A method and apparatus for identifying a cardiac rhythm abnormality

Abstract

The present disclosure relates to a method and device for identifying abnormal heart rhythm. The method comprises obtaining a heart shock wave signal of an identification object; based on each signal segment in the heart shock wave signal, extracting a multi-domain statistical feature of each signal segment and calculating a signal quality index of each signal segment. The method comprises based on each signal segment and its corresponding multi-domain statistical feature, combining the signal quality index corresponding to the signal segment to perform weighted fusion, and outputting a classification label and a prediction probability of each signal segment by a heart rhythm identification model. In addition, the method further comprises determining a heart rhythm identification result of the identification object based on the classification label and the prediction probability of all signal segments. The heart rhythm identification model according to the present disclosure dynamically allocates weights according to the signal quality index, so that the model outputs the heart rhythm classification probability distribution under different signal quality conditions based on the time sequence signal and the multi-domain statistical feature, and realizes high-precision, strong-robust, and interpretable heart rhythm abnormality identification.

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, and more specifically, to a method and apparatus for identifying arrhythmias. Background Technology

[0002] Cardiac rhythm monitoring is one of the core methods of cardiovascular risk assessment. Accurately identifying arrhythmias such as atrial fibrillation (AF) and premature ventricular beats (PB) can provide clinicians with crucial decision support information. Currently, clinical practice mainly relies on electrocardiography (ECG) for cardiac rhythm monitoring. However, ECG technology requires direct contact between electrodes and the skin to collect signals, leading to poor patient compliance during long-term use and frequent problems such as skin irritation and electrode detachment.

[0003] In some related technologies, a non-contact sensor placed under the mattress is used to collect the user's ballistocardiogram (BCG) signal. The RRI (RR interval) sequence is constructed based on the J wave or peak localization of the BCG signal, thereby enabling heart rhythm analysis. However, in real-world usage environments, BCG signals are highly susceptible to factors such as body movement, respiratory motion, and bed structure. The stability and robustness of peak detection are difficult to guarantee, directly limiting the reliability of downstream detection algorithms. Summary of the Invention

[0004] The embodiments of this disclosure provide a method and apparatus for identifying cardiac arrhythmias.

[0005] In a first aspect of this disclosure, a method for identifying cardiac arrhythmias is provided. The method includes acquiring a cardiac shockwave signal of the identified object. The method further includes extracting multi-domain statistical features for each signal segment of the cardiac shockwave signal and calculating a signal quality index for each signal segment. The multi-domain statistical features include multi-dimensional signal features characterizing the rhythm and force of the heartbeat, and the signal quality index is a continuous quality score characterizing the rhythmic periodicity, frequency domain energy concentration, and waveform stability of each signal segment. The method further includes weighted fusion of each signal segment and its corresponding multi-domain statistical features, combined with the corresponding signal quality index, and outputting a classification label and predicted probability for each signal segment by a cardiac rhythm identification model. Furthermore, the method includes determining the cardiac rhythm identification result of the identified object based on the classification labels and predicted probabilities of all signal segments.

[0006] In a second aspect of this disclosure, a cardiac arrhythmia identification device is provided. The device includes an acquisition module configured to acquire cardiac shockwave signals of the identified object.

[0007] The feature extraction module is configured to extract multi-domain statistical features for each signal segment in the cardiac shockwave signal. These features include multi-dimensional signal features characterizing the rhythm and force of the heartbeat. The index calculation module is configured to calculate the signal quality index for each signal segment in the cardiac shockwave signal. The signal quality index is a continuous quality score characterizing the rhythm periodicity, frequency domain energy concentration, and waveform stability of each signal segment. The heart rhythm recognition module is configured to perform weighted fusion based on each signal segment and its corresponding multi-domain statistical features, combined with the corresponding signal quality index. The heart rhythm recognition model outputs a classification label and predicted probability for each signal segment. Furthermore, the recognition determination module is configured to determine the heart rhythm recognition result for the target based on the classification labels and predicted probabilities of all signal segments.

[0008] It should be understood that the description in the Summary of the Invention section is not intended to limit the key or essential features of the embodiments of this disclosure, nor is it intended to restrict the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0009] The above and other objects, features, and advantages of embodiments of the present disclosure will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of the present disclosure are illustrated in the drawings by way of example and not limitation.

[0010] Figure 1 A schematic diagram of an example environment in which some embodiments of the present disclosure may be implemented is shown;

[0011] Figure 2 A flowchart of a method for identifying cardiac arrhythmias according to some embodiments of this disclosure is shown;

[0012] Figure 3 A schematic diagram illustrating the principle process of a heart rhythm recognition model according to some embodiments of the present disclosure during the training phase is shown;

[0013] Figure 4 A schematic diagram illustrating the process of heart rhythm recognition using a heart rhythm recognition model according to some embodiments of this disclosure is shown;

[0014] Figure 5 A schematic diagram illustrating the signal segmentation processing procedure of some embodiments of this disclosure is shown;

[0015] Figure 6 A schematic diagram of the heart rhythm recognition process according to some embodiments of this disclosure is shown;

[0016] Figure 7 A schematic diagram illustrating the cardiac rhythm risk score calculation process of some embodiments of this disclosure is shown;

[0017] Figure 8 An example block diagram of a cardiac arrhythmia identification device according to some embodiments of the present disclosure is shown.

[0018] In the various figures, the same or corresponding reference numerals indicate the same or corresponding parts. Detailed Implementation

[0019] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0020] In the description of embodiments of this disclosure, the term "comprising" and similar terms should be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" should be understood as "at least partially based on". The term "one embodiment" or "the embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc., may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0021] As mentioned earlier, clinical practice primarily relies on ECG technology for heart rhythm monitoring. However, this technology requires direct contact between electrodes and the skin to collect ECG signals, leading to poor patient compliance during long-term use. Issues such as skin irritation and electrode detachment frequently occur, limiting its widespread application in home settings and long-term continuous monitoring.

