2.1 Study design
This study used a traditional three-channel Holter device (GE) and a ConvNet algorithm-based single-lead AI Holter device (SWK801) to record inpatient or outpatient ECGs. All patients wore three-channel Holter and AI Holter devices simultaneously, and the Shuweikang AI center analyzed the data using independent software. Two professional cardiologists analyzed all the three-channel records. The First Affiliated Hospital of the Nanjing Medical University ethics committee approved this study (2020-SRFA-301).
Enrollment took place from the First Affiliated Hospital of Nanjing Medical University and the Maternal and Child Health Hospital of Jiangsu Province, and the data were collected from March 1, 2020, to December 1, 2020. All methods were performed in accordance with the relevant guidelines and regulations, as well as informed consent was obtained from all subjects and/or their legal guardian(s). The primary outcome included HRV comparison using traditional Holter and AI Holter recorders.
2.2 Participants and Descriptive Data
A total of 275 participants were included in this study. Due to a wear time of less than 22 h, exceeding rubbing disturbance, and a device patch falling out, 10 patients were excluded from the study, and 265 participants were finally analyzed. According to age and BMI, patients were divided into four groups: BMI < 24 kg/m2 and age < 65 years (n = 92), BMI < 24 kg/m2 and age ≥ 65 years (n = 51), BMI ≥ 24 kg/m2 and age < 65 years (n = 79), and BMI ≥ 24 kg/m2 and age ≥ 65 years (n = 43). The mean age of the groups was 55.08 ± 16.74 (the total), 44.54 ± 12.11, 74.10 ± 6.14, 46.27 ± 11.86, and 71.28 ± 6.09, and other baseline comparison analyses included diabetes, syncope, dyspnea, chest pain, HR, HTN, smoke, hyperlipemia and sex. All patient characteristics are presented in Table 1.
|
n = 265
|
n = 92
|
n = 51
|
n = 79
|
n = 43
|
|
Parameter
|
Total
|
BMI < 24&Age < 65
|
BMI < 24&Age ≥ 65
|
BMI ≥ 24&Age < 65
|
BMI ≥ 24&Age ≥ 65
|
P
|
Table 1
Clinical characteristics of the study subjects
Age, years
|
55.08 ± 16.74
|
44.54 ± 12.11
|
74.10 ± 6.14
|
46.27 ± 11.86
|
71.28 ± 6.09
|
<0.001
|
Gender(male, %)
|
30.94%
|
25.00%
|
35.29%
|
33.54%
|
30.23%
|
0.6399
|
HR, bpm
|
70.18 ± 11.49
|
72 ± 11.36
|
66 ± 9.59
|
73 ± 12.87
|
68 ± 8.70
|
0.0076
|
BMI, kg/m2
|
23.82 ± 3.19
|
21.41 ± 2.03
|
21.78 ± 1.46
|
26.65 ± 2.38
|
26.18 ± 1.84
|
<0.001
|
Diabetes
|
12
|
2
|
4
|
1
|
5
|
0.0240
|
HTN
|
45
|
4
|
12
|
12
|
17
|
<0.001
|
Syncope
|
0
|
2
|
1
|
1
|
2
|
0.6851
|
Dyspnea
|
4
|
0
|
1
|
1
|
2
|
0.2245
|
Chest pain
|
10
|
1
|
4
|
3
|
2
|
0.2361
|
Smoke(%)
|
106
|
42.39%
|
41.18%
|
34.18%
|
44.19%
|
0.7946
|
Hyperlipemia (%)
|
140
|
32.61%
|
50.98%
|
64.56%
|
76.74%
|
<0.001
|
HR: Heart Rates; BMI: Body Mass Index; HTN: Hypertension. BMI༚kg/m2;Age༚years. Hyperlipemia༚Low Density Lipoprotein Cholesterol (LDL-C) >2.5mmol/L or combined Total Cholesterol (TC), Triglycerides (TG) abnormal |
2.3 Setting and participants
Two general hospitals affiliated with Nanjing Medical University participated in this study. We recruited adults aged 19–90 years, exclude contraindications. Participants were excluded for fatal diseases, thoracic deformities, skin breaks, and acute diseases, such as chronic kidney disease, malignant arrhythmia, heart failure, acute coronary syndrome or high CVD risk, pregnancy, thoracocyllosis, cancer and so on. Metabolism-resourced obesity and other pathological overweightness were excluded. SWK801 real-time data transfer, which uses 250HZ sampling frequency, Bluetooth technology, and a smartphone were used.
2.4 ConvNet algorithm based single-lead AI Holter
Real-world ECG signals are often disturbed by noise (25). It is mandatory to perform denoising to remove or mitigate the baseline wander noise, frequency interference, and muscle movement. Following denoising, we identified ECG r-peaks and take one second of data centered at each r-peak, which generated 640 × 480 r-peak images as inputs to the AI model. The input is actually a 640 * 480 pixel ECG image. So the input tensor is 640 * 480 * 3 (because of 3 RGB channels). In this image, including 1 second of ECG data centered at its R-peak.
The AI model was built using convolutional neural networks (CNNs), a specialized neural network for processing data with a known grid-like topology. The network is composed of eight residual blocks. Each residual block includes the following five layers (Fig. 1):
- A convolution layer
- A batch normalization layer
- A ReLU (Rectified Linear Units) layer
- A convolution layer
- A batch normalization layer
-
Both convolution layers (1 and 4) use a 3 × 3 kernel and a stride of 1. After each convolution, batch normalization was performed to avoid the explosion of the parameters and the phenomenon of “vanishing gradients.” Following the batch layer, we added the ReLU activation function with output zero for negative inputs and identity output for positive inputs, the non-linearity of which allows the network to create a complex non-linear representation of the ECGs for automatic feature extraction. In addition to the above five layers, the residual blocks also include a residual connection to allow gradient propagation that is implemented using a 1 × 1 convolution layer between the input of the residual block and its output. Between any two neighboring residual blocks, the ReLU activation function is used to filter the output of the previous block to the next block. Following the last residual block, the data were fed to a global average pooling layer, followed by a fully connected layer to generate the probability of each type of arrhythmia (Fig. 2).
-
We split the dataset into training, validation, and test sets at a ratio of 8:1:1. There were no patient overlaps between the training, validation, and test sets.
2.5 Statistical Analysis
All analyses were performed using SPSS 22.0, and GraphPad Prism 5. Linear fitting and Bland–Altman analyses were used to illustrate the statistical results. Values are expressed as frequencies and percentages or as mean ± SD. Student’s t-test and Chi-squared test were used to analyze the population characteristics. ECG AI analysis used the ConvNet-related CNN algorithm (Shuweikang, Nanjing), a training database based on MIT-BIH, and the First Affiliated Hospital of Nanjing Medical University 1 + N ECG database.