neuralnet-mcg

Convolutional neural networks (CNNs), and their possible uses in electrocardiography and magnetocardiography are explored. A CNN that diagnoses myocardial infarction in electrocardiograms (ECGs) taken from the Physikalisch Technische Bundesanstalt (PTB) diagnostic database is described. This CNN has a diagnosis accuracy of 99.8% in unseen patients, on par with state of the art machine learning methods. CNNs that diagnose magnetocardiograms (MCGs) generated by a magnetocardiograph developed by Mooney et al are described. With a best diagnostic accuracy of $(88 \pm 3)%$, these CNNs outperform results obtained by Kangwanariyakul et al, Fenici et al, Tantimongcolwat et al, and Wilson. Through a moving window method, the diagnostic powers of different segments of ECG and MCG are found. Due to the platform agnosticity afforded by CNN’s automated feature extraction, the methods described here can be easily adapted to a wide range of vital signs, such as magnetoencephalography, electroencephalography, and ballistocardiography. The 3D CNN developed is highly portable; two MCG devices with different sensor arrangements can interpolate to an identically shaped output. Therefore, the CNN can be trained on one MCG device, and deployed on another dissimilar one.

Please see the full report for a more in depth explanation of the algorithm.