Performance measurements of seismocardiogram interpretation using neural networks

The authors describe the performance of a neural network architecture on a fraction of the total patient population after the network was trained on a separate fraction of the same patient population. This method allows prediction of how well the neural network will perform on cases it has never seen before. Exercise seismocardiographs were recorded in a population of 260 patients: 130 where diagnosed as having coronary artery disease (CAD), and the remaining 130 were diagnosed as low risk normals. 320 neural networks were implemented. The performance of these neural network architectures in detecting CAD was calculated on the test set as well as on the training set. The best performance was attained when the network was able to realize similar levels of sensitivity and specificity on the learning set and the test set. Overlearning was seen when the network performed exceedingly well on the training set, but with poor results on the test set. Neural network prediction of CAD based on seismocardiograms can reach 75% sensitivity and 75% specificity on data that is new to the neural network