Feature extraction of time series data on functional near-infrared spectroscopy and comparison of deep learning performance for classifying patients with Alzheimer's-related mild cognitive impairment: a post-hoc analysis of a diagnostic interventional trial.

Abstract

This study aimed to define a method of classifying patients with mild cognitive impairment caused by Alzheimer's disease by the retrieval of functional near-infrared spectroscopy (fNIRS) signal characteristics obtained during olfactory stimulation and the validation of deep learning findings. Participants were recruited for the study from March 02 and August 30, 2021. A total of 78 participants met the criteria for categorization. The Mini-Mental State Examination and the Seoul Neuropsychological Scale were used to distinguish between patients with mild Alzheimer's disease-related cognitive impairment and healthy controls. fNIRS data received during olfactory stimulation were used to create 1,680 time-series sample values. A total of 150 indices with a p-value ≤ 0.1 were used as deep learning features to construct the result values for 120 models accounting for all conceivable combinations of data ratios. For this trial, 78 participants were recruited for the original intervention trial. The average accuracy of the 120 deep-learning models for classifying patients with Alzheimer's-related mild cognitive impairment ranged from 0.78 to 0.90. Sensitivity ranged from 0.88 to 0.96 for the 120 models, while specificity ranged from 0.86 to 0.94. The F1 scores ranged from 0.74 to 0.88. At 0.78 to 0.90, the precision and recall were equivalent. This trial using a deep-learning model found that the representative value extracted from the time series data of each channel could distinguish between healthy people and patients with mild cognitive impairment caused by Alzheimer's disease.