A neural network approach for noninvasive detection of coronary artery disease

Abstract

The major cause of death in many cases is Coronary Artery Disease (CAD). Thisdisease can be detected by angiography. However, this technique is expensive, risky andinvasive. Another noninvasive technique for detecting coronary occlusion before theybecome serious enough to induce symptoms is based on the knowledge that coronarystenoses produce sounds due to the tulTbulent flow in partially occluded arteries. Recently,experimental systems that make use of the heart sounds for noninvasive detection of CADhave been the subject of active investigation by some research groups.In this study, we intended to improve on the previous studies concernmgnoninvasive detection of CAD, using some adaptive noise canceling schemes and artificialneural networks for automatizing detection. For this purpose, using a system developed inthe Institute, which includes a PC, two sensitive sound channels and an ECG channel, anumber of clinical studies have been performed.Heart sounds from 60 patients (22 healthy and 38 diseased) were recorded in arelatively quiet hospital room, while ambient sounds and patient's ECG were alsosimultaneously recorded. A sampling frequency of 4 kHz was used for data acquisition.Using ECG information, diastolic portions of the sound signals were isolated manually.The sound signals were first passed through an analog band-pass filter with 150 Hzand 1200 Hz cut-off frequencies and then an adaptive frequency domain filter was used toeliminate the background noise. Window functions of periodogram were employed toachieve better spectral estimation. Frequency regions that were related with the coronaryflow was defined. A two layer neural network with eight hidden nodes was trained usingdata from 20 patients. The neural network was then used for the diagnosis of theremaining 40 patients and gave correct classification rate of 62. 5%.Keywords : CAD, noninvasive techniques, angiography, diastolic heart sounds, adaptivefiltering, periodogram, artificial neural networks.