A cloud-edge collaborative electrocardio intelligent analysis device and method based on a multi-model pipeline

Abstract

The present application belongs to the technical field of intelligent medical equipment, and specifically relates to a cloud-edge collaborative electrocardio intelligent analysis device and method based on a multi-model pipeline. The problem of misjudgment under noise interference caused by the missing of signal quality evaluation links and the dependence on cloud end final determination due to weak autonomous decision-making ability of the edge end is solved. The device comprises an electrocardio signal acquisition unit, a signal quality evaluation unit containing a first special neural network model for real-time evaluation of original electrocardio signal quality, a preprocessing unit for filtering available signals, an analysis processing unit composed of a second and a third special neural network model in cascade, which sequentially performs multi-dimensional feature extraction and multi-label heart rhythm classification and outputs confidence, and a pre-warning decision unit which determines the pre-warning level according to preset rules in combination with the confidence and historical event statistics. According to the evaluation result and the pre-warning level, the data interaction content is dynamically decided, the desensitization health summary is uploaded in a regular state cycle, the encrypted original signal segment is uploaded when the trigger condition is met, and the cloud end model is updated.

Description

Technical Field

[0001] This invention belongs to the field of intelligent medical device technology, specifically relating to a cloud-edge collaborative intelligent ECG analysis device and method based on a multi-model pipeline. Background Technology

[0002] With an aging population and an increasing burden of cardiovascular disease, the demand for long-term, continuous, and convenient cardiac health monitoring is becoming increasingly urgent. Wearable ECG devices, due to their non-invasive and portable characteristics, have become an important tool for daily ECG monitoring. Traditional wearable ECG devices have relatively simple functions, mainly acquiring, filtering, and storing ECG signals, with subsequent analysis heavily reliant on smartphone applications or cloud servers. This architecture has several inherent drawbacks: First, the early warning delay is high, as the time from signal acquisition to cloud analysis and result return cannot meet the second-level response requirements for malignant arrhythmias (such as ventricular fibrillation and sustained ventricular tachycardia); second, the device's functionality is significantly limited or even malfunctions in environments with poor network conditions or no network access (such as elevators or in the wilderness); third, continuously uploading raw ECG data to the cloud poses a risk of user privacy leaks and results in high device communication power consumption.

[0003] In recent years, the integration of edge computing and artificial intelligence has provided a new solution to the aforementioned problems. By deploying lightweight AI models on the local microprocessors of wearable devices, some computational tasks can be offloaded, significantly improving response speed and reducing reliance on networks. Existing technologies have achieved single functions at the edge, such as R-wave detection and initial screening for specific arrhythmias (e.g., atrial fibrillation). However, a single model struggles to handle the complex chain from noisy raw data to the final warning decision. More advanced solutions propose the concept of cloud-edge collaboration, where preliminary processing and screening are performed at the edge, while complex calculations and confirmation are conducted in the cloud. However, most existing collaborative solutions adopt a simple linear triggering mode of "edge computing scores, cloud final decision," resulting in poor autonomy for edge devices. Essentially, it remains a cloud-centric computing model, failing to fully leverage the intelligent potential of edge devices and exhibiting insufficient robustness under complex signal quality interference.

[0004] Currently, patents and literature have proposed cloud-edge collaborative electrocardiogram (ECG) monitoring systems. For example, Chinese patent CN115151203A, entitled "A Cloud-Edge Collaborative Home-Use AI Agent for Myocardial Infarction Early Warning," describes a system comprising an edge agent, a cloud agent, and a collaborative decision-making engine. The edge agent is responsible for collecting signals and calculating a preliminary risk score using a "lightweight real-time analysis model." When the score exceeds a threshold, a local alert is triggered and the data is uploaded. The cloud agent utilizes high-precision models and user health profiles for in-depth analysis and comprehensive judgment, ultimately issuing instructions to the edge agent.

[0005] This solution represents the mainstream approach to current cloud-edge collaborative ECG analysis, but there are still areas for improvement:

[0006] 1. Limited autonomy at the edge: The core output of the edge is an abstract "risk score" rather than a specific pathological diagnosis. The final decision-making power for early warning is highly centralized in the cloud, which limits the real-time performance of early warnings to network round-trip latency, and the system cannot make effective judgments when the network is down.

[0007] 2. Single model and incomplete processing chain: Edge processing typically deploys only a single classification or regression model, lacking systematic integration of preprocessing steps such as signal quality assessment and multi-dimensional feature extraction. When signal quality is poor, the performance of a single model will drop sharply, potentially leading to false positives or false negatives.

[0008] 3. Rigid Collaboration Mechanism: The collaboration logic is mainly based on fixed threshold-triggered uploads, lacking intelligence. For example, it cannot dynamically adjust the strategy for interaction with the cloud (such as uploaded content and communication frequency) based on the device's judgment of signal quality and abnormal event types.

[0009] 4. Lack of on-device closed-loop processing capability: The system design goal is still to serve the deep analysis in the cloud, and it fails to achieve the full-process, closed-loop, real-time processing from high-quality signals to clear warning categories on the device side, which is the core requirement of wearable devices in mobile scenarios. Summary of the Invention

[0010] The purpose of this invention is to solve the technical problems of existing cloud-edge collaborative ECG analysis devices, which are caused by weak autonomous decision-making capabilities at the edge (relying on the cloud for final judgment), lack of signal quality assessment (high false alarm rate under noise interference), rigid cloud-edge collaborative strategies (fixed threshold triggering upload), high early warning response delay (unable to meet the second-level response requirements for malignant arrhythmias), poor analysis reliability in complex scenarios (false alarms / missed alarms caused by motion artifacts, etc.), high risk of privacy leakage of users' original physiological data, and high power consumption of device communication.

[0011] To achieve the above objectives, the present invention employs the following technical solution:

[0012] This invention provides a cloud-edge collaborative intelligent ECG analysis device based on a multi-model pipeline, comprising:

[0013] ECG signal acquisition unit: configured to acquire raw ECG signals;

[0014] Signal quality assessment unit: includes a first dedicated neural network model, configured to perform real-time quality assessment on the original ECG signal segment and output the assessment result. When the assessment result indicates that the signal is unavailable, the original ECG signal of that segment terminates the subsequent analysis process.

[0015] Preprocessing unit: configured to filter the signal when the evaluation result indication signal is available;

[0016] Analysis and processing unit: includes a second dedicated neural network model and a third dedicated neural network model cascaded in a pipeline manner. The second dedicated neural network model is configured to extract multi-dimensional features from the filtered signal, and the third dedicated neural network model is configured to perform multi-label heart rhythm classification based on the features and output classification confidence.

[0017] Early warning decision unit: configured to determine the early warning level and trigger the corresponding early warning action according to the classification confidence level and historical cardiac rhythm event statistics, based on preset grading rules;

[0018] Communication Management Unit: Configured to determine the data interaction content with the cloud according to the evaluation results and warning level and a preset strategy. The data interaction content includes: periodically uploading desensitized health summary data under normal conditions, uploading encrypted original signal fragments when preset trigger conditions are met, and receiving model update data sent from the cloud.

[0019] The device is configured to independently execute the entire process from signal acquisition to early warning triggering when there is no network connection.

