NIRS-Based BCIs: Reliability and Challenges

Previously we contributed to the development of a brain-computer interface (Brainput) using functional near infrared spectroscopy (NIRS). This NIRS-based BCI was designed to improve performance on a human-robot team task by dynamically adapting a robot’s autonomy based on the person’s multitasking state. Two multitasking states (corresponding to low and high workload) were monitored in real-time using an SVM-based model of the person’s hemodynamic activity in the prefrontal cortex. In the initial evaluation of Brainput’s efficacy, the NIRS-based adaptivity was found to significantly improve performance on the human-robot team task (from a baseline success rate of 45% to a rate of 82%). However, failure to find any performance improvements in an extension of the original evaluation prompted a reinvestigation of the system via: (1) a reanalysis of Brainput’s signal processing on a larger NIRS dataset and (2) a placebo-controlled replication using random (instead of NIRS-based) state classifications [1].

The reinvestigation revealed confounds responsible for the original performance improvements and underscored several challenges for NIRS-based BCIs in general. Specifically, it revealed the original performance improvements were due to a disparity in difficulty between experimental conditions of the original evaluation (i.e., the task being easier in the adaptive versus the baseline condition). Moreover, the reinvestigation showed Brainput’s model of user multitasking (trained on the n-back task) generalized to neither the human-robot team task (the classifications showed systemic violations of basic hemodynamic principles) nor to other workload-inducing tasks (classifications of brain activity while users performed arithmetic were better than chance for only 1/4 of the subject population). Hence, in in an effort to identify ways forward, we first summarize the methods and results of this reinvestigation and then explore the challenges for achieving more reliable NIRS-BCIs.