[0022] In some related technologies, BCG signals, which are the subtle mechanical vibration signals throughout the body caused by the heart's pumping of blood, are acquired using non-contact sensors. The J wave, representing cardiac contraction, is extracted from the BCG signal, and the time difference between two adjacent J waves is calculated to construct an RR interval sequence. BCG signals are highly susceptible to respiratory waves and subtle body movements, causing the morphology of the J wave to frequently change drastically or even disappear. If the J wave is not accurately located or not found at all, the accuracy of downstream detection algorithms will decrease, or even fail.

[0023] To address this, according to embodiments of this disclosure, a cardiac rhythm abnormality identification scheme is provided. This scheme utilizes a cardiac rhythm identification model to identify the cardiac rhythm in each signal segment of a cardiac shockwave signal, and determines the cardiac rhythm identification result based on all signal segments. During the identification process of each signal segment, the cardiac rhythm identification model dynamically allocates weights according to a signal quality index, enabling the model to output a probability distribution of cardiac rhythm classification under different signal quality conditions based on time-series signals and multi-domain statistical features.

[0024] This method divides continuous signals into segments. Using these segments as units, the heart rhythm recognition model dynamically and adaptively makes decisions based on the signal quality index of each segment, outputting a relatively accurate heart rhythm classification probability distribution. Finally, the heart rate classification probability distributions of all signal segments are aggregated, and the continuity of time is used to filter transient interference, outputting a comprehensive heart rhythm recognition result. This method does not rely on precise localization of every heartbeat cycle, especially J-wave localization; even if there are small signal quality deviations within a segment, the characteristics of the majority of effective signals are sufficient to support statistical regularities, allowing the model to make the correct classification result based on these characteristics. Furthermore, the statistical characteristics of this method are often relative, which gives the algorithm stronger generalization ability for users with different physical conditions and different sensor coupling states.

[0025] In the identification process of each signal segment, heart rhythm classification is performed based on features in both statistical and temporal dimensions. The model not only utilizes the smoothing effect of statistical features to combat noise but also learns waveform detail features, enabling it to output heart rhythm classification results with high robustness. Furthermore, a signal quality index is introduced into the model as a confidence assessment. Weights are dynamically allocated according to signal quality, focusing on statistical features for high-quality signals and on temporal signals for low-quality signals, thereby outputting heart rhythm classification results with higher confidence.

[0026] Furthermore, the embodiments disclosed herein support unified detection and probability output of various cardiac arrhythmias such as atrial fibrillation and premature beats. Relying on non-contact BCG acquisition, no electrode attachment is required, making it suitable for long-term continuous monitoring and automatic intelligent early warning in home settings.

[0027] Figure 1 A schematic diagram of an example environment 100 that may be implemented according to some embodiments of the present disclosure is shown. For example... Figure 1 As shown, a non-contact sensor, such as a vibration sensor 104, is installed on the bed 102 in environment 100 (e.g., under the mattress or between the bed board and the mattress). This sensor collects body surface vibration signals generated by the heartbeat of the user lying on the bed 102. The sampling frequency is typically set to 500Hz, and the sampling duration is no less than 8 hours (covering a complete sleep cycle). The vibration sensor 104 sends the collected signals to a computing device 106 for heart rhythm abnormality identification. This computing device can be a mobile terminal, wearable device, computing platform server, etc., with computing capabilities.

[0028] like Figure 1As shown, the computing device 106 includes an acquisition module 108, used to acquire the cardiac shock wave signal 108-1 of the identified object, and perform data processing on it, such as bandpass filtering, baseline drift correction, motion artifact detection (based on sliding window root mean square, RMS), and outlier removal, and to divide the continuous signal into multiple fixed-length segments. Heart rhythm identification is performed on each signal segment (i.e., signal segment i indicated by 108-2 in the figure, where i can be a natural number).

[0029] The computing device 106 also includes a feature extraction module 110, used to extract multi-domain statistical features for each signal segment (i.e., multi-domain statistical feature i indicated by label 110-1 in the figure, where i can be a natural number). Multi-domain statistical features include multi-dimensional signal features characterizing the rhythm and force of the heartbeat. In some implementations, based on each signal segment, at least frequency domain features, nonlinear features, autocorrelation features, and wavelet multi-scale features are extracted, and multi-domain statistical features are constructed by concatenating the extracted features across each dimension. The frequency domain features can be any one or more of the following: total power of the physiological information frequency band, spectral entropy, spectral flatness, high-frequency residual ratio, peak energy concentration, peak power proportion, spectral centroid, and spectral flux standard deviation. The nonlinear features can be any one or more of the following: sample entropy, maximum Lyapunov exponent, and Lempel-Ziv complexity. The autocorrelation features can be any one or more of the following: ACF periodicity intensity and AMDF depth. The wavelet multi-scale features can be any one or more of the following: core total energy, wavelet entropy, relative energy of each layer from D1 to D5, kurtosis of the heartbeat frequency band, skewness, and standard deviation. Furthermore, based on each signal segment, morphological statistical features and Hjorth parameters can be extracted, and the extracted features can be concatenated with the aforementioned features to form multi-domain statistical features. The morphological statistical features can be one or more of the following: peak factor, kurtosis, and several dimensions. The Hjorth parameter can be one or more of the following: activity, mobility, complexity, etc.