[0020] In the above scheme, the signal quality evaluation unit using the first dedicated neural network model is a bidirectional gated recurrent unit network. The bidirectional gated recurrent unit network includes, in sequence according to the signal processing order, an input layer, a normalization layer, a bidirectional GRU layer, a fully connected layer, and a Sigmoid output layer.

[0021] The input layer is configured to receive raw electrocardiogram signal segments divided according to a preset time window;

[0022] The bidirectional GRU layer includes parallel forward GRU sub-paths and backward GRU sub-paths, each of which is composed of multiple cascaded unidirectional GRU units; the unidirectional GRU unit is configured to operate at time step t as follows:

[0023] Reset Door:

[0024] Update Gate:

[0025] Candidate hidden state:

[0026] Final hidden state:

[0027] in, It is the sigmoid activation function. This indicates element-wise multiplication. , , , , , For trainable parameters, Includes time The sequence information up to this point, and the input vector at time step t;

[0028] The forward GRU subpath outputs the hidden state sequence:

[0029]

[0030] Backward GRU sub-path output of hidden state sequence:

[0031]

[0032] Hidden state at the final moment of splicing:

[0033]

[0034] The fully connected layer is configured to perform dimension mapping processing on the spliced ​​features and output a mapping value with a dimension of 1.

[0035] The Sigmoid output layer is configured to activate the mapped value and output a signal quality score (SQS) with a value range of [0,1]. The SQS operation configuration is as follows:

[0036]

[0037] in These are the weight parameters of the fully connected layer. These are the bias parameters for the fully connected layer;

[0038] When SQS < 0.8, an "unavailable" control signal is generated, which directly disables the enable port of the preprocessing unit and blocks the data flow to subsequent modules.

[0039] When SQS≥0.8, an "available" control signal is generated, and the original window signal is completely transmitted to the preprocessing unit for 50Hz power frequency notch filtering and baseline drift correction.

[0040] In the above scheme, the analysis and processing unit includes a second dedicated neural network model and a third dedicated neural network model cascaded in a pipeline manner:

[0041] The second neural network model is the front end of a one-dimensional SqueezeNet, and its structure, in order of signal processing, is as follows:

[0042] Initial convolutional layer: 16 convolutional kernels of size 3×1 are used, configured to extract local waveform features of electrocardiogram signals;

[0043] First max pooling layer: stride of 2, configured to reduce the temporal resolution of the feature map;

[0044] The first Fire module and the second Fire module, each Fire module contains a 1×1 convolutional squeezing layer and a parallel 1×1 and 3×1 convolutional expansion layer, configured to extract multi-scale morphological features with low parameter count.

[0045] The second max pooling layer, with a stride of 2, is configured to further expand the receptive field and reduce the feature map dimension.

[0046] The third neural network model is the backend part of a one-dimensional SqueezeNet, and its structure, in the order of signal processing, is as follows:

[0047] The third and fourth Fire modules are configured to perform non-linear transformations on front-end features to capture rhythmic patterns;

[0048] The global average pooling layer is configured to compress the temporal features of each channel into a fixed-length feature vector.

[0049] The fully connected layer is configured to map the feature vector into an N-dimensional classification confidence vector, corresponding to N preset arrhythmia types.

[0050] This architecture, based on the multi-scale feature capture characteristics of ECG signals, balances computational efficiency and real-time performance, while also addressing resource constraints on edge devices and overall computational efficiency. SqueezeNet itself is a lightweight network designed for mobile / embedded devices. The Fire module significantly reduces the number of parameters (using 1×1 convolutions for dimensionality reduction) through a "squeeze-expand" strategy, maintaining accuracy while significantly reducing memory usage and computational load. The initial 3×1 convolutions are used to capture local morphologies such as the P wave, QRS complex, and T wave. The parallel use of 1×1 and 3×1 convolutions in the Fire module is used to simultaneously extract point features and short-term patterns. The progressive pooling design is used to gradually expand the receptive field to identify long-term features such as the RR interval. The front-end pooling layer strategically downsamples, accelerating processing while preserving key morphologies, while global average pooling compresses temporal features to a fixed dimension, eliminating the influence of signal length differences, thus achieving a balance between efficiency and real-time performance.

[0051] In the above scheme, the Fire module in the second and third dedicated neural network models processes data according to the following formula:

[0052] Extrusion layer output:

[0053] Output of the first branch of the extended layer:

[0054] The second branch output of the extended layer:

[0055] Final output of the Fire module:

[0056] in For one-dimensional convolution operations with a kernel size of 1×1, For one-dimensional convolution operations with a kernel size of 3×1, Linear rectified activation function, For batch normalization operations;

[0057] First Fire module The input of the second Fire module As the input of the third Fire module, the third Fire module The fourth Fire module is used as the input; the output of the fourth Fire module is generated into a 64-dimensional feature vector F through global average pooling, and F is input into the fully connected layer to output the classification confidence vector C.

[0058] In the above scheme, the logic for determining the early warning level includes:

[0059] Level 1 alert: Occasional premature beats with low clinical risk are detected and recorded only in the device log or displayed as a faint icon on the screen;

[0060] Level 2 alert: When short runs of atrial fibrillation or paired premature beats are detected, a gentle vibration is triggered as a reminder, and a clear warning message is displayed on the screen;

[0061] Level 3 alarm: If persistent ventricular tachycardia, severe bradycardia, or prolonged signal loss is detected, a strong, continuous vibration and sound alarm will be triggered, and a red warning will be displayed on the screen.

[0062] In the above scheme, the data interaction strategy includes:

[0063] The encrypted upload of the original signal fragment will be automatically triggered when any of the following conditions are met:

[0064] a. Triggered a Level 3 alarm;

[0065] b. Multiple Level 2 warnings are triggered consecutively;

[0066] c. An unknown morphological heart rhythm was detected with low confidence in the classification model but high signal quality determined by the SQA model;

[0067] d. User manually triggers upload.

[0068] This invention also provides a cloud-edge collaborative intelligent ECG analysis method based on a multi-model pipeline, comprising:

[0069] ECG signal acquisition: Configured to acquire raw ECG signals;

[0070] Signal quality assessment: including a first dedicated neural network model, configured to perform real-time quality assessment on the original ECG signal segment and output the assessment result. When the assessment result indicates that the signal is unavailable, the original ECG signal of that segment terminates the subsequent analysis process.

[0071] Preprocessing: Configured to filter the signal when the evaluation result indication signal is available;

[0072] Analysis and processing: This includes a second dedicated neural network model and a third dedicated neural network model cascaded in a pipeline manner. The second dedicated neural network model is configured to extract multi-dimensional features from the filtered signal, and the third dedicated neural network model is configured to perform multi-label heart rhythm classification based on the features and output classification confidence.

[0073] Early warning decision: The configuration is to determine the early warning level and trigger the corresponding early warning action according to the classification confidence level and historical cardiac rhythm event statistics based on the preset classification rules;

[0074] Communication Management: Configured to determine the data interaction content with the cloud according to the evaluation results and warning level and a preset strategy. The data interaction content includes: periodically uploading desensitized health summary data under normal conditions, uploading encrypted original signal fragments when preset trigger conditions are met, and receiving model update data sent from the cloud.

[0075] The device is configured to independently execute the entire process from signal acquisition to early warning triggering when there is no network connection.

