An intelligent early warning system and method based on dynamic health signals of heart

Abstract

This invention relates to the field of cardiac health early warning technology, and discloses an intelligent early warning system and method based on dynamic cardiac health signals. The system includes: S1, acquiring dynamic cardiac health signals and obtaining single-modality or dual-modality signal segments; S2, converting PPG and ECG signals into a first signal representation and a second signal representation, respectively; calculating the quality of the first signal and the quality of the second signal; S3, independently determining the heart rhythm category of the first and second signal representations and calculating their respective probability distribution vectors; S4, determining the current signal combination mode within a preset correlation analysis time window and calling the corresponding fusion rules to calculate the fusion probability distribution vector; S5, comprehensively determining the early warning result and outputting it through the user terminal. This application improves robustness, enhances the robustness of multi-class arrhythmia classification, and improves the interpretability and self-evolution of early warnings through a dynamic soft fusion mechanism, a heterogeneous deep neural network architecture, and a three-terminal collaborative closed loop, achieving accurate identification and graded early warning.

Description

Technical Field

[0001] This invention relates to the field of cardiac health early warning technology, specifically to an intelligent early warning system and method based on dynamic cardiac health signals. Background Technology

[0002] Cardiovascular disease is a leading cause of death and disability worldwide, and cardiac arrhythmias are an important clinical manifestation and early warning indicator. Clinically, electrocardiogram (ECG) is considered the gold standard for diagnosing cardiac arrhythmias. However, long-term continuous ECG monitoring is limited by issues such as the invariable wearing of leads and electrodes, skin irritation, patient compliance, and reliance on equipment and professional operation, making large-scale, non-invasive daily monitoring and remote screening difficult. Complementing this is photoplethysmography (PPG) technology, which can be used for long-term, non-invasive data acquisition via wristband or patch-type wearable devices, facilitating widespread adoption and continuous monitoring. The morphology and rhythm of PPG waveforms contain cardiovascular physiological information and have the potential for detecting heart rate and arrhythmia abnormalities.

[0003] In recent years, attempts have begun to combine PPG and ECG signals for joint monitoring and early warning of heart health. However, in real-world dynamic wearable scenarios, existing technologies still have significant shortcomings and are insufficient to meet clinical-grade early warning needs. (1) Imbalance in modal utilization: Most schemes use ECG as the sole core criterion for judgment, while using PPG only as an auxiliary trigger signal for basic heart rate calculation or "wake-up" ECG acquisition. This leads to the neglect of the rhythm and morphological diagnostic value of PPG, making the system extremely prone to paralysis when ECG signals are missing.

[0004] (2) Limitations of heterogeneous signal processing: ECG reflects the strong temporal characteristics of cardiac electrophysiological activity, while PPG reflects the time-frequency characteristics of hemodynamic changes. The physical mechanisms of the two are significantly different. Existing methods mostly use homogeneous networks forcibly fused, failing to design differentiated extraction strategies based on the essence of the signal, resulting in feature dimension mismatch and information loss.

[0005] (3) Lack of robustness in dynamic scenarios: In daily activities, pulse wave conduction delay (50-200ms), motion artifacts and sensor displacement cause asynchronous mismatch between the two signals or a sharp drop in signal-to-noise ratio. Traditional fusion strategies rely on static weights or simple splicing, lacking a real-time assessment and dynamic adjustment mechanism for signal quality. In complex environments, the false alarm rate is as high as 35% or more, which seriously affects clinical trust.

[0006] (4) Insufficient classification granularity: Most methods only support binary classification or single disease detection, and cannot distinguish fine-grained arrhythmia types such as atrial premature beats and ventricular tachycardia, which cannot meet the needs of clinical diagnosis.

[0007] (5) Lack of clinical applicability and iterative mechanism: The early warning results lack traceable evidence such as original waveforms and feature maps, making it difficult for doctors to review the decision-making basis; the static open-loop model cannot be optimized using clinical correction data, resulting in long-term performance drift.

[0008] The aforementioned shortcomings result in insufficient reliability of existing systems in real-world wearable scenarios. They cannot balance the accuracy of ECG with the convenience of PPG, nor can they meet the clinical needs for early screening, accurate classification, and long-term monitoring of arrhythmias. Ultimately, this limits the clinical translation and large-scale application of wearable cardiac health technologies. Summary of the Invention

[0009] The present invention aims to provide an intelligent early warning system and method based on dynamic cardiac health signals, which solves the problems of poor reliability and difficulty in achieving both accuracy and convenience in the existing technology.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent early warning method based on dynamic cardiac health signals, comprising the following steps: S1, Signal Acquisition: The user's dynamic cardiac health signals are acquired synchronously or asynchronously through wearable or contact sensors to obtain single-modal or dual-modal signal segments. The dynamic cardiac health signals include at least one of the original PPG signal and the original ECG signal. S2, Heterogeneous signal representation: The acquired raw signal is preprocessed, and the fixed-length segment of the PPG signal is converted into a two-dimensional time-frequency diagram as the first signal representation through continuous wavelet transform; the fixed-length segment of the ECG signal is processed to output a fixed-length time sequence representation as the second signal representation; the first signal quality of the PPG mode and the second signal quality of the ECG mode are calculated respectively through the set mapping mechanism. S3, Classification Reasoning: Call the set deep neural network to perform independent feature extraction and heart rhythm category determination on the first signal representation and the second signal representation respectively, and calculate the probability distribution vector of each heart rhythm category; S4, Dynamic Fusion: Within the preset correlation analysis time window, determine the overlap between PPG and ECG signals, combine the first signal quality with the second signal quality to determine the current signal combination mode, and calculate the fusion probability distribution vector by calling the corresponding fusion rules according to the signal combination mode. S5, Tiered Early Warning: The early warning result is determined by combining the fusion probability distribution vector and the stepped threshold, and then output through the user terminal.

[0011] Meanwhile, this solution also provides an intelligent early warning system based on dynamic cardiac health signals, which is applied to the aforementioned intelligent early warning method based on dynamic cardiac health signals, including a user terminal, a cloud server terminal, and an advisor terminal. The user terminal is used to acquire single-mode or dual-mode signal segments; The cloud server includes a signal processing module, a classification and reasoning module, a fusion and early warning module, and an advisor support module. The signal processing module is used to perform heterogeneous signal representation on the received raw signal. The classification and reasoning module is used to call a preset deep neural network model that matches the available signal modes to independently determine the heart rhythm category of the processed signal representation and output the probability distribution vector of each mode classification result. The fusion early warning module is used to perform fusion analysis and hierarchical early warning judgment on the classification judgment results within a preset correlation analysis time window, combining real-time signal quality and the missing status of signals in different modes. When the judgment is low confidence or an early warning is triggered, the module uses the correlation analysis time window to extract the corresponding heterogeneous features and original signals to construct a cross-modal structured evidence package. The advisor support module is used to push structured evidence packages to the advisor, receive manually corrected labels and medical orders returned by the advisor, and use the manually corrected labels as real-world samples to trigger incremental learning in order to dynamically update the neural network parameters of the classification reasoning module.

