Date of Graduation


Document Type



This document reports a research effort to define criteria which can be used for early indications of right ventricular hypertrophy (RVH) in association with coal miners' pneumoconiosis. Selected measurements of the vectorcardiograms (VCG's) of 18 miners with pneumoconiosis are compared with corresponding measurements of a normal group of 32 non-miners. The components of the vectors associated with the MRSV, SMRSV, MLSV, and the MAZ are the selected measurements. MAZ is a vector occurring before the MLSV and which has a maximum anterior component. Each of these vectors has been reported as an indicator of RVH. The major methods of analysis were the approximate t statistic and pattern recognition techniques. The approximate t statistic was used to evaluate the differences in estimated means, to project directions of population mean changes, and as a feature selection procedure. The following pattern recognition procedures were used: 1. Fixed Increment Adaptive Method (FI) 2. Multiple Linear Regression (MLR) 3. Discriminant Analysis (DA) 4. Polynomial Discriminant Method (PDM) 5. Multivariate Normal (MN) The components of the VCG's were normalized and the first four pattern recognition procedures listed above were used to define the weights of linear separating surfaces. The absolute magnitude of the weights were ordered and assigned a rank number corresponding to the ordering. With the pattern recognition procedures, it is shown that the two groups are highly separable using: 1. All 15 components in linear classifiers (Multivariate Normal is inherently quadric) 2. All planar pairs (xy, xz, or yz) in linear classifiers 3. The top five features of each of the ranking tests in quadric classifiers 4. Only the x, y, or z components in quadric classifiers 5. The vector magnitudes in quadric classifiers. The MLR and DA procedures are shown to be basically equivalent and also to give the best subset of features for the data used. The t test is shown to give the poorest feature subset. The FI method gives best classification results and the RDM the poorest with the constraints imposed. The Generalized Mahalanobis statistic is used to evaluate the worth of the feature subsets. It is shown that the means of the 15 components differ significantly and that the xz plane accounts for most of this difference.