[0076] In the above scheme, the signal quality evaluation unit using the first dedicated neural network model is a bidirectional gated recurrent unit network. The bidirectional gated recurrent unit network includes, in sequence according to the signal processing order, an input layer, a normalization layer, a bidirectional GRU layer, a fully connected layer, and a Sigmoid output layer.

[0077] The input layer is configured to receive raw electrocardiogram signal segments divided according to a preset time window;

[0078] The bidirectional GRU layer includes parallel forward GRU sub-paths and backward GRU sub-paths, each of which is composed of multiple cascaded unidirectional GRU units; the unidirectional GRU unit is configured to operate at time step t as follows:

[0079] Reset Door:

[0080] Update Gate:

[0081] Candidate hidden state:

[0082] Final hidden state:

[0083] in, It is the sigmoid activation function. This indicates element-wise multiplication. , , , , , For trainable parameters, Includes time The sequence information up to this point, and the input vector at time step t;

[0084] The forward GRU subpath outputs the hidden state sequence:

[0085]

[0086] Backward GRU sub-path output of hidden state sequence:

[0087]

[0088] Hidden state at the final moment of splicing:

[0089]

[0090] The fully connected layer is configured to perform dimension mapping processing on the spliced ​​features and output a mapping value with a dimension of 1.

[0091] The Sigmoid output layer is configured to activate the mapped value and output a signal quality score (SQS) with a value range of [0,1]. The SQS operation configuration is as follows:

[0092]

[0093] in These are the weight parameters of the fully connected layer. These are the bias parameters for the fully connected layer;

[0094] When SQS < 0.8, an "unavailable" control signal is generated, which directly disables the enable port of the preprocessing unit and blocks the data flow to subsequent modules.

[0095] When SQS≥0.8, an "available" control signal is generated, and the original window signal is completely transmitted to the preprocessing unit for 50Hz power frequency notch filtering and baseline drift correction.

[0096] In the above scheme, the analysis and processing includes a second dedicated neural network model and a third dedicated neural network model cascaded in a pipeline manner:

[0097] The second neural network model is the front end of a one-dimensional SqueezeNet, and its structure, in order of signal processing, is as follows:

[0098] Initial convolutional layer: 16 convolutional kernels of size 3×1 are used, configured to extract local waveform features of electrocardiogram signals;

[0099] First max pooling layer: stride of 2, configured to reduce the temporal resolution of the feature map;

[0100] The first Fire module and the second Fire module, each Fire module contains a 1×1 convolutional squeezing layer and a parallel 1×1 and 3×1 convolutional expansion layer, configured to extract multi-scale morphological features with low parameter count.

[0101] The second max pooling layer, with a stride of 2, is configured to further expand the receptive field and reduce the feature map dimension.

[0102] The third neural network model is the backend part of a one-dimensional SqueezeNet, and its structure, in the order of signal processing, is as follows:

[0103] The third and fourth Fire modules are configured to perform non-linear transformations on front-end features to capture rhythmic patterns;

[0104] The global average pooling layer is configured to compress the temporal features of each channel into a fixed-length feature vector.

[0105] The fully connected layer is configured to map the feature vector into an N-dimensional classification confidence vector, corresponding to N preset arrhythmia types.

[0106] In the above scheme, the Fire module in the second and third dedicated neural network models processes data according to the following formula:

[0107] Extrusion layer output:

[0108] Output of the first branch of the extended layer:

[0109] The second branch output of the extended layer:

[0110] Final output of the Fire module:

[0111] in For one-dimensional convolution operations with a kernel size of 1×1, For one-dimensional convolution operations with a kernel size of 3×1, Linear rectified activation function, For batch normalization operations;

[0112] First Fire module The input of the second Fire module As the input of the third Fire module, the third Fire module The fourth Fire module is used as the input; the output of the fourth Fire module is generated into a 64-dimensional feature vector F through global average pooling, and F is input into the fully connected layer to output the classification confidence vector C.

[0113] In the above scheme, the logic for determining the early warning level includes:

[0114] Level 1 alert: Occasional premature beats with low clinical risk are detected and recorded only in the device log or displayed as a faint icon on the screen;

[0115] Level 2 alert: When short runs of atrial fibrillation or paired premature beats are detected, a gentle vibration is triggered as a reminder, and a clear warning message is displayed on the screen;

[0116] Level 3 alarm: If persistent ventricular tachycardia, severe bradycardia, or prolonged signal loss is detected, a strong, continuous vibration and sound alarm will be triggered, and a red warning will be displayed on the screen.

[0117] Because the present invention adopts the above-described technical solution, it has the following beneficial effects:

[0118] 1. This invention solves the problems of high early warning latency and high dependence on the network in existing technologies by constructing a complete processing pipeline on the device side (steps 2 to 5), achieving the effect of real-time detection and multi-level local early warning of various common arrhythmias without cloud involvement. This design fundamentally eliminates the impact of network latency, ensuring the immediacy of critical early warnings, especially enabling rapid response when malignant arrhythmias occur.

[0119] 2. This invention addresses the problem of low-quality, high-noise data interfering with analysis results by introducing a specialized signal quality assessment model as the "first checkpoint" in the processing pipeline (step 2.1), enabling the identification and discarding of unreliable data segments. This mechanism significantly increases the likelihood that subsequent analysis will be based solely on reliable signals, thereby significantly reducing the false alarm rate and improving the accuracy and reliability of the system.

[0120] 3. This invention addresses the rigidity and lack of flexibility of traditional cloud-edge collaboration solutions by changing the single "threshold upload" mode to an intelligent collaboration strategy based on multi-dimensional information such as event type, signal quality, and historical context (step 5). This strategy enables the device to autonomously decide when and with what data granularity to interact with the cloud, achieving an optimal balance between efficiency and performance while reducing unnecessary communication overhead.

[0121] 4. This invention reduces unnecessary raw data uploads through local closed-loop processing and selectively uploads high-value information (step 5), solving the device's battery life problem and enhancing user privacy protection. This design not only reduces the power consumption of the communication module but also maximizes the protection of the user's raw physiological data privacy, improving user experience and trust.

[0122] 5. This invention integrates the entire process of signal acquisition, quality assessment, feature extraction, classification decision-making, and early warning triggering at the device end (steps 1 to 4), and forms a bidirectional, intelligently collaborative organic whole with the cloud platform (steps 4.1 to 4.2), generating a synergistic effect and solving the drawbacks of poor autonomy of edge devices and centralized cloud computing in traditional architectures. This design ensures both real-time performance and independence, while fully utilizing the powerful computing resources and data analysis capabilities of the cloud, achieving continuous optimization and adaptive improvement of system performance. Furthermore, by using abstract feature vectors as a key intermediary, it perfectly combines "edge autonomous decision-making capabilities" with "cloud continuous optimization capabilities," effectively addressing the privacy and bandwidth challenges of traditional cloud diagnostics, while also overcoming the dilemma of pure edge solutions being unable to iteratively optimize. Attached Figure Description

[0123] Figure 1 : A simplified flowchart of the invention;

[0124] Figure 2 : Overall architecture diagram of a miniature classification network;

[0125] Figure 3 Fire module structure diagram;

[0126] Figure 4 : An electrocardiogram signal segment containing motion artifacts and its probability curve output by an RNN quality assessment model. Detailed Implementation

[0127] The embodiments of the present invention will be described in detail below. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, any modifications or equivalent substitutions made to the present invention should be covered within the scope of the claims of the present invention.