[0030] The computing device 106 also includes an index calculation module 112, used to calculate the signal quality index 1 (i.e., the signal quality index i indicated by 112-1 in the figure, where i can be a natural number) for each signal segment in the cardiac impulse wave signal. This signal quality index is a continuous quality score characterizing the rhythmic periodicity, frequency domain energy concentration, and waveform stability of each signal segment. The continuous quality score SQI ∈ [0,1]. In some implementations, the rhythmic periodicity score, frequency domain energy concentration score, and waveform stability score are calculated separately for each signal segment. The scores are normalized and then weighted to obtain the continuous quality score. This rhythmic periodicity score is used to characterize the similarity and rhythmic regularity between adjacent heartbeat waveforms. In some examples, the normalized autocorrelation function of each signal segment is calculated, and the maximum sidelobe peak value (excluding the origin) is extracted within a specific time delay window corresponding to a normal physiological heartbeat cycle. This peak value is then mapped to the interval using a nonlinear monotonically increasing function (such as the Sigmoid function) to obtain the rhythmic periodicity score. Frequency domain energy concentration scoring can effectively quantify and identify non-physiological high-frequency spikes or low-frequency baseline drift. In some examples, frequency domain transformation is performed on each signal segment to calculate the proportion of energy within the core physiological information's main frequency band to the total energy of the entire frequency band. This is then mapped using a nonlinear monotonically increasing function to obtain the frequency domain energy concentration score. In other examples, based on each signal segment, amplitude jumps between adjacent sampling points are detected, such as calculating the first-order difference variance. An analytical envelope is extracted for each signal segment, and the standard deviation or coefficient of variation of the envelope sequence is calculated. The micro-zero-crossing rate is also statistically analyzed for each signal segment, and the segmented energy fluctuations are calculated for each signal segment (e.g., energy ratio or entropy). Finally, the above four indicators are normalized and weighted to obtain the waveform stability score.

[0031] like Figure 1 As shown, the computing device 106 also includes a heart rhythm recognition module 114. This module is configured with a heart rhythm recognition model, which can perform weighted fusion based on each signal segment and its corresponding multi-domain statistical features, combined with the signal quality index corresponding to the signal segment, to output a classification label and predicted probability for each signal segment (i.e., the classification label i and predicted probability i indicated by mark 114-1 in the figure, where i can be a natural number). In some implementations, the heart rhythm recognition model dynamically assigns different weights to each signal segment (i.e., the time-series signal mentioned above) and multi-domain statistical features according to the signal quality index of each signal segment, so that the heart rhythm recognition model outputs a classification label and predicted probability based on the signal segment and its corresponding multi-domain statistical features. The classification label may include a normal heart rate label, an atrial fibrillation label, a premature beat label, etc.

[0032] Furthermore, the computing device 106 also includes an identification and determination module 116, used to determine the heart rhythm identification result 116-1 of the identified object based on the classification labels and predicted probabilities of all signal segments. In some implementations, the predicted probabilities of all valid signal segments for a given day (24 hours) are statistically aggregated according to classification, and the time proportion of each abnormal category segment to all valid segments is calculated. Then, a comprehensive heart rhythm abnormality risk score is calculated based on the aggregated probability and time proportion of each abnormal category. Finally, the heart rhythm identification result is displayed to the monitoring party or the assessment object through the visual interface of the computing device.

[0033] Through the above method, the heart rhythm recognition model of this disclosure dynamically allocates weights according to the signal quality index, so that the model outputs the probability distribution of heart rhythm classification under different signal quality conditions based on time-series signals and multi-domain statistical features, thereby achieving high-precision, robust, and interpretable heart rhythm abnormality recognition.

[0034] Figure 2 A flowchart of a cardiac arrhythmia identification method 200 according to some embodiments of this disclosure is shown. Figure 2As shown in block 202, method 200 can acquire the cardiac shockwave signal of the identified object. In some embodiments, the cardiac shockwave signal is acquired using a non-contact sensor. In block 204, method 200 can extract multi-domain statistical features for each signal segment in the cardiac shockwave signal and calculate the signal quality index for each signal segment. The multi-domain statistical features include multi-dimensional signal features characterizing the rhythm and force of the heartbeat, and the signal quality index is a continuous quality score characterizing the rhythmic periodicity, frequency domain energy concentration, and waveform stability of each signal segment. In some embodiments, the acquired cardiac shockwave signal is preprocessed to obtain a continuous and effective signal. Based on the processed cardiac shockwave signal, it is segmented into multiple equal-length signal segments according to a fixed length. For example, with a window length of 20 seconds, a 2-minute cardiac shockwave signal is segmented according to the window length to obtain six 20-second signal segments. In some embodiments, based on each signal segment, at least frequency domain features, nonlinear features, autocorrelation features, and wavelet multi-scale features are extracted, and multi-domain statistical features are constructed by splicing the extracted features of each dimension. For each signal segment, a continuous quality score is calculated for its rhythm periodicity, frequency domain energy concentration, and waveform stability. The multi-domain statistical characteristics and signal quality indices of each signal segment correspond one-to-one. In some implementations, before calculating the continuous quality score for each signal segment, the signal quality score of a signal segment that meets any rejection criterion is assigned a value of 0. After discarding signal segments with a signal quality score of 0, continuous quality scores for rhythm periodicity, frequency domain energy concentration, and waveform stability are calculated for the remaining signal segments. The rejection criterion can be a peak count exceeding the limit. In one example, a peak detection strategy based on an adaptive height threshold for Hilbert analytic signals is used. If the number of valid peaks does not meet the heart rate range, the peak count is determined to be exceeding the limit, and the signal segment is considered to have no pulse or noise bursts, with the signal quality score set to 0. Another rejection criterion is excessively long beginning and end edges. In one example, if the edge before the first peak or after the last peak of a signal segment exceeds a preset time threshold, it indicates a segment splicing error or loss of the signal start segment, and is thus identified as a beginning and end edge process, with the signal quality score set to 0. Another rejection condition is that the spacing between adjacent peaks is too large. In one example, if the interval between any two adjacent peaks in a signal segment exceeds a preset time threshold, indicating severe motion artifacts or signal interruption, then it is determined that the spacing between adjacent peaks is too large, and the signal quality score is set to 0.

