Design of statistical pattern recognition methods for improved cardiac rhythm specificity by implantable cardioverter defibrillators.

Sudden cardiac death (SCD) claims the lives of over 400,000 people annually. A modern medical device called the implantable cardioverter defibrillator (ICD) is used to forestall SCD by providing incessant electrocardiographic monitoring and immediate electrical therapy. Specifications for this device require efficient power consumption for ensuring device longevity and conservation of power for delivery of therapy. For this reason, detection algorithms that sense potentially lethal arrhythmia are based on simplistic measures of heart rate and rate related criteria. The use of these algorithms has resulted in high levels of sensitivity but low levels of specificity (0.59-0.73) with numerous false shocks. The focal point of the research described in this dissertation is two new designs for increasing the specificity of current ICD detection algorithms while maintaining acceptable levels of sensitivity and device longevity. Correlation waveform analysis (CWA) that employs time domain template (mean vector) matching is a favored technique for morphological arrhythmia recognition. Despite high levels of detection specificity and sensitivity, CWA has not been used in ICD detection because of its high computation requirement. In an effort to design a computationally efficient version of CWA, a classic detection theory approach was used whereby a minimum effectual bandpass and sampling rate were rigorously identified, followed by an optimal eigenvector-based signal transformation and dimensionality reduction scheme using the Karhunen-Loeve Transform. Results yielded a significant decrease in the computational requirement of CWA (53.6% to 79.4% fewer multiplies, 55.5% to 82.2% fewer adds) without any sacrifice of diagnostic accuracy. A second approach, Samples Above the Median (SAM), was derived from the sign-detector (a nonparametric method of binary hypothesis testing) for morphometric signal classification. SAM was designed, developed and rigorously tested to determine its performance level with regard to sensitivity and specificity of diagnosis. Following the identification of a median value for normal cardiac conduction, SAM requires minimal hardware (two counters and a comparator) to classify each cardiac event. This requirement is well within current device technology. Correct classification of 103 of 106 (sensitivity = 0.906, specificity = 1.00) intraventricular electrocardiographic episodes was obtained using a fully automated version of SAM.Ph.D.Applied SciencesBiomedical engineeringElectrical engineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/130310/2/9722045.pd