[0128] Furthermore, to better illustrate the present invention, numerous specific details are set forth in the following detailed embodiments. Those skilled in the art will understand that the present invention can be practiced without these specific details.

[0129] This invention aims to solve the following key technical problems existing in cloud-edge collaborative electrocardiogram analysis devices:

[0130] 1. Solve the problems of early warning delay and network dependence: By building a complete processing pipeline on the device, it can realize real-time detection and multi-level local early warning of various common arrhythmias without cloud involvement, fundamentally eliminating the impact of network delay and ensuring the timeliness of critical early warnings.

[0131] 2. Improve the reliability of analysis in complex scenarios: By introducing a dedicated signal quality assessment model as the "first checkpoint" of the processing pipeline, low-quality and high-noise data can be identified and discarded, ensuring that subsequent analysis is based only on reliable signals, thus significantly reducing the false alarm rate.

[0132] 3. Achieve intelligent and differentiated cloud-edge collaboration: Move beyond the single "threshold upload" model and establish an intelligent collaboration strategy based on multi-dimensional information such as event type, signal quality, and historical context. This enables devices to autonomously decide when and with what data granularity (such as raw signal fragments, extracted features, and model intermediate layer outputs) to interact with the cloud, achieving an optimal balance between efficiency and effectiveness.

[0133] 4. Optimize device battery life and privacy: Reduce unnecessary raw data uploads through local closed-loop processing, and selectively upload high-value information only when collaborating with the cloud, thereby reducing the power consumption of the communication module and protecting the user's raw physiological data privacy to the greatest extent.

[0134] This invention provides a cloud-edge collaborative intelligent ECG analysis device and method based on a multi-model pipeline. The device architecture is as follows: Figure 1 As shown, the core lies in the construction of an embedded intelligent processing pipeline on the device side, which consists of multiple micro neural network models, and forms an organic whole with the cloud intelligent platform that optimizes in both directions and collaborates intelligently.

[0135] Embedded intelligent processing pipeline on the device side

[0136] The device side is the core of this invention for achieving real-time and offline early warning, and its processing flow strictly follows... Figure 1 The production line shown.

[0137] Step 1: ECG signal acquisition

[0138] By integrating an analog front-end (AFE) and electrodes into wearable devices such as smartwatches and chest patches, raw single-lead or multi-lead electrocardiogram signals from users are acquired at a certain sampling rate (such as 125 Hz or 250 Hz).

[0139] Step 2: Signal quality assessment and preprocessing

[0140] The raw signal is first input into the "Signal Quality Assessment (SQA) model." This model is used to determine in real time whether the signal is usable for analysis within the current time window (e.g., 5 seconds). Its judgment criteria include baseline drift, motion artifacts, and the severity of power frequency interference. If the assessment is "low quality," the data segment is marked and discarded, not entering the subsequent analysis chain, and a slight notification of "poor signal" will be displayed on the device interface. If the assessment is "usable," the signal enters a fixed digital filter bank for power frequency filtering and baseline correction.

[0141] Step 2.1: Detailed Design of the Signal Quality Assessment (SQA) Model

[0142] This proposal presents an SQA model that employs a recurrent neural network structure based on gated recurrent units (GRUs) to fully leverage the time dependence of ECG signals and accurately identify noise interference patterns. Its specific network structure and signal flow processing logic are as follows:

[0143] Input layer: Receives a raw electrocardiogram signal within a fixed-length time window, denoted as... ,in This represents the number of sampling points. The input signal is processed without any preprocessing to preserve the original noise characteristics.

[0144] Normalization layer: Normalizes the input signal by performing maximum and minimum values, mapping its amplitude to the interval [−1,1] to accelerate model convergence and improve generalization ability. The normalization layer is only used with the initial input signal.

[0145]

[0146] Bidirectional GRU layer: Normalized sequence Send in a with A bidirectional GRU layer with hidden units. The bidirectional GRU captures past and future contextual information through two independent GRU networks, one for forward and one for backward. The computation process for each time step t is as follows:

[0147]

[0148] This represents the input vector at time step t. This represents the computation function of the feedforward GRU network, which processes the sequence from the beginning to the end, capturing information from "past to present". The computation function of the feedforward GRU network is represented. The vector concatenation operation concatenates the forward and backward hidden states along the feature dimension to form a complete hidden state containing bidirectional contextual information.

[0149] The gating mechanism within the GRU unit is defined by the following formula:

[0150]

[0151]

[0152]

[0153]

[0154] It is the reset door, which determines how new inputs are combined with past memories.

[0155] It is the update gate, which determines the proportion of past memories retained in the current state.

[0156] It is a candidate hidden state. Combined with the reset gate information, the candidate state of the current time step is calculated.

[0157] The final hidden state, combined with the update gate, updates the final state at the current moment.

[0158] in, It is the sigmoid activation function. This indicates element-wise multiplication. , , , , , For trainable parameters, Includes time The sequence information up to this point, and the input vector at time step t;

[0159] The output of the bidirectional GRU is a concatenation of the hidden states at each time step. We take the hidden state at the last time step. As a global context representation of the entire sequence.

[0160] Fully connected layer: The input is passed to a fully connected layer, which reduces the dimension from... Mapping to 1:

[0161]

[0162] in , .

[0163] Output layer: Activated by the Sigmoid function Convert to quality score :

[0164] S

[0165] Preset threshold (Usually taken as 0.8), when If the signal is high quality, it is considered a high-quality signal; otherwise, it is considered a low-quality signal.

[0166] Model Training and Deployment: This SQA model is trained in the cloud using a large-scale labeled ECG noise dataset (labeled as "high-quality" or "low-quality"), employing a binary classification cross-entropy loss function. After training, the model is compressed to a size suitable for microcontroller deployment (typically less than 50KB) through model pruning and 8-bit quantization, and then embedded in the device's firmware. This model has the following innovative features:

[0167] 1. Temporal noise modeling: GRU is used to model the temporal dynamics of noise, which can distinguish between transient disturbances (such as loose electrodes) and continuous noise (such as motion artifacts).

[0168] 2. Lightweight design: Hidden number of units It can be set to 32 or 64, resulting in a very low number of model parameters, which meets the power consumption and memory constraints of wearable devices.

[0169] 3. Independent of subsequent analysis: As the first checkpoint in the pipeline, it ensures that only high-quality data enters downstream analysis, fundamentally improving the reliability of the system.

[0170] Step 3: Miniature Neural Network Analysis Engine

[0171] This is the core of the pipeline, consisting of multiple lightweight models with clearly defined functions, cascaded or paralleled:

[0172] Feature extraction model: Receives the pre-processed clean waveform and extracts multi-dimensional feature vectors in parallel, including time domain, frequency domain, and nonlinear features.

[0173] Multi-label heart rhythm classification model: It takes feature vectors as input and outputs a multi-dimensional vector, which represents the confidence score of various preset arrhythmias (such as normal sinus rhythm, atrial fibrillation, premature ventricular contractions, bradycardia, tachycardia, etc.) within the current time window.

