Session S41.4

A Reconstructed Phase Space/Gaussian Mixture Model for the Characterization of Myocardial Infarcts

MA Mneimneh*, RJ Povinelli

Marquette University
Milwaukee, WI, USA

The 2007 Physionet/Computers in Cardiology challenge focuses on the ability to identify the segment, extent, and centroid of infarcts through ECG signals and body surface maps. The results from the participants are compared to a gold standard that consists of expert analysis of gadolinium-enhanced MRI data. The scoring procedure for the challenge consists of the percentage discrepancy of the estimated infarct size, the overlap between the infarct segments, and the distance between the centroids as compared to the gold standard. The challenge dataset consists of two records for training and two for testing. As a result, the infarcted records from the PTB Diagnostics database are used as training data. The PTB Diagnostics database consists of 549 records from 294 subjects of which 367 records taken from 148 patients are infarcted. The main hypothesis in this work is that the ordinary 12 leads and 3 frank leads ECG signals contain the information regarding the segment of the infarct. In order to study the proposed hypothesis, this work proposes the use of a reconstructed phase space and Gaussian Mixture Model approach in order to classify the infarcted segments. The segments under consideration are anterior, inferior, posterior, septal, lateral, and any combination of those 5 types of infarcts. The records with multiple types of infarcts are labeled as distinct classes. The number of classes under consideration is 22. The approach is applied to all 15 leads, and a majority vote is taken to determine the segments of the infarct. As applied to the PTB diagnostics dataset, the accuracy for identifying the 22 classes is about 30%, where chance is 4.5%. The result of applying the proposed approach to the testing set of the challenge yielded in a score of 0.69 out of 2.

(Abstract Control Number: 299)