Motion Artifact Contaminated Functional Near-infrared Spectroscopy Signals Classification using K-Nearest Neighbor (KNN)

Abstract

This research reports the use of K-Nearest Neighbor (KNN) method for classifying clean and motion artifact contaminated functional near-infrared spectroscopy (fNIRS) signals. fNIRS is one of the non-invasive methods to monitor brain activity by looking at the oxy-Hemoglobin (oxy-Hb) and deoxy-Hemoglobin (deoxy-Hb) levels. fNIRS data consists of 18 clean signals and 18 contaminated signals which needs to be classified by extracting its features. The feature extraction process uses a first-order statistical method such as kurtosis, skewness, mean, variance, standard deviation, and interquartile range. These features are then processed with weighted K-Nearest Neighbor (wKNN), one of the methods under KNN, to classify the artifact contaminated and clean fNIRS signals. The performance of the classification measured by accuracy, sensitivity, specificity, and AUC. Based on the trial, the highest performance result was obtained when using five features except for skewness with k=5, also when using four features except for skewness and kurtosis with k=9, which result in an accuracy of 88.9%, a sensitivity of 85%, specificity of 93.7%, and AUC of 0.90