Cardiogram Analysis - Feature Extraction and Clustering Studies.
Abstract
This work is concerned with the problem of morphology analysis, with emphasis on the problems of feature extraction and classification via clustering. The problem of lead reduction of twelve lead electrocardiograms is considered using factor analysis. Experiments with a limited amount of data indicate that eight factors are sufficient for preserving the information present in the electrocardiogram. A general approach to linear feature extraction is presented. An alternate approach to feature extraction using piecewise continuous polynomial approximations is introduced, which has desirable local optimality properties. Optimal sampling times are determined via dynamic programming. A dynamical heart model interpretation is given. The diagnosis problem is treated as an unsupervised clustering problem. After a rather general treatment of cluster analysis, a decision directed algorithm using a Mahalanobis distance measure is formulated. Tests on actual data using 20 classes indicate that clustering may be a superior method in comparison to statistical likelihood ratio methods. The problem of visualizing high-dimensional clusters in a two-dimensional manifold is considered and a systematic method is developed for its solution.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 1975
- Accession Number
- ADA024827
Entities
People
- Adnan Akant
- Donald E. Gustafson
- Sanjoy K. Mitter
Organizations
- Charles Stark Draper Laboratory