[0174] Step 3.1: Refinement of the Micro Neural Network Analysis Engine: A Variant Based on 1D SqueezeNet

[0175] The "feature extraction model" and "multi-label heart rhythm classification model" in this proposal are not independent, but together constitute an end-to-end micro-classification network, the overall architecture of which is as follows: Figure 2 As shown, this network takes pre-processed clean ECG waveforms as input and directly outputs confidence scores for various arrhythmias. The entire network follows the lightweight design philosophy of SqueezeNet, greatly compressing the number of parameters while maintaining high accuracy by stacking multiple 1D Fire Modules.

[0176] (1) Overall architecture and signal flow

[0177] The network input is a fixed-length electrocardiogram (ECG) segment. ,in This represents the number of sampling points. The signal stream passes through the following layers in sequence:

[0178] Input layer: Receives the clean ECG waveform after preprocessing in step 2, with the following shape: .

[0179] Convolutional layer 1 (Conv1): Uses 16 layers of size... One-dimensional convolution is performed using a convolutional kernel with a stride of 1 and padding of 1, resulting in an output feature map of size [size missing]. This layer aims to extract low-level local features of the ECG signal, such as the slope and width of the QRS wave. The activation function used is ReLU, followed by batch normalization (BatchNorm).

[0180] Max Pooling Layer 1 (First Max Pooling Layer MaxPool1): Uses a size of A pooling window with a stride of 2 is used to downsample the feature map, and the output size is [size missing]. This layer reduces feature dimensionality, expands the receptive field, and enhances the translation invariance of features.

[0181] Fire Module 1 (First Fire Module): 16 input channels, 32 output channels (output by the Expand layer). This module compresses the number of channels using a Squeeze layer (1x1 convolution) and then extracts multi-scale features using an Expand layer (parallel 1x1 and 3x1 convolutions). Detailed structure is shown below. Output size is... .

[0182] Fire Module 2 (First Fire Module): 32 input channels, 48 ​​output channels. Further abstracted features, dimensions unchanged.

[0183] MaxPool2 (Second MaxPool2): Same as MaxPool1, with an output size of... .

[0184] Fire Module 3 (First Fire Module): Input 48, Output 64.

[0185] Fire Module 4 (First Fire Module): Input 64, Output 80.

[0186] Global AvgPool1D: Averaging the feature maps of each channel along the time dimension to obtain a fixed-length feature vector. This vector is the output of the "feature extraction model," which aggregates global time-frequency information of the electrocardiogram signal, including morphological features, rhythm features, etc.

[0187] Fully connected layer (FC): converts feature vectors Mapping to the number of categories (For example, C=6 corresponds to normal, atrial fibrillation, premature ventricular contractions, bradycardia, tachycardia, and others). The output is unnormalized logits.

[0188] Output layer: The Softmax function is used to convert logits into confidence scores for each category. ,satisfy .

[0189] The signal flow of the entire network can be summarized as: original waveform - local feature extraction - multi-scale feature abstraction - global feature aggregation - classification output.

[0190] (2) Core Module: Detailed Structure of the Fire Module

[0191] Each Fire module consists of two sub-layers: "Squeeze" and "Expand," with the structure as follows: Figure 3 As shown, the mathematical process is as follows:

[0192] Squeeze layer: Use Convolution kernel on input feature map Perform dimensionality reduction and output feature maps. ,in (usually taken) (or smaller). This layer significantly reduces the computational cost of subsequent Expand layers by compressing the number of channels. The activation function is ReLU.

[0193]

[0194] Expand layer: Processes the output of the Squeeze layer in parallel using two sets of convolutions.

[0195] Branch 1: Convolution, number of output channels (usually taken) )

[0196] Branch 2: Convolution, number of output channels (usually taken) Both convolutional branches are followed by batch normalization and ReLU activation. The outputs of the two branches are then concatenated along the channel dimension to obtain the final output.

[0197]

[0198] in .

[0199] E1 represents a 1×1 convolution operation on the output S of the Squeeze layer, followed by normalization and activation. E2 represents a 3×1 convolution operation on the output S of the Squeeze layer, followed by normalization and activation. The activation function introduces non-linearity. The Concat function concatenates the outputs of the two branches along the channel dimension. The purpose of this design is to fuse ECG signal features at multiple scales, significantly reducing the number of parameters while better adapting to ECG signals.

[0200] Inside each Fire module, input and output Skip connections can be introduced (when the number of input and output channels is the same) to alleviate the gradient vanishing problem and promote information flow.

[0201] This design is optimized for the characteristics of ECG signals: 1) Multi-scale convolutions (1x1 and 3x1) are used to simulate the superposition of different frequency components in the ECG waveform; 2) The receptive field is gradually expanded by layer-by-layer dimensionality reduction, enabling the network to simultaneously perceive local heartbeat morphology and overall rhythm changes; 3) Global average pooling replaces the flattening operation before fully connected layers, greatly reducing the number of parameters and preventing overfitting. Finally, the model can achieve real-time inference on a microcontroller (single inference time <10ms), meeting the power consumption constraints of wearable devices.

[0202] Step 4: Multi-level Early Warning Decision Module. This module is a rule-based decision maker that receives the confidence score output by the classification model and makes a comprehensive judgment based on historical context (the frequency of similar events in the previous 10 minutes). Its decision logic includes:

[0203] Early warning tiers: Multiple early warning levels are defined based on the type and persistence of the anomaly. For example:

[0204] Level 1 alert: Occasional, low-clinical-risk premature beats are detected and recorded only in the device log or displayed as a faint icon on the screen.

[0205] Level 2 alert: When short runs of atrial fibrillation or paired premature beats are detected, a gentle vibration is triggered as a reminder, and a clear warning message is displayed on the screen.

[0206] Level 3 alarm: If persistent ventricular tachycardia, severe bradycardia, or prolonged signal loss is detected, a strong, continuous vibration and sound alarm will be triggered, and a red warning will be displayed on the screen.

[0207] Contextual awareness: Avoid overreacting to isolated, sporadic events. For example, only escalate a ventricular premature beat from "recorded" to "alert" when a high frequency of premature ventricular beats is detected within a given time period.

[0208] Step 5: Device-side communication management module

[0209] This module is responsible for managing the intelligent collaboration between devices and the cloud, and its strategy is driven by the outputs of steps 2 and 4:

[0210] Normal mode: When there are no abnormalities or low-level abnormalities, the device only periodically (e.g., once a day) uploads an encrypted "health summary" such as de-identified feature vector summaries and event type statistical logs to the cloud to update the user's health profile, without uploading any original waveforms.

[0211] Intelligent Triggered Upload Mode: The encrypted upload of the original signal fragment is automatically triggered when any of the following conditions are met:

[0212] a. Triggered a Level 3 alarm.

[0213] b. Multiple Level 2 warnings are triggered consecutively.

[0214] c. An unknown morphological rhythm was detected, but the SQA model determined the signal quality to be high despite the low confidence level of the classification model.

[0215] d. User manually triggers upload.

[0216] Model Update Reception: Securely receive and verify incremental update packages of model parameters sent from the cloud, used to update one or more micro neural network models on the device.

