A Deep Learning-Based Smartwatch Arrhythmia Early Warning Method and System

Abstract

This invention discloses a method and system for early warning of arrhythmias in smartwatches based on deep learning. The method first acquires PPG signals using a photoplethysmography (PPG) sensor and performs noise reduction and standardization using an adaptive filtering algorithm. Then, it uses a wavelet transform peak detection algorithm to identify heartbeat locations and obtain a heart rate variability (HRV) feature sequence. Next, it uses an improved Gramian angular field transform algorithm to convert the one-dimensional HRV feature sequence into a two-dimensional heart rhythm image. Then, it uses a ResNet-50 convolutional neural network to train an arrhythmia classification model, achieving automatic identification of different types of arrhythmias. Finally, it uses sliding window technology and an early warning mechanism to achieve continuous monitoring and timely alarms. This invention effectively solves key technical problems in existing arrhythmia detection technologies, such as low detection accuracy, poor real-time performance, high equipment costs, and unsatisfactory user experience, providing an innovative solution for heart rate health monitoring in smart wearable devices.

Description

Technical Field

[0001] This invention relates to the technical field of arrhythmia early warning, and more particularly to a method and system for early warning of arrhythmia in smartwatches based on deep learning. Background Technology

[0002] Cardiac arrhythmias refer to abnormalities in the origin, frequency, rhythm, conduction velocity, and order of cardiac impulses. They are among the most complex cardiovascular diseases, exhibiting significant variations, a wide range of types, and posing a serious threat to human health. With the accelerating pace of modern life and the increasing aging of the population, the incidence of arrhythmias is rising year by year. Dangerous arrhythmias such as atrial fibrillation and ventricular tachycardia, if not detected and treated promptly, can lead to serious consequences such as sudden cardiac death. Therefore, continuous monitoring and early warning of arrhythmias have significant clinical and social value.

[0003] In recent years, with the rapid development of sensor technology, communication technology, and artificial intelligence algorithms, smart wearable devices have been widely used in the field of heart rate monitoring. Smartwatches, as the most representative wearable heart rate monitoring devices, have gradually become important tools for personal health management and disease prevention due to their advantages such as ease of wear, continuous monitoring, and real-time feedback. Currently, smartwatches on the market mainly use photoplethysmography (PPG) sensors for heart rate detection. Compared with traditional electrocardiogram (ECG) monitoring methods, PPG technology has advantages such as non-invasiveness, low cost, and ease of integration, laying the technological foundation for the widespread adoption of heart rate monitoring.

[0004] However, existing smart wearable devices still suffer from numerous technical shortcomings and application limitations in arrhythmia monitoring. Traditional arrhythmia detection methods are mainly based on simple threshold judgments or basic statistical feature analysis, resulting in low accuracy in identifying complex arrhythmia types and a tendency to produce false alarms and false negatives. In particular, during users' daily activities, factors such as motion artifacts, changes in skin contact, and ambient light interference can severely affect the quality of PPG signals, leading to a significant decrease in the reliability of arrhythmia detection.

[0005] While existing technologies, exemplified by patent CN202411576388.X, offer some solutions for environmental interference identification, they still suffer from significant limitations. This technology primarily relies on location-based interference labeling methods, using active positioning signals within medical settings to identify environmental interference factors and labeling electrocardiograms (ECGs) upon detection of interference. However, this method has the following key drawbacks: First, it can only identify known interference sources in specific locations, lacking adaptability to unpredictable interference environments and dynamically changing interference factors, thus limiting its application in complex real-world environments. Second, the simple interference labeling strategy cannot fundamentally improve the accuracy of arrhythmia detection; it merely identifies potentially interfered data without effectively extracting and utilizing valid cardiac rhythm feature information. Finally, this technology lacks intelligent signal processing and pattern recognition algorithms, failing to achieve accurate classification and identification of different types of arrhythmias, thus limiting its practical value in clinical applications.

[0006] Furthermore, existing technologies also have significant shortcomings in feature extraction and pattern recognition. Traditional methods often employ single-dimensional features in the time or frequency domains, failing to fully extract the complex physiological information contained in PPG signals and exhibiting limited feature representation capabilities. Simultaneously, classification methods based on traditional machine learning algorithms perform poorly when dealing with high-dimensional and complex data, struggling to handle nonlinear feature relationships and temporal dependencies in arrhythmia detection. Regarding real-time performance, existing systems often require lengthy data acquisition and processing times, failing to meet the timeliness requirements of emergency arrhythmia warnings and potentially missing the optimal window for medical intervention.