[0035] In box 206, method 200 can perform weighted fusion based on each signal segment and its corresponding multi-domain statistical features, combined with the signal quality index corresponding to the signal segment, and output the classification label and predicted probability of each signal segment by the heart rhythm recognition model. In some implementations, the heart rhythm recognition model includes a time-series signal branch module, a statistical feature branch module, a normalization module, a feature fusion module, and a classification module. Based on each signal segment, the time-series signal branch module extracts a deep time-series representation. Based on the multi-domain statistical features corresponding to each signal segment, the statistical feature branch module extracts a statistical embedding vector. Based on the deep time-series representation and the statistical embedding vector, the normalization module outputs the normalized deep time-series representation and the normalized statistical embedding vector. Based on the normalized deep time-series representation and the normalized statistical embedding vector, combined with the signal quality index corresponding to the signal segment, the feature fusion module outputs a weighted concatenated fusion vector. Based on the fusion vector, the classification module outputs the classification label and predicted probability of each signal segment.

[0036] In box 208, method 200 can determine the heart rhythm recognition result of the identified object based on the classification labels and predicted probabilities of all signal segments. In some implementations, the predicted probabilities of all valid signal segments on a given day (24 hours) are statistically aggregated according to classification to calculate an average probability value.

[0037] In this way, embodiments of this disclosure extract multi-dimensional physiological information from signal fragments to construct a multi-domain statistical feature input model, providing the model with richer pathological discrimination criteria. Simultaneously, by combining the temporal signal characteristics of the signal fragments themselves, the model can consider both global features and local dynamic details to output more accurate classification results. Furthermore, a signal quality index is introduced for weighted processing during the model fusion stage. The model can perceive the signal quality of the current signal fragment and dynamically adjust the weights of the temporal and statistical dimensions, automatically reducing the decision weights of low-quality fragments and focusing on high-quality data to output more robust classification results. In addition, focusing on the small-granularity identification process of signal fragments not only reduces computational complexity but also captures local transient lesions (such as the transient characteristics of premature beats and atrial fibrillation), improving identification accuracy. Finally, combining the model output results of all signal fragments for heart rhythm analysis not only identifies the distribution pattern of abnormal events on the time axis but also combines global information from long-term series to smooth out local random noise, significantly reducing the false positive rate and resulting in more accurate output results.

[0038] Figure 3 A schematic diagram of the principle process 300 of a heart rhythm recognition model according to some embodiments of this disclosure during the training phase is shown. For example... Figure 3As shown, the process 300 can acquire training samples 302, which includes multiple sample signal segments 302-1 of the cardiac shockwave signals of the sample objects. Each sample signal segment corresponds to a multi-domain statistical feature 302-2, a sample signal quality index 302-3, and a sample classification label 302-4. In some implementations, static pressure and physiological vibration of the human body are collected by a non-contact sensor located at the bottom of the mattress to obtain cardiac shockwave signals of multiple sample objects. Correspondingly, electrocardiogram (ECG) signals are also collected by electrode pads attached to the chest of the human body, which are then manually interpreted by a specialist physician based on the ECG. The interpretation results are used as the gold standard label (i.e., the sample classification label). The sample classification labels are matched one-to-one with the sample signal segments of the corresponding time period. In some implementations, the cardiac shockwave signals of the sample objects are preprocessed to obtain continuous and effective signals. Based on the processed cardiac shockwave signals, they are segmented into multiple equal-length sample signal segments according to a fixed length. For each sample signal segment, at least frequency domain features, nonlinear features, autocorrelation features, and wavelet multi-scale features are extracted. Based on the extracted features of each dimension, multi-domain statistical features are constructed. For each sample signal segment, a continuous quality score is calculated for that sample signal segment regarding rhythmic periodicity, frequency domain energy concentration, and waveform stability.

[0039] like Figure 3As shown, process 300 can perform supervised training on heart rhythm recognition model 310 based on training samples 302. Process 300 includes extracting a deep temporal representation 312-1 for each sample signal segment by a temporal signal branch module; and extracting a statistical embedding vector 314-1 for each sample signal segment by a statistical feature branch module based on the multi-domain statistical features corresponding to each sample signal segment. Based on the deep temporal representation 312-1 and the statistical embedding vector 314-1, normalization module 316 outputs a normalized deep temporal representation 316-1 and a normalized statistical embedding vector 316-2. In some implementations, normalization module 316 employs a domain-adaptive normalization module. Based on the deep temporal representation and the statistical embedding vector, the normalization module calculates the batch normalization term of the deep temporal representation (i.e., the normalized deep temporal representation) and the batch normalization term of the statistical embedding vector (i.e., the normalized statistical embedding vector). The process 300 further includes, based on the normalized deep temporal representation of the sample 316-1 and the normalized statistical embedding vector of the sample 316-2, combined with the sample signal quality index 320-3 corresponding to the sample signal segment, the feature fusion module 318 outputs a weighted concatenated sample fusion vector 318-1. In some implementations, the feature fusion module 318 includes a multilayer perceptron and a fusion unit. Based on the sample signal quality index 320-3 corresponding to the sample signal segment, the multilayer perceptron outputs a first sample weight and a second sample weight. Based on the normalized deep temporal representation of the sample 316-1 and the normalized statistical embedding vector of the sample 316-2, the first sample weight and the second sample weight, the fusion unit outputs a weighted concatenated sample fusion vector, i.e., first sample weight * normalized deep temporal representation of the sample + second sample weight * normalized statistical embedding vector of the sample. The process 300 also includes, based on the sample fusion vector 318-1, the classification module 320 outputs the predicted classification label and predicted probability 320-1 for each sample signal segment. The heart rhythm recognition model 310 is trained using the sample classification labels 302-4 of each sample signal segment as supervision signals. Based on the sample classification labels and the predicted classification labels, the loss module 322 calculates the loss value. The loss calculation can use the cross-entropy loss function. If the calculated loss value is larger, it is backpropagated to multiple models to guide the optimization of the adjustment parameters of each module within the model. This training process can be executed iteratively until the model can generate more accurate classification labels.

