Method of atrial fibrillation and apnea-hypopnea prediction model establishment and system of atrial fibrillation and apnea-hypopnea prediction
Patent Information
- Authority / Receiving Office
- TW · TW
- Patent Type
- Patents
- Current Assignee / Owner
- NAT CHENG KUNG UNIV
- Filing Date
- 2025-09-15
- Publication Date
- 2026-08-01
AI Technical Summary
Current methods fail to simultaneously detect obstructive sleep apnea (OSA) and atrial fibrillation (AF), despite their shared risk factors and differing treatment approaches.
A predictive model is established using a computer system that processes multiple training electrocardiogram (ECG) signals, dividing them into segments, extracting features, and applying a machine learning algorithm to identify atrial fibrillation and respiratory arrest, utilizing techniques like zero-mean normalization, regularization, and reconstructive independent component analysis.
Enables simultaneous detection of atrial fibrillation and respiratory arrest using a single-lead ECG, facilitating appropriate treatment strategies and improving patient prognosis by reducing the need for expensive equipment and lowering operator skill barriers.
Smart Images
Figure TWG2TB001904124_001 
Figure TWG2TB001904124_002 
Figure TWG2TB001904124_003
Abstract
Description
[Technical Field]
[0001] This invention relates to a prediction method, and more particularly to a method for establishing a prediction model for atrial fibrillation and respiratory arrest, and a method for predicting atrial fibrillation and respiratory arrest. [Previous Technology]
[0002] Obstructive sleep apnea (OSA) is the most common sleep-related breathing disorder. It is primarily caused by excessive relaxation, abnormal proliferation, or hypertrophy of the upper airway muscles and soft tissues of the mouth and throat, leading to repeated complete or partial obstruction of the upper airway during sleep. The symptom is apnea-hypopnea (AH). Besides its association with breathing disorders, OSA is also a significant risk factor for atrial fibrillation (AF). AF is a cardiac arrhythmia caused by irregular atrial excitation, which can be due to various pathological reasons. Symptoms include irregular rhythm and loss of effective atrial contraction. OSA patients are four times more likely to develop AF than the general population, and the severity of OSA significantly affects the risk of AF. With the increasing aging population and the rising incidence of cardiovascular complications such as obesity, hypertension, metabolic disorders, and heart failure, OSA and AF are becoming increasingly common. OSA and AF share common risk factors and related complications, but their preferred treatment approaches differ. Common treatments for OSA include weight control, mechanical ventilation, or ENT surgery, while AF can be treated with medication, cardiac defibrillation, or radiofrequency ablation. However, there is currently no effective method to simultaneously detect both OSA and AF.
[0003] Therefore, how to simultaneously detect OSA and AF to improve patient prognosis has become one of the issues that related technical fields want to address. [Summary of the Invention]
[0004] Therefore, one of the objectives of the present invention is to provide a method for establishing a predictive model for atrial fibrillation and respiratory arrest, which can overcome at least one disadvantage of the prior art.
[0005] Therefore, the present invention provides a method for establishing a predictive model for atrial fibrillation and respiratory arrest, which is executed by a computer system storing multiple training electrocardiogram (ECG) signals and includes the following steps: (A) For each training ECG signal, the training ECG signal is divided into multiple training signal segments with the same duration; (B) For each training ECG signal, multiple symptom category labels are generated by a user input operation, each corresponding to a training signal segment of the training ECG signal, and each symptom category label indicates whether the corresponding training signal segment corresponds to atrial fibrillation or respiratory arrest; (C) For each training ECG signal, feature extraction is performed on the corresponding training signal segments to obtain multiple feature information corresponding to the training signal segments; and (D) Based on all feature information and all symptom category labels of the training ECG signals, a machine learning algorithm is used to establish a predictive model for atrial fibrillation and respiratory arrest.
[0006] In some embodiments, before step (A), the following step is further included: (E) zero-mean normalization and regularization of the training electrocardiogram signals.
