavioral dynamics of the personnel and teams is invaluable. In this work, a portable wearable fNIRS system (WearLight) developed by the author was upgraded, and an experimental protocol was designed to image the prefrontal cortex (PFC) area of the brain of 25 healthy homogeneous participants in a naturalistic environment while participants performed n-back working memory (WM) tasks with four difficulty levels. The raw fNIRS signals were processed using a signal processing pipeline to derive the brain's hemodynamic responses. An unsupervised k-means machine learning (ML) clustering approach, utilizing the task-induced hemodynamic responses as input variables, suggested three unique participant groups. Task performance in terms of % correct, % missing, reaction time, inverse efficiency score (IES), and a proposed IES was extensively evaluated for each participant and the three groups. Results showed that, on average, brain hemodynamic response increased, whereas task performance degraded, with increasing WM load. However, the regression and correlation analysis of WM task, performance, and the brain's hemodynamic responses (TPH) revealed interesting hidden characteristics and the variation in the TPH relationship between groups. The proposed IES also served as a better scoring method that had distinct score ranges for different load levels as opposed to the overlapping scores of the traditional IES method. Results showed that the k-means clustering has the potential to find groups of individuals in an unsupervised manner using the brain's hemodynamic responses and to study the underlying relationship between the TPH in groups. Using the method presented in this paper, real-time monitoring of cognitive and task performance of soldiers, and preferentially forming small units to accomplish tasks based on the insights and goals may be helpful. The results showed that WearLight can image PFC, and this study also suggests future directions for the multi-modal body sensor network (BSN) combining advanced ML algorithms for real-time state classification, cognitive and physical performance prediction, and the mitigation of performance degradation in the high-stakes environment.