Method and system for atrial fibrillation warning based on multi-time scale trend energy modeling

Abstract

The application discloses an atrial fibrillation early warning method and system based on multi-time scale trend energy modeling, and the method comprises the following steps: collecting an ECG signal and performing pretreatment; performing sliding segmentation on the ECG signal at different time scales after pretreatment, and calculating a trend energy function value; constructing a risk probability evaluation model and a risk prediction model; outputting a risk prediction probability and an advance prediction interval; combining the risk prediction probability and the trend energy function value to calculate a risk score; and when the risk integral continuously exceeds a preset threshold and the advance prediction interval is less than a preset early warning time, outputting an AF early warning prompt. The system comprises a signal acquisition unit, a trend energy modeling unit, a risk prediction model training unit, a risk integral calculation unit and an early warning unit. By using the application, the accuracy and stability of AF early warning can be improved. The application can be widely applied to the field of health monitoring.

Description

Technical Field

[0001] This invention relates to the field of health monitoring, and in particular to a method and system for early warning of atrial fibrillation based on multi-timescale trend energy modeling. Background Technology

[0002] Atrial fibrillation (AF) is one of the most common cardiac arrhythmias, with its prevalence and recurrence rate increasing significantly with age. AF is highly insidious and sudden in its onset, with some patients experiencing no obvious symptoms before an attack, making early identification and warning difficult. Therefore, effectively capturing potential warning signs before an AF attack has become an important research direction in electrocardiogram signal analysis and intelligent health monitoring in recent years.

[0003] Traditional atrial fibrillation (AF) detection primarily relies on ECG signal feature analysis. For example, features such as RR interval variability, P wave disappearance or morphological abnormalities, and irregular heart rhythm, combined with traditional machine learning algorithms like support vector machines, decision trees, or random forests, can be used to identify AF episodes. The advantage of these methods is their high interpretability; however, most of these methods focus on classification within the "epidemic state," lacking sufficient ability to "predict the pre-epidemic state" and sensitivity to early prodromal stages. Summary of the Invention

[0004] In view of this, in order to address the technical problem that existing atrial fibrillation early warning methods lack comprehensive multi-timescale modeling and fail to effectively capture the gradual trend characteristics before an attack, thus resulting in the inability to achieve early warning, this invention proposes an atrial fibrillation early warning method based on multi-timescale trend energy modeling. This method includes the following steps: Single-lead or multi-lead ECG signals from the subjects were collected, and bandpass filtering, baseline drift correction, power line interference suppression, and normalization were performed to obtain high-quality continuous time-series signals. The preprocessed ECG signal is divided into sliding segments according to short window, medium window and long window, and the first difference, second difference and cross-scale consistency index are calculated, and the trend energy function value is calculated. A risk probability assessment model was constructed, which has two output branches, used for AF state classification (normal, aura, onset) and advance regression prediction, respectively. The risk probability assessment model is trained using labeled data; the risk prediction model is trained using the data output by the risk probability assessment model. The real-time input ECG signal is fed into the trained risk prediction model to obtain the risk prediction probability; Calculate the risk score by combining the risk prediction probability and the trend energy function value; When the risk score continuously exceeds the preset threshold and the lead time prediction interval is less than the preset warning time, an AF early warning prompt is output.

[0005] In addition to the above method, the present invention also proposes an atrial fibrillation early warning system based on multi-timescale trend energy modeling. The system includes a signal acquisition unit, a trend energy modeling unit, a risk prediction model training unit, a risk integral calculation unit, and an early warning unit.

[0006] Based on the above scheme, this invention provides an atrial fibrillation (AF) early warning method and system based on multi-timescale trend energy modeling. By introducing a multi-timescale trend energy modeling mechanism, it can characterize the dynamic change trend of ECG signals under different time windows; combined with a dynamic risk discrimination strategy, it can achieve continuous quantitative assessment of AF prodromal states and provide an operable early warning time through advance prediction. This invention can improve the accuracy and stability of early AF warning and provide new technical support for intelligent health monitoring. Attached Figure Description

[0007] Figure 1 This is a flowchart of the steps of an atrial fibrillation early warning method based on multi-timescale trend energy modeling according to the present invention; Figure 2 This is a flowchart of the risk probability assessment model training method of the present invention; Figure 3 This is a schematic diagram of the data flow in the method of the present invention; Figure 4 This is a structural block diagram of an atrial fibrillation early warning system based on multi-timescale trend energy modeling according to the present invention. Detailed Implementation