[0007] In summary, existing arrhythmia monitoring technologies for smart wearable devices have significant shortcomings in signal processing algorithms, feature extraction methods, pattern recognition accuracy, and real-time early warning capabilities. There is an urgent need to develop more intelligent, high-precision, and robust arrhythmia early warning technologies to meet the growing demand for personal health monitoring and clinical applications. Summary of the Invention

[0008] In view of this, the present invention provides a method for early warning of arrhythmia in smartwatches based on deep learning, aiming to solve the technical problems of low detection accuracy, poor real-time performance, high equipment cost and poor user experience in existing arrhythmia detection technologies.

[0009] To achieve the above objectives, this invention provides a deep learning-based method for early warning of cardiac arrhythmia in smartwatches, comprising the following steps:

[0010] S1: The raw PPG signal is collected by the photoplethysmography (PPG) sensor built into the smartwatch. The collected raw PPG signal is then denoised and standardized to obtain a standardized PPG signal.

[0011] S2: A wavelet transform-based peak detection algorithm is used to identify the heartbeat positions contained in the standardized PPG signal. After outlier detection and smoothing, a heart rate variability feature sequence is obtained.

[0012] S3: The heart rate variability feature sequence is subjected to min-max normalization, and then an improved Gramian angle field transform algorithm is used to convert the one-dimensional heart rate variability feature sequence into a two-dimensional heart rhythm image matrix. The improved Gramian angle field transform algorithm is to perform polar coordinate transformation on the heart rate variability feature sequence after min-max normalization, converting the normalized sequence into a point representation on the unit circle. The improved Gramian angle field transform algorithm calculates the Gramian angle and field and angle difference field between any two sequence points in the transformed sequence, respectively capturing the similarity and difference features of the time series data. Finally, the average value of the two angle field information is taken and fused to generate a two-dimensional heart rhythm image matrix, realizing an effective conversion from one-dimensional time series to two-dimensional image.

[0013] S4: The ResNet-50 convolutional neural network architecture is adopted, and the arrhythmia classification model is trained through supervised learning to achieve automatic identification and classification of different types of arrhythmias.

[0014] S5: Through sliding window technology and early warning mechanism, it realizes continuous monitoring of the user's heart rhythm status and timely alarm of abnormal situations.

[0015] As a further improvement of the present invention:

[0016] Optionally, in step S1, the raw PPG signal is acquired using the photoplethysmography (PPG) sensor built into the smartwatch, and an adaptive filtering algorithm is used to denoise and standardize the acquired raw PPG signal to obtain a standardized PPG signal, including:

[0017] During the PPG signal acquisition process, the sensor operates at a fixed sampling frequency. Photoelectric detection is performed on the user's wrist to obtain the raw PPG signal. ,in Indicates the sampling time point;

[0018] For the acquired raw PPG signal Bandpass filtering is performed to remove low-frequency drift and high-frequency noise interference, resulting in a preliminary clean signal. The cutoff frequency of the bandpass filter is designed based on the normal human heart rate range to ensure that effective heart rate-related frequency components are retained.

[0019] An adaptive noise cancellation algorithm is used to further process the initial cleaning signal to eliminate artifact interference and obtain the denoised PPG signal. ;

[0020] For the denoised PPG signal Amplitude normalization is performed to obtain a standardized PPG signal. .

[0021] Optionally, in step S2, a wavelet transform-based peak detection algorithm is used to identify the heartbeat locations contained in the standardized PPG signal, including:

[0022] For standardized PPG signals Perform continuous wavelet transform processing to obtain the wavelet transform result. ,in This represents the scaling parameter, used to control the frequency characteristics of the wavelet. This represents the displacement parameter, used to determine the wavelet's position on the time axis;

[0023] In wavelet transform results The system searches for local extrema, which correspond to the peak positions of heartbeats in the standardized PPG signal. By using predefined threshold conditions, heartbeat events are selected, and spurious peaks caused by noise or abnormal fluctuations are eliminated, resulting in a heartbeat time series. ,in Indicates the first The timing of the second heartbeat This indicates the total number of heartbeats detected;

[0024] Based on the detected adjacent heartbeat time points, the time interval between consecutive heartbeats is calculated to form an RR interval sequence. ,in Indicates the first The duration of each RR interval;

[0025] The extracted RR interval sequences are subjected to outlier detection and smoothing to eliminate extreme values ​​caused by detection errors or physiological abnormalities, thereby obtaining heart rate variability characteristic sequences. ,in The first sequence representing the characteristic sequence of heart rate variability One effective RR interval value, This indicates the number of valid RR intervals after outlier detection and smoothing.

