Deep learning-based atrial fibrillation detection system using PPG signal-sensing rings

Abstract

A deep learning-based atrial fibrillation determination system using a photoplethysmogram (PPG) signal sensing ring is provided. The system includes a server, the server including a signal quality classification component configured to classify the quality of a PPG signal as good or poor using a first deep learning model, and an atrial fibrillation determination component configured to determine whether atrial fibrillation occurs from the PPG signal using a second deep learning model, the PPG signal sensing ring includes a plurality of sensors configured to simultaneously measure a plurality of PPG signals at different positions, each of the plurality of sensors including a light source and a photoelectric conversion device, the PPG signal being measured using the PPG signal sensing ring, the server receiving the PPG signal from the PPG signal sensing ring via a terminal, and the terminal including a sensor selection component configured to select a sensor that measures a test PPG signal having the highest signal quality among a plurality of test PPG signals among the plurality of sensors as a sensor for measuring the PPG signal.

Description

Technical Field

[0001] The present invention relates to an atrial fibrillation determination system, and more specifically, to a deep learning-based atrial fibrillation determination system using a photoplethysmography (PPG) signal sensing ring.

Background Art

[0002] Atrial fibrillation is an arrhythmia disease that causes irregular pulses in the atria. The atrial fibrillation occurs intermittently, and in order to diagnose atrial fibrillation by using an electrocardiogram (ECG), which has been widely used conventionally, it has been difficult to make an early diagnosis because the patient has to visit a hospital. However, if reliable level of continuous monitoring is accompanied, the patient can be detected and managed early, and the risk can be reduced. Therefore, there is a need for a system and method for monitoring atrial fibrillation in daily life in real time.

Summary of the Invention

Problems to be Solved by the Invention

[0003] The problem to be solved by the present disclosure is to provide a deep learning-based atrial fibrillation determination system using a photoplethysmography (PPG) signal sensing ring.

Means for Solving the Problems

[0004] To solve the aforementioned problems, a deep learning-based atrial fibrillation detection system using a PPG (photoplethysmography) signal sensing ring according to one embodiment of the present disclosure includes a server, the server includes a signal quality classification component configured to classify the quality of the PPG signal into good or bad, and an atrial fibrillation detection component configured to determine whether atrial fibrillation has occurred from the PPG signal using a deep learning model, the PPG signal is measured using the PPG signal sensing ring, the server receives the PPG signal from the PPG signal sensing ring via a terminal, the PPG signal sensing ring includes a plurality of sensors configured to simultaneously measure a plurality of PPG signals at different locations, each of the plurality of sensors includes a light source and a photoelectric converter, and the terminal also includes a sensor selection component configured to select the sensor that measured the test PPG signal with the highest signal quality among a plurality of test PPG signals from the plurality of sensors as the sensor for measuring the PPG signal.

[0005] In some embodiments, the signal quality of the plurality of test PPG signals is also evaluated by at least one of the following: the magnitude of the acceleration signal, the signal-to-noise ratio, and the AC (alternating current) component magnitude to the DC (direct current) component magnitude ratio.

[0006] In some embodiments, the server further includes an atrial fibrillation index calculation component configured to calculate an atrial fibrillation index, which may be defined as the ratio of the time during which the signal quality classification component classifies the PPG signal as good quality and the atrial fibrillation determination component determines that atrial fibrillation has occurred to the time during which the signal quality classification component classifies the PPG signal as good quality.

[0007] In some embodiments, the terminal further includes a light source control component configured to control the light source of each of the multiple sensors such that the DC component of each of the multiple test PPG signals measured using the multiple sensors falls within a predetermined range.

[0008] In some embodiments, the control of the light source by the light source control component and the selection of the sensor by the sensor selection component may be performed sequentially, or the control of the light source by the light source control component and the selection of the sensor by the sensor selection component may be performed periodically.

[0009] A deep learning-based atrial fibrillation detection method using a PPG signal sensing ring according to one embodiment of the present disclosure includes the steps of receiving a PPG signal received from the PPG signal sensing ring via a terminal, classifying the quality of the PPG signal into good or bad, and determining whether atrial fibrillation has occurred from the PPG signal using a deep learning model, wherein the PPG signal sensing ring includes a plurality of sensors configured to simultaneously measure a plurality of PPG signals at different locations, each of the plurality of sensors includes a light source and a photoelectric converter, the PPG signal is measured using the PPG signal sensing ring, the server receives the PPG signal from the PPG signal sensing ring via a terminal, and the terminal also includes a sensor selection component configured to select the sensor that measured the test PPG signal with the highest signal quality among the plurality of test PPG signals as the sensor for measuring the PPG signal.

