Use of fuzzy logic in the interpretation of electrocardiograms

Interpretations of the electrocardiogram by expert systems are subject to error because of the presence of noise in the electrocardiogram, and extracted electrocardiographic features. The noise has several sources: physiological, electrical, and algorithmic. The electrocardiographic diagnostic HELP frames use Classically formulated deterministic rules to interpret electrocardiographic data for myocardial infarction. The rules are written with absolute or "crisp" boundary conditions, and are evaluated according to binary logic. If the electrocardiographic wave data have significant amounts of noise, the rules are subject to errors in their interpretations (classification errors). Using Fuzzy Logic and Set formalisms, Fuzzy rules can be created. Fuzzy rules use Fuzzy subsets to describe the electrocardiographic features, and a multivalent logic. Fuzzy membership functions ascribe the possibility or confidence that a datum is a member of a fuzzy subset, such as a significant q-wave or inferior infarction. This particular project compares a Fuzzy rule to the Absolute Criterion from the HELP electrocardiographic analysis frames for relative classification error rates in the interpretation of inferior infarction. This particular project compares a Fuzzy rule to the Absolute Criterion from the HELP electrocrdiographic analysis frames for relative classification error rates in the interpretation of inferior infarction. The Fuzzy and Absolute Criteria were compares to each other in a study of serial interpretation from sequential electrocardiograms from computerized patient records of LDS Hospital. The Absolute Criterion was more consistent in its interpretational patters than was the Fuzzy Criterion when compared to the physicians' overread statements. A Monte Carlo Simulation was developed to study the effects of noise on the interpretative behavior and stability of the Absolute and Fuzzy Criteria. If the Absolute (viz, the Relaxed Criterion) and Fuzzy Criteria have similar sensitivities and specificities, and data with the same amount of noise, then their interpretational behavior and stability profiles are not significantly different. If the gain is equal to zero, the Fuzzy Criterion mimicks exactly the Absolute Criterion in both interpretational behavior and stability. The gain of the Fuzzy Criterion alters the behavior profile of the Fuzzy Criterion. The Fuzzy Criterion appeared to be less prone to classification errors for inferior infarction. In a second patient study, sensitivities, specificities, and positive predictive values were calculated for the criteria. Altering the gain of the Fuzzy Criterion altered the sensitivity and specificity of the Fuzzy Criterion. The Fuzzy Criterion was less prone to classification errors if the data were indicative of inferior infarction as shown by the higher sensitivity for the Fuzzy Criterion. However, the Fuzzy Criterion was prone to classification errors if the data were not indicative of inferior infarction as shown by a higher specificity for the Absolute Criterion. This project has demonstrated the interesting possibility for the use of Fuzzy formalisms to augment the Bayesian and Classical formalism employed in the HELP system for the interpretation of electrocardiograms.