Binary Classification Methods for Movement Analysis from Functional Near-Infrared Spectroscopy Signals

This paper investigates different techniques for binary classification in a multi-participant setting, with a focus on complex movement tasks. It uses statistical methods to extract features from preprocessed biosignals acquired by functional near-infrared spectroscopy (fNIRS) from real participants, obtained from the validated finger-tapping dataset. Unique approaches are used to process the fNIRS signals, including attenuation of short channel contributions and various filtering and other pre-processing techniques. For this investigation, a number of algorithms are used to optimise hyperparameters and model topologies in six different models: four conventional machine learning methods and two artificial neural networks. Among these models, the support vector machine classifier emerges as the top performer, achieving the highest average accuracy, precision, recall and F1-score (89.17%, 91.44%, 86.67% and 88.92%, respectively). However, the multi-layer perceptron classifier shows superior performance in terms of area under the ROC curve (92.56%), closely followed by the convolutional neural network classifier (91.70%), suggesting their slightly better ability to discriminate between classes. This study highlights the potential of using different classification methods to improve the accuracy of biosignal analysis obtained from fNIRS devices.