[0026] Optionally, in step S3, the heart rate variability feature sequence is subjected to minimum-maximum normalization, and then an improved Gramian angle field transformation algorithm is used to convert the one-dimensional heart rate variability feature sequence into a two-dimensional heart rhythm image. The improved Gramian angle field transformation algorithm maps the heart rate variability feature sequence to the 0-1 interval through minimum-maximum normalization, and then performs polar coordinate transformation to convert the normalized sequence into a point representation on the unit circle. The algorithm calculates the Gramian angle and field and the angle difference field between any two sequence points in the sequence, respectively capturing the similarity and difference features of the time series data. Finally, the average value of the two angle field information is fused to generate a two-dimensional heart rhythm image matrix, realizing an effective conversion from one-dimensional time series to two-dimensional image, including:

[0027] Heart rate variability feature sequence Min-max normalization is performed to map the numerical range to the interval [0, 1], resulting in a standardized heart rate variability feature sequence. ;

[0028] The standardized heart rate variability feature sequence is mapped to polar coordinates, converting the standardized heart rate variability feature sequence into points on the unit circle;

[0029] An improved Gramian angle field transformation algorithm is used to calculate the angular relationship between any two standardized effective RR interval values ​​in a standardized heart rate variability feature sequence. The angular relationship includes the Gramian angle sum field and the Gramian angle difference field.

[0030] The average value of the Gramian angle and field and the Gramian angle difference field is used as the cardiac rhythm image matrix. Output, where and These represent the row and column coordinates of the image, respectively.

[0031] Optionally, step S4 employs a ResNet-50 convolutional neural network architecture to train an arrhythmia classification model through supervised learning, thereby achieving automatic identification and classification of different types of arrhythmias, including:

[0032] Construct the input layer of the ResNet-50 network, using the heart rhythm image matrix. The standard input size was adjusted to 224×224×3, and the single-channel heart rhythm image was expanded into a three-channel RGB format using an image copying method.

[0033] Load the parameters of the ResNet-50 model pre-trained on the ImageNet dataset; freeze all convolutional layer parameters of the ResNet-50 network and fine-tune the training only for the fully connected layers;

[0034] The last fully connected layer of the ResNet-50 model was replaced, and a new classification head structure was designed, including a global average pooling layer, a Dropout regularization layer, and a fully connected output layer. The number of neurons in the output layer was set to correspond to the number of arrhythmia types, and a softmax activation function was used to output the probability distribution vectors of each category. ,in Indicates the first Predictive probability of arrhythmias This indicates the total number of arrhythmia categories;

[0035] The network was fine-tuned and trained using a labeled cardiac arrhythmia image dataset. The error between the predicted results and the true labels was calculated using the cross-entropy loss function, and the parameters were updated using the Adam optimizer. After several iterations of training, a fully trained arrhythmia classification model is obtained. .

[0036] Optionally, step S5 utilizes sliding window technology and an early warning mechanism to achieve continuous monitoring of the user's heart rhythm and timely alarm for abnormal situations, including:

[0037] Establish a real-time data stream processing mechanism to continuously receive new PPG signal data collected by the smartwatch and process it at fixed time intervals. The new data is segmented and processed.

[0038] The newly received PPG data segments are converted in real time according to the processing flow from steps S1 to S3 to generate the corresponding cardiac rhythm image matrix. ;

[0039] The real-time generated heart rhythm image matrix Input into the trained arrhythmia classification model Inference calculations are performed to obtain the predicted probability vector of arrhythmia for the current time period. ,in Indicates the currently detected number Probability value of arrhythmia-like events;

[0040] Design a multi-level early warning judgment mechanism based on the predicted probability vector. The system compares the maximum probability value with a preset risk threshold for judgment; when a high-risk arrhythmia type is detected and the probability exceeds the safety threshold, the system automatically generates a warning signal and immediately notifies the user through the vibration, sound or screen prompt of the smartwatch.