[0040] In some implementations, training samples are divided into high-quality training subsets and low-quality training subsets based on the numerical relationship between the sample signal quality index and a first signal quality threshold. This training process can be performed on the heart rhythm recognition model based on these subsets. In one example, sample signal segments with a sample signal quality index greater than or equal to the first signal quality threshold, along with their multi-domain statistical features, sample signal quality index, and sample classification labels, are assigned to the high-quality training subset. Conversely, sample signal segments with a sample signal quality index less than the first signal quality threshold, along with their multi-domain statistical features, sample signal quality index, and sample classification labels, are assigned to the low-quality training subset. The first signal quality threshold is set between 0.5 and 0.7. Correspondingly, the loss function is a weighted classification loss function based on the signal quality index, as shown below:

[0041]

[0042] Where CE is the cross-entropy loss, γ is the quality weight index (hyperparameter), which applies a larger gradient to high-quality samples and a smaller parameter update magnitude to low-quality samples, SQI is the sample signal quality index, and y is the sample classification label. To predict classification labels, a label smoothing strategy (configuring smoothing coefficients for labels) is used for low-quality samples to further reduce the risk of overfitting, thereby mitigating the contribution of uncertain labels to the gradient.

[0043] To further improve the accuracy of model training, this embodiment also introduces a misjudgment penalty mechanism. The improved loss function is as follows:

[0044] Where, λ FP •FP Penalty is a penalty for false positives, λ FP To predetermine a penalty coefficient, a penalty is imposed on cases that are correctly identified as AF. reg Entropy Regularization is entropy regularization, γ reg This is the regularization coefficient, used to prevent the model from overconfident in its output.

[0045] Furthermore, the training process includes removing sample signal segments below an initial signal quality threshold before constructing the training set, and then using the remaining sample signal segments and their corresponding multi-domain statistical features, signal quality indices, and classification labels to form the training set. In one example, the initial signal quality threshold is lower than a first signal quality threshold to remove low-quality sample signal segments before training, ensuring that the model primarily learns from samples with acceptable signal quality during the training phase, thus avoiding contamination of model weights by low-quality noisy samples.

[0046] In this way, the sample labels determined by the ECG acquisition signal are used as supervision signals, and the model is trained based on the ECG shockwave BCG signal, resulting in a heart rhythm recognition model with higher accuracy and more in line with clinical judgment.

[0047] Figure 4 A schematic diagram of a process 400 for heart rhythm recognition using a heart rhythm recognition model, according to some embodiments of this disclosure, is shown. Figure 4 As shown, process 400 can extract deep temporal representations 412-1 based on each signal segment 402 by the temporal signal branch module 412. In some implementations, the temporal signal branch module 412 includes an entry layer composed of Stem convolution, batch normalization, and ReLU activation functions, a backbone layer (such as stacked ResBlock1D layers) for capturing multi-scale temporal dependencies, and an output layer for global feature embedding. Process 400 can extract statistical embedding vectors 414-1 based on the multi-domain statistical features 404 corresponding to each signal segment 402 by the statistical feature branch module 414. In some implementations, the statistical feature branch module 414 is composed of an MLP and activation functions, adaptively learning a soft mask for the input features to suppress noise or redundant descriptors; the masked vector is normalized by the activation function and then multiplied element-wise with the original feature vector to achieve adaptive filtering and noise reduction at the feature level; the masked vector is linearly projected to a fixed dimension by ReLU to obtain the statistical embedding vector.

[0048] The process 400 can also output normalized deep time series representation 416-1 and normalized statistical embedding vector 416-2 from the normalization module 416 based on deep time series representation and statistical embedding vector. In some implementations, the normalization module 416 employs a domain-adaptive normalization module. This module calculates the batch normalization term and sample-level normalization term of the deep time series representation, as well as the batch normalization term and sample-level normalization term of the statistical embedding vector. Based on the batch normalization term and sample-level normalization term of the deep time series representation, combined with the learning parameters of the heart rhythm recognition model, the normalized deep time series representation is output. And based on the batch normalization term and sample-level normalization term of the statistical embedding vector, combined with the learning parameters of the heart rhythm recognition model, the normalized statistical embedding vector is output.

[0049]

[0050] Where BN(x) is the batch normalization term of the deep temporal representation and Norm_sample(x) is the sample-level normalization term of the deep temporal representation, y is the normalized deep temporal representation. When BN(x) is the batch normalization term of the statistical embedding vector and Norm_sample(x) is the sample-level normalization term of the statistical embedding vector, y is the normalized statistical embedding vector. The learning parameter α is updated along with the network weights during backpropagation, enabling the model to automatically learn the optimal normalization strategy. σ(α) maps α to the (0,1) interval. In this way, combining batch normalization and sample-level normalization during the inference phase enables the model to have the ability to generalize across devices. Even if the collected BCG signals come from different hardware and the application environment is different (such as different mattress thicknesses or even different electromagnetic environments at home), the absolute amplitude and baseline of the signal will not deviate significantly from the hospital's benchmark, and normal human data will not cause medical ethics issues such as misdiagnosis due to different hardware gains.