[0007] In some embodiments, in step (B), each symptom category label indicates that the corresponding training signal segment belongs to one of the following categories: atrial fibrillation and respiratory arrest, atrial fibrillation only without respiratory arrest, respiratory arrest only without atrial fibrillation, and respiratory arrest without atrial fibrillation.
[0008] In some embodiments, in step (C), reconstructive independent component analysis is used to extract features from the corresponding training signal segments.
[0009] In some embodiments, in step (D), the atrial fibrillation and respiratory arrest prediction model includes an input layer, sixteen convolutional layers, fifteen batch normalization layers, fifteen leaky linear rectified excitation layers, five max pooling layers, a global average pooling layer, a normalized exponential layer, and an output layer.
[0010] Another object of the present invention is to provide a method for predicting atrial fibrillation and respiratory arrest, which can overcome at least one disadvantage of the prior art.
[0011] Therefore, the present invention provides a method for predicting atrial fibrillation and respiratory arrest, which is executed by a computer system storing an atrial fibrillation and respiratory arrest prediction model, and includes the following steps: (I) obtaining a test electrocardiogram signal corresponding to a test patient; (II) dividing the test electrocardiogram signal into multiple test signal segments with the same time length; (III) extracting features from the test signal segments to obtain multiple feature information corresponding to the test signal segments respectively; and (IV) for each test signal segment feature information, using the atrial fibrillation and respiratory arrest prediction model based on the feature information, obtaining a symptom category prediction result corresponding to the test signal segment, wherein the symptom category prediction result indicates whether the test signal segment corresponds to atrial fibrillation or respiratory arrest.
[0012] In some embodiments, in step (III), reconstructive independent component analysis is used to extract features from the corresponding training signal segments.
[0013] In some embodiments, in step (IV), the symptom category prediction result indicates that the signal segment to be tested belongs to one of the following categories: atrial fibrillation and respiratory arrest, atrial fibrillation only without respiratory arrest, respiratory arrest only without atrial fibrillation, and respiratory arrest without atrial fibrillation.
[0014] In some embodiments, after step (IV), the following step is further included: (V) obtaining a disease risk assessment result corresponding to the patient under test based on the prediction results of all symptom categories of the pending test signal segment, the disease risk assessment result indicating whether the patient under test may have atrial fibrillation or respiratory arrest.
[0015] In some embodiments, in step (V), the category indicated by the symptom category prediction results with the most corresponding numbers is taken as the disease risk determination result.
[0016] In some embodiments, the atrial fibrillation and respiratory arrest prediction model includes an input layer, sixteen convolutional layers, fifteen batch normalization layers, fifteen leaky linear rectified excitation layers, five max pooling layers, a global average pooling layer, a normalized exponential layer, and an output layer.
[0017] The advantages of this invention are as follows: By utilizing this atrial fibrillation and respiratory arrest prediction model, it is only necessary to collect the patient's single-lead electrocardiogram signal to simultaneously detect whether the patient experiences atrial fibrillation or respiratory arrest during sleep. This allows for the determination of whether the patient may be at risk of atrial fibrillation or respiratory arrest, helping medical staff to prescribe appropriate treatment strategies and improve the patient's prognosis. Furthermore, using machine learning models and single-lead electrocardiogram signals for discrimination avoids the inconvenience caused by using expensive or bulky equipment and lowers the barrier to entry for operators, making it suitable for widespread application in clinics or homes.
Implementation Method
[0019] Before the present invention is described in detail, it should be noted that similar elements are represented by the same numbers in the following description.
[0020] Referring to Figure 1, an embodiment of the atrial fibrillation and respiratory arrest prediction method of the present invention is executed by a computer system 1, which includes a storage module 11 and a processing module 12 connected to the storage module 11.
[0021] The storage module 11 can be implemented as, for example, but not limited to, any type of fixed or removable random access memory (RAM), hard disk drive (HDD), solid state drive (SSD) or similar element or combination of the above elements, and stores multiple training electrocardiogram signals and an atrial fibrillation and respiratory arrest prediction model established based on the training electrocardiogram signals.