[0008] Existing studies often use fixed window lengths (e.g., 5 or 30 seconds) for feature extraction, neglecting the dynamic changes in pre-AF (atrial fibrillation) warning signs across different time scales. For example, short windows help capture abrupt changes in heart intervals, while long windows better reflect rhythm complexity and waveform gradation. Single-time-scale analysis struggles to capture this complementary information. Many methods rely on static thresholds or single prediction results for discrimination, leading to false positives or false negatives in early prediction scenarios. In reality, pre-AF risk signals often exhibit a gradually increasing trend; without combining trends and historical accumulation for dynamic discrimination, it's difficult to form a stable early warning. Some studies introduce "warning" labels, but these remain at the classification level, failing to perform regression modeling on the possible advance time of AF. This limits the application of these methods in clinical early warning.

[0009] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0010] It should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0011] It should be understood that the terms "system," "apparatus," "unit," and / or "module" used in this application are a method of distinguishing different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0012] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "a," and / or "the" are not specifically singular and may include the plural. Generally, the terms "comprising" and "including" only indicate the inclusion of expressly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements. An element defined by the phrase "comprising an..." does not exclude the presence of other identical elements in the process, method, product, or apparatus that includes the element.

[0013] In the description of the embodiments of this application, "a plurality of" refers to two or more. The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0014] Furthermore, flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, the steps can be processed in reverse order or simultaneously. Additionally, other operations can be added to these processes, or one or more steps can be removed from them.

[0015] Reference Figure 1 This is a flowchart illustrating an optional example of the atrial fibrillation early warning method based on multi-timescale trend energy modeling proposed in this invention. The method can be applied to computer devices, and the atrial fibrillation early warning method proposed in this embodiment may include, but is not limited to, the following steps: Step S1: Acquire ECG signal; Step S2: Perform multi-timescale trend energy modeling based on ECG signals and generate trend energy function values; Step S3: Process the ECG signal based on the pre-trained risk probability assessment model to generate risk probability training data and lead time prediction interval; Step S4: Train a risk prediction model based on the risk probability training data; Step S5: Output the risk prediction probability according to the risk prediction model, and generate the risk integral by combining the trend energy function value; Step S6: Combine the risk score, lead time prediction range, and preset threshold to trigger an early warning.

[0016] In some feasible embodiments, step S1 specifically includes: S1.1 Acquire ECG signal; S1.2 Perform bandpass filtering on the original ECG signal to remove low-frequency drift and high-frequency noise.

[0017] Specifically, a Butterworth filter of 0.5–45 Hz is used to ensure that key frequency bands related to heart rhythm changes are preserved.

[0018] S1.3. Baseline drift is eliminated for each sample segment using the local mean sliding method; S1.4. Normalize the amplitude of the signal and cut the continuous signal into 5-second non-overlapping segments (corresponding to 1250 points). Calculate the mean of each segment independently. ) and standard deviation ( z-score standardization is performed. , The range of standardized values ​​is limited to [-1, 1] to eliminate the influence of differences in signal strength between individuals.

[0019] S1.5 Sliding window segmentation and label alignment: The sliding window segmentation adopts an overlapping window strategy, with a window length of... Set to 1024 sampling points, window interval to 512 points, overlap rate 50%. Sample generation: in, For the first A signal segment; For a complete ECG sequence, This is the index of the starting time point of the current window.

[0020] For AF episode segments in the labeled dataset, the labels within each time window are aligned to the state with the largest proportion in the window, achieving consistent label mapping. For unlabeled ECG data, after preprocessing and window segmentation, its structured time series segments are retained as unlabeled input samples, and they are not involved in label alignment or any model inference operations during the preprocessing stage.

[0021] In some feasible embodiments, step S2 specifically includes: S2.1 Perform multi-time-scale sliding segmentation processing on the preprocessed ECG signal. The time scales include, but are not limited to: short time window: length of 5 seconds, step size of 1 second; medium time window: length of 30 seconds, step size of 5 seconds; long time window: length of 120 seconds, step size of 10 seconds.

