Diagnosis of Mild Cognitive Impairment Using Cognitive Tasks: A Functional Near-Infrared Spectroscopy Study

Background: Early diagnosis of Alzheimer’s disease (AD) is essential in preventing its progression to dementia. Mild cognitive impairment (MCI) can be indicative of early-stage AD. In this study, we propose a channel-wise feature extraction method of functional near-infrared spectroscopy (fNIRS) data to diagnose MCI when performing cognitive tasks, including two-back, Stroop, and semantic verbal fluency tasks (SVFT). Methods: A new channel-wise feature extraction method is proposed as follows: A region-of-interest (ROI) channel is defined as such channel having a statistical difference (p <0.05) in t-values between two groups. For each ROI channel, features (the mean, slope, skewness, kurtosis, and peak value of oxy- and deoxy-hemoglobin) are extracted. The extracted features for the two classes (MCI, HC) are classified using the linear discriminant analysis (LDA) and support vector machine (SVM). Finally, the classifiers are validated using the area under curve (AUC) of the receiver operating characteristics. Furthermore, the suggested feature extraction method is compared with the conventional approach. Fifteen MCI patients and fifteen healthy controls (HCs) participated in the study. Results: In the two-back and Stroop tasks, HCs showed activation in the ventrolateral prefrontal cortex (VLPFC). However, in the case of MCI, the VLPFC was not activated. Instead, Ch. 30 was activated. In the SVFT task, the PFC was activated in both groups, but the t-values of HCs were higher than those of MCI. For the SVFT, the classification accuracies using the proposed feature extraction method were 80.77% (LDA) and 83.33% (SVM), showing the highest among the three tasks; for the Stroop task, 79.49% (LDA) and 73.08% (SVM); and for the two-back task, 73.08% (LDA) and 69.23% (SVM). Conclusion: The cognitive disparities between the MCI and HC groups were detected in the ventrolateral prefrontal cortex using fNIRS. The proposed feature extraction method has shown an improvement in the classification accuracies, see Subsection 3.3. Most of all, the suggested method contains a groupdistinction information per cognitive task. The obtained results successfully discriminated MCI patients from HCs, which reflects that the proposed method is an efficient tool to extract features in fNIRS signals.