[0012] The principles and advantages of this scheme are: The fundamental contradiction addressed by this solution lies in how to ensure both monitoring accuracy and ease of clinical application in dynamic wearable scenarios. The reason existing technologies have failed to overcome this contradiction is that traditional solutions treat signal acquisition, feature extraction, and fusion decision-making as independent modules, failing to construct a systematic architecture based on the essential differences in physiological signals and the closed-loop nature of clinical needs.

[0013] Specifically, the limitations of existing technologies are reflected in three aspects: First, there is insufficient understanding of the physical differences between ECG and PPG, and the forced splicing of isomorphic networks leads to feature dimension mismatch. For example, the timing characteristics of the QRS complex in ECG cannot be aligned with the pulse wave conduction delay in PPG, and the complementarity of bimodal information is masked. Second, the fusion strategy in dynamic scenarios is rigid, relying on static weights or a single signal source, which cannot cope with signal quality fluctuations caused by motion artifacts (such as a sudden drop in the signal-to-noise ratio of PPG) and asynchronous mismatches (such as a pulse wave conduction delay of 50-200ms), resulting in a high false alarm rate or system paralysis. Third, there is a lack of clinical trust mechanisms. The "black box" decision-making of AI models is difficult to trace, and a closed-loop system of expert review and model iteration has not been built, which cannot meet the rigorous requirements of arrhythmia early warning.

[0014] This solution achieves breakthroughs through cross-modal collaborative design and dynamic closed-loop optimization: First, a heterogeneous deep neural network architecture driven by physiological and physical characteristics is constructed, enabling multi-dimensional, zero-misalignment extraction of arrhythmia features. Addressing the fundamental differences in the dimensions of the two signal streams, this solution abandons the conventional method of forcibly splicing homogeneous networks. For the strong temporal electrophysiological characteristics of ECG, multi-branch one-dimensional convolution is used to extract waveform morphology distortions of different spans in parallel; for the hemodynamic characteristics implicit in PPG, continuous wavelet transform is used to convert it into a two-dimensional time-frequency map, and multi-scale convolution and channel rearrangement operations are employed to accurately capture the abnormal distribution of low-frequency energy. This heterogeneous feature extraction strategy, "customized according to physical characteristics," avoids the information loss and gradient explosion problems caused by forced dimensional alignment, greatly enhancing the system's robustness in classifying multiple categories of fine-grained arrhythmias such as atrial premature beats, ventricular premature beats, and atrial fibrillation.

[0015] Secondly, a dynamic soft fusion mechanism based on signal-to-noise ratio (SNR) gating and associated time windows is proposed, significantly improving the system's robustness against motion artifacts, signal mismatch, and quality fluctuations in real-world wearable scenarios. Unlike traditional dual-signal alignment strategies that rely solely on a single signal source or fixed weights, this solution fully considers the inherent physiological delay in pulse wave propagation time between ECG and PPG, as well as the asynchronous sensor mismatch caused by daily activities. By introducing real-time SNR as a gating condition, the system can dynamically calculate and allocate fusion weights for heterogeneous signals. Especially when strenuous exercise causes a sharp drop in the quality of a single signal, the system can automatically trigger a soft shielding mechanism to downgrade to single-signal mode, or execute a logic judgment prioritizing severe warning intervals in asynchronous dual-signal mode. This mechanism fundamentally solves the pain point of extremely high false positives (false alarms) caused by signal interference in traditional wearable devices, achieving a perfect balance between monitoring accuracy and user-friendly, seamless wearing convenience while ensuring high sensitivity in intercepting malignant arrhythmias.

[0016] Finally, a three-terminal collaborative closed-loop system based on structured evidence packages—user-cloud-consultant—was constructed to enhance the interpretability and self-evolution capabilities of AI early warning. When an early warning is triggered, this solution not only provides classification results but also automatically backtracks and constructs a structured evidence package including the original waveform, time-frequency graph, classification probability distribution, and trigger threshold. This mechanism provides complete judgment criteria for manual review by consultants, resolving the clinical trust issues arising from the "black box" nature of AI algorithms. Simultaneously, the system supports sending the consultant-corrected labels back to the cloud server as the gold standard for model fine-tuning. This labeling-training-deployment closed loop based on real business flows ensures that the early warning algorithm can continuously evolve for specific populations or rare cases, achieving dynamic incremental optimization of system performance. Attached Figure Description

[0017] Figure 1This is a flowchart illustrating an intelligent early warning method based on dynamic cardiac health signals according to the present invention. Figure 2 This is a schematic diagram of the two-dimensional convolutional neural network structure based on composite feature extraction blocks in the PPG classification process of this invention; Figure 3 This is a flowchart illustrating the composite feature extraction block in the PPG classification process of this invention. Figure 4 This is a schematic diagram of the ECG classification process network of the present invention.

[0018] Figure 5 This is a schematic diagram of the structure of an intelligent early warning system based on dynamic cardiac health signals according to the present invention; Figure 6 This is a data flow diagram of an intelligent early warning system based on dynamic cardiac health signals according to the present invention; Detailed Implementation The following detailed description illustrates the specific implementation method: Example This embodiment presents an intelligent early warning system and method based on dynamic cardiac health signals. It collects PPG and ECG dual-modal signals using wearable sensors, employs a heterogeneous feature extraction network to match the physical characteristics differences between the two types of signals, and combines a dynamic fusion strategy based on correlation analysis time windows to achieve robust multimodal fusion. Simultaneously, it constructs a structured evidence package containing the original signal, feature maps, and probability vectors to support expert review. The review results are fed back into the model through an incremental learning mechanism for optimization. This addresses issues in existing technologies such as uneven modality utilization, mismatched heterogeneous feature dimensions, insufficient robustness of dynamic scene fusion, and lack of system closure, enabling accurate identification and graded early warning of six types of heart rhythms, including sinus rhythm and premature atrial contractions.

