Finding disease similarity by combining ECG with heart auscultation sound

Abstract

Heart auscultation and ECG are two very important and commonly used diagnostic aids in cardiovascular disease diagnosis. Physicians routinely perform diagnosis from simple heart auscultation and visual examination of ECG waveform shapes. It is common knowledge to physicians that patients with the same disease have similar-looking ECG shapes and comparable heart sounds. A key idea explored in this paper is to automatically capture such shape similarity in the ECG and audio signals, which are combined to find disease similarity. Specifically, we present a general method of capturing the perceptual shape similarity of the ECG and audio waveforms by modeling the morphological variations in the signals representing the same disease across patients. Differences in shape corresponding to the same disease are modeled as a constrained non-rigid translation. Patients with similar diseases are retrieved by recovering the non-rigid alignment transform using a variant of dynamic time warping. Results are presented that demonstrate the method on audio shape-based discrimination of various cardiovascular diseases.