Exercise seismocardiograms (SCG) were recorded in a population of 114 patients, 57 of which were diagnosed as having coronary artery disease. The remaining 57 were diagnosed as low-risk normals. Three 60-s SCG recordings were performed at rest, immediately after exercise, and after recovery. A multilayered neural network has been developed to accept the 48 input parameters derived from the three SCG recordings, and to produce an output value of zero if the input parameters correspond to a normal patient and an output value of one if the input parameters correspond to a diseased patient. Three SCG parameters have been identified as most sensitive indicators of heart disease. The relative performance of several neural-network architectures in detecting heart disease based on the selected SCG parameters is discussed. Optimal learning and overlearning issues are discussed. The work suggests that using artificial neural networks can be a useful tool in the interpretation of subtle changes in exercise SCGs necessary for the early detection of heart disease.