[0051] like Figure 4 As shown, process 400 can be based on the normalized deep temporal representation 416-1 and the normalized statistical embedding vector 416-2, combined with the signal quality index 406 corresponding to the signal segment, and the feature fusion module 418 outputs a weighted concatenated fusion vector 418-1. In some implementations, based on the signal quality index corresponding to the signal segment, the multilayer perceptron outputs a first weight and a second weight. Based on the normalized deep temporal representation, the normalized statistical embedding vector, the first weight, and the second weight, the fusion unit outputs a weighted concatenated fusion vector. This embodiment is based on the Signal Quality Index (SQI). A lightweight MLP maps the scalar SQI to two non-negative weights (w_cnn, w_stat). After Softmax normalization, the contribution ratio of the two branches is dynamically adjusted: when the signal is of high quality (SQI→1), the waveform branch is gated to make full use of the temporal morphological information; when the signal contains artifacts (SQI→0), the weight of the more robust statistical branch is increased to achieve adaptive signal quality perception fusion. The final fusion representation is z=[h_cnn·w_cnn;s·w_stat], where h_cnn is the normalized deep temporal representation output by the temporal signal branch module and the normalization module in sequence, and s is the normalized statistical embedding vector output by the statistical feature branch module and the normalization module in sequence.

[0052] The process 400 can also be based on the fusion vector 418-1, with the classification module 420 outputting a classification label and predicted probability 420-1 for each signal segment. In some implementations, the classification label can be a normal classification or an abnormal classification. The abnormal classification can be further subdivided into abnormal categories such as atrial fibrillation and premature beats.

[0053] Furthermore, during the system deployment phase, when new unlabeled data is collected, it is not necessary to follow the steps outlined above. Figure 3 The example undergoes full training with only minor adjustments. In some implementations, freezing the model backbone weights and updating the BN statistics of each layer using a small amount of newly collected unlabeled data allows for rapid, lightweight adaptation to device and individual differences.

[0054] Figure 5 A schematic diagram of a signal segmentation processing procedure 500 according to some embodiments of this disclosure is shown. For example... Figure 5 As shown, in this process 500, the BCG signal can be acquired in box 502, and then saturation detection is performed on the raw BCG data. In box 504, it is determined whether the signal amplitude is greater than 95% of the dynamic range. If so, it is determined to be an invalid saturation segment, and the invalid saturation segment is removed in box 506; otherwise, RMS detection is performed in box 508. An adaptive sliding window root mean square (RMS) detection based on short-time energy distribution is used. In box 510, it is determined whether the RMS is greater than the energy threshold. If so, it is determined to be a non-steady-state data segment, and the non-steady-state data segment affected by severe body motion artifacts is removed in box 512; otherwise, it proceeds to box 514 for filtering and denoising. For example, a high-order IIR bandpass filter is used, which effectively covers the main physiological information frequency band of the BCG signal, and a soft threshold denoising method is used to suppress broadband random noise. In this process 500, baseline drift correction can also be performed in box 516. For example, low-frequency baseline drift caused by respiratory motion and sensor offset can be removed by polynomial fitting or high-pass filtering to ensure that the signal maintains a stable zero-mean characteristic during long-term monitoring. This process 500 can also be further refined in box 518 by resampling. Continuous effective signal segments are downsampled to the processed sampling rate, reducing computational redundancy while fully matching the effective energy bandwidth of the BCG signal, providing a unified input format for subsequent feature extraction and deep learning model training. Then, in box 520, the processed signal is segmented. Considering the temporal resolution and computational efficiency of the heart rhythm features, the window length is determined, which determines the length of each signal segment. A preset number of sampling points is set within each window, for example, 1000 sampling points within a 2-second window. To increase sample diversity, a certain proportion of overlap can be set between adjacent windows. For example, with a window length of 20 seconds and an overlap of 10 seconds, a 1-minute BCG signal can be divided into 6 signal segments: signal segment 1 (0~20s), signal segment 2 (10~30s), signal segment 3 (20~40s), signal segment 4 (30~50s), and signal segment 5 (40s~60s).

[0055] Figure 6 A schematic diagram of a heart rhythm recognition process 600 according to some embodiments of this disclosure is shown. Figure 6As shown, in box 602, process 600 acquires continuous BCG signals from a non-contact sensor, preprocesses the BCG signals according to the processing method mentioned above, and segments them into fixed-length segments (see box 604). The signal quality index (SQI) is calculated for each signal segment (see box 606). In box 608, process 600 determines whether the SQI is 0. For example, if the hard rejection condition is met, the SQI is 0, and the signal segment needs to be discarded (see box 610). If it is not 0, the process proceeds to box 612 to determine whether the SQI is greater than or equal to a first signal threshold for high-quality signal segment selection. If not, low-quality signal segments are discarded (see box 614). If yes, the process proceeds to box 616 to calculate multi-domain statistical features, and then to box 618 for heart rhythm recognition model identification.

[0056] Figure 7 A schematic diagram of a cardiac rhythm risk score calculation process 700 according to some embodiments of this disclosure is shown. For example... Figure 7 As shown, this process combines the time-series signal x_sig and multi-domain statistical features x_stat of high-quality signal segments with the signal quality index and inputs them into the heart rhythm recognition model, outputting the predicted probability and classification label for different signal segments. Based on the classification label of the signal segment, statistical classification is performed, as shown in the figure (sets 702-1, 702-2, ..., 702-m), each set representing a category. Then, the daily average probability value and the time proportion of signal segments belonging to that category among all signal segments are calculated for each set (sets 704-1, 704-2, ..., 704-m). This process 700 can also determine the heart rhythm score 706-1, 706-2, ..., 706-m under each classification label based on the daily average probability value and time proportion. Finally, this process 700 can calculate a comprehensive risk score 708 for abnormal classifications. In some implementations, a comprehensive heart rhythm abnormality risk score is calculated by combining the daily average probability and time proportion of each category.

