Study on The Identification of Normal/Abnormal Conditions by Applying MFCC Standard Deviation Characteristics of Heart Rate and Respiration Signals

Abstract

This study investigates a neural network with clustering capabilities designed to identify whether a person is in a normal physiological condition using heart rate and respiration signals. In particular, the study extracts MFCC (Mel-Frequency Cepstral Coefficients) features from the bio-signals such as heart rate and respiration and performs clustering using an Autoencoder and K-Means algorithm. Based on the resulting clusters, a simulated experiment was conducted to determine whether the method could distinguish between healthy individuals and those in abnormal states. Through these experiments, it was confirmed that the classification performance varies depending on the number of clusters (states), which is a key parameter of the K-Means algorithm. Based on further analysis and discussion, the study proposes a method to automatically calculate the optimal number of K-Means states by reflecting the standard deviation characteristics of the MFCC distributions of healthy individuals and patients with heart failure. To validate the proposed approach, simulated experiments were conducted using heart rate and respiration data of healthy individuals and heart failure patients from Physio.Net, applying both training and non-training datasets. As a result, the proposed method successfully estimated the optimal number of states, enabling reliable identification performance of heart rate and respiration signals using the automatically determined number of states.