Efficient performance of neural network models as artificial intelligence prediction tools in cardiology

Pattern generalization capabilities of neural networks are evaluated in two models: (1) prediction of coronary lesions from myocardial perfusion SPECT data; and (2) determination of comments to stress test results (EST). The SPECT network correctly predicted the stenosed/occluded vessel in all single-vessel disease cases, while the success rate was 83% for multi-vessel diseases. The EST network provided 106/125 correct interpretations. It is concluded that the functional success on software simulated neural networks is derived from the underlying computing model. Input values are comparable to multidimensional vectors since they are sequences of values wherein value position in the sequence is as important as the value itself. Interneural connections are numerical matrices instead. Network knowledge is included in the matrices that define each particular network. This structure allows some generalization ability since the response can always be computed for unexpected or incomplete data clusters