[0041] This invention also discloses a deep learning-based smartwatch arrhythmia early warning system, comprising:

[0042] Signal Acquisition and Standardization Module: The raw PPG signal is acquired through the photoplethysmography (PPG) sensor built into the smartwatch, and the acquired raw PPG signal is denoised and standardized using an adaptive filtering algorithm.

[0043] Heart rate variability feature extraction module: The peak detection algorithm based on wavelet transform is used to identify the heart beat positions contained in the standardized PPG signal. After outlier detection and smoothing, the heart rate variability feature sequence is obtained.

[0044] Image conversion module: Performs minimum-maximum normalization on the heart rate variability feature sequence, and then uses an improved Gramian angular field transform algorithm to convert the one-dimensional heart rate variability feature sequence into a two-dimensional heart rhythm image;

[0045] Classification module: The ResNet-50 convolutional neural network architecture is used to train the arrhythmia classification model through supervised learning, so as to realize the automatic identification and classification of different types of arrhythmias;

[0046] Monitoring module: Through sliding window technology and early warning mechanism, it realizes continuous monitoring of the user's heart rhythm status and timely alarm of abnormal situations.

[0047] Compared with the prior art, the present invention has at least the following beneficial effects:

[0048] This invention combines Gramian angles and fields with Gramian angle difference fields. Through polar coordinate mapping and angle relationship calculation, it simultaneously encodes the temporal dependence, periodic features, and local variation patterns of a one-dimensional heart rate variability time series into different dimensions of a two-dimensional image. Compared to traditional direct time-domain or frequency-domain feature extraction methods, this algorithm can preserve the global temporal structure of the original signal while highlighting local anomalous changes, effectively solving the feature loss problem in PPG signals caused by motion artifacts, contact changes, and other factors.

[0049] This invention employs a ResNet-50 network pre-trained on the ImageNet dataset as a feature extractor. By freezing all convolutional layer parameters and fine-tuning only the fully connected layers, it effectively utilizes the general feature representation capabilities learned from large-scale natural image datasets. This transfer learning method not only significantly reduces the amount of arrhythmia-labeled data required for model training, lowering data collection costs, but also significantly shortens training time and improves the model's generalization ability.

[0050] The multi-level early warning judgment mechanism designed in this invention intelligently compares the maximum probability value in the predicted probability vector with a preset risk threshold. This not only accurately identifies high-risk arrhythmia events but also effectively reduces the false alarm rate, avoiding the impact of frequent false alarms on user experience. When a dangerous arrhythmia is detected and its probability exceeds the safety threshold, the system can immediately issue an early warning to the user simultaneously through multiple sensory channels, ensuring timely response at critical moments. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating a method for early warning of cardiac arrhythmia in a smartwatch based on deep learning, according to an embodiment of the present invention.

[0052] Figure 2 This is a schematic diagram of the original PPG according to an embodiment of the present invention;

[0053] Figure 3 This is a schematic diagram of a heart rhythm image according to an embodiment of the present invention. Detailed Implementation

[0054] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0055] Example 1: A deep learning-based method for early warning of cardiac arrhythmia in smartwatches, such as... Figure 1 As shown, it includes the following steps:

[0056] S1: The raw PPG signal is acquired through the photoplethysmography (PPG) sensor built into the smartwatch, such as... Figure 2 As shown, the acquired raw PPG signal is denoised and standardized to obtain a standardized PPG signal. In this embodiment, the adaptive filtering algorithm is the minimum mean square error adaptive filtering algorithm.

[0057] During the PPG signal acquisition process, the sensor operates at a fixed sampling frequency. Photoelectric detection is performed on the user's wrist to obtain the raw PPG signal. ,in Indicates the sampling time point;

[0058] For the acquired raw PPG signal Bandpass filtering is performed to remove low-frequency drift and high-frequency noise interference, resulting in a preliminary clean signal. The cutoff frequency of the bandpass filter is designed according to the normal human heart rate range to ensure that effective heart rate-related frequency components are retained. In this embodiment, the cutoff frequency of the bandpass filter is 2Hz.