[0022] In this embodiment, the training ECG signals were taken from the overnight single-lead ECGs of 101 patients in the SHHS1 (Sleep Heart Health Study Visit 1) database. These patients included 36 patients with prevalent atrial fibrillation (PAF), 25 patients with incident atrial fibrillation (IAF), and 40 patients with obstructive sleep apnea (OSA). The atrial fibrillation and sleep apnea prediction model comprises an input layer, sixteen convolutional layers, fifteen batch normalization layers, fifteen leaky ReLU activation layers, five max pooling layers, a global average pooling layer, a softmax layer, and an output layer. The process of establishing this atrial fibrillation and sleep apnea prediction model will be described in detail below.
[0023] The processing module 12 is electrically connected to the storage module 11 and can be implemented as, for example, but not limited to, a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor, or other similar elements or combinations thereof.
[0024] The operation of each component in the computer system 1 will be described below by way of an embodiment of the atrial fibrillation and respiratory arrest prediction method of the present invention. The embodiment includes a model building program and a prediction program.
[0025] Referring to Figures 1 and 2, the model building procedure illustrates how to build the atrial fibrillation and respiratory arrest prediction model, and includes the following steps S21 to S25.
[0026] In step S21, the processing module 12 performs zero-mean subtraction and normalization on the training electrocardiogram signals.
[0027] In step S22, for each training ECG signal, the processing module 12 divides the training ECG signal into multiple training signal segments with the same duration. For example, the training ECG signal may be divided into a training signal segment every 60 seconds.
[0028] In step S23, for each training ECG signal, the processing module 12 generates multiple symptom category labels corresponding to the training signal segments of the training ECG signal via a user input operation. Each symptom category label indicates whether the corresponding training signal segment corresponds to atrial fibrillation or respiratory arrest.
[0029] In this embodiment, each symptom category label indicates that the corresponding training signal segment belongs to one of the following categories: atrial fibrillation and respiratory arrest (labeled "AF AH"), atrial fibrillation only without respiratory arrest (labeled "AF no-AH"), respiratory arrest only without atrial fibrillation (labeled "no-AF AH"), and no atrial fibrillation without respiratory arrest (labeled "no-AF no-AH").
[0030] It is worth mentioning that the symptom category labels of these training signal segments for PAF patients mainly include "AF AH" and "AF no-AH"; for IAF patients, they include "AF AH", "AF no-AH", "no-AF AH" and "no-AF no-AH"; while for OSA patients, they include "AF AH" and "no-AF AH".
[0031] In step S24, for each training ECG signal, the processing module 12 performs feature extraction on the corresponding training signal segment to obtain multiple feature information corresponding to the training signal segment. In this embodiment, the processing module 12 uses Reconstruction Independent Component Analysis (RICA) technology to extract features from the corresponding training signal segment, and sets the number of independent components (ICs) to 500.
[0032] In step S25, the processing module 12 establishes the atrial fibrillation and respiratory arrest prediction model using a machine learning algorithm based on all feature information and all symptom category labels of the training electrocardiogram signals. In this embodiment, the processing module 12 establishes the atrial fibrillation and respiratory arrest prediction model based on the DarkNet machine learning algorithm. The performance of the atrial fibrillation and respiratory arrest prediction model is shown in Table 1 below. Table 1 category accuracy Sensitivity Specificity F1 Score AUC AF AH 99.7 99.3 99.8 99.3 0.9999 AF no-AH 97 96.6 97.2 96.2 0.9957 no-AF AH 99.7 98.7 99.9 98.9 0.9999 no-AF no-AH 97 93.9 98.2 94.4 0.9948 average 98.4 97.1 98.8 97.2 0.9973
[0033] Referring to Figures 1 and 3, the prediction procedure illustrates how to predict whether a patient to be tested may be at risk of having atrial fibrillation or respiratory arrest, and includes the following steps S31 to S35.
