Unsupervised Learning Using Time-Domain and Frequency-Domain Features of Audio Signals for the Classification of Mild Cognitive Impairment

Abstract

In this research, we conducted a study on MCI detection with unsupervised learning. We collected a total of 104 audio samples from Mandarin-speaking test subjects. The study sample contains 72 MCI patients and 32 normal test subjects diagnosed by a clinical psychiatric doctor. Time-domain and frequency-domain features of the mid- and short-term audio signals are extracted and their accuracy performances are analyzed. With unsupervised learning based on frequency-domain features, the accuracy of MCI detection can reach 73%.