Decoding fNIRS Neural Responses: A Machine Learning Approach

Authors: Dalton Aaker $^1$ , Maureen Shader $^2$

Affiliations:

$^1$ Weldon School of Biomedical Engineering, Purdue University, West Lafayette, IN, USA

$^2$ College of Health and Hearing Sciences, Purdue University, West Lafayette, IN, USA

Abstract: Using machine learning (ML) to interpret neural data from brain-computer interfaces (BCIs) is an approach that has gained popularity as models have become more robust. This project aims to compare ML classifiers on their ability to predict the auditory stimulus applied to a subject using only the neural activity data. Human neuroimaging data collected via fNIRS from a single subject twice daily for five consecutive days was analyzed. The data followed a block-design paradigm with two conditions: meaningful auditory speech and silence serving as a baseline control. Hemoglobin concentration data was collected using a continuous-wave fNIRS system (NIRx NIRSport2) with specific source-detector pairs optimized for the regions associated with sound acquisition and language comprehension. Standard fNIRS data cleaning and preprocessing practices were applied, and Python's Sci-kit learn library was utilized for decoding and prediction on the extracted datasets. Estimators were trained on hemoglobin concentrations and applied stimuli, with cross-validation using leave one subject out. Some estimators required training on both physiological and fNIRS datasets, and it was discovered that physiological signals decrease the performance of the classifiers. To mitigate this, features were extracted and selected based on their classification ability. This feature selection technique not only increased prediction ability for models trained on fNIRS data and the combined fNIRS and physiology data, but also decreased the computational load of all models tested. This study provides insights applicable to machine learning, neuroscience, and optical engineering and the ability to combine cofactors for maximum prediction capabilities in machine learning models is a key area of ongoing research.

Preprint

Preprint

Decoding fNIRS Neural Responses: A Machine Learning Approach