[0034] In step S31, the processing module 12 obtains a test electrocardiogram signal corresponding to the patient to be tested. The test electrocardiogram signal is, for example, from the patient's nocturnal single-lead electrocardiogram.
[0035] In step S32, the processing module 12 divides the ECG signal to be tested into multiple signal segments with the same duration. For example, the ECG signal to be tested may be divided into a signal segment every 60 seconds.
[0036] In step S33, the processing module 12 performs feature extraction on the waiting signal segment to obtain multiple feature information corresponding to the waiting signal segment. In this embodiment, the processing module 12 uses RICA technology to extract features from the corresponding training signal segments.
[0037] In step S34, for each signal segment to be tested, the processing module 12 uses the atrial fibrillation and respiratory arrest prediction model based on the feature information to obtain a symptom category prediction result corresponding to the signal segment to be tested. The symptom category prediction result indicates whether the signal segment to be tested corresponds to atrial fibrillation or respiratory arrest.
[0038] In this embodiment, the symptom category prediction result indicates that the signal segment to be tested belongs to one of the following categories: atrial fibrillation and respiratory arrest, atrial fibrillation only without respiratory arrest, respiratory arrest only without atrial fibrillation, and respiratory arrest without atrial fibrillation.
[0039] In step S35, the processing module 12 obtains a disease risk assessment result corresponding to the patient under test based on the prediction results of all symptom categories of the pending test signal segment. This disease risk assessment result indicates whether the patient under test may have atrial fibrillation or respiratory arrest. For example, the processing module 12 may use the category indicated by the symptom category prediction results with the highest number of results as the disease risk assessment result.
[0040] It is worth mentioning that after obtaining the risk assessment result of the disease, corresponding treatment recommendations can be generated based on the risk assessment result. For example, for the patient to be tested who is classified as having only atrial fibrillation without respiratory arrest, the treatment recommendation is to perform cardiac shock or radiofrequency ablation. For the patient to be tested who is classified as having only respiratory arrest without atrial fibrillation, the recommendation is to wear a ventilator or perform ENT surgery. For the patient to be tested who is classified as having both atrial fibrillation and respiratory arrest, the recommendation is to prioritize the treatment of respiratory arrest (i.e., wear a ventilator or perform ENT surgery) to assist medical staff in making the most suitable treatment for the patient to be tested.
[0041] It should be added that, since atrial fibrillation is not often included in the routine treatment of OSA, and existing literature has indicated that if atrial fibrillation can be screened from OSA patients and their respiratory arrest is treated, the chance of curing both diseases at the same time will be higher. Therefore, it is recommended to prioritize the treatment of respiratory arrest for this category of atrial fibrillation and respiratory arrest.
[0042] In summary, by utilizing this atrial fibrillation and respiratory arrest prediction model, only a single-lead electrocardiogram (ECG) signal from the patient is needed to simultaneously detect whether the patient experiences atrial fibrillation or respiratory arrest during sleep. This allows for the determination of whether the patient is at risk of atrial fibrillation or respiratory arrest, helping medical staff to prescribe appropriate treatment strategies and improve patient prognosis. Furthermore, using machine learning models and single-lead ECG signals avoids the inconvenience of using expensive or bulky equipment and lowers the barrier to entry for operators, facilitating widespread application in clinics or homes. Therefore, the objectives of this invention are indeed achieved.
[0043] However, the above description is only an embodiment of the present invention and should not be construed as limiting the scope of the present invention. Any simple equivalent changes and modifications made in accordance with the scope of the patent application and the contents of the patent specification shall still fall within the scope of the patent of the present invention. [Simplified Explanation of the Diagram]
[0018] Other features and effects of the present invention will be clearly presented in the embodiments with reference to the drawings, wherein: Figure 1 is a block diagram illustrating, by way of example, a computer system used in an embodiment of the atrial fibrillation and respiratory arrest prediction method of the present invention; Figure 2 is a flowchart illustrating, by way of example, a model building procedure of the embodiment; and Figure 3 is a flowchart illustrating, by way of example, a prediction procedure of the embodiment.