[0217] 4.1 Cloud-based intelligent platform

[0218] The role of the cloud platform has completely transformed from a "decision and diagnosis center" in the traditional architecture to a "capability support and optimization center." Its core function is to empower edge devices, rather than replace their decision-making. No raw ECG waveforms are transmitted between the cloud and the device; all interactions are based on highly abstracted and anonymized data, ensuring user privacy and data transmission efficiency. The following details the implementation steps of each functional module in the cloud to ensure its technical feasibility.

[0219] 4.1.1 Model Training, Optimization and Performance Evaluation Center

[0220] This module is responsible for continuously improving the performance of the device-side model using de-identified group data, and specifically includes the following sub-steps:

[0221] Step 1: Basic Teacher Model Training

[0222] Data preparation: A large-scale, high-quality electrocardiogram (ECG) database is maintained on the cloud platform (which may include publicly available datasets such as MIT-BIH, as well as clinical data that has been explicitly authorized and anonymized by the user). ECG signals in this database are labeled with multiple arrhythmia categories.

[0223] Model selection: A deep learning model with a large number of parameters and a complex structure was selected as the teacher model, based on the ECG analysis network of Transformer.

[0224] Training process: The teacher model was trained end-to-end using the aforementioned database, with the optimization objective being multi-class cross-entropy loss. After training, the teacher model achieved extremely high classification accuracy and can be used as a knowledge source.

[0225] Step 2: Knowledge distillation to generate a lightweight student model

[0226] Distillation Dataset Construction: A large number of unlabeled electrocardiogram signal fragments were randomly extracted from a cloud database to form the distillation dataset. .

[0227] Soft label generation: Distill each sample in the dataset Input the teacher model and obtain its output probability distribution. (i.e., "soft label"), by setting higher temperature parameters Soften the probability distribution to highlight the similarity between categories.

[0228] Student model initialization: The student model is the 1D SqueezeNet variant (structure as described in 3.1) that is finally deployed on the device, and the weights are randomly initialized.

[0229] Distillation Training: The training objective of the student model is to jointly optimize the hard label loss and the soft label loss.

[0230]

[0231] in, The cross-entropy loss is used if the distilled samples have true labels. KL divergence loss is used to make the softened output of the student model approximate the softened output of the teacher model. Hyperparameters Balance the two (e.g., set it to 0.3). The training temperature is consistent with the teacher model.

[0232] Post-training compression: After distillation training, the student model undergoes further lightweighting.

[0233] Weight quantization: The model weights are quantized from 32-bit floating-point to 8-bit integers (INT8). Layer-by-layer or channel-by-channel quantization calibration can be used to reduce model size and inference memory usage.

[0234] Model pruning: Based on the L1 norm or gradient information of the weights, remove convolutional kernels with lower importance, and then fine-tune to restore accuracy.

[0235] Step 3: Model Performance Evaluation

[0236] Offline validation: Use an independent test set to evaluate the accuracy, recall, and other metrics of the compressed student model, and compare it with the previous version of the model to ensure that performance has not degraded.

[0237] Online monitoring: The cloud receives anonymized classification confidence scores (i.e., the probability distribution output by the student model) and corresponding event logs (such as whether users have confirmed warnings) reported by the devices. By analyzing the consistency between the confidence scores and actual events, systematic biases in the model are identified at the group level. For example, if the average confidence score of a certain type of event is low but the reporting frequency is high, it indicates that the model may be underfitting to that type of event, triggering the model optimization process.

[0238] 4.1.2 Group Health Profiling and Trend Analysis Platform

[0239] This platform utilizes massive amounts of desensitized feature vectors (rather than the original signal) for population-level health analysis. The specific steps include:

[0240] Feature clustering: Unsupervised clustering (such as K-Means or DBSCAN) is performed on the 80-dimensional feature vector (from the global average pooling layer) uploaded by the device to identify user groups with similar ECG morphology.

[0241] Time series analysis: Aggregate the characteristic distribution changes of each group according to the time dimension, and extract macro indicators such as the group mean of heart rate variability (HRV) and the incidence of abnormal events.

[0242] Risk pattern discovery: Using association rule mining or time series pattern discovery algorithms, we identify characteristic patterns or external factors (such as time and season) associated with specific abnormal events (such as high incidence of atrial fibrillation) and generate health insight reports.

[0243] Adaptive Threshold Adjustment: Based on population analysis results, the cloud can provide dynamic adjustment suggestions for the warning thresholds on the device. For example, if it is found that the incidence of bradycardia at night is generally high among users in a certain age group, it can be suggested that the warning threshold for bradycardia be appropriately relaxed to reduce false alarms. This suggestion is issued in the form of a parameter package, and users decide whether to adopt it.

[0244] 4.1.3 Security Model Distribution and Management Hub

[0245] This module is responsible for securely and reliably distributing the optimized model update package to a massive number of devices. The specific process is as follows:

[0246] Model Encryption and Signing: The lightweight model parameters (typically a few hundred KB) to be distributed are encrypted using the device's public key, ensuring that only the corresponding device can decrypt them. Simultaneously, the encrypted package is digitally signed using a private key from the cloud to prevent tampering.

[0247] Version management: A cloud-based model version database is maintained, recording each version number, applicable device type, update log, etc. Devices periodically (e.g., daily) query the cloud for new version availability.

[0248] Phased distribution: To avoid network congestion, a phased distribution strategy is adopted: first push the update to a small number of devices, monitor their feedback (such as update success rate, model performance), and expand the distribution scope after confirming that there are no errors.

[0249] Device-side update: During idle periods (such as when charging and connected to Wi-Fi), the device downloads the encrypted package, verifies the signature, decrypts it, replaces the local model, and the update takes effect upon the next startup. The update process is seamless for the user and supports rollback (if the new model causes performance degradation, it can be restored to the previous version).

[0250] Through the detailed design described above, the cloud platform has achieved continuous optimization and evolution of the edge device model without being completely removed from the original waveform, ensuring the long-term reliability and adaptability of the system.

[0251] Existing technologies employ a "cloud-based diagnostic" model, requiring the uploading of raw ECG waveforms to the cloud for analysis. This poses risks of privacy breaches and bandwidth waste. This patent addresses this by eliminating the transmission of raw ECG waveforms between the cloud and the device. Instead, encrypted signal waveforms are transmitted only when specific conditions are met, achieving a separation of roles: "decision-making power at the edge, optimization power in the cloud." Existing technologies rely on manual annotation for model iteration updates. This patent achieves a closed loop of "deployment-monitoring-discovery-optimization" by analyzing the consistency between classification confidence reported by the device and user confirmation behavior, eliminating the need for manual intervention. Existing technologies typically only perform model compression without considering the extreme resource constraints of edge devices. This patent achieves extreme compression through a three-stage collaborative process of distillation quantization and pruning.

[0252] This patent is characterized by the following: the device independently completes the entire process of signal acquisition, quality assessment, feature extraction, classification decision-making, and early warning triggering, uploading only the fixed-dimensional feature vectors of the intermediate layers of the neural network and the classification confidence distribution to the cloud; the cloud, without ever touching the original signal, automatically identifies model defects and generates a lightweight model update package optimized through three stages of distillation, quantization, and pruning by statistical analysis of the correlation between massive anonymous feature vectors and user feedback, and securely distributes it to the device using a phased gray-scale push strategy, achieving a closed-loop system with privacy protection, low bandwidth consumption, and adaptive optimization. This design perfectly combines "edge autonomous decision-making capability" and "cloud continuous optimization capability" through the key intermediary of "abstract feature vectors," solving both the privacy and bandwidth issues of traditional cloud diagnostics and the dilemma of pure edge solutions being unable to iteratively optimize.

