# US 11,298,067 B2 · Method for classifying type of heartbeat and apparatus using the same

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

- Number: US 11,298,067 B2
- Document ID: US11298067B2
- Status: Granted
- Kind: B2
- Application number: 16/439,949
- Filed: 2019-06-13
- Published: 2020-09-17
- Granted: 2022-04-12
- Earliest claimed: 2019-03-14
- Assignees: University of Seoul Industry Cooperation Foundation
- Applicants: University of Seoul Industry Cooperation Foundation
- Inventors: Heasoo Hwang; Sellami Ali
- Web page: https://patentgrep.com/patent/US11298067B2

## CPC

- A61B 5/352: Detecting R peaks, e.g. for synchronising diagnostic apparatus; Estimating R-R interval
- A61B 5/346: Analysis of electrocardiograms
- A61B 5/35: by template matching
- A61B 5/364: Detecting abnormal ECG interval, e.g. extrasystoles, ectopic heartbeats
- A61B 5/7267: {involving training the classification device}
- G06N 3/045: Combinations of networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/08: Learning methods
- G06N 3/082: modifying the architecture, e.g. adding, deleting or silencing nodes or connections
- G06N 3/09: Supervised learning
- A61B 5/02438: {with portable devices, e.g. worn by the patient}
- A61B 5/361: Detecting fibrillation
- A61B 5/363: Detecting tachycardia or bradycardia
- A61B 5/366: Detecting abnormal QRS complex, e.g. widening
- A61B 5/681: {Wristwatch-type devices}
- A61B 5/7203: {for noise prevention, reduction or removal}
- A61B: DIAGNOSIS; SURGERY; IDENTIFICATION (usefulness limited to only animals A61D)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A method for classifying types of heartbeats includes obtaining a dataset including multiple heartbeat waveform data, generating, from the dataset, input data regarding a heartbeat waveform for training and generating a learning model to which the generated input data is input and from which a heartbeat type of the heartbeat waveform for training is output, training the learning model by determining a loss weight of each batch sampled from the dataset and determining a loss function based on the loss weight of each batch, and inputting a heartbeat waveform for test to the learning model and classifying a heartbeat type of the heartbeat waveform for test.

## Claims

### Claim 1

A method for detecting abnormal heart rhythm from classified types of heartbeats, comprising:
obtaining a dataset including multiple heartbeat waveform data;
generating, from the dataset, input data including characteristics of a heartbeat waveform for training and generating a learning model to which the generated input data is input and from which a heartbeat type of the heartbeat waveform for training is output;
minimizing an imbalance in the dataset between the classified types of heartbeats by training the learning model by determining a loss weight of each batch sampled from the dataset and determining a loss function based on the loss weight of each batch;
inputting a heartbeat waveform for test to the learning model and classifying a heartbeat type of the heartbeat waveform for test,
wherein the loss weight of each batch is calculated based on the following equation:
c
⁢
w
i
,
class
k
=
1
-
Σ
j
=
1
M
⁢
1
y
i
,
j
=
class
k
M
+
ɛ
wherein M is a batch size, ε is a predetermined constant, cwi,class k is a loss weight of a kth class in an ith batch; and
using at least a single lead that produces an electrocardiogram waveform as an input to the trained learning model which outputs the heartbeat type to detect an abnormal heart rhythm, wherein the electrocardiogram waveform is used without any data preprocessing.

### Claim 2 (depends on 1)

The method for classifying types of heartbeats of claim 1, wherein the input data includes a target heartbeat waveform, a heartbeat type of the target heartbeat waveform, and a neighboring heartbeat waveform adjacent to the target heartbeat waveform.

### Claim 3 (depends on 1)

The method for classifying types of heartbeats of claim 1, wherein the heartbeat waveform for training includes a heartbeat waveform of a predetermined number of samples around an R-peak which is extracted from electrocardiography (ECG) data.

### Claim 4 (depends on 1)

The method for classifying types of heartbeats of claim 1,
wherein the loss function is calculated based on the following equation:
L
i
=
⁢
-
∑
j
=
1
M
⁢
c
⁢
w
i
,
y
i
,
j
⁢
y
i
,
j
⁢
log
⁢
⁢
y
^
i
,
j
+
λ
⁢

W

2
2
wherein ŷi,j is a predicted probability of a jth training instance in the ith batch, W is weight matrices of all the layers, λ is L2 regularization parameter and Li is the loss function.

### Claim 5 (depends on 1)

The method for classifying types of heartbeats of claim 1, wherein the learning model includes multiple convolutional layers, at least one fully connected layer, and a softmax function layer.

### Claim 6 (depends on 5)

The method for classifying types of heartbeats of claim 5, wherein the learning model further includes batch normalization layers, each normalization layer connected to a respective one of the multiple convolutional layers.

### Claim 7 (depends on 6)

The method for classifying types of heartbeats of claim 6, wherein the batch normalization layers perform a batch normalization based on the following equation:
B
⁢
N
C
i
=
x
C
i
-
μ
β
⁡
(
x
C
i
)
σ
β
2
⁡
(
x
C
i
)
wherein XCi is an output of an ith convolutional layer, and μβ(XCi) is the mean and σβ(XCi) is the variance of output values after values in a batch pass through the ith convolutional layer, and BNCi is an output after applying the batch normalization.

### Claim 8 (depends on 5)

The method for classifying types of heartbeats of claim 5, wherein the learning model further includes at least one skip connection.

### Claim 9

An apparatus for detecting abnormal heart rhythm from classified types of heartbeats, comprising:
a single lead configured to produce an electrocardiogram waveform, wherein the electrocardiogram waveform is used without any data preprocessing;
a storage unit configured to store a dataset including multiple heartbeat waveform data;
a learning model generation unit configured to generate, from the dataset, input data regarding a heartbeat waveform for training and generate a learning model to which the generated input data is input and from which a heartbeat type of the heartbeat waveform for training is output;
a training unit configured to train the learning model by determining a loss weight of each batch sampled from the dataset and determining a loss function based on the loss weight of each batch; and
a classification unit configured to input a heartbeat waveform for test to the learning model and classify a heartbeat type of the heartbeat waveform for test,
wherein the loss weight of each batch is calculated based on the following equation:
c
⁢
w
i
,
class
k
=
1
-
Σ
j
=
1
M
⁢
1
y
i
,
j
=
class
k
M
+
ɛ
wherein M is a batch size, E is a predetermined constant, cwi,class k is a loss weight of a kth class in an ith batch,
wherein the classification unit is also configured to use the electrocardiogram waveform produced by the single lead to detect an abnormal heart rhythm based on the heartbeat type output from the trained learning model.

### Claim 10 (depends on 9)

The apparatus for classifying types of heartbeats of claim 9, wherein the input data includes a target heartbeat waveform, a heartbeat type of the target heartbeat waveform, and a neighboring heartbeat waveform adjacent to the target heartbeat waveform.

### Claim 11 (depends on 9)

The apparatus for classifying types of heartbeats of claim 9, wherein the loss function is calculated based on the following equation:
L
i
=
⁢
-
∑
j
=
1
M
⁢
c
⁢
w
i
,
y
i
,
j
⁢
y
i
,
j
⁢
log
⁢
⁢
y
^
i
,
j
+
λ
⁢

W

2
2
where ŷi,j is a predicted probability of a jth training instance in the ith batch, W is weight matrices of all the layers, λ is L2 regularization parameter and Li is the loss function.