[0022] S2.2 Calculate the following trend indicators for each scale window segment: S2.2.1 Calculate the first-order difference at each scale: , in Representing scale The ECG window sequence below. This indicator reflects the rate of change of the signal between adjacent sampling points. If the ECG waveform fluctuates drastically in a short period of time, the amplitude of the first difference will increase significantly, which can be used to detect signs of rhythm instability before an atrial fibrillation (AF) attack.

[0023] S2.2.2 Calculate the second-order difference: It describes the acceleration of the rate of change, i.e., whether the fluctuation trend is rapidly increasing or decreasing. In the prodromal stage of arrhythmia (AF), an abnormal increase in the second-order difference often indicates that the rhythm disorder is accelerating, and therefore it is an important reference for prodromal detection.

[0024] S2.2.3 Calculate the differential energy: To improve stability, the first and second order differences are converted into energy forms: Energy forms can comprehensively measure the fluctuation range over the entire time window, reducing the impact of single-point anomalies.

[0025] S2.2.4 Calculate cross-scale consistency: Calculate the correlation between the first-order difference sequences of short window and medium window, medium window and long window, and short window and long window respectively to characterize the degree of dynamic coordination at different time scales.

[0026] in This is the inner product. The correlations between the three scale pairs (short-medium, medium-long, and short-long) are calculated and summed: The overall consistency index was obtained. Low cross-scale consistency indicates inconsistent rhythmic behavior across different time scales, which is more likely to indicate atrial prodromal abnormalities.

[0027] S2.3. Combining differential energy and cross-scale consistency, define the trend energy function: in is the weighting coefficient. This function amplifies the precursor trend signal by "enhancing fluctuation energy and suppressing consistency," resulting in a significant increase in the 𝐸 value under abnormal conditions.

[0028] S2.4 Normalize the trend energy function value 𝐸 Map it to the [0,1] interval to obtain the trend energy score. It can be combined with the model's predicted probabilities to improve the reliability and robustness of risk scoring.

[0029] In some feasible embodiments, the risk probability assessment model in step S3 is constructed in the following ways: A ResNet-based 1D residual neural network structure is constructed. The network consists of 34 layers with learnable parameters, mainly including an initial convolutional layer, four residual module stacking stages (Conv2_x to Conv5_x), a global average pooling layer, and an output layer.

[0030] The input is long-window ECG data, which is fused from the first and second-order differences of short-window, medium-window, and long-window ECGs, as well as cross-scale consistency features. The data passes through an initial convolutional layer, which is a 1D convolution with a kernel size of 7, a stride of 2, and 64 channels, and is processed with batch normalization and ReLU activation function. Subsequently, the signal is downsampled through a max-pooling layer (kernel size of 3, stride of 2). Afterward, the signal sequentially enters the following four residual modules:

[0031] The Conv2_x module contains three residual units, each consisting of two 1D convolutional layers with a kernel size of 3 and a stride of 1, and a fixed number of channels of 64. This module does not perform downsampling, and all residual connections are identity mappings. Its expression can be represented as: in, , There are 3 residual units, keeping the time dimension and number of channels unchanged;

[0032] The Conv3_x module contains four residual units. The first unit uses a convolutional layer with a stride of 2 for temporal downsampling and increases the number of channels to 128. The remaining three residual units maintain a stride of 1 and a fixed channel dimension of 128. Due to the change in input and output dimensions, a 1×1 convolution is introduced into the residual connections for channel matching and downsampling. The formula is expressed as follows: in, Use 1×1 convolution for downsampling and channel expansion.

[0033] The Conv4_x module contains 6 residual units. The first unit uses a convolutional layer with a stride of 2 for further downsampling and expands the number of channels to 256; the remaining residual units maintain the same number of channels. All residual connections maintain a consistency mapping or use a 1×1 convolution.

[0034] Using the same downsampling method, the final feature length is T / 4.

[0035] The Conv5_x module contains three residual units. The first unit uses a convolution with a stride of 2 to further reduce the temporal dimension and increases the number of channels to 512. Subsequent residual units maintain the original number of channels. The temporal dimension of the output feature map is compressed to approximately 1 / 32 of the original input length, and the channel dimension is 512.

[0036] The final feature length is .

