Neural network based multi sensor heart sound analysis

It is demonstrated how neural nets, used for nonlinear modeling and signal classification, can estimate from a continuous recording of an acoustical stethoscope signal the beginning of a cardiac cycle, the systolic and diastolic time intervals and suggest suitable diagnoses in case of heart failures and arrhythmia. While during the training phase multisensor signals and catheterized verifications of heart failures have been utilized, the PC-based diagnostic system analyzes heart sound signals quasi-on-line in a way similar to how physicians perform their subjective learning and interpretation of stethoscopic signals. The systolic and diastolic time intervals of healthy probands as well as patients with heart failures estimated by the neural network were found to be in excellent correspondence with measurements. Also, the sound pattern recognition results demonstrate the high accuracy, though training and test sets were based on different sources. Using this method a valuable documentation of heart failure development or treatment with the auscultation technique is possible.<<ETX>>