Claims
1. A method for establishing a predictive model for atrial fibrillation and respiratory arrest, executed by a computer system storing multiple training electrocardiogram (ECG) signals, comprising the following steps: (A) for each training ECG signal, dividing the training ECG signal into multiple training signal segments with the same duration; (B) for each training ECG signal, generating multiple symptom category labels corresponding to the training signal segments respectively via a user input operation, each symptom category label indicating whether the corresponding training signal segment corresponds to atrial fibrillation or respiratory arrest; (C) for each training ECG signal, performing feature extraction on the corresponding training signal segments to obtain multiple feature information corresponding to the training signal segments respectively; and (D) based on all feature information and all symptom category labels of the training ECG signals, using a machine learning algorithm to establish a predictive model for atrial fibrillation and respiratory arrest.
2. The method for establishing atrial fibrillation and respiratory arrest prediction models as described in claim 1, before step (A), further includes the following step: (E) zero-mean normalization and regularization of the training electrocardiogram signals.
3. The method for establishing atrial fibrillation and respiratory arrest prediction models as described in claim 1, wherein, In step (B), each symptom category label indicates that the corresponding training signal segment belongs to one of the following categories: atrial fibrillation and respiratory arrest, atrial fibrillation only without respiratory arrest, respiratory arrest only without atrial fibrillation, and respiratory arrest without atrial fibrillation.
4. The method for establishing atrial fibrillation and respiratory arrest prediction models as described in claim 1, wherein, In step (C), reconstructive independent component analysis is used to extract features from the corresponding training signal segments.
5. The method for establishing atrial fibrillation and respiratory arrest prediction models as described in claim 1, wherein, In step (D), the atrial fibrillation and respiratory arrest prediction model includes an input layer, sixteen convolutional layers, fifteen batch normalization layers, fifteen leaky linear rectified excitation layers, five max pooling layers, a global average pooling layer, a normalized exponential layer, and an output layer.
6. An atrial fibrillation and respiratory arrest prediction system, comprising: a storage module; and a processing module connected to the storage module, configured to obtain a test electrocardiogram (ECG) signal corresponding to a patient to be tested, segment the test ECG signal into multiple test signal segments with the same duration, extract features from the test signal segments to obtain multiple feature information corresponding to the test signal segments respectively, and for each test signal segment feature information, using the atrial fibrillation and respiratory arrest prediction model based on the feature information to obtain a symptom category prediction result corresponding to the test signal segment, the symptom category prediction result indicating whether the test signal segment corresponds to atrial fibrillation or respiratory arrest.
7. The atrial fibrillation and respiratory arrest prediction system as described in claim 6, wherein, The processing module uses reconstructive independent component analysis to extract features from the corresponding waiting signal segment.
8. The atrial fibrillation and respiratory arrest prediction system as described in claim 6, wherein, The symptom category prediction results indicate that the signal segment to be tested belongs to one of the following categories: atrial fibrillation and respiratory arrest, atrial fibrillation only without respiratory arrest, respiratory arrest only without atrial fibrillation, and atrial fibrillation without respiratory arrest.
9. The atrial fibrillation and respiratory arrest prediction system as described in claim 8, wherein, Based on the prediction results of all symptom categories of the waiting signal segment, the processing module obtains a disease risk assessment result corresponding to the patient to be tested, which indicates whether the patient to be tested may have atrial fibrillation or respiratory arrest.
10. The atrial fibrillation and respiratory arrest prediction system as described in claim 9, wherein, The processing module uses the category indicated by the prediction results of the symptom category with the most corresponding results as the risk assessment result for the disease.
11. The atrial fibrillation and respiratory arrest prediction system as described in any one of claims 6 to 10, wherein the atrial fibrillation and respiratory arrest prediction model comprises an input layer, sixteen convolutional layers, fifteen batch normalization layers, fifteen leaky linear rectified excitation layers, five max pooling layers, a global average pooling layer, a normalized exponential layer, and an output layer.