[0059] An adaptive noise cancellation algorithm is used to further process the initial cleaning signal to eliminate artifact interference caused by factors such as user movement and changes in skin contact, thus obtaining a denoised PPG signal. In this embodiment, the adaptive noise cancellation algorithm is a recursive least squares algorithm;

[0060] For the denoised PPG signal Amplitude normalization is performed to obtain a standardized PPG signal. .

[0061] S2: A wavelet transform-based peak detection algorithm is used to identify the heartbeat positions contained in the standardized PPG signal.

[0062] For standardized PPG signals Perform continuous wavelet transform processing to obtain the wavelet transform result. ,in This represents the scaling parameter, used to control the frequency characteristics of the wavelet. This represents the displacement parameter, used to determine the wavelet's position on the time axis;

[0063] In wavelet transform results The search involves identifying local extrema, which correspond to the peak heartbeat position in the standardized PPG signal. Heartbeat events are then selected based on a predefined threshold condition; in this embodiment, the threshold condition is... By eliminating spurious peaks caused by noise or abnormal fluctuations, the heartbeat time series is obtained. ,in Indicates the first The timing of the second heartbeat This indicates the total number of heartbeats detected;

[0064] Based on the detected adjacent heartbeat time points, the time interval between consecutive heartbeats is calculated to form an RR interval sequence. ,in Indicates the first The duration of each RR interval;

[0065] The extracted RR interval sequence undergoes outlier detection and smoothing to eliminate extreme values ​​caused by detection errors or physiological abnormalities. In this embodiment, outlier detection and smoothing are based on an outlier monitoring algorithm using interquartile range and a median filtering smoothing algorithm, respectively, to obtain the heart rate variability feature sequence. ,in The first sequence representing the characteristic sequence of heart rate variability One effective RR interval value, This indicates the number of valid RR intervals after outlier detection and smoothing.

[0066] This step utilizes continuous wavelet transform technology to achieve high-precision identification of the center beat position of PPG signals, offering significant advantages over traditional differential thresholding or template matching methods. Wavelet transform can simultaneously analyze signal features in the time and frequency domains, effectively capturing local features of the heartbeat peak through multi-scale analysis, maintaining stable detection performance even in situations with baseline drift or morphological changes. Particularly in cases of user motion or poor signal quality, the multi-resolution characteristics of wavelet transform can identify heartbeat features at different scales, greatly improving the robustness and accuracy of peak detection.

[0067] S3: The heart rate variability feature sequence is subjected to min-max normalization, and then an improved Gramian angle field transform algorithm is used to convert the one-dimensional heart rate variability feature sequence into a two-dimensional heart rhythm image matrix. The improved Gramian angle field transform algorithm is to perform polar coordinate transformation on the heart rate variability feature sequence after min-max normalization, converting the normalized sequence into a point representation on the unit circle. The improved Gramian angle field transform algorithm calculates the Gramian angle and field and angle difference field between any two sequence points in the transformed sequence, respectively capturing the similarity and difference features of the time series data. Finally, the average value of the two angle field information is taken and fused to generate a two-dimensional heart rhythm image matrix, realizing an effective conversion from one-dimensional time series to two-dimensional image.

[0068] Heart rate variability feature sequence Min-max normalization is performed to map the numerical range to the interval [0, 1], resulting in a standardized heart rate variability feature sequence. ,in Represents the standardized first digit in the heart rate variability characteristic sequence. One effective RR interval value;

[0069] The standardized heart rate variability feature sequence is mapped to polar coordinates, transforming it into points on the unit circle, specifically as follows:

[0070]

[0071] in express The corresponding polar angle;

[0072] An improved Gramian angle field transformation algorithm is used to calculate the angular relationship between any two standardized effective RR interval values ​​in the standardized heart rate variability feature sequence. This angular relationship includes the Gramian angle sum field and the Gramian angle difference field, specifically:

[0073]

[0074]

[0075] in, and These represent the index positions of the standardized heart rate variability feature sequence; For Gramian angles and fields The value at; For Gramian angle difference field The value at; and Represents the standardized first digit in the heart rate variability characteristic sequence. The and the first One effective RR interval value;

[0076] The average value of the Gramian angle and field and the Gramian angle difference field is used as the cardiac rhythm image matrix. Output, such as Figure 3 As shown, where and These represent the row and column coordinates of the image, respectively.