[0019] Option 1 A smart early warning method based on dynamic cardiac health signals is provided, as shown in the attached document. Figure 1 As shown, it includes the following steps: S1, signal acquisition.

[0020] The user's dynamic cardiac health signals are collected synchronously or asynchronously by wearable or contact sensors to obtain single-modal or dual-modal signal segments. In this embodiment, the dynamic cardiac health signals include at least one of the original PPG signal and the original ECG signal.

[0021] S2, heterogeneous signal representation.

[0022] The acquired raw signals are preprocessed, and fixed-length segments of the PPG signal are converted into two-dimensional time-frequency diagrams through continuous wavelet transform as the first signal representation; fixed-length segments of the ECG signal are processed to output fixed-length time sequence representations as the second signal representation; the first signal quality of the PPG mode and the second signal quality of the ECG mode are calculated respectively through a set mapping mechanism.

[0023] In this embodiment, the acquired raw one-dimensional PPG signal is preprocessed and subjected to time-frequency transformation, and the acquired raw one-dimensional ECG signal is preprocessed and a fixed-length time sequence representation is output.

[0024] Specifically, the acquired raw PPG sequence is downsampled to 100Hz; then a bandpass digital filter is used to remove baseline drift and high-frequency noise. The filter is preferably a front-phase to back-phase zero-phase fourth-order Butterworth bandpass with a passband set to 0.5Hz-10Hz to retain the main frequency bands related to heartbeat.

[0025] The filtered sequence is then segmented according to a fixed-length sliding window, with a preferred window length of 10s. For each segment, amplitude normalization is first performed using the minimum-maximum normalization method.

[0026] After completing the basic preprocessing described above, the first signal quality of the current segment is calculated. This metric is used to quantify the degree to which the current PPG signal is affected by motion artifacts, ambient light, or sensor slippage. In practice, the preferred metric for calculating the first signal quality is the signal-to-noise ratio (SNR).

[0027] The normalized fixed-length segment is converted into a two-dimensional time-frequency diagram using the continuous wavelet transform method. In this embodiment, the Morlet mother wavelet is preferably used with a scale range of 1-128 to cover the spectral information of the PPG signal on both short-time and long-time scales, and the two-dimensional time-frequency diagram is used as the first signal representation.

[0028] The acquired raw ECG signal is downsampled, with a sampling rate preferably of 250Hz. The ECG signal is then subjected to noise suppression and bandgap constraint processing. In this embodiment, power frequency interference suppression is achieved by setting a notch filter with a center frequency of 50Hz or 60Hz to eliminate the influence of power supply noise on the ECG signal. Bandpass filtering uses a digital bandpass filter to limit the ECG signal to the effective ECG frequency band, preferably 0.5 Hz-40 Hz, to remove baseline drift, high-frequency electromyography noise, and other non-cardiac interference. The above filtering process preferably uses a forward-reverse zero-phase filtering method to avoid phase distortion introduced by the filter and ensure the integrity of the ECG waveform.

[0029] Baseline drift correction is performed on the filtered ECG signal. Specifically, baseline drift correction can be achieved through sliding median filtering, morphological filtering, or low-order polynomial fitting, in order to further eliminate low-frequency baseline fluctuations caused by factors such as breathing and changes in body position.

[0030] The ECG signal is segmented into fixed lengths. Specifically, the continuous ECG signal is divided into multiple ECG segments of fixed length according to a preset time window, preferably with each segment having a duration of 10 seconds. After segmentation, amplitude standardization is performed on each fixed-length ECG segment, preferably using Z-score standardization.

[0031] Simultaneously, the quality of the second signal extracted from the current segment is calculated to objectively assess the clinical diagnostic usability of the current ECG waveform. In this embodiment, the preferred metric for calculating the quality of the second signal is the signal-to-noise ratio of a specific effective frequency band. After completing the above steps, the fixed-length ECG waveform that retains the ECG timing morphology is directly used as the second signal representation.

[0032] Because the raw signal-to-noise ratio (SNR) values ​​calculated by the underlying signal processing module have a large range, they cannot be directly used for subsequent probability weighting and threshold comparison. Therefore, in this embodiment, a normalization mapping mechanism is introduced to uniformly convert the raw SNR of the two modes into scalar features within the [0, 1] interval.

[0033] Specifically, the mapping mechanism unifies the original signal quality indices of the two modes into a normalized mapping mechanism within the [0, 1] interval, expressed as follows: (1) In the formula, This represents the original signal-to-noise ratio of the current segment; The lower limit of the signal-to-noise ratio for extremely poor signals preset by the system; This represents the upper limit of the ideal signal-to-noise ratio.

[0034] In this embodiment, regarding the first signal quality Since photocapacitors are highly susceptible to interference from ambient light and motion artifacts, it is preferable to set... =10dB, =30dB.

[0035] For the second signal quality Since the electrophysiological signals are relatively stable, it is preferable to set... =5dB, =20dB.

[0036] In this embodiment, the first signal quality and the second signal quality will be used as real-time signal quality in the subsequent fusion warning step.

[0037] S3, Categorical Reasoning.

[0038] The predefined deep neural network is invoked to independently extract features and determine the heart rhythm category from the first and second signal representations, and calculate the corresponding PPG probability distribution vector P. ppg and ECG probability distribution vector P ecg There are six types of heart rhythms: sinus rhythm, atrial premature beats, ventricular premature beats, ventricular tachycardia, supraventricular tachycardia, and atrial fibrillation.

[0039] In this embodiment, a time-frequency graph classification model based on a two-dimensional convolutional neural network is used for the first signal representation. A time-series classification model based on a one-dimensional convolutional neural network is used for the second signal representation; both independently output probability distribution vectors containing six heart rhythm categories.

[0040] Considering the varying characteristics of hemodynamic energy distribution across different time spans and frequency scales inherent in PPG time-frequency spectra, therefore, as shown in the appendix... Figure 2 As shown, the hidden layer structure of a two-dimensional convolutional neural network contains at least one or more stacked composite feature extraction blocks. After features are extracted by one or more composite feature extraction blocks, they are further mapped by global average pooling and fully connected layers, and then the PPG probability distribution vector P is output through Softmax. ppg .

[0041] As attached Figure 3 As shown, the composite feature extraction block includes multi-scale feature extraction and channel swapping.

[0042] Multi-scale feature extraction is used to apply channel-wise convolutions to the input feature map on parallel branches. The dilation rate of the parallel branches contains at least three different values ​​to form a multi-scale receptive field. After channel-wise convolution, a 1×1 convolution is performed to restore and adjust the number of channels.