[0057]

[0058] in and The weighting coefficients for each category can be adjusted according to clinical significance, with higher weights assigned to high-risk categories such as atrial fibrillation. This represents the daily average probability of each category. This represents the proportion of time spent on segments of the abnormal category (AF, PB) out of all valid segments.

[0059] Furthermore, this process 700 can also issue early warnings based on a comprehensive risk score. The risk level is categorized according to the comprehensive risk score: Low risk (Risk < θ_low): The system is normal and requires no intervention; Medium risk (θ_low ≤ Risk < θ_high): An alert is triggered, advising the user to pay attention to and record symptoms; High risk (Risk ≥ θ_high): An early warning is triggered, advising the user to seek medical attention promptly and sending a notification to a guardian or medical institution. The thresholds θ_low and θ_high can be configured by the system administrator according to the application scenario.

[0060] Figure 8 Example block diagrams of a cardiac arrhythmia recognition device 800 according to some embodiments of the present disclosure are shown. Figure 8 As shown, the device 800 includes an acquisition module 802 configured to acquire the cardiac shockwave signal of the object to be identified. The device 800 includes a feature extraction module 804 configured to extract multi-domain statistical features for each signal segment in the cardiac shockwave signal, including multi-dimensional signal features characterizing the rhythm and force of the heartbeat. The device 800 includes an index calculation module 806 configured to calculate a signal quality index for each signal segment in the cardiac shockwave signal, whereby the signal quality index is a continuous quality score characterizing the rhythmic periodicity, frequency domain energy concentration, and waveform stability of each signal segment. The device 800 includes a heart rhythm recognition module 808 configured to perform weighted fusion based on each signal segment and its corresponding multi-domain statistical features, combined with the corresponding signal quality index, and output a classification label and predicted probability for each signal segment by the heart rhythm recognition model. Furthermore, the device 800 includes a recognition determination module 110 configured to determine the heart rhythm recognition result of the object to be identified based on the classification labels and predicted probabilities of all signal segments.

[0061] In some implementations, the heart rhythm recognition module 808 is configured with a heart rhythm recognition model, which includes a statistical feature branch module, a time-series signal branch module, a normalization module, and a feature fusion module. The heart rhythm recognition module is configured to extract a deep time-series representation for each signal segment using the time-series signal branch module; and to extract a statistical embedding vector using the statistical feature branch module based on the multi-domain statistical features corresponding to each signal segment. Based on the deep time-series representation and the statistical embedding vector, the normalization module outputs a normalized deep time-series representation and a normalized statistical embedding vector. Based on the normalized deep time-series representation and the normalized statistical embedding vector, combined with the signal quality index corresponding to the signal segment, the feature fusion module outputs a weighted concatenated fusion vector; and based on the fusion vector, the classification module outputs the classification label and predicted probability for each signal segment.

[0062] In some implementations, the identification and determination module 810 includes a first determination unit and a second determination unit. The first determination unit is configured to determine the daily average probability value under different classification labels and the time proportion of signal segments whose classification labels belong to the abnormal category within the statistical day, based on the classification labels and predicted probabilities of all signal segments within the statistical day. The second determination unit is configured to determine the arrhythmia risk score of the identified object under different classification labels, based on the daily average probability value and time proportion under different classification labels.

[0063] In some implementations, the device 800 further includes a training module configured to acquire training samples, which include multiple sample signal segments of the cardiac shockwave signal of the sample object. Each sample signal segment corresponds to a multi-domain statistical feature, a sample signal quality index, and a sample classification label. Based on each sample signal segment, a time-series signal branch module extracts a deep temporal representation of the sample. Based on the multi-domain statistical feature corresponding to each sample signal segment, a statistical feature branch module extracts a sample statistical embedding vector. Based on the deep temporal representation and the sample statistical embedding vector, a normalization module outputs a normalized deep temporal representation and a normalized sample statistical embedding vector. Based on the normalized deep temporal representation and the normalized sample statistical embedding vector, combined with the sample signal quality index corresponding to the sample signal segment, a feature fusion module outputs a weighted concatenated sample fusion vector. Based on the sample fusion vector, a classification module outputs a predicted classification label and a predicted probability for each sample signal segment. The heart rhythm recognition model is trained using the sample classification label of each sample signal segment as a supervision signal.

[0064] While several specific implementation details are included in the foregoing discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations.

[0065] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

[0066] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for identifying cardiac arrhythmias, characterized in that, include: Acquire the cardiac shockwave signal of the identified object; Based on each signal segment in the cardiac shock wave signal, multi-domain statistical features of each signal segment are extracted and the signal quality index of each signal segment is calculated. The multi-domain statistical features include multi-dimensional signal features that characterize the rhythm and force of the heartbeat. The signal quality index is a continuous quality score that characterizes the rhythm periodicity, frequency domain energy concentration and waveform stability of each signal segment. Based on each signal segment and its corresponding multi-domain statistical features, a weighted fusion is performed using the signal quality index corresponding to that signal segment. The heart rhythm recognition model then outputs the classification label and predicted probability for each signal segment. as well as Based on the classification labels and predicted probabilities of all signal segments, the heart rhythm recognition result of the identified object is determined.

