Abstract

Arrhythmia Classification by Multinomial Logistic Regression

Author: Jianwei Zheng

An arrhythmia is a problem with the rate or rhythm of the heartbeat. During an arrhythmia, the heart can beat too fast, too slow, or with an irregular rhythm. Almost all of serious arrhythmias often can be successfully treated, and all kinds of arrhythmia can be detected by ECG data. But the arrhythmia happened randomly during the initial stage, and patients do not have strong uncomfortable feeling. Technicians have to continuously collect ECG data, maybe for couple days, if they intend to confirm certain type of arrhythmia. Nowadays, more and more vendors provide wearable tools that can collect ECG data over 3 days, such as Vital Connect, AliveCor and so on. After the ECG is available, the challenge to help on automatically diagnostic is as following, automatically recognizing P, Q, R, S, T, U waves from normal and abnormal ECG graph, and fault tolerance algorithm to classify or discriminate arrhythmia. To extract P, Q, R, S, T, and U wave information as features of classification model, the multi-resolution Wavelets analysis are applied. The raw ECG data will be de-noised at first, and decomposed to 8 level by Symlet5 wavelet. There are four common threshold selection methods: universal, minimax, Steins unbiased estimate of risk (SURE) and minimum description length (MDL). In the proposed scheme, we choose the universal threshold selection method. After feature data extraction, the following 9 features will be used in the subsequent implementation of multinomial logistic regression classification model, (a) normalized PR interval (b) normalized RR interval (c) normalized QRS interval (d) normalized PR interval (e) normalized QT interval (f) normalized ST interval (g) normalized R amplitude (h) normalized P amplitude (i) normalized T amplitude. We use exhaustive best subset variable selection for multinomial logistic regression with the 37 types of arrhythmia as the outcome variable of interest to derive the best predictive model. Thus, we develop a novel statistical approach for automatic arrhythmia classification that assigns predictive probabilities for each patient and each type of irregularity allowing doctors to save time and only manually assess the most demanding cases.

Co Author/Co-Investigator Names/Professional Title: Dr. Cyril Rakovski Associate Professor Schmid College of Science and Technology; Mathematics and Computer Science Dr. Mohamed Allali Associate Professor Schmid College of Science and Technology; Mathematics and Computer Science