An AI-based full-process ECG detection and diagnosis integrated intelligent auxiliary analysis system and method

Abstract

This invention belongs to the field of medical testing and diagnostic technology, and discloses an AI-based, end-to-end intelligent auxiliary analysis method for electrocardiogram (ECG) testing and diagnosis, comprising: S1, constructing a three-dimensional mapping database, which includes a labeled dataset of samples; S2, positioning of a multimodal acquisition terminal and acquisition of multimodal data; S3, signal analysis and diagnosis through a fusion diagnostic model, including a two-level fusion decision model. In the first-level fusion part, the two-level fusion decision model forms a multimodal fusion feature vector; in the second-level fusion part, it analyzes and obtains diagnostic results and abnormality diagnostic results for ECG signals, and then weightedly fuses them to output the final diagnostic result. This invention can accurately distinguish between electrode displacement and actual pathological abnormalities, reducing the false positive rate, decreasing clinical misdiagnosis, and improving the reliability and efficiency of diagnostic results.

Description

Technical Field

[0001] This invention belongs to the field of medical testing and diagnostic technology, specifically relating to an AI-based integrated intelligent auxiliary analysis system and method for the entire process of electrocardiogram testing and diagnosis. Background Technology

[0002] Electrocardiogram (ECG) monitoring is a core technology for screening, diagnosing, and following up on cardiovascular diseases. It collects electrical activity signals from the human heart, providing crucial information for clinical diagnosis of arrhythmias, myocardial ischemia, and myocardial infarction. Existing ECG diagnostic algorithms are mostly based on fixed feature extraction rules or single deep learning models. For example, using only convolutional neural networks (CNNs) to process ECG signals makes it difficult to consider the complex correlation between temporal signal characteristics and pathological features. Relying solely on a single ECG signal for analysis, without combining it with imaging information such as the patient's thoracic anatomy and body posture, can easily lead to the inability to distinguish between "signal abnormalities caused by electrode displacement" and "true pathological abnormalities," resulting in false positive reports. The article "Research Progress in Multimodal ECG Diagnostic Technology" in *China Medical Equipment*, Vol. 38, No. 8, 2023, points out that diagnostic algorithms based solely on ECG signals have an average false positive rate as high as 18%. Furthermore, traditional ECG monitoring relies on medical staff manually placing lead electrodes, and the accuracy of this placement is easily affected by factors such as operational experience and the patient's chest morphology (e.g., obesity, scoliosis). Studies have shown that a 5mm offset in the chest lead electrodes can reduce the diagnostic sensitivity of myocardial ischemia by more than 30%; some automated lead devices are based solely on surface markers for localization and have not established a correlation model between chest morphology and electrocardiogram signals, thus failing to adapt to individual anatomical differences.

[0003] In summary, existing technologies need to address how to establish a precise correlation between thoracic morphology, cardiac anatomy, and electrocardiogram (ECG) signals to accommodate individual anatomical differences, obtain ECG signals under optimal lead localization, and effectively apply them to compatible detection and diagnostic models to achieve accurate disease detection and diagnosis, thereby reducing the false positive rate and improving the reliability and efficiency of diagnostic results. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-based, end-to-end intelligent auxiliary analysis method for ECG detection and diagnosis, which addresses the technical problem of how to establish a precise correlation between chest morphology, cardiac anatomy, and ECG signals to adapt to individual anatomical differences, obtain ECG signals under optimal lead positioning, and effectively apply them to a compatible detection and diagnosis model to achieve accurate disease detection and diagnosis, thereby reducing the false positive rate and improving the reliability and efficiency of diagnostic results.

