# US 10,869,610 B2 · System and method for identifying cardiac arrhythmias with deep neural networks

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 10,869,610 B2
- Document ID: US10869610B2
- Status: Granted
- Kind: B2
- Application number: 16/210,688
- Filed: 2018-12-05
- Published: 2020-06-11
- Granted: 2020-12-22
- Earliest claimed: 2018-12-05
- Assignees: General Electric Company
- Applicants: General Electric Company
- Inventors: Weijia Lu; Joel Xue; Shuyan Gu; Jie Shuai; Lifei Hu
- Web page: https://patentgrep.com/patent/US10869610B2

## CPC

- A61B 5/361: Detecting fibrillation
- G06N 3/088: Non-supervised learning, e.g. competitive learning
- A61B 5/02405: {Determining heart rate variability}
- A61B 5/0245: by using sensing means generating electric signals, {i.e. ECG signals}
- A61B 5/346: Analysis of electrocardiograms
- A61B 5/363: Detecting tachycardia or bradycardia
- A61B 5/7264: {Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems}
- A61B 5/7275: {Determining trends in physiological measurement data; Predicting development of a medical condition based on physiological measurements, e.g. determining a risk factor}
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/045: Combinations of networks
- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/047: Probabilistic or stochastic networks
- G06N 3/08: Learning methods
- G06N 3/09: Supervised learning
- A61B 5/0006: {ECG or EEG signals}
- A61B 5/364: Detecting abnormal ECG interval, e.g. extrasystoles, ectopic heartbeats
- A61B 5/7257: {using Fourier transforms}
- G06N 7/01: Probabilistic graphical models, e.g. probabilistic networks
- A61B: DIAGNOSIS; SURGERY; IDENTIFICATION (usefulness limited to only animals A61D)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A system for identifying arrhythmias based on cardiac waveforms includes a storage system storing a trained deep neural network system, wherein the trained deep neural system includes a trained representation neural network and a trained classifier neural network. A processing system is communicatively connected to the storage system and configured to receive cardiac waveform data for a patient, identify a time segment in the cardiac waveform data, and transform the time segment into a spectrum image. The processing system is further configured to generate, with the representation neural network, a latent representation from the spectrum image, and then to generate, with the classifier neural network, an arrhythmia classifier from the latent representation.

## Claims

### Claim 1

A system for identifying arrhythmias based on cardiac waveforms, the system comprising:
a storage system storing a trained deep neural network system, wherein the trained deep neural network system includes a trained representation neural network and a trained classifier neural network;
a processing system communicatively connected to the storage system and configured to:
receive cardiac waveform data for a patient;
identify a time segment of the cardiac waveform data;
transform the time segment of the cardiac waveform data into a spectrum image;
generate, with the representation neural network, a latent representation from the spectrum image; and
generate, with the classifier neural network, an arrhythmia classifier from the latent representation.

### Claim 2 (depends on 1)

The system of claim 1, wherein the arrhythmia classifier includes a classifier value for each of a predetermined list of rhythm types.

### Claim 3 (depends on 2)

The system of claim 2, wherein the predetermined list of rhythm types includes at least two of a normal sinus rhythm, asystole, supraventricular tachycardia, ventricular fibrillation, ventricular tachycardia, and atrial fibrillation.

### Claim 4 (depends on 3)

The system of claim 3, wherein the representation neural network includes a variational autoencoder (VAE) encoder neural network, wherein the VAE encoder neural network outputs the latent representation.

### Claim 5 (depends on 4)

The system of claim 4, wherein the neural network system includes the trained representation neural network and the trained classifier neural network organized sequentially such that the latent representation output of the VAE encoder is fed directly to the classifier neural network to generate the arrhythmia classifier.

### Claim 6 (depends on 4)

The system of claim 4, wherein the representation neural network further includes a convolutional neural network.

### Claim 7 (depends on 1)

The system of claim 1, wherein the classifier neural network comprises a bidirectional recurrent neural network (RNN).

### Claim 8 (depends on 1)

The system of claim 1, wherein the representation neural network and the classifier neural network are sequentially arranged and sequentially trained neural networks with no overlap.

### Claim 9 (depends on 1)

The system of claim 1, wherein the spectrum image is a two-dimensional image representing transient frequency distribution on one axis and time on a second axis.

### Claim 10 (depends on 1)

The system of claim 1, wherein the time segment is identified in a limb lead of a multi-lead ECG.

### Claim 11

A method for identifying arrhythmias based on cardiac waveforms, the method comprising:
identifying a time segment of a cardiac waveform data for a patient;
transforming the time segment of cardiac waveform data into a spectrum image;
generating, by a trained representation neural network, a latent representation from the spectrum image; and
generating, by a trained classifier neural network, an arrhythmia classifier from the latent representation.

### Claim 12 (depends on 11)

The method of claim 11, wherein the arrhythmia classifier includes a classifier value for each of a predetermined list of rhythm types, including two or more of a normal sinus rhythm, asystole, supraventricular tachycardia, ventricular fibrillation, ventricular tachycardia, and atrial fibrillation.

### Claim 13 (depends on 11)

The method of claim 11, wherein the representation neural network further comprises a variational autoencoder (VAE) encoder neural network, wherein the VAE encoder neural network outputs the latent representation used by the classifier neural network to generate the arrhythmia classifier.

### Claim 14 (depends on 13)

The method of claim 13, further comprising training the representation neural network to minimize mean error and to minimize standard deviation error.

### Claim 15 (depends on 11)

The method of claim 11, wherein the classifier neural network comprises a bidirectional recurrent neural network (RNN), and further comprising training the classifier neural network using a cross entropy loss function.

### Claim 16 (depends on 11)

The method of claim 11, wherein the representation neural network and the classifier neural network are sequentially arranged and trained neural networks with no overlap.

### Claim 17 (depends on 11)

The method of claim 11, wherein the spectrum image is a two-dimensional image representing transient frequency distribution on one axis and time on a second axis.

### Claim 18 (depends on 11)

The method of claim 11, wherein the time segment is identified in a limb lead of a multi-lead ECG.

### Claim 19 (depends on 18)

The method of claim 18, wherein a length of the time segment is defined based on a predetermined list of rhythm types that the classifier neural network is trained to identify.

### Claim 20 (depends on 11)

The method of claim 11, further comprising pre-processing the time segment of cardiac waveform data to remove baseline wander normalize sample rate.