[0253] 4.2 Cloud-Edge Collaboration Workflow

[0254] Independent device operation: In most cases, the embedded pipeline at the device end independently completes the entire process of "signal acquisition - quality assessment - feature extraction - heart rhythm classification - early warning decision", realizing true offline real-time early warning.

[0255] Intelligent collaboration trigger: When the device-side communication management module determines that cloud intervention is required based on preset intelligent conditions, the collaboration process is initiated.

[0256] Two-way data exchange: The device uploads encrypted raw data fragments and context information; after the cloud completes in-depth analysis, it sends the verification results, health trend reports or model update packages to the device or associated smartphone APP.

[0257] Silent model updates: The device safely receives and applies model updates during idle periods (such as when charging), continuously improving the accuracy of subsequent analyses.

[0258] To achieve high-performance real-time analytics on resource-constrained wearable devices, this proposal features a carefully designed core AI module within the embedded processing pipeline, employing and modifying a specific lightweight neural network architecture.

[0259] 4.2.1 Refinement of the signal quality assessment module: RNN-based sequence modeling

[0260] The quality of electrocardiogram (ECG) signals (e.g., stability, noise contamination) is a problem strongly correlated with the time context. This proposal employs recurrent neural networks (RNNs), specifically gated recurrent units (GRUs), to construct a signal quality assessment model, leveraging their powerful temporal modeling capabilities.

[0261] (1) Model Input and Processing: The model takes the raw electrocardiogram (ECG) signal as input through a sliding window (e.g., 250 sampling points corresponding to a 2-second duration). This signal sequence It is fed into a bidirectional GRU layer for encoding.

[0262] (2) Mathematical principles of GRU units

[0263] GRU controls the flow of information by introducing update gates and reset gates, and its cells operate at time steps. The calculation formula is as follows:

[0264] Reset Gate: Determines how to combine new input with past memories.

[0265]

[0266] Update Gate: Determines the proportion of past memories retained in the current state.

[0267]

[0268] Candidate Hidden State: Calculate the candidate state for the current time step by combining the reset gate information.

[0269]

[0270] Final Hidden State: Combined with the update gate, update the final state at the current moment.

[0271]

[0272] in, It is the sigmoid activation function. This indicates element-wise multiplication. and These are trainable parameters. Includes time The sequence information up to this point.

[0273] (3) Architecture and Output

[0274] This module uses a bidirectional GRU layer, which contains both forward and backward GRUs, to capture the contextual dependencies at each time step. Finally, the hidden states of the last time step of the bidirectional GRU are concatenated and passed through a fully connected layer and a sigmoid activation function to output a scalar SQS (Signal Quality Score) between 0 and 1.

[0275]

[0276] When the SQS is higher than the set threshold (e.g., 0.8), the signal quality of that segment is determined to be "reliable"; otherwise, it is determined to be "unreliable".

[0277] Figure 4 This paper presents an electrocardiogram (ECG) signal fragment containing motion artifacts and its probability curve output by an RNN quality assessment model.

[0278] As shown in the figure, during periods of signal purity, the model outputs a high confidence level ( When the signal is interfered with by sudden noise (as shown in the figure) to Between [a certain point in time], the model output probability drops rapidly. This dynamic evaluation capability enables the system to accurately identify and discard low-quality data segments, ensuring the reliability of subsequent analysis.

[0279] By concretizing the micro neural network into a task-adapted 1D SqueezeNet variant and the signal quality assessment model into a GRU-based RNN, the technical solution proposed in this paper is highly feasible and professional, and fully meets the needs of achieving high-precision and high-real-time intelligent ECG analysis in extremely resource-constrained environments.

Claims

1. A cloud-edge collaborative intelligent ECG analysis device based on a multi-model pipeline, characterized in that, include: ECG signal acquisition unit: configured to acquire raw ECG signals; Signal quality assessment unit: includes a first dedicated neural network model, configured to perform real-time quality assessment on the original ECG signal segment and output the assessment result. When the assessment result indicates that the signal is unavailable, the original ECG signal of that segment terminates the subsequent analysis process. Preprocessing unit: configured to filter the signal when the evaluation result indication signal is available; Analysis and processing unit: includes a second dedicated neural network model and a third dedicated neural network model cascaded in a pipeline manner. The second dedicated neural network model is configured to extract multi-dimensional features from the filtered signal, and the third dedicated neural network model is configured to perform multi-label heart rhythm classification based on the features and output classification confidence. Early warning decision unit: configured to determine the early warning level and trigger the corresponding early warning action according to the classification confidence level and historical cardiac rhythm event statistics, based on preset grading rules; Communication Management Unit: Configured to determine the data interaction content with the cloud according to the evaluation results and warning level and a preset strategy. The data interaction content includes: periodically uploading desensitized health summary data under normal conditions, uploading encrypted original signal fragments when preset trigger conditions are met, and receiving model update data sent from the cloud. The device is configured to independently execute the entire process from signal acquisition to early warning triggering when there is no network connection.

2. The apparatus according to claim 1, characterized in that, The signal quality evaluation unit using the first dedicated neural network model is a bidirectional gated recurrent unit network. The bidirectional gated recurrent unit network includes, in sequence, an input layer, a normalization layer, a bidirectional GRU layer, a fully connected layer, and a Sigmoid output layer, according to the signal processing order. The input layer is configured to receive raw electrocardiogram signal segments divided according to a preset time window; The bidirectional GRU layer includes parallel forward GRU sub-paths and backward GRU sub-paths, each of which is composed of multiple cascaded unidirectional GRU units; the unidirectional GRU unit is configured to operate at time step t as follows: Reset Door: Update Gate: Candidate hidden state: Final hidden state: in, It is the sigmoid activation function. This indicates element-wise multiplication. , , , , , For trainable parameters, Includes time The sequence information up to this point, and the input vector at time step t; The forward GRU subpath outputs the hidden state sequence: Backward GRU sub-path output of hidden state sequence: Hidden state at the final moment of splicing: The fully connected layer is configured to perform dimension mapping processing on the spliced ​​features and output a mapping value with a dimension of 1. The Sigmoid output layer is configured to activate the mapped value and output a signal quality score (SQS) with a value range of [0,1]. The SQS operation configuration is as follows: in These are the weight parameters of the fully connected layer. These are the bias parameters for the fully connected layer; When SQS < 0.8, an "unavailable" control signal is generated, which directly disables the enable port of the preprocessing unit and blocks the data flow to subsequent modules. When SQS≥0.8, an "available" control signal is generated, and the original window signal is completely transmitted to the preprocessing unit for 50Hz power frequency notch filtering and baseline drift correction.