[0010] In some embodiments, the signal quality of the plurality of test PPG signals is also evaluated by at least one of the following: the magnitude of the acceleration signal, the signal-to-noise ratio, and the AC (alternating current) component magnitude to the DC (direct current) component magnitude ratio.

[0011] In some embodiments, the method further includes the step of calculating an atrial fibrillation index, which may be defined as the ratio of the time during which the signal quality classification component classifies the PPG signal as good quality and the atrial fibrillation determination component determines that atrial fibrillation has occurred to the time during which the signal quality classification component classifies the PPG signal as good quality.

[0012] In some embodiments, the terminal further includes a light source control component configured to control the light source of each of the multiple sensors such that the DC component of each of the multiple test PPG signals measured using the multiple sensors falls within a predetermined range.

[0013] In some embodiments, the control of the light source by the light source control component and the selection of the sensor by the sensor selection component may be performed sequentially, or the control of the light source by the light source control component and the selection of the sensor by the sensor selection component may be performed periodically. [Effects of the Invention]

[0014] A deep learning-based atrial fibrillation detection system utilizing a PPG (photoplethysmography) signal-sensing ring is provided. According to this invention, atrial fibrillation can be continuously and easily monitored in daily life. According to this invention, the PPG signal-sensing ring also includes multiple sensors. Among the multiple sensors, a sensor with superior PPG signal quality can be selected. Therefore, a sensor can be selected to suit the different location and shape of blood vessels for each user. [Brief explanation of the drawing]

[0015] [Figure 1] This is a three-dimensional view of a PPG signal sensing ring according to one embodiment of the present disclosure. [Figure 2] This is an exploded three-dimensional view of a PPG signal sensing ring according to one embodiment of the present disclosure. [Figure 3]Block diagram of a deep learning-based atrial fibrillation determination system using a PPG signal sensing ring according to an embodiment of the present disclosure. [Figure 4A] Graph showing the PPG signal before preprocessing. [Figure 4B] Graph showing the PPG signal before preprocessing. [Figure 4C] Graph showing the PPG signal of FIG. 4A after preprocessing. [Figure 4D] Graph showing the PPG signal of FIG. 4B after preprocessing. [Figure 5A] Graph showing the PPG signal classified as good quality. [Figure 5B] Graph showing the PPG signal classified as good quality. [Figure 5C] Graph showing the PPG signal classified as poor quality. [Figure 5D] Graph showing the PPG signal classified as poor quality. [Figure 6A] Graph showing the PPG signal determined not to be atrial fibrillation. [Figure 6B] Graph showing the PPG signal determined not to be atrial fibrillation. [Figure 6C] Graph showing the PPG signal determined to be atrial fibrillation. [Figure 6D] Graph showing the PPG signal determined to be atrial fibrillation. [Figure 7] Flowchart showing a deep learning-based atrial fibrillation determination method using a PPG signal sensing ring. [Figure 8] Flowchart showing a method for setting the sensor of a PPG signal sensing ring.

Mode for Carrying Out the Invention

[0016] Figure 1 is a three-dimensional view of a PPG (photoplethysmography) signal-sensing ring 100 according to one embodiment of the present disclosure. Figure 2 is an exploded three-dimensional view of the PPG signal-sensing ring 100 according to one embodiment of the present disclosure.

[0017] Referring to Figures 1 and 2, the PPG signal sensing ring 100 also includes an external electrode 130, an internal electrode 140, an insulating unit 150, a top cover 110, an operation indicator unit 120, and multiple sensors 160.

[0018] The external electrode 130 may have an arc shape. The external electrode 130 is made of a conductor and can function as an electrode for measuring an electrocardiogram (ECG). The external electrode 130 also forms the appearance of the PPG signal sensing ring 100 and can come into contact with the user's body.

[0019] The internal electrode 140 has a ring shape and may have multiple openings 145 for multiple sensors 160. The internal electrode 140 is made of a conductor and may function as an electrode for measuring an electrocardiogram. The internal electrode 140 also forms the interior of the PPG signal sensing ring 100 and may come into contact with the user's finger.

[0020] The insulating unit 150 may be placed between the external electrode 130 and the internal electrode 140. The insulating unit 150 can enable electrical insulation between the external electrode 130 and the internal electrode 140.