[0005] The AI-based integrated intelligent auxiliary analysis method for the entire process of electrocardiogram detection and diagnosis includes the following steps: S1. Construct a three-dimensional mapping database, which includes a labeled dataset of samples. Each sample is associated with a thoracic morphology feature vector, a cardiac anatomy feature vector, an electrocardiogram signal feature vector, and optimal lead position parameters. S2. The positioning and multimodal data acquisition of the multimodal acquisition terminal first extracts the features of the thoracic contour, the features of the heart structure, and the features of the initial electrocardiogram signal. Then, it calls the three-dimensional mapping database to perform three-dimensional electrocardiogram lead positioning, and then performs electrocardiogram adaptive optimization positioning to acquire the final electrocardiogram signal.

[0006] S3. Signal analysis and diagnosis are performed using a fusion diagnostic model, which includes a two-level fusion decision model. The two-level fusion decision model includes: In the first-level fusion section, the thoracic morphology feature vector, cardiac structure feature vector, final electrocardiogram signal feature vector, and clinical information feature vector are concatenated to form a multimodal fusion feature vector; The secondary fusion section includes an ECG signal diagnosis module and a multimodal diagnosis module. The ECG signal diagnosis module analyzes and diagnoses the final acquired ECG signal to obtain the ECG signal diagnosis result. The multimodal diagnosis module performs correlation anomaly diagnosis based on the multimodal fusion feature vector output from the primary fusion section to obtain the anomaly diagnosis result. The ECG signal diagnosis result and the anomaly diagnosis result are then weighted and fused to output the final diagnosis result.

[0007] Preferably, in step S2, the extraction of thoracic contour features includes: 3D point cloud data of the thoracic contour collected by the infrared scanning unit, fitting the equation of the thoracic surface through an algorithm, extracting key feature points, calculating the Euclidean distance, angle and surface curvature between each marker point, and forming a thoracic morphological feature vector. Cardiac structural feature extraction includes: short-axis and long-axis cross-sectional images of the heart acquired by an ultrasound probe, segmenting the left ventricle, right ventricle, atrium and main coronary artery region using the U-Net semantic segmentation model, and extracting cardiac structural feature vectors; The initial ECG signal feature extraction includes: acquiring the initial ECG signal using a 12-lead dry electrode, removing electromyographic interference and power frequency noise through bandpass filtering, and extracting the initial ECG signal feature vector.

[0008] Preferably, in step S2, the three-dimensional ECG lead localization includes: inputting the initial ECG signal and chest feature data into the three-dimensional ECG lead localization module, calling the three-dimensional mapping database, and performing matching retrieval using a similarity retrieval algorithm; the similarity retrieval algorithm introduces a weighted distance calculation model based on the K-nearest neighbor algorithm, quickly matching several most similar samples during retrieval, extracting the lead position parameters of the matched samples as a candidate set, calculating the initial predicted value of the lead position parameters of the candidate set using a weighted average method, and adjusting the electrode position based on the initial predicted position.

[0009] Preferably, in step S2, after three-dimensional ECG lead localization, ECG signals are acquired based on the initial predicted position, and judgment indicators are calculated. The judgment indicators include signal-to-noise ratio and QRS complex identification. If any judgment indicator is lower than a set threshold, the electrode position is iteratively optimized. ECG signals are acquired after each adjustment of the electrode position, and calibration is performed based on real-time signal feedback until the judgment indicators reach the comprehensive optimum. Finally, the optimal lead position parameters are output, and accurate localization is completed based on real-time signal feedback calibration. Then, the final ECG signal is acquired.

[0010] Preferably, in step S3, the electrocardiogram signal diagnosis module includes: The first model introduces an attention mechanism on the basis of the DS-CNN model. The attention mechanism focuses on extracting key frequency bands and local features of the final electrocardiogram signal and outputs local pathological features. The second model uses a bidirectional LSTM model to simultaneously capture the positive temporal dependence and negative correlation of the final electrocardiogram signal and output long-term rhythm features. The ECG signal diagnosis module introduces a dynamic weight adjustment mechanism to analyze and diagnose local pathological features and long-term rhythm features after fusing them in the fusion layer, thereby obtaining ECG signal diagnosis results.

