• Rune Kongsgaard Rasmussen
Over 17.5 million people die ev- ery year due to cardiovascular disease. The most widespread used to detect heart disease is the 12- Lead ECG. In ECG research there has been a lot of research in understanding its features and the underlying grammar of the ECG not only for diag- nostics but also for identification of a specific per- son or even mood. Hierarchical Temporal Mem- ory (HTM) seemed to be able to identify the underlying grammar automatically in a robust matter. It was therefore evaluated to which extent HTM could identify and use biometric grammar of the ECG in the context of identifying subjects based on their heartbeat from unfiltered ECG.
Data subsets with different number of subjects were created from an ECG database with 25,000 subjects. Two sessions were included from each subject of 500 Hz 10 seconds ECG. Lead II was used from these ECGs and the dataset contained healthy and unhealthy subjects. A HTM model was configured and built using Numenta NuPic software and exposed to the different subsets. Subsets were applied with different number of iterations spanning from 1 to 1,000.
The maximum accuracy was achieved from using only 10 subjects where the accuracy was found to be 31.3 % and down to 0.07 % for 1,250 subjects.
The number of subjects decreased this accuracy and the number of iterations had no effect. It was not possible to determine if the accuracy found was due to limitation of intersubject variability of the ECG or to the configuration of HTM. The authors of the software used in this current study now provide tools that might be a logical next step in improving the results.
Publication date2017
Number of pages64
External collaboratorNew Zealand College of Chiropractic
Rasmus Wiberg Nedergaard Rasmus.Nedergaard@nzchiro.co.nz
ID: 259117858