[0021] The top cover 110 has an arc shape and can form a ring shape together with the external electrode 130. The top cover 110 can form the appearance of the PPG signal sensing ring 100.

[0022] The operation indicator unit 120 may be coupled to the top cover 110. The operation indicator unit 120 may also include multiple LEDs (light-emitting diodes), such as a green LED and a red LED. The operation indicator unit 120 may use multiple LEDs to display the operation of the PPG signal sensing ring 100. For example, the green LED may turn on for 2 seconds to indicate the start of measurement, and the green LED may turn on for 1 second to indicate the end of measurement. In addition, the red LED may blink repeatedly at a 1-second interval to indicate a malfunction.

[0023] Multiple sensors 160 may be positioned to contact the user's fingers. Each of the multiple sensors 160 may be located within a plurality of openings 145 of the internal electrode 140 and may protrude from the surface of the internal electrode 140. The multiple sensors 160 may be configured to acquire different PPG signals at different positions from each other. Each sensor 160 also includes a light source and a photoelectric converter.

[0024] In some embodiments, although not shown in Figures 1 and 2, the PPG signal sensing ring 100 further includes an acceleration sensor positioned between the external electrode 130 and the internal electrode 140.

[0025] Figure 3 is a block diagram of a deep learning-based atrial fibrillation detection system 1000 using a PPG signal sensing ring according to one embodiment of the present disclosure. Figures 4A and 4B are graphs showing the PPG signals before preprocessing. Figures 4C and 4D are graphs showing the PPG signals from Figures 4A and 4B after preprocessing, respectively. Figures 5A and 5B are graphs showing PPG signals classified as good quality. Figures 5C and 5D are graphs showing PPG signals classified as poor quality. Figures 6A and 6B are graphs showing PPG signals determined not to be atrial fibrillation. Figures 6C and 6D are graphs showing PPG signals determined to be atrial fibrillation.

[0026] Referring to Figures 3, 4A to 4D, 5A to 5D, and 6A to 6D, the deep learning-based atrial fibrillation detection system 1000 utilizing a PPG signal sensing ring also includes a PPG signal sensing ring 100, a first terminal 200, a server 300, and a second terminal 400.

[0027] The PPG signal sensing ring 100 may also include multiple sensors, for example, a first sensor 160A and a second sensor 160B. Although Figure 3 illustrates the PPG signal sensing ring 100 including two sensors, the number of sensors included in the PPG signal sensing ring 100 is not limited to two. In some embodiments, which are not shown in Figure 3, the PPG signal sensing ring 100 may further include an accelerometer.

[0028] The first sensor 160A also includes a first light source 161A and a first photoelectric converter 164A. The first sensor 160A can sense a first PPG signal using the first light source 161A and the first photoelectric converter 164A. The first light source 161A may include, for example, a green, red, or infrared LED, and the first photoelectric converter 164A may include a photodiode.

[0029] The second sensor 160B also includes a second light source 161B and a second photoelectric converter 164B. The second sensor 160B can sense a second PPG signal using the second light source 161B and the second photoelectric converter 164B. The second light source 161B may include, for example, a green, red, or infrared LED. The second light source 161B may emit light of the same wavelength as the first light source 161A. The second photoelectric converter 164B may include a photodiode.

[0030] The accelerometer can detect the user's movement by measuring the acceleration of the PPG signal sensing ring 100.

[0031] The first terminal 200 can be used by a user. The first terminal 200 may include, for example, a smartphone or a tablet PC (personal computer). The first terminal 200 may be connected to the PPG signal sensing ring 100 by wire or wireless. For example, the first terminal 200 may be connected to the PPG signal sensing ring 100 using Bluetooth or Wi-Fi (wireless fidelity). The first terminal 200 may be connected to the server 300 by wire or wireless. For example, the first terminal 200 may be connected to Wi-Fi or, for example, 3G (3 rd generation), LTE (long term evolution), 5G (5 th It can be connected to server 300 via mobile telecommunication technology such as generation.

[0032] The first terminal 200 also includes a light source control component 210, a sensor selection component 220, a measurement control component 230, and a display component 240. An application may be installed on the first terminal 200. The light source control component 210, the sensor selection component 220, the measurement control component 230, and the display component 240 may be embodied by the application. Alternatively, the first terminal 200 may connect to a website, and the light source control component 210, the sensor selection component 220, the measurement control component 230, and the display component 240 may be embodied on the website.

[0033] The light source control component 210 can control the first light source 161A and the second light source 161B so that the DC components of each of the two test PPG signals received from the first sensor 160A and the second sensor 160B satisfy a predetermined range.