[0043] In this embodiment, the kernel size for channel-by-channel convolution is preferably 7×7 to ensure sufficient local context acquisition. The dilation rates of the three branches take different values ​​to construct a multi-scale receptive field, with the value set being {1, 2, 3}. The technical significance of this large-size convolution kernel combined with a stepped dilation rate design is that the branch with a dilation rate of 1 is used to accurately capture high-frequency transient changes in blood flow; while the branches with dilation rates of 2 and 3 have a larger span and are used to extract long-term vascular compliance changes.

[0044] Channel swapping is used to first perform linear mapping on the fused features according to the channel dimension to expand the channels, divide the expanded channels into several groups and perform channel shuffling to facilitate cross-group channel information exchange, and apply activation and linear shrinkage to the shuffled features to restore the original channel dimension.

[0045] For ECG signals, corresponding features are first extracted using a predefined multi-branch convolution, as shown in the attached figure. Figure 4 As shown, a multi-branch system includes at least three branches: a short-term branch, a medium-term branch, and a long-term branch.

[0046] The short-time branch has two kernel sizes of 3, focusing on capturing morphological changes in high-frequency, narrow-pulse-width QRS complexes; the medium-time branch has two kernel sizes of 5, used to extract P-wave features with medium spans; and the long-time branch has two kernel sizes of 9, used to cover and capture the evolution trend of low-frequency, wide-pulse-width T-waves. Max pooling is used after each stage to achieve downsampling.

[0047] Then, the final outputs of each parallel branch are fused in the time dimension and then concatenated in the channel dimension to form a joint feature map. Subsequently, channel fusion and dimensionality compression are achieved by pointwise one-dimensional convolution. After the fusion layer, normalization and activation are applied.

[0048] The fused features are subjected to global average pooling in the time dimension and then mapped through at least one fully connected layer before being output as the ECG probability distribution vector P via Softmax. ecg .

[0049] S4, Dynamic Fusion.

[0050] Within a preset correlation analysis time window, the overlap between PPG and ECG signals is determined. The current signal combination mode is determined by combining the first signal quality with the second signal quality. Based on the signal combination mode, the corresponding fusion rules are called to calculate the fusion probability distribution vector.

[0051] In this embodiment, the signal combination modes include synchronous dual-signal mode, asynchronous dual-signal mode, single-signal mode, and invalid signal mode; the determination method is as follows: Synchronous dual-signal mode: The two signals overlap on the time axis and the quality of the first signal and the quality of the second signal are both greater than or equal to the signal quality threshold. Asynchronous dual-signal mode: Two signals exist within a time window but do not completely overlap, and the quality of the first signal and the quality of the second signal are both greater than or equal to the signal quality threshold.

[0052] Single-signal mode: Modal masking is performed when there is only one signal input, or when the signal quality of one of the signals is lower than the signal quality threshold.

[0053] Invalid signal mode: The signal quality of both signals is below the signal quality threshold. In this case, the user should be prompted to remeasure.

[0054] In this embodiment, a correlation analysis time window is set. The length is 10 seconds. By extracting the high-precision timestamps of the two signals after alignment in the cloud, the time between them is calculated. The time overlap rate within the time frame. Preferably, the overlap rate threshold is set to 80%: when the overlap rate is >80%, it is considered "overlapping on the time axis"; when the overlap rate is ≤80%, it is considered "incomplete overlap", thus accommodating physiological pulse wave conduction delay or slight sensor asynchrony. The signal quality threshold is preferably set to 0.60.

[0055] In this embodiment, the fusion probability distribution vector is calculated by calling the corresponding fusion rules according to the current signal combination mode.

[0056] The fusion rules are as follows: (1) If it is a synchronous dual-signal mode, then perform linear weighted fusion based on real-time signal quality, expressed as follows: (2) In the formula, , These are dynamic weight variables generated based on real-time signal quality mapping, and .

[0057] The calculation method for generating dynamic weight variables through real-time signal quality mapping is as follows: (3) (4) In the formula, These are the prior bias coefficients for the ECG modes; This refers to the second signal quality corresponding to the ECG signal; Let be the prior bias coefficients of the PPG mode; and satisfy . ; This represents the first signal quality corresponding to the PPG signal.

[0058] In this embodiment, considering that ECG signals directly reflect cardiac electrophysiological characteristics and their confidence level in clinical diagnosis is naturally higher than that of PPG signals, which reflect peripheral hemodynamics, the prior bias coefficient of the ECG modality is... The preferred setting is 1.2, which is the prior bias coefficient for the PPG mode. The preferred setting is 1.0.

[0059] (2) If it is an asynchronous dual-signal mode, then perform a logical consistency-based discrimination fusion, including, a. If the two types of signals are determined to be the same heart rhythm type, the fusion probability value is increased by adjusting the gain coefficient, which is expressed as follows: (5) In the formula, The gain coefficient is preferably set to 1.1, which can play a role in cross-validation and amplify the warning signal in weak feature scenarios.

[0060] b. If the two types of signals are determined to be different heart rhythm types, then according to the ECG probability distribution vector... With PPG probability distribution vector The dominant mode is determined by the largest probability component in the matrix, and the probability distribution vector of the dominant mode is used as the final fusion probability distribution vector.

[0061] In this embodiment, the six cardiac rhythm categories are ranked from highest to lowest clinical urgency as follows: ventricular tachycardia > atrial fibrillation > supraventricular tachycardia > premature ventricular contractions (PVCs) > premature atrial contractions (PACs) > sinus rhythm. When faced with high-confidence conflicts in asynchronous signals, the system will strictly follow this priority list and prioritize the prediction result with higher risk.

[0062] In this embodiment, the dominant mode is determined as follows: ① If the maximum probability component of at least one mode triggers a severe warning level alert: The mode that triggers a severe warning level alert is determined as the dominant mode; if both modes trigger severe warning level alerts, the dominant mode is determined in order of higher clinical urgency level and larger maximum probability component value of the corresponding heart rhythm category; if the above conditions are equal, the ECG mode is designated as the dominant mode. ② If neither of the maximum probability components of the two modes triggers a severe warning level alert, but the maximum probability component of at least one mode triggers a warning level alert: The mode that triggers a warning-level alert is identified as the dominant mode; if both modes trigger a warning-level alert, the mode with the larger maximum probability component value is identified as the dominant mode; if the values ​​are equal, the ECG mode is designated as the dominant mode. ③ If neither of the maximum probability components of the two modes triggers a severe warning or a warning: The ECG mode is designated as the dominant mode.