[0077] This step achieves intelligent conversion from one-dimensional time-series data to two-dimensional images by improving the Gramian angular field transform algorithm, providing rich feature representations for deep learning models. Compared to traditional direct time-frequency analysis or simple numerical feature extraction methods, the Gramian angular field transform can completely preserve the temporal dependencies and local variation patterns of the original time series, while encoding this complex temporal information into the spatial structure of the two-dimensional image. This transformation not only preserves the global trend information in the heart rate variability sequence, but more importantly, it can highlight local abnormal changes and periodic patterns, providing an intuitive and information-rich input format for subsequent convolutional neural network recognition.

[0078] S4: Employing a ResNet-50 convolutional neural network architecture, a cardiac arrhythmia classification model is trained using supervised learning to achieve automatic identification and classification of different types of cardiac arrhythmias.

[0079] Construct the input layer of the ResNet-50 network, using the heart rhythm image matrix. The standard input size was adjusted to 224×224×3, and the single-channel heart rhythm image was expanded into a three-channel RGB format using an image copying method.

[0080] Load the parameters of the ResNet-50 model pre-trained on the ImageNet dataset; freeze all convolutional layer parameters of the ResNet-50 network and fine-tune the training only for the fully connected layers;

[0081] The last fully connected layer of the ResNet-50 model was replaced, and a new classification head structure was designed, including a global average pooling layer, a Dropout regularization layer, and a fully connected output layer. The number of neurons in the output layer was set to correspond to the number of arrhythmia types, and a softmax activation function was used to output the probability distribution vectors of each category. ,in Indicates the first Predictive probability of arrhythmias This represents the total number of arrhythmia categories. In this embodiment, the arrhythmia categories include no abnormality, atrial fibrillation, ventricular fibrillation, and ventricular tachycardia.

[0082] The network was fine-tuned and trained using a labeled cardiac arrhythmia image dataset. The error between the predicted results and the true labels was calculated using the cross-entropy loss function, and the parameters were updated using the Adam optimizer. After several iterations of training, a fully trained arrhythmia classification model is obtained. In this embodiment For training data with imbalanced classes, the system uses an average loss function of Focal loss and cross-entropy loss to improve the recognition accuracy of the minority classes.

[0083] This step leverages the powerful advantages of deep residual networks in complex pattern recognition tasks by employing the ResNet-50 deep convolutional neural network architecture. The residual connection structure of ResNet-50 effectively solves the gradient vanishing problem in deep network training, enabling the network to learn richer and more abstract feature representations. Compared to traditional shallow machine learning methods or simple neural networks, ResNet-50 possesses powerful nonlinear modeling capabilities, enabling it to capture complex spatial patterns and texture features in cardiac rhythm images.

[0084] S5: Through sliding window technology and an early warning mechanism, continuous monitoring of the user's heart rhythm status and timely alarms for abnormal situations are achieved.

[0085] Establish a real-time data stream processing mechanism to continuously receive new PPG signal data collected by the smartwatch and process it at fixed time intervals. The new data is segmented in this embodiment. Second;

[0086] The newly received PPG data segments are converted in real time according to the processing flow from steps S1 to S3 to generate the corresponding cardiac rhythm image matrix. ;

[0087] The real-time generated heart rhythm image matrix Input into the trained arrhythmia classification model Inference calculations are performed to obtain the predicted probability vector of arrhythmia for the current time period. ,in Indicates the currently detected number The probability value of arrhythmia; when the confidence of a single prediction result is low, the system adopts a sliding window multi-frame fusion decision mechanism, which combines historical detection results for weighted averaging to improve the stability and reliability of the prediction. The multi-frame refers to the previous 5 frames of the current frame.

[0088] Design a multi-level early warning judgment mechanism based on the predicted probability vector. The system compares the maximum probability value with a preset risk threshold for judgment. When a high-risk arrhythmia type is detected and the probability exceeds the safety threshold, the system automatically generates a warning signal and immediately notifies the user through the vibration, sound, or screen prompt of the smartwatch. In this embodiment, the safety threshold is 0.8.

[0089] Example 2: This invention also discloses a deep learning-based smartwatch arrhythmia early warning system, comprising the following five modules:

[0090] Signal Acquisition and Standardization Module: The raw PPG signal is acquired through the photoplethysmography (PPG) sensor built into the smartwatch, and the acquired raw PPG signal is denoised and standardized using an adaptive filtering algorithm.