[0034] After the light source control component 210 controls the first light source 161A and the second light source 161B, the sensor selection component 220 may select one of the first sensor 160A and the second sensor 160B as the sensor for subsequently measuring the PPG signal. The sensor selection component 220 may select the sensor that measured the test PPG signal with the highest signal quality from among the two test PPG signals from the first sensor 160A and the second sensor 160B. Signal quality is also evaluated by at least one of the following: the magnitude of the acceleration signal, the signal-to-noise ratio (SNR), and the AC (alternating current) component magnitude to DC (direct current) component magnitude ratio. The smaller the acceleration signal magnitude, the higher the signal-to-noise ratio, and the higher the AC component magnitude to DC component magnitude ratio, the higher the signal quality. Subsequently, the PPG signal may be measured by the sensor selected by the sensor selection component 220.

[0035] The control of the first light source 161A and the second light source 161B by the light source control component 210, and the sensor selection by the sensor selection component 220, can be performed sequentially. That is, after the light source control component 210 controls the first light source 161A and the second light source 161B, the sensor selection component 220 can select a sensor. The control of the first light source 161A and the second light source 161B by the light source control component 210, and the sensor selection by the sensor selection component 220, can be performed periodically. Since the PPG signal sensing ring 100 can move or rotate relative to the finger, the light source control component 210 can periodically control the light source, and the sensor selection component 220 can periodically select a sensor.

[0036] The measurement control component 230 can select a measurement mode or start or stop measurements based on user input. For example, the user can use the measurement control component 230 to select a measurement mode from self-check mode and background mode. In self-check mode, the user can use the measurement control component 230 to start and stop measurements. In background mode, measurements start and continue regardless of user input. In background mode, the user can set or change the measurement cycle. In some embodiments, the measurement control component 230 may be included in the PPG signal sensing ring 100. That is, the user can use the measurement control component 230 of the PPG signal sensing ring 100 to select a measurement mode or start or stop measurements. In some embodiments, the measurement control component 230 is also included in the server 300. That is, the service provider can use the measurement control component 230 of the server 300 to select a measurement mode or start or stop measurements.

[0037] The display component 240 can display at least one of the following: the PPG signal measured by the PPG signal sensing ring 100 and stored in the PPG signal storage component 381 of the server 300; the date and time of measurement of the PPG signal stored in the PPG signal storage component 381 of the server 300; the signal quality classification result classified by the signal quality classification component 320 of the server 300 and stored in the atrial fibrillation judgment result storage component 382 of the server 300; the atrial fibrillation judgment result determined by the atrial fibrillation judgment component 350 of the server 300 and stored in the atrial fibrillation judgment result storage component 382 of the server 300; the judgment confidence calculated by the judgment confidence calculation component 355 of the server 300 and stored in the atrial fibrillation judgment result storage component 382 of the server 300; and the atrial fibrillation index calculated by the atrial fibrillation index calculation component 360 of the server 300 and stored in the atrial fibrillation index storage component 383 of the server 300.

[0038] The server 300 also includes a PPG signal preprocessing component 310, a signal quality classification component 320, an atrial fibrillation determination component 350, a determination confidence calculation component 355, a PPG signal storage component 381, and an atrial fibrillation determination result storage component 382. In some embodiments, the server 300 further includes an atrial fibrillation index calculation component 360 and an atrial fibrillation index storage component 383. In some embodiments, the server 300 further includes an alarm component 370.

[0039] The PPG signal preprocessing component 310 can preprocess the PPG signal received by the server 300 via the first terminal 200 from a selected sensor of the PPG signal sensing ring 100. For example, a low-pass filter, a high-pass filter, and normalization may be used to preprocess the PPG signal. Figures 4A and 4B show the PPG signal before preprocessing, and Figures 4C and 4D show the PPG signal from Figures 4A and 4B after preprocessing. In some embodiments, the PPG signal preprocessing component 310 can preprocess the acceleration signal received by the server 300 via the first terminal 200 from the acceleration sensor of the PPG signal sensing ring 100.

[0040] The signal quality classification component 320 can classify the signal quality of a PPG signal into good or bad. Figures 5A and 5B show PPG signals classified as good quality by the signal quality classification component 320. Figures 5C and 5D show PPG signals classified as poor quality by the signal quality classification component 320. In some embodiments, the signal quality classification component 320 may refer to an acceleration signal to classify the signal quality of a PPG signal into one of good or bad.

