Session S24.6

Retrieval of Similar ECG Images from a Database

T Syeda-Mahmood*

IBM Almaden Research Center
San Jose, CA, USA

An electrocardiogram (ECG) is an important and commonly used diagnostic aid in cardiovascular disease diagnosis. Physicians routinely perform diagnosis by a simple visual examination of ECG waveform shapes recorded on paper printouts produced by ECG recorders. In addition, most hospitals have a large collection of historical ECG data per patient that still exists in paper form. The paper printout or their scanned versions are in fact, a desired form for physicians to interpret, being easier to view than digital recordings. Digital ECG recordings are sampled very finely (1000 samples/sec) and contain a large amount of noisy data with baseline wandering problems, whereas the paper printouts are actually cleaner in appearance. Further, the relevant heart beats that illustrate a disease are carefully screened by technicians before taking the recording, so that the paper version shows only 6 seconds of data of highly relevant data needed for diagnosis. Further, all 12 lead information is displayed compactly in a standardized (lead order is fixed) 3 row by 4 column format. Thus it is reasonable to expect that such ECG representation when scanned in an image can be directly useful for finding similar ECG image reports.
In this paper, we address, for the first time, the problem of retrieval of similar ECG image from a database of scanned ECG images. Unlike other approaches to scanned ECG that extract the multi-channel ECG signals from the scanned images, we directly characterize the ECG image pattern. Segmentation and extraction of single channel ECG waveforms from the scanned ECG image is difficult particularly when the diseased state of the patient gives rise to large fluctuations in recording causing signal cross-over on the page. By directly recognizing the pattern, we avoid such lead separation errors. In our approach, we remove background text and paper grid using color and text thresholding. The resulting waveform image depicting the 12 channels is retained directly as a pattern to be recognized. Given a query image, the pattern is similarly extracted from query image and matched to the stored pattern using a variant of shape-based dynamic time warping.
We tested the method by creating a database of scanned ECG recordings. Our dataset came from hospitals in India where we collected patients with known diseases where ECG is a good indicator of the disease. Another set was assembled from medical textbooks illustrating sample ECG interpretation per disease. Our collection has over 500 ECG image whose ground truth disease labels were known. Each of the images in the database was used as a query for testing the system, and the number of times the top matches corresponded to ECG images of the same disease class was used to derive the precision and recall numbers. Overall our system was found to achieve 87% recall and 73% precision.

(Abstract Control Number: 82)