[0037] In all the residual modules mentioned above, batch normalization (BatchNorm) and ReLU activation functions are sequentially applied after each convolutional layer to enhance the stability of network training and its nonlinear modeling capabilities. The input of each residual unit is added to the output of the main branch through residual connections and then activated by ReLU to prevent gradient vanishing and improve the trainability of model depth.

[0038] The output layer has two branches: Classification branches: used to output the state category of the electrocardiogram signal, including three categories: normal, prodromal, and epileptic. Lead time regression branch: used to output the amount of time (in seconds or minutes) that the predicted lead time of AF may be. This dual-branch design enables the model to capture both category and temporal information simultaneously, improving the comprehensiveness of predictions.

[0039] In some feasible embodiments, the training process of the risk probability assessment model specifically includes: Input: Annotated multi-timescale ECG window segments; corresponding state category labels (normal / prodromal / attack); The corresponding lead time label (the possible lead time of AF).

[0040] Joint loss function: The training objective of the risk probability assessment model is to simultaneously optimize classification performance and lead time prediction performance; therefore, a joint loss function is designed. The classification cross-entropy loss and the mean squared error (MSE) loss from lead time regression are weighted and summed, with the following form: in, Cross-entropy loss is used to optimize AF state classification; Mean squared error loss is used to optimize lead time forecasting; These are weighting coefficients used to balance the contributions of classification and regression tasks.

[0041] Training mechanism: Multi-task learning: Through dual-branch joint optimization, the classification task provides a regularization effect for the regression task, and vice versa; Trend-enhanced input: Optionally, the trend energy score obtained in step S2 is used as an additional input feature and concatenated with the deep features to enhance the sensitivity to aura states; Early stopping strategy: The classification accuracy and regression error are monitored on the validation set, and training is stopped when the performance no longer improves to avoid overfitting.

[0042] The data flow in this specific training process is referenced. Figure 2 .

[0043] In some feasible embodiments, it also includes: For unlabeled ECG data, construct high-confidence pseudo-labels.

[0044] First, calculate the trend energy score. Then, by combining the prediction results of the risk probability assessment model within this window, a dynamic risk score is constructed: in, The AF risk probability is the output of the risk probability assessment model. This represents the normalized trend energy score. The weighting coefficients are used. This score integrates "data-driven probabilistic prediction" with "signal-level trend characteristics," thereby improving sensitivity and robustness to precursor anomalies.

[0045] Continuous trend monitoring: To avoid false alarms caused by single-point fluctuations, a continuous window cumulative score is defined: in This is the length of the sliding window. Continuously exceeding the adaptive threshold And it also satisfies the condition of increasing trend: The sample fragment is then determined to be in a high-risk precursor state.

[0046] High-confidence pseudo-label generation For unlabeled samples that meet the above conditions, their risk probability assessment model prediction results are used as pseudo-labels and assigned confidence weights: in This represents the variance of the risk probability assessment model's predictions under perturbation input. A smaller variance indicates more stable model predictions and higher pseudo-label confidence. Ultimately, the resulting set of high-confidence pseudo-labels can be used together with labeled samples to train the risk prediction model.

[0047] The labeled samples and high-confidence pseudo-labeled samples obtained through trend enhancement and memory retrieval are jointly input into the risk prediction model. Memory retrieval discrimination: A memory bank of typical precursor trends is pre-established. The trend features of unlabeled samples are compared with the templates in the memory bank to calculate the similarity score. High-confidence pseudo-label screening: Only when the dynamic risk score exceeds an adaptive threshold and the memory similarity score is greater than a preset threshold, the risk probability assessment model prediction result of the corresponding sample is marked as a high-confidence pseudo-label and included in the subsequent risk prediction model training.

[0048] In some feasible embodiments, the risk prediction model in step S4 specifically includes: A lightweight 1D risk prediction model based on the MobileNet V2 architecture is constructed. Its input and output are the same as those of the risk assessment model. The input is long-window ECG data, which is a fusion of the first and second differences of short-window, medium-window, and long-window ECGs and cross-scale consistency features. The output is a two-branch model: a classification branch and a lead regression branch. The classification branch uses the softmax activation function and outputs the classification probability of the window belonging to the three categories of "normal / prodrome / AF onset". The lead branch uses the ReLU activation function and outputs the regression prediction value of the remaining time (early amount) from the window to the onset of AF.