[0041] If the PPG signal is classified as of poor quality, the atrial fibrillation detection result is deemed unreliable and is not displayed by the display component 240 of the first terminal 200. If the quality is good, the atrial fibrillation detection result from the PPG signal is deemed reliable and may be displayed by the display component 240 of the first terminal 200. In an alternative embodiment, regardless of the signal quality classification result, the atrial fibrillation detection result from the PPG signal may be displayed together with the signal quality classification result by the display component 240 of the first terminal 200. Also, regardless of the signal quality classification result, the atrial fibrillation detection result from the PPG signal may be displayed together with the signal quality classification result by the display component 420 of the second terminal 400.

[0042] In some embodiments, the signal quality classification component 320 may use a deep learning model to classify the quality of the PPG signal into good or bad. In some embodiments, the signal quality classification component 320 may refer to an acceleration signal to classify the quality of the PPG signal into good or bad. The deep learning model used in the signal quality classification component 320 may include a convolutional neural network (CNN), long short-term memory (LSTM), a fully connected network (FCN), an encoder, a decoder, or a combination thereof. The method for classifying signal quality is not limited to the methods described herein. The atrial fibrillation detection component 350 may use a deep learning model to determine whether atrial fibrillation has occurred from the PPG signal. Figures 6A and 6B are graphs in which the atrial fibrillation detection component 350 has determined that atrial fibrillation has not occurred. Figures 6C and 6D are graphs in which the atrial fibrillation detection component 350 has determined that atrial fibrillation has occurred. The atrial fibrillation detection component 350 calculates the probability of atrial fibrillation occurring from the PPG signal. If the probability of atrial fibrillation occurring exceeds a critical value (e.g., 0.5), it determines that atrial fibrillation has occurred. If the probability of atrial fibrillation occurring does not exceed a critical value (e.g., 0.5), it can determine that atrial fibrillation has not occurred.

[0043] The deep learning model used in the atrial fibrillation detection component 350 may include a convolutional neural network (CNN), long short-term memory (LSTM), fully connected network (FCN), encoder, decoder, or a combination thereof. The deep learning model may be trained using guided learning. Data augmentation may be used to ensure sufficient datasets for such guided learning.

[0044] The diagnostic confidence calculation component 355 can calculate the confidence level of atrial fibrillation diagnosis. For example, the diagnostic confidence calculation component 355 can calculate the diagnostic confidence level by applying temperature scaling to the atrial fibrillation diagnosis probability calculated by the atrial fibrillation diagnosis component 350 (Guo C, Pleiss G, Sun Y, Weinberger KQ, "On Calibration of Modern Neural Networks", Proceedings of the 34 th International Conference on Machine Learning, PMLR 70, 1321-1330, 2017).

[0045] The atrial fibrillation index calculation component 360 can calculate an atrial fibrillation index. The atrial fibrillation index may be defined based on the time during which the PPG signal quality is classified as good by the signal quality classification component 320, and the time during which the PPG signal quality is classified as good by the signal quality classification component 320 and the atrial fibrillation determination component 350 determines that atrial fibrillation has occurred. For example, the atrial fibrillation index may be defined as the ratio of the time during which the PPG signal quality is classified as good by the signal quality classification component 320 and the atrial fibrillation determination component 350 determines that atrial fibrillation has occurred to the time during which the PPG signal quality is classified as good by the signal quality classification component 320. In other words, the atrial fibrillation index may represent the ratio of the time during which atrial fibrillation occurred to the time during which a reliable atrial fibrillation determination result can be obtained.

[0046] The alarm component 370 can transmit an alarm to at least one of the first terminal 200 and the second terminal 400 if at least one of the atrial fibrillation determination result determined by the atrial fibrillation determination component 350 and the atrial fibrillation index calculated by the atrial fibrillation index calculation component 360 satisfies the alarm condition set by the alarm condition setting component 410 of the second terminal 400.

[0047] The PPG signal storage component 381 can store the PPG signal received by the server 300 via the first terminal 200 from the selected sensor of the PPG signal sensing ring 100. The PPG signal storage component 381 can further store the measurement date and time of the PPG signal. The PPG signal storage component 381 can further store the PPG signal that has been preprocessed by the PPG signal preprocessing component 310. The PPG signal storage component 381 can further store the acceleration signal received by the server 300 via the first terminal 200 from the acceleration sensor of the PPG signal sensing ring 100. The PPG signal storage component 381 can further store the acceleration signal that has been preprocessed by the PPG signal preprocessing component 310. The atrial fibrillation determination result storage component 382 can store the atrial fibrillation determination result determined by the atrial fibrillation determination component 350. The atrial fibrillation diagnosis result storage component 382 can further store the signal quality classification results classified by the signal quality classification component 320, and the diagnosis confidence calculated by the diagnosis confidence calculation component 355. The atrial fibrillation index storage component 383 can store the atrial fibrillation index calculated by the atrial fibrillation index calculation component 360.