[0063] (3) If it is a single signal mode, the probability distribution vector of the currently available signal source is directly used as the fusion probability distribution vector.

[0064] S5, graded early warning.

[0065] The early warning result is determined by combining the fusion probability distribution vector and the stepped threshold, and then output through the user terminal.

[0066] In this embodiment, based on the fusion probability distribution vector maximum component The corresponding heart rhythm type C and step threshold The warning results are determined comprehensively and the following conditions are met. The specific judgment process is as follows: If the heart rhythm type C is "sinus rhythm", it is considered normal and no warning is triggered.

[0067] If the heart rhythm type C is another type, the maximum component will be matched with the stepped threshold range and graded processing will be performed.

[0068] In this embodiment, the hierarchical processing is as follows: Low confidence handling: when In [0, When the result is determined to be a "low confidence identification result", the system will not push a warning to the user and will mark it as "signal quality uncertain". Warning level: When In [ , When the heart rhythm pattern deviates from the normal range, a prompt-level warning is triggered to indicate that the system has initially identified the heart rhythm pattern as deviating from the normal range, reminding the user to pay attention to the current status and providing a reference for subsequent continuous monitoring; Warning level alert: When In [ , When a heart rhythm disorder occurs, a warning level alert is triggered. This alert indicates that the system has identified the arrhythmia event with high confidence. The system recommends that the user stop strenuous exercise, remain still to obtain a higher quality signal for verification, and consult professional medical advice as needed. Severe warning level alert: When In [ When [1] is triggered, a severe warning level alert is activated, which indicates that the system has identified an arrhythmia event with extremely high confidence, reminding the user that there may be a high risk and that immediate medical consultation or emergency measures are required.

[0069] In this embodiment, These can be set to 0.5, 0.75, and 0.9 respectively.

[0070] In this embodiment, the method also includes backtracking and extracting the correlation analysis time window when a low-confidence processing or early warning event is triggered. The original signal fragments, time-frequency graphs, probability distribution vectors, timestamp indices, and trigger thresholds related to the warning event are used to construct a structured evidence package.

[0071] Option 2 In this embodiment, an intelligent early warning system based on dynamic cardiac health signals is also provided, applied to the aforementioned intelligent early warning method based on dynamic cardiac health signals, as shown in the attached figure. Figure 5 As shown, it includes the user terminal, the cloud server terminal, and the consultant terminal.

[0072] The user terminal is used to acquire single-mode or dual-mode signal segments.

[0073] In this embodiment, the user terminal is composed of physiological signal acquisition hardware and a mobile interactive terminal, including a signal acquisition module, a signal transmission module, and a result output module.

[0074] Among them, combined with the appendix Figure 6 As shown, the physiological signal acquisition hardware is preferably a wearable device with skin-contact properties, including but not limited to smartwatches, smart bracelets, smart rings, chest patch-type dynamic electrocardiogram recorders, or smart fabrics. It is mainly used for the unobtrusive, continuous, or on-demand acquisition of physiological signals during the user's daily life. The mobile interactive terminal is preferably a smart communication device carried by the user, serving as a data relay gateway and human-computer interaction interface.

[0075] The signal acquisition module is mainly integrated within the physiological signal acquisition hardware. It includes a PPG acquisition unit for acquiring photoplethysmography (PPG) signals and an ECG acquisition unit for acquiring electrocardiogram (ECG) signals.

[0076] The PPG acquisition unit includes at least one light-emitting diode and one photodetector. It acquires the pulse wave volume signal by detecting minute changes in the intensity of reflected or transmitted light of a specific wavelength beam in the capillary bed. The ECG acquisition unit includes electrode patches or dry electrodes integrated on the device surface, a front-end analog amplifier circuit, and a filter circuit. It is used to acquire the raw one-dimensional ECG time-series signal reflecting changes in myocardial electrical activity. In this embodiment, the PPG acquisition unit and the ECG acquisition unit can be integrated into the same wearable device or deployed separately in different wearable or contact devices, achieving synchronous or asynchronous acquisition through timestamp information.

[0077] The signal transmission module is used to establish a two-way communication link between the user end and the cloud server.

[0078] The results output module is used to receive and display the heart rhythm abnormality warning results and medical orders returned by the cloud server. It mainly relies on the screen display, audio output and vibration components of the mobile interactive terminal.

[0079] The cloud server includes a signal processing module, a classification reasoning module, a fusion early warning module, and an advisor support module.

[0080] The signal processing module is used to perform heterogeneous signal representation on the received raw signal. In this embodiment, the signal processing module is used to perform targeted preprocessing on the PPG signal and ECG signal respectively, specifically including a PPG processing unit and an ECG processing unit.

[0081] The PPG processing unit performs preprocessing and time-frequency transformation on the acquired raw one-dimensional PPG signal to obtain a first signal representation, and calculates the first signal quality through a set mapping mechanism. The ECG processing unit performs preprocessing on the acquired raw one-dimensional ECG signal to obtain a second signal representation, and calculates the second signal quality through a set mapping mechanism.

[0082] The classification reasoning module is used to call a preset deep neural network model that matches the available signal modes to independently determine the heart rhythm category of the processed first signal representation and second signal representation, and output the probability distribution vector of each mode classification result.

[0083] In this embodiment, the classification reasoning module includes a PPG classification unit and an ECG classification unit, which independently output probability distribution vectors containing six heart rhythm categories.

[0084] Specifically, the PPG classification unit employs a time-frequency graph classification model based on a two-dimensional convolutional neural network for the representation of the first signal. The hidden layer structure of this two-dimensional convolutional neural network contains at least one or more stacked composite feature extraction blocks. After feature extraction by one or more composite feature extraction blocks, the data is further processed by global average pooling and a fully connected layer, and then the PPG probability distribution vector P is output via Softmax. ppg .

[0085] The composite feature extraction block includes a multi-scale feature extraction unit and a channel exchange unit. The multi-scale feature extraction unit applies channel-wise convolution to the input feature map on parallel branches; the channel exchange unit facilitates cross-group channel information exchange.

[0086] The ECG classification unit uses a one-dimensional convolutional neural network temporal classification model to represent the second signal. The one-dimensional convolutional neural network includes a multi-branch convolution extraction unit, a branch output fusion unit, and a classification output unit.

[0087] Multi-branch convolutional extraction units are used to extract corresponding features according to branch type.