[0049] A trend enhancement submodule and a risk discrimination submodule are introduced within the risk prediction model. The trend enhancement submodule is used to fuse trend energy features across multiple time scales to improve sensitivity to precursor signals; its input is the intermediate feature representation obtained from the first two layers of the risk prediction model to the residual unit. Where 𝐵 is the batch size, 𝑇 is the time step, d is the number of channels; and the trend energy score is calculated from step S2. The trend features and convolutional features are weighted and fused using a gating mechanism. in For the Sigmoid function, For element-wise multiplication, These are learnable weights. As a feature representation enhanced by trend enhancement, it is more sensitive to precursor segments and can strengthen the risk prediction model's response to weak anomalies.

[0050] The risk discrimination submodule integrates classification probabilities and trend indicators in the model output stage, enhancing the final discrimination capability against precursor states. Its classification branch outputs the AF risk probability. and trend energy score Construct a fusion score: in These are the weighting coefficients. An attention mechanism can be optionally introduced to adaptively adjust the weights based on the dynamic characteristics of the input segment. The fused risk assessment results can serve as a supplementary signal for the final classification decision, thereby improving the detection rate of early warning signs and reducing false alarms. In some feasible embodiments, the joint loss function can also be designed as follows: A joint loss function is constructed, which consists of three parts: first, the supervision loss calculated for labeled samples, using the cross-entropy function; second, the consistency loss calculated for pseudo-labeled samples, using mean squared error to encourage the risk prediction model to output category predictions consistent with pseudo-labels; and third, the early regression loss function, using mean squared error.

[0051] The joint loss function is expressed as follows: in, This represents the cross-entropy loss for labeled samples. This is the early mean square error loss. This represents the consistency loss of pseudo-labeled samples. The distillation loss is used as the trend memory distribution for the risk prediction probability assessment model. Stochastic gradient descent or Adam optimization algorithms are employed to iteratively train the risk prediction model. During the process, a validation set can be set to monitor model performance and avoid overfitting.

[0052] In some feasible implementations, the following mechanisms are employed during training: pseudo-label weighting: the loss term of pseudo-label samples is weighted according to confidence level; trend memory distillation: in addition to the classification results, the trend distribution of the risk probability assessment model is distilled; regularization and early stopping: overfitting is reduced through L2 regularization and dropout, and the generalization is improved by adopting an early stopping strategy; lightweight optimization: optional pruning and quantization compression are performed to further reduce the model size.

[0053] In some feasible embodiments, step S6 specifically includes: S6.1 To avoid false alarms caused by fluctuations in a single window, cumulative risk is calculated using an integral form: in This is the output of the risk discrimination submodule of the risk prediction model. The length of the integration window. As a time decay factor, the recent window has a higher weight, emphasizing the importance of recent signals; S6.2, Individualized threshold setting.

[0054] To accommodate the physiological differences among individuals, dynamic thresholds Defined as: in Based on the threshold, The standard deviation of recent heart rate variability. This is an adjustment coefficient. Individualized threshold settings for each employee avoid a rigid, one-size-fits-all approach.

[0055] S6.3, Continuous overthreshold discrimination.

[0056] When accumulated risk K consecutive windows exceeding the dynamic threshold And advance forecast Less than the preset warning threshold If the signal is detected, the system is determined to be in a high-risk zone for early warning signs (AF). This continuous discrimination mechanism ensures the robustness of the early warning system.

[0057] S6.4, Lead time correction and output.

[0058] To reflect the impact of trend changes on lead time, a trend acceleration correction term is introduced: in For the second-order difference at the medium window scale, This is a correction factor. If the trend changes rapidly, the lead time will decrease accordingly, making the warning more timely.

[0059] S6.5 Early Warning Triggering and Feedback.

[0060] When the revised lead If the risk falls within the set time range, the system will trigger an early warning. The warning results, including the risk level and expected lead time, can be fed back in real time via mobile terminals or wearable devices.

[0061] In summary, the data processing and flow reference of the method of this invention Figure 3 .

[0062] like Figure 4 As shown, an atrial fibrillation early warning system based on multi-timescale trend energy modeling includes: The signal acquisition unit is used to execute step S1; Trend energy modeling unit, used to perform step S2; The risk prediction model training unit is used to perform steps S3 and S4; The risk score calculation unit is used to execute step S5; the early warning unit is used to execute step S6.

