US 12,268,527 B2
System and method for automated analysis and detection of cardiac arrhythmias from electrocardiograms
Vignesh Kalidas, Richardson, TX (US); and Lakshman S. Tamil, Plano, TX (US)
Assigned to Board of Regents, The University of Texas System, Austin, TX (US)
Filed by Board of Regents, The University of Texas System, Austin, TX (US)
Filed on Jul. 20, 2021, as Appl. No. 17/380,765.
Claims priority of provisional application 63/054,166, filed on Jul. 20, 2020.
Prior Publication US 2022/0015711 A1, Jan. 20, 2022
Int. Cl. A61B 5/00 (2006.01); A61B 5/352 (2021.01); A61B 5/361 (2021.01); A61B 5/364 (2021.01); G06N 3/045 (2023.01); G06N 3/088 (2023.01); G06N 20/20 (2019.01); G16H 15/00 (2018.01); G16H 50/30 (2018.01)
CPC A61B 5/7203 (2013.01) [A61B 5/0006 (2013.01); A61B 5/0022 (2013.01); A61B 5/352 (2021.01); A61B 5/361 (2021.01); A61B 5/364 (2021.01); G06N 3/045 (2023.01); G06N 3/088 (2013.01); G06N 20/20 (2019.01); G16H 15/00 (2018.01); G16H 50/30 (2018.01)] 21 Claims
OG exemplary drawing
 
1. An arrhythmia analysis method comprising:
training, by an at least one computing device, a noise suppression model with training data comprising electrocardiogram (ECG) waveform signals;
generating, by the at least one computing device, model weights for the noise suppression model during the training of the noise suppression model;
initializing, by the at least one computing device, the noise suppression model with the model weights for the noise suppression model generated during the training of the noise suppression model;
training, by the at least one computing device, a noise detection model with training data comprising noisy ECG waveform signals;
generating, by the at least once computing device, model weights for the noise detection model during the training of the noise detection model;
initializing, by the at least one computing device, the noise detection model with the model weights for the noise detection model generated during the training of the noise detection model;
training, by the at least one computing device, a beat extraction model with training data comprising denoised ECG waveform signals;
generating, by the at least once computing device, model parameters for the beat extraction model during the training of the beat extraction model;
initializing, by the at least one computing device, the beat extraction model with the model parameters generated during the training of the beat extraction model;
acquiring, by the at least one computing device, an electrocardiogram (ECG) waveform signal of a subject at a set sampling frequency rate;
applying, by the at least one computing device, the acquired ECG waveform signal as input to the noise suppression model and processing, by the at least one computing device using the noise suppression model, the acquired ECG waveform signal to remove low-frequency noise and high-frequency noise artifacts and form a denoised ECG waveform signal;
applying, by the at least one computing device, the denoised ECG waveform signal as input to the noise detection model and processing, by the at least one computing device using the noise detection model, the denoised ECG waveform signal to remove low-quality segments and form a high-quality ECG waveform signal;
analyzing, by the at least one computing device, the high-quality ECG waveform signal to detect a presence of a beat-independent ventricular arrhythmia within the high-quality ECG waveform signal;
processing, by the at least one computing device using the beat extraction model, the denoised ECG waveform signal to extract beat (R-peak) locations corresponding to QRS complexes from the denoised ECG waveform signal;
analyzing, by the at least one computing device using the beat extraction model, the denoised ECG waveform signal to detect a presence of a beat-dependent ventricular arrhythmia within the denoised ECG waveform signal based on the extracted beat (R-peak) locations of the denoised ECG waveform signal;
analyzing, by the at least one computing device using the beat extraction model, the denoised ECG waveform signal to detect a presence of one or more supraventricular arrhythmias within the denoised ECG waveform signal based on the extracted beat (R-peak) locations of the denoised ECG waveform signal, wherein the analysis of the denoised ECG waveform signal to detect the presence of one or more supraventricular arrhythmias comprises performing supraventricular ectopic beats (SVEBs) classifications of the denoised ECG waveform signal using semi-supervised autoencoder and random forest machine learning techniques; and
outputting, by the at least one computing device, a report containing one or more arrhythmias detected by the analyzing steps.