[0088] The branch output fusion unit is used to align the features of parallel branches in the time dimension and then stitch them together in the channel dimension to form a joint feature map.

[0089] The classification output unit is used to output the ECG probability distribution vector P from the fused features through Softmax. ecg .

[0090] The fusion early warning module is used to perform fusion analysis and hierarchical early warning judgment on the classification judgment results within a preset correlation analysis time window, combining real-time signal quality and the missing status of signals in different modes. When the judgment is low confidence or an early warning is triggered, the module uses the correlation analysis time window to extract the corresponding heterogeneous features and original signals to construct a cross-modal structured evidence package.

[0091] In this embodiment, the fusion early warning module includes an input receiving unit, a weighted fusion unit, a threshold judgment unit, and an evidence construction unit.

[0092] The input receiving unit is used to perform correlation analysis within a preset time window. Internally, it receives and caches the heart rate category probability distribution vectors output by the PPG classification unit and the ECG classification unit, respectively. and It simultaneously acquires the first signal quality corresponding to the PPG signal and the second signal quality corresponding to the ECG signal, calculated by the signal processing module, and determines the currently available signal combination mode.

[0093] The weighted fusion unit is used to calculate the fusion probability distribution vector by calling the corresponding fusion rules according to the signal combination pattern.

[0094] The threshold judgment unit is used to determine the early warning result based on the maximum component of the fusion probability distribution vector, the corresponding heart rhythm type C, and the step threshold.

[0095] The evidence construction unit is used to backtrack and extract the original signal segments, time-frequency graphs, probability distribution vectors, timestamp indexes and trigger thresholds related to the warning event within the correlation analysis time window when a low-confidence processing or warning event is triggered, and to construct a structured evidence package.

[0096] The consultant support module is used to push structured evidence packages to the consultant, receive manually corrected labels and medical orders returned by the consultant, and use the manually corrected labels as real-world samples to trigger incremental learning in order to dynamically update the neural network parameters of the classification reasoning module.

[0097] The advisor support module includes a task management unit, a feedback interaction unit, and a model optimization unit.

[0098] The task management unit is used to prioritize and queue tasks to be reviewed based on the user's alert level and alert time, and to extract and distribute the corresponding structured evidence packages from cloud storage according to the task requests from the consultant.

[0099] The feedback interaction unit is used to synchronously receive and parse the manual correction tags, waveform marker coordinates and medical order information returned from the consultant terminal, associate and encapsulate the manual correction tags with the original signal segments in the structured evidence package, and realize the real-time feedback of the warning results to the result output module of the user terminal.

[0100] The model optimization unit is used to store the manually corrected labels and corresponding original signal segments encapsulated by the feedback interaction unit as gold standard samples into the training database, and trigger the incremental learning or parameter fine-tuning of the corresponding neural network in the classification inference module, so as to achieve self-evolution of classification accuracy and personalized recognition performance.

[0101] The consultant module includes a task retrieval module, an early warning review module, and a conclusion submission module.

[0102] The task retrieval module is used to establish communication with the task management unit on the cloud server, obtain and display the list of tasks to be reviewed in real time and in a hierarchical manner, and support filtering based on user warning level, warning type or timestamp, and retrieve the structured evidence package of a specified user.

[0103] The early warning and review module provides a multi-view synchronous playback interface, aligning and rendering the original ECG time-series waveforms, PPG time-series waveforms, and two-dimensional time-frequency graphs from the structured evidence package distributed by the cloud server. This unit supports dynamic scaling, translation, and anomaly marking of specific cardiac segments on the waveforms.

[0104] The conclusion submission module is used by consultants to perform overwrite corrections on the classification tags automatically generated by the system and add medical order information, and then send the overwritten correction tags and the medical order information back to the feedback interaction unit on the cloud server.

[0105] The following is a further explanation using specific embodiments.

[0106] Scenario 1: Linear weighted fusion scenario based on synchronous dual-signal mode.

[0107] This scenario primarily describes a linear weighted fusion scenario based on a synchronous dual-signal mode. In this scenario, a stepped threshold is used. The preset values ​​are 0.5, 0.75, and 0.9, respectively.

[0108] (1) Synchronization signal acquisition: Synchronously acquire the original one-dimensional PPG signal and the original one-dimensional ECG signal from the same source. Extract the signal quality of the current segment. =0.8, =0.6. Parallel output of probability distribution vectors for both modes. , .

[0109] but ; ; (2) Mode determination and fusion: If two types of signals are found to be overlapping and both signal quality feature values ​​are ≥ threshold value 0.6, it is determined to be a synchronous dual-signal mode. Then, fusion rule (1) is called to set the ECG bias coefficient. The PPG bias coefficient is 1.2. The value is 1.0. Substituting this value into equations (3) and (4), the weights are calculated. ≈0.615, ≈0.385, then calculate the fusion result using equation (2): ; (3) Early warning classification and evidence consolidation: extraction The maximum component in the value is 0.792, corresponding to the category of "premature ventricular contractions". A value of 0.792 is considered to be within the range of […]. , The warning interval will output the warning result as: "Ventricular premature beats - warning".

[0110] Backtrack and extract the original ECG waveform, PPG waveform, and time-frequency graph aligned to the current time point to construct a structured evidence package.

[0111] (4) Task retrieval and review: The task management unit pushes the "warning" task to the consultant. The doctor opens the multi-view alignment interface through the warning review module, observes and confirms that the given "ventricular premature beats" result is accurate.

[0112] (5) Conclusion Submission and Feedback: Doctors can perform the "direct confirmation" operation through the conclusion submission module and add medical advice suggestions. The conclusion is transmitted back to the feedback interaction unit on the cloud server in real time and displayed on the user's end by the result output module.

[0113] Scenario 2: Same-category gain fusion scenario based on asynchronous dual-signal mode.

[0114] Unlike scenario 1, in this scenario, in step (1) ; .

[0115] In step (2), two types of signals are identified as not completely overlapping and both signal quality characteristic values ​​are ≥ threshold value 0.6, thus determining it as an asynchronous dual-signal mode. Since the maximum components of both signals point to "ventricular tachycardia", the class a gain formula (5) in fusion rule (2) is called to calculate the fusion result, and its gain coefficient is... ;but ; In step (3), because In [ , Within the specified interval, the output warning result is: "Ventricular tachycardia - Warning". The evidence construction unit backtracks and extracts the PPG waveform and time-frequency graph within the correlation analysis time window and the current ECG raw waveform to construct a structured evidence package.

