Application of Deep Learning Techniques to Diagnose Mild Cognitive Impairment: Functional Near-Infrared Spectroscopy Study

Abstract

In this paper, we propose a novel deep learning algorithm (i.e., multi-scale convolutional neural network (CNN) accompanied by long short-term memory (LSTM) layers) for diagnosing patients with functional near-infrared spectroscopy (fNIRS). Mild cognitive impairment (MCI) is associated with aging and describes early symptoms of severe cognitive impairment known as Alzheimer's disease (AD). Meanwhile, early detection of MCIs can prevent progression to AD. Many studies have been investigated on MCI diagnose over the past decade. In this work, we measure three mental tasks (i.e., N-back, Stroop, semantic verbal fluency tasks) for 38 MCI patients and 25 healthy people. The proposed multiclass CNN-LSTM network improves classification accuracy and reduces the classification time. By classifying the fNIRS data directly and minimizing the preprocessing process without finding the region of interest channels, the reduction of data processing time was achieved. The proposed algorithm was compared with a single LSTM algorithm to validate the performance and compare the accuracy, precision, recall, and F1 score with standard metrics. For the proposed multiclass CNN-LSTM, the maximum classification accuracy was 83.33% when performing N-back tasks, with an average accuracy of 77.77% for all tasks. For the LSTM only, the maximum classification accuracy is 73.33% when performing N-back tasks. The average classification accuracy for all tasks is 57.77%. The results, therefore, demonstrate that the proposed algorithm is superior to the LSTM, indicating the potential to be used for online classification.