3. The apparatus according to claim 1, characterized in that, The analysis and processing unit includes a second and a third dedicated neural network model cascaded in a pipeline manner: The second neural network model is the front end of a one-dimensional SqueezeNet, and its structure, in order of signal processing, is as follows: Initial convolutional layer: 16 convolutional kernels of size 3×1 are used, configured to extract local waveform features of electrocardiogram signals; First max pooling layer: stride of 2, configured to reduce the temporal resolution of the feature map; The first Fire module and the second Fire module, each Fire module contains a 1×1 convolutional squeezing layer and a parallel 1×1 and 3×1 convolutional expansion layer, configured to extract multi-scale morphological features with low parameter count. The second max pooling layer, with a stride of 2, is configured to further expand the receptive field and reduce the feature map dimension. The third neural network model is the backend part of a one-dimensional SqueezeNet, and its structure, in the order of signal processing, is as follows: The third and fourth Fire modules are configured to perform non-linear transformations on front-end features to capture rhythmic patterns; The global average pooling layer is configured to compress the temporal features of each channel into a fixed-length feature vector. The fully connected layer is configured to map the feature vector into an N-dimensional classification confidence vector, corresponding to N preset arrhythmia types.

4. The apparatus according to claim 1, characterized in that, The Fire module in the second and third dedicated neural network models processes data according to the following formula: Extrusion layer output: Output of the first branch of the extended layer: Output of the second branch of the extended layer: Final output of the Fire module: in For one-dimensional convolution operations with a kernel size of 1×1, For one-dimensional convolution operations with a kernel size of 3×1, Linear rectified activation function, For batch normalization operations; First Fire module The input of the second Fire module As the input of the third Fire module, the third Fire module The fourth Fire module is used as the input; the output of the fourth Fire module is generated into a 64-dimensional feature vector F through global average pooling, and F is input into the fully connected layer to output the classification confidence vector C.

5. The apparatus according to claim 1, characterized in that, The logic for determining the warning level includes: Level 1 alert: Occasional premature beats with low clinical risk are detected and recorded only in the device log or displayed as a faint icon on the screen; Level 2 alert: When short runs of atrial fibrillation or paired premature beats are detected, a gentle vibration is triggered as a reminder, and a clear warning message is displayed on the screen; Level 3 alarm: If persistent ventricular tachycardia, severe bradycardia, or prolonged signal loss is detected, a strong, continuous vibration and sound alarm will be triggered, and a red warning will be displayed on the screen.

6. The apparatus according to claim 1, characterized in that, The data interaction strategy includes: The encrypted upload of the original signal fragment will be automatically triggered when any of the following conditions are met: a. Triggered a Level 3 alarm; b. Multiple Level 2 warnings are triggered consecutively; c. An unknown morphological heart rhythm was detected with low confidence in the classification model but high signal quality determined by the SQA model; d. User manually triggers upload.

7. A cloud-edge collaborative intelligent ECG analysis method based on a multi-model pipeline, characterized in that, include: ECG signal acquisition: Configured to acquire raw ECG signals; Signal quality assessment: including a first dedicated neural network model, configured to perform real-time quality assessment on the original ECG signal segment and output the assessment result. When the assessment result indicates that the signal is unavailable, the original ECG signal of that segment terminates the subsequent analysis process. Preprocessing: Configured to filter the signal when the evaluation result indication signal is available; Analysis and processing: This includes a second dedicated neural network model and a third dedicated neural network model cascaded in a pipeline manner. The second dedicated neural network model is configured to extract multi-dimensional features from the filtered signal, and the third dedicated neural network model is configured to perform multi-label heart rhythm classification based on the features and output classification confidence. Early warning decision: The configuration is to determine the early warning level and trigger the corresponding early warning action according to the classification confidence level and historical cardiac rhythm event statistics based on the preset classification rules; Communication Management: Configured to determine the data interaction content with the cloud according to the evaluation results and warning level and a preset strategy. The data interaction content includes: periodically uploading desensitized health summary data under normal conditions, uploading encrypted original signal fragments when preset trigger conditions are met, and receiving model update data sent from the cloud. The device is configured to independently execute the entire process from signal acquisition to early warning triggering when there is no network connection.

8. The method according to claim 7, characterized in that, The first dedicated neural network model is used for signal quality evaluation, which employs a bidirectional gated recurrent unit network. The bidirectional gated recurrent unit network includes, in sequence, an input layer, a normalization layer, a bidirectional GRU layer, a fully connected layer, and a Sigmoid output layer, according to the signal processing order. The input layer is configured to receive raw electrocardiogram signal segments divided according to a preset time window; The bidirectional GRU layer includes parallel forward GRU sub-paths and backward GRU sub-paths, each of which is composed of multiple cascaded unidirectional GRU units; the unidirectional GRU unit is configured to operate at time step t as follows: Reset Door: Update Gate: Candidate hidden state: Final hidden state: in, It is the sigmoid activation function. This indicates element-wise multiplication. , , , , , For trainable parameters, Includes time The sequence information up to this point, and the input vector at time step t; The forward GRU subpath outputs the hidden state sequence: Backward GRU sub-path output of hidden state sequence: Hidden state at the final moment of splicing: The fully connected layer is configured to perform dimension mapping processing on the spliced ​​features and output a mapping value with a dimension of 1. The Sigmoid output layer is configured to activate the mapped value and output a signal quality score (SQS) with a value range of [0,1]. The SQS operation configuration is as follows: in These are the weight parameters of the fully connected layer. These are the bias parameters for the fully connected layer; When SQS < 0.8, an "unavailable" control signal is generated, which directly disables the enable port of the preprocessing unit and blocks the data flow to subsequent modules. When SQS≥0.8, an "available" control signal is generated, and the original window signal is completely transmitted to the preprocessing unit for 50Hz power frequency notch filtering and baseline drift correction.

9. The method according to claim 7, characterized in that, Analysis and processing are performed by cascading a second and a third dedicated neural network model in a pipeline manner: The second neural network model is the front end of a one-dimensional SqueezeNet, and its structure, in order of signal processing, is as follows: Initial convolutional layer: 16 convolutional kernels of size 3×1 are used, configured to extract local waveform features of electrocardiogram signals; First max pooling layer: stride of 2, configured to reduce the temporal resolution of the feature map; The first Fire module and the second Fire module, each Fire module contains a 1×1 convolutional squeezing layer and a parallel 1×1 and 3×1 convolutional expansion layer, configured to extract multi-scale morphological features with low parameter count. The second max pooling layer, with a stride of 2, is configured to further expand the receptive field and reduce the feature map dimension. The third neural network model is the backend part of a one-dimensional SqueezeNet, and its structure, in the order of signal processing, is as follows: The third and fourth Fire modules are configured to perform non-linear transformations on front-end features to capture rhythmic patterns; The global average pooling layer is configured to compress the temporal features of each channel into a fixed-length feature vector. The fully connected layer is configured to map the feature vector into an N-dimensional classification confidence vector, corresponding to N preset arrhythmia types.

10. The method according to claim 1, characterized in that, The Fire module in the second and third dedicated neural network models processes data according to the following formula: Extrusion layer output: Output of the first branch of the extended layer: Output of the second branch of the extended layer: Final output of the Fire module: in For one-dimensional convolution operations with a kernel size of 1×1, For one-dimensional convolution operations with a kernel size of 3×1, Linear rectified activation function, For batch normalization operations; First Fire module The input of the second Fire module As the input of the third Fire module, the third Fire module The fourth Fire module is used as the input; the output of the fourth Fire module is generated into a 64-dimensional feature vector F through global average pooling, and F is input into the fully connected layer to output the classification confidence vector C.

Images