[0063] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0064] An atrial fibrillation early warning device based on multi-timescale trend energy modeling: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements an atrial fibrillation early warning method based on multi-timescale trend energy modeling as described above.

[0065] The content of the above method embodiments is applicable to the device embodiments. The specific functions implemented by the device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0066] A storage medium storing processor-executable instructions, which, when executed by a processor, are used to implement an atrial fibrillation early warning method based on multi-timescale trend energy modeling as described above.

[0067] The content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0068] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for early warning of atrial fibrillation based on multi-timescale trend energy modeling, characterized in that, Includes the following steps: Acquire ECG signal; Based on the ECG signal, multi-timescale trend energy modeling is performed, and trend energy function values ​​are generated; The ECG signal is processed based on a pre-trained risk probability assessment model to generate risk probability training data and lead prediction intervals. Train a risk prediction model based on risk probability training data; The risk prediction probability is output based on the risk prediction model, and a risk integral is generated by combining the trend energy function value. By combining the risk score, the lead time prediction range, and the preset threshold, an early warning prompt is triggered. The step of performing multi-timescale trend energy modeling based on the ECG signal and generating trend energy function values ​​specifically includes: By setting up a multi-timescale sliding window and performing sliding segmentation on the ECG signal, a multi-scale segment sequence is obtained. Based on the multi-scale segment sequence, the first-order difference and the second-order difference are calculated respectively; Based on the multi-scale segment sequence, the similarity between trend sequences at different time scales is calculated to obtain a cross-scale consistency index; Based on the first-order difference, the second-order difference, and the cross-scale consistency index, a trend energy function value is generated; The formula for calculating the trend energy function value is as follows: in, This represents the trend energy function value. Represents the first-order difference. Indicates the second-order difference. Indicators representing cross-scale consistency. This represents the corresponding weighting coefficient.

2. The atrial fibrillation early warning method based on multi-timescale trend energy modeling according to claim 1, characterized in that, Following the step of acquiring the ECG signal, the following are also included: The ECG signal is subjected to bandpass filtering, baseline drift correction, power frequency interference suppression, and normalization.

3. The atrial fibrillation early warning method based on multi-timescale trend energy modeling according to claim 1, characterized in that, The risk probability assessment model includes an initial convolutional layer, a residual module, a global average pooling layer, and an output layer. The output layer includes a classification branch and an advance regression branch.

4. The atrial fibrillation early warning method based on multi-timescale trend energy modeling according to claim 1, characterized in that, The training process of the risk probability assessment model specifically includes: Acquire training signals and construct a training set; Generate multi-timescale trend features based on the training signals in the training set; Construct a joint loss function; The risk probability assessment model is iteratively trained by combining the multi-timescale trend features and corresponding labels until the joint loss function converges.

5. The atrial fibrillation early warning method based on multi-timescale trend energy modeling according to claim 1, characterized in that, The formula for calculating the risk score is as follows: in, To score the risk, As a weighting factor, Indicates the probability of risk. This is a normalization function for the trend energy function value. The time decay factor, Indicates risk score, Indicates the lead time for forecasting. Indicates the index of the current time window. This is expressed as the length of the sliding window. This represents a local index within the sliding window.

6. An atrial fibrillation early warning system based on multi-timescale trend energy modeling, characterized in that, A method for performing atrial fibrillation early warning based on multi-timescale trend energy modeling as described in claim 1 includes: The signal acquisition unit is used to acquire ECG signals; The trend energy modeling unit is used to perform multi-timescale trend energy modeling based on the ECG signal and generate trend energy function values. The risk prediction model training unit processes the ECG signal based on a pre-trained risk probability assessment model to generate risk probability training data and lead prediction intervals; and trains the risk prediction model based on the training data. The risk integral calculation unit is used to output the risk prediction probability based on the risk prediction model and generate a risk integral by combining the trend energy function value. The early warning unit, in conjunction with the risk score, the lead time prediction range, and the preset threshold, triggers an early warning notification.

7. An atrial fibrillation early warning device based on multi-timescale trend energy modeling, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the atrial fibrillation early warning method based on multi-timescale trend energy modeling as described in any one of claims 1-5.

Images