[0116] In step (4), the doctor opens the multi-view alignment interface through the early warning review module, observes and confirms that the given result of "ventricular tachycardia" is accurate.

[0117] Scenario 3: Cross-interval arbitration scenario based on asynchronous dual-signal mode.

[0118] Unlike scenario 1, in this scenario, step (1) outputs the probability distribution vectors of the two modes: ; .

[0119] In step (2), two types of signals were identified as being in a state of incomplete overlap, and both signal quality characteristic values ​​were ≥ the threshold value of 0.6, thus determining it to be an asynchronous dual-signal mode. Since the maximum components of the two signals point to ventricular premature beats and atrial fibrillation respectively, and the maximum component (0.92) of the PPG mode falls within […]. In the severe warning interval of [1], according to rule b, the system directly determines PPG as the dominant mode and outputs... .

[0120] In step (3), the warning result is output as "Atrial fibrillation - serious warning", and a structured evidence package is constructed.

[0121] In step (4), the task management unit pushes the "serious warning" task to the consultant. The doctor's review found that the PPG captured transient hemodynamic deterioration, while the ECG failed to report the highest risk due to signal amplitude attenuation caused by the lead position, confirming that the given "atrial fibrillation" result was accurate.

[0122] Scenario 4: Arbitration scenario within the same interval based on asynchronous dual-signal mode.

[0123] Unlike scenario 1, in this scenario, step (1) outputs the probability distribution vectors of the two modes: ; .

[0124] In step (2), the maximum components of the two signals point to ventricular premature beats and atrial premature beats, respectively, and the maximum components of the two modes are in the same warning interval. , And the values ​​are completely identical (both are 0.85). According to rule b, ECG is determined to be the dominant mode, and the output is... .

[0125] In step (3), the warning result is output as "ventricular premature beats - warning", and a structured evidence package is constructed.

[0126] In step (4), the task management unit pushes the "warning" task to the consultant. The doctor observes the aligned waveform through the review module and finds that the waveform has a wide and distorted QRS wave but is accompanied by an incomplete compensatory pause. After careful manual identification, it is confirmed that it should actually be "atrial premature beat".

[0127] In step (5), the doctor performs a "correction overwrite" operation, changing the label to "atrial premature beats". This corrected label serves as the incremental learning of the gold standard trigger model optimization unit.

[0128] Scenario 5: Silent interception scenario based on single-signal mode.

[0129] Unlike scenario 1, in this scenario, in step (1), the signal quality of the current segment is extracted. =0.8, =0.4. Vigorous exercise caused the PPG signal quality to be less than the threshold of 0.6; therefore, the classification inference module relies solely on the probability distribution vector of the effective ECG output. .

[0130] In step (2), the input receiving unit triggers soft shielding, determining it to be in single-signal mode. According to rule (3), the weighted fusion unit directly uses the currently available ECG signal source as its output. .

[0131] In step (3), extract The maximum component is 0.45 (atrial fibrillation). Since 0.45 falls within the range [0, ... If the result falls within the specified range, the system will determine it as a "low confidence identification result" and implement silent processing (without sending a warning) on ​​the user end. However, the evidence construction unit will still silently extract this fuzzy waveform to construct a noise floor evidence package.

[0132] In step (4), the task management unit pushes the "low confidence" task to the consultant. The doctor confirms that it is actually "sinus rhythm" and no conclusion needs to be submitted; the system does not push the task to the consultant.

[0133] In this embodiment, an innovative dynamic soft fusion mechanism based on signal-to-noise ratio gating and associated time windows is proposed. This mechanism fully considers the physiological delay of pulse wave propagation time between ECG and PPG and the problem of sensor asynchronous mismatch. By dynamically allocating fusion weights through real-time signal-to-noise ratio, a soft shielding mechanism is automatically triggered to degrade to single-signal mode when the quality of a single signal deteriorates. In asynchronous scenarios, the logic of prioritizing severe warning intervals is executed. This fundamentally solves the problem of high false positives caused by signal interference in traditional wearable devices. While ensuring high sensitivity interception of malignant arrhythmias, it achieves a balance between monitoring accuracy and user-friendly wearing convenience, significantly improving the robustness against motion artifacts in real wearable scenarios.

[0134] Secondly, a heterogeneous deep neural network architecture driven by physiological and physical characteristics is constructed, abandoning the forced splicing scheme of homogeneous networks. For the strong temporal electrophysiological characteristics of ECG, multi-branch one-dimensional convolution is used to extract waveform morphology distortions of different spans in parallel. For the hemodynamic characteristics of PPG, after converting it into a two-dimensional time-frequency map through continuous wavelet transform, multi-scale convolution and channel rearrangement operations are used to capture the abnormal distribution of low-frequency energy. This avoids the information loss and gradient explosion problems caused by forced dimension alignment, and realizes multi-dimensional, zero-misalignment extraction of arrhythmia features. It greatly enhances the robustness of classification of fine-grained arrhythmias such as atrial premature beats, ventricular premature beats, and atrial fibrillation.

[0135] Furthermore, an innovative three-terminal collaborative closed-loop system based on structured evidence packages—"user-cloud-consultant"—is constructed. When an alert is triggered, the system automatically backtracks and constructs a structured evidence package containing the original waveform, time-frequency graph, classification probability distribution, and trigger threshold. This provides complete judgment criteria for expert review, addressing the clinical trust issue of the "black box" nature of AI algorithms. Simultaneously, consultant-corrected labels are sent back to the cloud as the gold standard for model fine-tuning, forming a labeling-training-deployment closed loop based on real business flows. This ensures that the alert algorithm continuously evolves for specific populations or rare cases, achieving dynamic incremental optimization of system performance and enhancing the interpretability and self-evolution capabilities of AI alerts.

[0136] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. An intelligent early warning method based on dynamic cardiac health signals, characterized in that, Includes the following steps: S1, Signal Acquisition: The user's dynamic cardiac health signals are acquired synchronously or asynchronously through wearable or contact sensors to obtain single-modal or dual-modal signal segments. The dynamic cardiac health signals include at least one of the original PPG signal and the original ECG signal. S2, Heterogeneous signal representation: The acquired raw signal is preprocessed, and the fixed-length segment of the PPG signal is converted into a two-dimensional time-frequency diagram as the first signal representation through continuous wavelet transform; the fixed-length segment of the ECG signal is processed to output a fixed-length time sequence representation as the second signal representation; the first signal quality of the PPG mode and the second signal quality of the ECG mode are calculated respectively through the set mapping mechanism. S3, Classification Reasoning: Call the set deep neural network to perform independent feature extraction and heart rhythm category determination on the first signal representation and the second signal representation respectively, and calculate the probability distribution vector of each heart rhythm category; S4, Dynamic Fusion: Within the preset correlation analysis time window, determine the overlap between PPG and ECG signals, combine the first signal quality with the second signal quality to determine the current signal combination mode, and calculate the fusion probability distribution vector by calling the corresponding fusion rules according to the signal combination mode. S5, Tiered Early Warning: The early warning result is determined by combining the fusion probability distribution vector and the stepped threshold, and then output through the user terminal.

