Spatial estimation of the 12-lead electrocardiogram using linear, least-squares modeling techniques.

Abstract

The objective of this dissertation research was to investigate accurate spatial estimation of the 12-lead electrocardiogram (ECG) from a three-signal basis set using linear least-squares techniques. Four key issues were germane to accurate estimation: (1) continuity of a spatial autocorrelation function among the ECG waveforms, (2) stationarity of the autocorrelation function within an ECG cycle, (3) redundancy of signal information, and (4) stationarity of the autocorrelation over time and in the presence of cardiac disease. The target waveforms of leads V$\sb1$, V$\sb3$, V$\sb4$, V$\sb5$, and V$\sb6$, were estimated by three independent signals, leads I, II and V$\sb2$, using linear models. The research project was divided into two phases--modeling and synthesis. The purpose of modeling was to derive a set of techniques which accurately estimated ECG waveforms. Standard methods of linear estimation, including orthonormalization and block and adaptive processing, are combined with time and frequency segmentation and patient-specificity to devise new piecewise-linear approaches. The accuracy of the estimated ECGs was evaluated on over 400 recordings using error measurements, diagnostic measures, and clinical assessment. The aim of the synthesis phase was to evaluate the robustness of the ECG-estimation parameters over time. The concept of robustness referred to the capability of accurately reconstructing an ECG waveform using model parameters derived from another recording. Multiple-day synthesis was evaluated in 50 patients with inter-recording periods up to eleven days. The accuracy of the synthesized ECGs was assessed using error measurements and clinical evaluation. The results from the ECG modeling estimation indicated that accurate reconstruction of the V$\sb1$, V$\sb3$, V$\sb4$, V$\sb5$, and V$\sb6$ waveforms was achieved with patient-specific models and time segmentation optimized for a given waveform. Under some circumstances, the accuracy was also improved by frequency segmentation of the ECG waveforms. The new approach of time and frequency segmentation combined with linear least-squares models was a very effective mechanism for deriving piecewise-linear solutions in the presence of inherent nonstationarities. The results from multiple-day ECG synthesis suggested that the simple three-parameter linear models were not sufficiently robust in the presence of inter-recording nonstationarities. However, new insights into the ECG-synthesis problem were afforded by these results. The need to incorporate additional waveform information into the linear models was documented. The dependence of the parameter robustness on the stationarity of the spatial autocorrelation was also highlighted. It was shown that a better understanding of the cause-and-effect relationship between changes in the autocorrelation function and the presence of cardiac disease was required for an accurate and practical solution.