Automated EKG Annotation with Neural Networks

Abstract

We have developed two automated systems using 11,810 physician-annotated single lead EKG signals. A focused tap delay line multilayer neural network followed by a threshold detection module is used to develop the QRS detection system. Approximately, a miss rate of 5.6%, false detection rate of 5.5%, error rate of 0.049%, and relative error rate of 10.6% is observed with a precision of 10 ms. With this same precision, the Pan and Tompkins QRS detection algorithm is tested on these sequences and encountered 20.3% of false detection rate, 19.8% of miss rate, 0.2\% error rate, and 33.4% of relative error rate. For a precision of 167 ms, the neural network has only 55% of the error made by the Pan and Tompkins QRS detection algorithm. The beat classification system uses the correctly detected beats along with the physician indication of normal and abnormal for the beat as an input for the multilayer neural network. The training and validation error for this system is 24.7% error rate, 24% false detection rate and miss rate. The inconsistency in the physician annotation has a significant effect on these systems. Sixty percent of the error in the QRS detection system is due to the inconsistency. About 41.4% of the false negative and 30.5% of the false positive errors of the beat classification system are also due to this inconsistency. A more accurate data set will augment the performance of these systems.