2. The intelligent early warning method based on dynamic cardiac health signals according to claim 1, characterized in that: In S2, the mapping mechanism calculates a normalized quality score based on the signal-to-noise ratio of the corresponding signal segments, expressed as follows: ； In the formula, This represents the original signal-to-noise ratio of the current segment; The lower limit of the signal-to-noise ratio for extremely poor signals preset by the system; This represents the upper limit of the ideal signal-to-noise ratio.

3. The intelligent early warning method based on dynamic cardiac health signals according to claim 1, characterized in that: The heart rhythm categories include six types: sinus rhythm, atrial premature beats, ventricular premature beats, ventricular tachycardia, supraventricular tachycardia, and atrial fibrillation.

4. The intelligent early warning method based on dynamic cardiac health signals according to claim 3, characterized in that: In S3, a time-frequency graph classification model based on a two-dimensional convolutional neural network is used for the first signal representation; a time-series classification model based on a one-dimensional convolutional neural network is used for the second signal representation; and the two independently output probability distribution vectors containing six heart rhythm categories.

5. The intelligent early warning method based on dynamic cardiac health signals according to claim 1, characterized in that, In S4, the signal combination modes include synchronous dual-signal mode, asynchronous dual-signal mode, single-signal mode, and invalid signal mode; The determination method is as follows: Synchronous dual-signal mode: The two signals overlap on the time axis and the quality of the first signal and the quality of the second signal are both greater than or equal to the signal quality threshold. Asynchronous dual-signal mode: Two signals exist within a time window but do not completely overlap, and the quality of the first signal and the quality of the second signal are both greater than or equal to the signal quality threshold. Single-signal mode: Mode masking is performed when there is only one signal input, or when the signal quality of one of the signals is lower than the signal quality threshold. Invalid signal mode: The signal quality of both signals is lower than the signal quality threshold.

6. The intelligent early warning method based on dynamic cardiac health signals according to claim 5, characterized in that, The fusion rule is as follows: (1) If it is a synchronous dual-signal mode, then perform linear weighted fusion based on real-time signal quality, expressed as follows: ； In the formula, , These are dynamic weight variables generated based on real-time signal quality mapping. ; (2) If it is an asynchronous dual-signal mode, then perform a logical consistency-based discrimination fusion, including, a. If the two types of signals are determined to be the same heart rhythm type, the fusion probability value is increased by adjusting the gain coefficient, which is expressed as follows: ； In the formula, This is the gain coefficient; b. If the two types of signals are determined to be different heart rhythm types, then according to the ECG probability distribution vector... With PPG probability distribution vector The dominant mode is determined by the largest probability component in the matrix, and the probability distribution vector of the dominant mode is used as the final fusion probability distribution vector. (3) If it is a single signal mode, the probability distribution vector of the currently available signal source is directly used as the fusion probability distribution vector.

7. The intelligent early warning method based on dynamic cardiac health signals according to claim 6, characterized in that: The calculation method for generating dynamic weight variables through real-time signal quality mapping is as follows: ； ； In the formula, These are the prior bias coefficients for the ECG modes; This refers to the second signal quality corresponding to the ECG signal; For the prior bias coefficients of the PPG mode; ; This represents the first signal quality corresponding to the PPG signal.

8. The intelligent early warning method based on dynamic cardiac health signals according to claim 1, characterized in that: In S5, the warning result is determined by comprehensively considering the maximum component of the fusion probability distribution vector, the corresponding heart rhythm type C, and the step-based threshold. The judgment process is as follows: If the heart rhythm type C is "sinus rhythm", it is considered normal and no warning is triggered; If the heart rhythm type C is another type, the maximum component will be matched with the stepped threshold range and graded processing will be performed.

9. The intelligent early warning method based on dynamic cardiac health signals according to claim 6, characterized in that, The dominant mode is determined as follows: (1) If the maximum probability component of at least one mode triggers a severe warning level alert: The mode that triggers a severe warning level alert is determined as the dominant mode; if both modes trigger severe warning level alerts, the dominant mode is determined in order of higher clinical urgency level and larger maximum probability component value of the corresponding heart rhythm category; if the above conditions are equal, the ECG mode is designated as the dominant mode. (2) If neither of the maximum probability components of the two modes triggers a severe warning level alert, and at least one of the maximum probability components of the mode triggers a warning level alert: The mode that triggers a warning-level alert is identified as the dominant mode; if both modes trigger a warning-level alert, the mode with the larger maximum probability component value is identified as the dominant mode; if the values ​​are equal, the ECG mode is designated as the dominant mode. (3) If neither of the maximum probability components of the two modes triggers a severe warning or a warning: The ECG mode is designated as the dominant mode.

10. An intelligent early warning system based on dynamic cardiac health signals, characterized in that: The intelligent early warning method based on dynamic cardiac health signals, applied to any one of claims 1-9, includes a user terminal, a cloud server terminal, and an advisor terminal; The user terminal is used to acquire single-mode or dual-mode signal segments; The cloud server includes a signal processing module, a classification and reasoning module, a fusion and early warning module, and an advisor support module. The signal processing module is used to perform heterogeneous signal representation on the received raw signal. The classification and reasoning module is used to call a preset deep neural network model that matches the available signal modes to independently determine the heart rhythm category of the processed signal representation and output the probability distribution vector of each mode classification result. The fusion early warning module is used to perform fusion analysis and hierarchical early warning judgment on the classification judgment results within a preset correlation analysis time window, combining real-time signal quality and the missing status of signals in different modes. When the judgment is low confidence or an early warning is triggered, the module uses the correlation analysis time window to extract the corresponding heterogeneous features and original signals to construct a cross-modal structured evidence package. The advisor support module is used to push structured evidence packages to the advisor, receive manually corrected labels and medical orders returned by the advisor, and use the manually corrected labels as real-world samples to trigger incremental learning in order to dynamically update the neural network parameters of the classification reasoning module.

Images