2. The method according to claim 1, characterized in that, The heart rhythm recognition model includes a time-series signal branch module, a statistical feature branch module, a normalization module, a feature fusion module, and a classification module. Based on each signal segment and its corresponding multi-domain statistical features, and combined with the signal quality index corresponding to the signal segment, a weighted fusion is performed. The heart rhythm recognition model outputs the classification label and predicted probability of each signal segment, including: Based on each signal segment, the time-series signal branching module extracts a deep time-series representation; Based on the multi-domain statistical features corresponding to each signal segment, the statistical feature branch module extracts the statistical embedding vector; Based on the deep temporal representation and the statistical embedding vector, the normalization module outputs the normalized deep temporal representation and the normalized statistical embedding vector. Based on the normalized deep temporal representation and the normalized statistical embedding vector, combined with the signal quality index corresponding to the signal segment, the feature fusion module outputs a weighted concatenated fusion vector; and Based on the fusion vector, the classification module outputs the classification label and predicted probability for each signal segment.

3. The method according to claim 2, characterized in that, Based on the deep time series representation and the statistical embedding vector, the normalization module outputs the normalized deep time series representation and the normalized statistical embedding vector, including: Calculate the batch normalization term and sample-level normalization term of the deep temporal representation, as well as the batch normalization term and sample-level normalization term of the statistical embedding vector; Based on the batch normalization term and sample-level normalization term of the deep temporal representation, combined with the learning parameters of the heart rhythm recognition model, the normalized deep temporal representation is output; and Based on the batch normalization term and sample-level normalization term of the statistical embedding vector, and combined with the learning parameters of the heart rhythm recognition model, the normalized statistical embedding vector is output.

4. The method according to claim 2, characterized in that, The feature fusion module includes a multilayer perceptron and a fusion unit. Based on the normalized deep temporal representation and the normalized statistical embedding vector, combined with the signal quality index corresponding to the signal segment, the feature fusion module outputs a weighted concatenated fusion vector, including: Based on the signal quality index corresponding to the signal segment, the multilayer perceptron outputs a first weight and a second weight; and Based on the normalized deep temporal representation, the normalized statistical embedding vector, the first weight, and the second weight, the fusion unit outputs a weighted concatenated fusion vector.

5. The method according to claim 1, characterized in that, The heart rhythm recognition model includes a time-series signal branch module, a statistical feature branch module, a normalization module, a feature fusion module, and a classification module. The method also includes: Acquire training samples, which include multiple sample signal segments of the cardiac shock wave signal of the sample object, and each sample signal segment corresponds to sample multi-domain statistical features, sample signal quality index and sample classification label; Based on each sample signal segment, the time-series signal branching module extracts the deep time-series representation of the sample. Based on the multi-domain statistical features of each sample signal segment, the statistical feature branch module extracts the sample statistical embedding vector. Based on the deep temporal representation of the sample and the statistical embedding vector of the sample, the normalization module outputs the normalized deep temporal representation of the sample and the normalized statistical embedding vector of the sample. Based on the normalized deep temporal representation of the sample and the normalized statistical embedding vector of the sample, combined with the sample signal quality index corresponding to the sample signal segment, the feature fusion module outputs a weighted and concatenated sample fusion vector. Based on the sample fusion vector, the classification module outputs the predicted classification label and predicted probability for each sample signal segment; and The heart rhythm recognition model is trained using the sample classification label of each sample signal segment as a supervision signal.

6. The method according to claim 5, characterized in that, Also includes: Based on the numerical relationship between the sample signal quality index and the first signal quality threshold, the training samples are divided into a high-quality training subset and a low-quality training subset. Based on the sample classification labels and the predicted classification labels, a loss value is calculated using a weighted classification loss function based on the sample signal quality index; and Based on the loss value, the parameters of the heart rhythm recognition model are adjusted.

7. The method according to claim 5, characterized in that, Based on the deep temporal representation of the samples and the statistical embedding vector of the samples, the normalization module outputs the normalized deep temporal representation of the samples and the normalized statistical embedding vector of the samples, including: Based on the deep temporal representation of the sample and the statistical embedding vector of the sample, the normalization module calculates the batch normalization term of the deep temporal representation and the batch normalization term of the statistical embedding vector.

8. The method according to claim 1, characterized in that, The acquisition process for each signal segment includes: Based on the acquired cardiac shockwave signal, preprocessing is performed to obtain a continuous and effective signal; and According to the preset window length and the overlap ratio of adjacent windows, the processed cardiac shock wave signal is divided into multiple signal segments of equal length.

9. The method according to claim 1, characterized in that, Based on the classification labels and predicted probabilities of all signal segments, the heart rhythm recognition result of the identified object is determined as follows: Based on the classification labels and predicted probabilities of all signal segments within a statistical day, determine the daily average probability value under different classification labels and the time proportion of signal segments whose classification labels belong to the abnormal category within the statistical day; and Based on the daily average probability value and time proportion under different classification labels, the risk score of cardiac arrhythmia of the identified object under different classification labels is determined.

10. A cardiac arrhythmia identification device, characterized in that, include: The acquisition module is configured to acquire the cardiac shockwave signal of the identified object; The feature extraction module is configured to extract multi-domain statistical features for each signal segment in the cardiac shock wave signal, the multi-domain statistical features including multi-dimensional signal features characterizing the rhythm and force of the heartbeat. The index calculation module is configured to calculate the signal quality index of each signal segment based on each signal segment in the cardiac shock wave signal. The signal quality index is a continuous quality score characterizing the rhythmic periodicity, frequency domain energy concentration, and waveform stability of each signal segment. The heart rhythm recognition module is configured to perform weighted fusion based on each signal segment and its corresponding multi-domain statistical features, combined with the signal quality index corresponding to the signal segment, and output the classification label and predicted probability of each signal segment by the heart rhythm recognition model; as well as The identification and determination module is configured to determine the heart rhythm identification result of the identified object based on the classification labels and predicted probabilities of all signal segments.

Images