[0048] The second terminal 400 may be connected to the server 300 by wire or wireless connection. For example, the second terminal 400 may be connected to the server 300 via Wi-Fi (wireless fidelity) or mobile communication technology such as 3G, LTE, or 5G. The second terminal 400 may be used by a physician. The second terminal 400 may also include, for example, a smartphone, tablet PC, computer, or laptop computer.

[0049] The second terminal 400 also includes an alarm condition setting component 410 and a display component 420. The second terminal 400 may be connected to a website or have an application installed. The alarm condition setting component 410 and the display component 420 may be implemented using a website or application.

[0050] Using the alarm condition setting component 410, a physician can set alarm conditions. For example, an alarm may be set to sound if atrial fibrillation persists for more than 10 minutes.

[0051] The display component 420 can display at least one of the following: the PPG signal from the PPG signal storage component 381, the measurement date and time of the PPG signal stored in the PPG signal storage component 381 of the server 300, the pre-processed PPG signal stored in the PPG signal storage component 381, the signal quality classification result from the atrial fibrillation judgment result storage component 382, ​​the atrial fibrillation judgment result from the atrial fibrillation judgment result storage component 382, ​​the judgment confidence from the atrial fibrillation judgment result storage component 382, ​​and the atrial fibrillation index from the atrial fibrillation index storage component 383.

[0052] Figure 7 is a flowchart illustrating a deep learning-based atrial fibrillation detection method using a PPG signal sensing ring (2000).

[0053] Referring to Figures 3 and 7, the server 300 may receive a PPG signal from a selected sensor of the PPG signal sensing ring 100 via the first terminal 200 (S2050). Next, the PPG signal preprocessing component 310 may preprocess the PPG signal (S2100). For example, a low-pass filter, a high-pass filter, and normalization may be used to preprocess the PPG signal. Figures 4A and 4B show the PPG signal before preprocessing, and Figures 4C and 4D show the PPG signal from Figures 4A and 4B, respectively, after preprocessing. In some embodiments, the PPG signal preprocessing component 310 may further preprocess the acceleration signal received by the server 300 from the acceleration sensor of the PPG signal sensing ring 100 via the first terminal 200.

[0054] Next, the signal quality classification component 320 can classify the signal quality of the PPG signal into good or bad (S2200). Figures 5A and 5B show PPG signals classified as good quality by the signal quality classification component 320. Figures 5C and 5D show PPG signals classified as poor quality by the signal quality classification component 320. Atrial fibrillation detection results from PPG signals classified as poor quality are deemed unreliable and are not displayed by the display component 240 of the first terminal 200. If the quality is good, the atrial fibrillation detection results from the PPG signal are deemed reliable and can be displayed by the display component 240 of the first terminal 200.

[0055] In some embodiments, a deep learning model may be used in the signal quality classification component 320. The deep learning model used in the signal quality classification component 320 may include a convolutional neural network (CNN), long short-term memory (LSTM), fully connected network (FCN), encoder, decoder, or a combination thereof. The method for classifying signal quality is not limited to the methods described herein. Next, the atrial fibrillation detection component 350 may use a deep learning model to determine from the PPG signal whether atrial fibrillation has occurred (S2500). Figures 6A and 6B are graphs in which the atrial fibrillation detection component 350 has determined that atrial fibrillation has not occurred. Figures 6C and 6D are graphs in which the atrial fibrillation detection component 350 has determined that atrial fibrillation has occurred. The atrial fibrillation detection component 350 calculates the probability of atrial fibrillation occurring from the PPG signal. If the probability of atrial fibrillation occurring exceeds a critical value (e.g., 0.5), it determines that atrial fibrillation has occurred. If the probability of atrial fibrillation occurring does not exceed a critical value (e.g., 0.5), it can determine that atrial fibrillation has not occurred.

[0056] The deep learning model used in the atrial fibrillation detection component 350 may include a convolutional neural network (CNN), long short-term memory (LSTM), fully connected network (FCN), encoder, decoder, or a combination thereof. The deep learning model may be trained using guided learning. Data augmentation may be used to ensure sufficient datasets for such guided learning.