[0091] Heart rate variability feature extraction module: The peak detection algorithm based on wavelet transform is used to identify the heart beat positions contained in the standardized PPG signal. After outlier detection and smoothing, the heart rate variability feature sequence is obtained.

[0092] Image conversion module: Performs minimum-maximum normalization on the heart rate variability feature sequence, and then uses an improved Gramian angular field transform algorithm to convert the one-dimensional heart rate variability feature sequence into a two-dimensional heart rhythm image;

[0093] Classification module: The ResNet-50 convolutional neural network architecture is used to train the arrhythmia classification model through supervised learning, so as to realize the automatic identification and classification of different types of arrhythmias;

[0094] Monitoring module: Through sliding window technology and early warning mechanism, it realizes continuous monitoring of the user's heart rhythm status and timely alarm of abnormal situations.

[0095] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0096] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0097] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for early warning of cardiac arrhythmia in smartwatches based on deep learning, characterized in that, Includes the following steps: S1: The raw PPG signal is collected by the photoplethysmography (PPG) sensor built into the smartwatch. The collected raw PPG signal is then denoised and standardized to obtain a standardized PPG signal. S2: A wavelet transform-based peak detection algorithm is used to identify the heartbeat positions contained in the standardized PPG signal. After outlier detection and smoothing, a heart rate variability feature sequence is obtained. S3: The heart rate variability feature sequence is subjected to min-max normalization, and then an improved Gramian angle field transform algorithm is used to convert the one-dimensional heart rate variability feature sequence into a two-dimensional heart rhythm image matrix. The improved Gramian angle field transform algorithm is to perform polar coordinate transformation on the heart rate variability feature sequence after min-max normalization, converting the normalized sequence into a point representation on the unit circle. The improved Gramian angle field transform algorithm calculates the Gramian angle and field and angle difference field between any two sequence points in the transformed sequence, respectively capturing the similarity and difference features of the time series data. Finally, the average value of the two angle field information is taken and fused to generate a two-dimensional heart rhythm image matrix, realizing an effective conversion from one-dimensional time series to two-dimensional image. S4: The ResNet-50 convolutional neural network architecture is adopted, and the arrhythmia classification model is trained through supervised learning to achieve automatic identification and classification of different types of arrhythmias. S5: Through sliding window technology and early warning mechanism, it realizes continuous monitoring of the user's heart rhythm status and timely alarm of abnormal situations.

2. The method for early warning of cardiac arrhythmia in smartwatches based on deep learning according to claim 1, characterized in that, Step S1 includes: During the PPG signal acquisition process, the sensor operates at a fixed sampling frequency. Photoelectric detection is performed on the user's wrist to obtain the raw PPG signal. ,in Indicates the sampling time point; For the acquired raw PPG signal Bandpass filtering is performed to remove low-frequency drift and high-frequency noise interference, resulting in a preliminary clean signal. The cutoff frequency of the bandpass filter is designed based on the normal human heart rate range to ensure that effective heart rate-related frequency components are retained. An adaptive noise cancellation algorithm is used to further process the initial cleaning signal to eliminate artifact interference and obtain the denoised PPG signal. ; For the denoised PPG signal Amplitude normalization is performed to obtain a standardized PPG signal. .

3. The method for early warning of cardiac arrhythmia in smartwatches based on deep learning according to claim 2, characterized in that, Step S2 includes: For standardized PPG signals Perform continuous wavelet transform processing to obtain the wavelet transform result. ,in This represents the scaling parameter, used to control the frequency characteristics of the wavelet. This represents the displacement parameter, used to determine the wavelet's position on the time axis; In wavelet transform results The system searches for local extrema, which correspond to the peak positions of heartbeats in the standardized PPG signal. By using predefined threshold conditions, heartbeat events are selected, and spurious peaks caused by noise or abnormal fluctuations are eliminated, resulting in a heartbeat time series. ,in Indicates the first The timing of the second heartbeat This indicates the total number of heartbeats detected; Based on the detected adjacent heartbeat time points, the time interval between consecutive heartbeats is calculated to form an RR interval sequence. ,in Indicates the first The duration of each RR interval; The extracted RR interval sequences are subjected to outlier detection and smoothing to eliminate extreme values ​​caused by detection errors or physiological abnormalities, thereby obtaining heart rate variability characteristic sequences. ,in The first sequence representing the characteristic sequence of heart rate variability One effective RR interval value, This indicates the number of valid RR intervals after outlier detection and smoothing.