[0057] Next, the judgment confidence calculation component 355 can calculate the confidence of the atrial fibrillation diagnosis (S2550). The judgment confidence calculation component 355 can calculate the judgment confidence, for example, by applying temperature scaling to the atrial fibrillation diagnosis probability calculated by the atrial fibrillation diagnosis component 350 (Guo C, Pleiss G, Sun Y, Weinberger KQ, "On Calibration of Modern Neural Networks", Proceedings of the 34 th International Conference on Machine Learning, PMLR 70, 1321-1330, 2017).

[0058] Next, the PPG signal, signal quality classification result, atrial fibrillation determination result, and determination confidence level can be stored (S2570). The PPG signal may be stored in the PPG signal storage component 381. The atrial fibrillation determination result, signal quality classification result, and determination confidence level may be stored in the atrial fibrillation determination result storage component 382.

[0059] Next, the steps of receiving the PPG signal (S2050), pre-processing the PPG signal (S2100), classifying the quality of the PPG signal (S2220), determining whether atrial fibrillation has occurred (S2500), calculating the reliability of the determination (S2550), and saving the PPG signal, signal quality classification result, atrial fibrillation determination result, and determination reliability (S2570) may be repeated.

[0060] Next, the atrial fibrillation index calculation component 360 can calculate the atrial fibrillation index (S2600). The atrial fibrillation index may be defined based on the time during which the PPG signal quality is classified as good by the signal quality classification component 320, and the time during which the PPG signal quality is classified as good by the signal quality classification component 320 and the atrial fibrillation determination component 350 determines that atrial fibrillation has occurred. For example, the atrial fibrillation index may be defined as the ratio of the time during which the PPG signal quality is classified as good by the signal quality classification component 320 and the atrial fibrillation determination component 350 determines that atrial fibrillation has occurred to the time during which the PPG signal quality is classified as good by the signal quality classification component 320. That is, the atrial fibrillation index may mean the time during which atrial fibrillation occurred within the time during which a reliable atrial fibrillation determination result can be obtained.

[0061] Next, the atrial fibrillation index can be stored in the atrial fibrillation index storage component 383 (S2700).

[0062] Figure 8 is a flowchart showing how to configure the sensor of the PPG signal sensing ring (3000).

[0063] Referring to Figures 8 and 3, the light source control component 210 can utilize multiple sensors to control each light source so that the DC component of each measured test PPG signal satisfies a predetermined range (S3100). This step (S3100) is to set each sensor, for example, the first sensor 160A and the second sensor 160B, to the best possible state for measuring the PPG signal.

[0064] Next, the sensor selection component 220 may select one of several sensors to be used to measure the PPG signal (S3200). The sensor selection component 220 may select the sensor that measured the test PPG signal with the highest signal quality from among the two test PPG signals from the first sensor 160A and the second sensor 160B. The signal quality is also evaluated by at least one of the following: the magnitude of the acceleration signal, the signal-to-noise ratio, and the AC component magnitude to DC component magnitude ratio. The smaller the magnitude of the acceleration signal, the higher the signal-to-noise ratio, and the higher the AC component magnitude to DC component magnitude ratio, the higher the signal quality. Subsequently, the PPG signal measured by the sensor selected by the sensor selection component 220 may be transmitted to the server 300 via the first terminal 200.

[0065] The steps of controlling the light source for each of the multiple sensors (S3100) and selecting a sensor (S3200) can be performed sequentially. Furthermore, the method for setting the sensors of the PPG signal sensing ring (3000) can be performed periodically. Since the PPG signal sensing ring 100 can be moved or rotated relative to the finger, the light source control component 210 can periodically control the light source (S3100), and the sensor selection component 220 can periodically select a sensor (S3200).

[0066] The embodiments disclosed herein are for illustrative purposes only, and not to limit the technical idea of ​​the disclosure, and such embodiments do not limit the scope of the technical idea of ​​the disclosure. The scope of protection of this disclosure should be interpreted as defined by the following claims, and all technical ideas within an equivalent scope should be interpreted as being included within the scope of the rights of this disclosure.

Claims

1. This is a deep learning-based atrial fibrillation detection system that utilizes a photoplethysmography (PPG) signal sensing ring, and the system includes a server, the server is A signal quality classification component configured to classify the quality of PPG signals into good or bad, It includes an atrial fibrillation detection component configured to determine whether atrial fibrillation has occurred from the PPG signal using a deep learning model, The PPG signal sensing ring includes a plurality of sensors configured to simultaneously measure a plurality of biosignals at different locations. Each of the plurality of sensors includes a light source and a photoelectric converter, and the terminal wirelessly connected to the PPG signal sensing ring includes a sensor selection component configured to select, from among the plurality of sensors, the sensor that measured the test PPG signal with the highest signal quality from among a plurality of test PPG signals obtained from each of the plurality of sensors, in order to select the sensor from which the PPG signal should be acquired, as the sensor for measuring the PPG signal. A deep learning-based atrial fibrillation detection system utilizing a PPG signal sensing ring, characterized in that the PPG signal is measured using only the selected sensor of the PPG signal sensing ring that comes into contact with the user's finger, and the terminal receives the PPG signal measured by only the selected sensor from the PPG signal sensing ring and transmits it to the server.

2. The deep learning-based atrial fibrillation detection system using a PPG signal sensing ring according to claim 1, characterized in that the signal quality of the plurality of test PPG signals is evaluated by at least one of the acceleration signal magnitude, signal-to-noise ratio, and AC component magnitude to DC component magnitude ratio.

3. The server further includes an atrial fibrillation index calculation component configured to calculate an atrial fibrillation index, The atrial fibrillation index is defined as the ratio of the time at which the signal quality classification component classifies the PPG signal as having good quality and the atrial fibrillation determination component determines that atrial fibrillation has occurred to the time at which the signal quality classification component classifies the PPG signal as having good quality, thus providing a deep learning-based atrial fibrillation determination system using a PPG signal sensing ring according to claim 1.

4. The deep learning-based atrial fibrillation detection system using a PPG signal sensing ring according to claim 1, further comprising a light source control component configured to control the light source of each of the multiple sensors so that the DC component of each of the multiple test PPG signals measured using the multiple sensors falls within a predetermined range.

5. The light source control by the light source control component and the sensor selection by the sensor selection component are performed sequentially. The deep learning-based atrial fibrillation detection system using a PPG signal sensing ring according to claim 4, characterized in that the light source control by the light source control component and the sensor selection by the sensor selection component are performed periodically.

6. The process involves measuring the user's PPG signal using a photoplethysmography (PPG) signal-sensing ring that comes into contact with the user's finger, and receiving the PPG signal via a terminal wirelessly connected to the PPG signal-sensing ring. The server classifies the quality of the PPG signal into good or bad, The server includes the step of using a deep learning model to determine whether or not atrial fibrillation has occurred from the PPG signal, The PPG signal sensing ring includes a plurality of sensors configured to simultaneously measure a plurality of biosignals at different locations. Each of the plurality of sensors includes a light source and a photoelectric converter, and the terminal includes a sensor selection component configured to select, from among the plurality of sensors, the sensor that measured the test PPG signal with the highest signal quality from among a plurality of test PPG signals obtained from each of the plurality of sensors, in order to select the sensor from which the PPG signal should be acquired, as the sensor for measuring the PPG signal. A deep learning-based atrial fibrillation detection method using a PPG signal sensing ring, characterized in that the PPG signal is measured using only selected sensors of the PPG signal sensing ring that come into contact with the user's finger, and the terminal device receives the PPG signal measured by only the selected sensors from the PPG signal sensing ring and transmits it to the server.

7. The deep learning-based atrial fibrillation detection method using a PPG signal sensing ring according to claim 6, characterized in that the signal quality of the plurality of test PPG signals is evaluated by at least one of the acceleration signal magnitude, signal-to-noise ratio, and AC component magnitude to DC component magnitude ratio.

8. The server further comprises the step of calculating an atrial fibrillation index, The atrial fibrillation index is defined as the ratio of the time during which the PPG signal quality is classified as good by the signal quality classification component and atrial fibrillation is determined to have occurred by the atrial fibrillation determination component to the time during which the PPG signal quality is classified as good by the signal quality classification component, and a deep learning-based atrial fibrillation determination method using a PPG signal sensing ring according to claim 6.

9. The deep learning-based atrial fibrillation detection method using a PPG signal sensing ring according to claim 6, further comprising a light source control component configured to control the light source of each of the multiple sensors so that the DC component of each of the multiple test PPG signals measured using the multiple sensors falls within a predetermined range.

10. The light source control by the light source control component and the sensor selection by the sensor selection component are performed sequentially. The deep learning-based atrial fibrillation detection method using a PPG signal sensing ring according to claim 9, characterized in that the light source control by the light source control component and the sensor selection by the sensor selection component are performed periodically.

Images