4. The method for early warning of cardiac arrhythmia in smartwatches based on deep learning according to claim 3, characterized in that, Step S3 includes: Heart rate variability feature sequence Min-max normalization is performed to map the numerical range to the interval [0, 1], resulting in a standardized heart rate variability feature sequence. ; The standardized heart rate variability feature sequence is mapped to polar coordinates, converting the standardized heart rate variability feature sequence into points on the unit circle; An improved Gramian angle field transformation algorithm is used to calculate the angular relationship between any two standardized effective RR interval values ​​in a standardized heart rate variability feature sequence. The angular relationship includes the Gramian angle sum field and the Gramian angle difference field. The average value of the Gramian angle and field and the Gramian angle difference field is used as the cardiac rhythm image matrix. Output, where and These represent the row and column coordinates of the image, respectively.

5. The method for early warning of cardiac arrhythmia in smartwatches based on deep learning according to claim 4, characterized in that, Step S4 includes: Construct the input layer of the ResNet-50 network, using the heart rhythm image matrix. The standard input size was adjusted to 224×224×3, and the single-channel heart rhythm image was expanded into a three-channel RGB format using an image copying method. Load the parameters of the ResNet-50 model pre-trained on the ImageNet dataset; freeze all convolutional layer parameters of the ResNet-50 network and fine-tune the training only for the fully connected layers; The last fully connected layer of the ResNet-50 model was replaced, and a new classification head structure was designed, including a global average pooling layer, a Dropout regularization layer, and a fully connected output layer. The number of neurons in the output layer was set to correspond to the number of arrhythmia types, and a softmax activation function was used to output the probability distribution vectors of each category. ,in Indicates the first Predictive probability of arrhythmias This indicates the total number of arrhythmia categories; The network was fine-tuned and trained using a labeled cardiac arrhythmia image dataset. The error between the predicted results and the true labels was calculated using the cross-entropy loss function, and the parameters were updated using the Adam optimizer. After several iterations of training, a fully trained arrhythmia classification model is obtained. .

6. The method for early warning of cardiac arrhythmia in smartwatches based on deep learning according to claim 5, characterized in that, Step S5 includes: Establish a real-time data stream processing mechanism to continuously receive new PPG signal data collected by the smartwatch and process it at fixed time intervals. The new data is segmented and processed. The newly received PPG data segments are converted in real time according to the processing flow from steps S1 to S3 to generate the corresponding cardiac rhythm image matrix. ; The real-time generated heart rhythm image matrix Input into the trained arrhythmia classification model Inference calculations are performed to obtain the predicted probability vector of arrhythmia for the current time period. ,in Indicates the currently detected number Probability value of arrhythmia-like events; Design a multi-level early warning judgment mechanism based on the predicted probability vector. The system compares the maximum probability value with a preset risk threshold for judgment; when a high-risk arrhythmia type is detected and the probability exceeds the safety threshold, the system automatically generates a warning signal and immediately notifies the user through the vibration, sound or screen prompt of the smartwatch.

7. A deep learning-based smartwatch arrhythmia early warning system, characterized in that, include: Signal Acquisition and Standardization Module: The raw PPG signal is acquired through the photoplethysmography (PPG) sensor built into the smartwatch, and the acquired raw PPG signal is denoised and standardized using an adaptive filtering algorithm. Heart rate variability feature extraction module: The peak detection algorithm based on wavelet transform is used to identify the heart beat positions contained in the standardized PPG signal. After outlier detection and smoothing, the heart rate variability feature sequence is obtained. Image conversion module: Performs minimum-maximum normalization on the heart rate variability feature sequence, and then uses an improved Gramian angular field transform algorithm to convert the one-dimensional heart rate variability feature sequence into a two-dimensional heart rhythm image; Classification module: The ResNet-50 convolutional neural network architecture is used to train the arrhythmia classification model through supervised learning, so as to realize the automatic identification and classification of different types of arrhythmias; Monitoring module: Through sliding window technology and early warning mechanism, it realizes continuous monitoring of the user's heart rhythm status and timely alarm of abnormal situations; To achieve the arrhythmia early warning method for smartwatches based on deep learning as described in any one of claims 1-6.

Images