[0011] Preferably, in step S3, the dynamic weight adjustment mechanism includes: constructing a clinical information feature vector, mapping the clinical information feature vector to weight coefficients through a fully connected layer, and dynamically adjusting the weight ratio of local pathological features and long-term rhythm features in the fusion layer.

[0012] Preferably, the first model adopts a lightweight structure that combines grouped convolution and pointwise convolution. At the same time, the convolution module uses multi-scale convolution kernels, which correspond to signal features with short, medium and long periods, respectively.

[0013] Preferably, the final electrocardiogram signal is divided into several cardiac cycle segments based on duration and RR intervals. Each segment serves as the input unit of LSTM, and rhythm abnormalities are identified by combining the temporal correlation coefficients between segments.

[0014] Preferably, in step S3, the Bayesian fusion rule in the secondary fusion part performs weighted fusion of the ECG signal diagnosis results and the abnormal diagnosis results. The Bayesian fusion rule is used to calculate and adjust the confidence level, and the fusion weight of the diagnosis results is set based on the confidence level.

[0015] This invention also provides an AI-based integrated intelligent auxiliary analysis system for the entire process of electrocardiogram (ECG) detection and diagnosis, used to implement the AI-based integrated intelligent auxiliary analysis method for the entire process of ECG detection and diagnosis as described above, and includes: The multimodal data acquisition module integrates the ECG electrode array, ultrasound probe, and infrared scanning unit into the same monitoring terminal, and is equipped with a lead adaptive adjustment structure, which includes a miniature electric adjustment mechanism configured below the ECG electrode array. The three-dimensional ECG lead positioning module includes a three-dimensional mapping database, a three-dimensional ECG positioning unit, and a data preprocessing unit. The data preprocessing unit is used to preprocess the collected data. The three-dimensional mapping database stores the labeled dataset of the samples. The three-dimensional ECG positioning unit calls the three-dimensional mapping database to optimize the position of the motor and outputs position parameters to control the adaptive adjustment structure of the lead to complete the precise positioning of the ECG electrode array. Fusion Diagnosis Module: This module includes the fusion diagnosis model used for analysis and diagnosis.

[0016] The technical advantages of this invention are as follows: This invention establishes a mapping relationship between chest morphology and ECG signals through a three-dimensional ECG lead localization database, which can adapt to special chest morphologies such as obesity and scoliosis, improves the signal-to-noise ratio, and reduces the invalid monitoring rate caused by lead displacement. After the collected final ECG signal is input into the fusion diagnostic model, local pathological features and long-term rhythm features are extracted through the first and second models respectively. The two are then weighted and fused for ECG signal diagnosis, which can significantly improve the comprehensive recognition accuracy of common cardiovascular diseases such as myocardial infarction, atrial fibrillation, and premature ventricular contractions. On the other hand, in the multimodal two-level fusion decision model, Bayesian fusion rules are also used to weightedly fuse the abnormal diagnosis of the multimodal fusion feature vector and the diagnostic results of the ECG signal to output the final diagnostic result. This can accurately distinguish between electrode displacement and true pathological abnormalities, reduce the false positive rate, and reduce clinical misdiagnosis. Attached Figure Description

[0017] Figure 1 This is a flowchart of step S2 in the AI-based integrated intelligent auxiliary analysis method for full-process electrocardiogram detection and diagnosis of the present invention.

[0018] Figure 2 This is a flowchart of step S3 in the AI-based integrated intelligent auxiliary analysis method for full-process electrocardiogram detection and diagnosis of the present invention. Detailed Implementation

[0019] The following detailed description of the embodiments, with reference to the accompanying drawings, will further illustrate the specific implementation of the present invention, in order to help those skilled in the art to have a more complete, accurate, and in-depth understanding of the inventive concept and technical solution of the present invention.

[0020] like Figure 1-2 As shown, this invention provides an AI-based, end-to-end intelligent auxiliary analysis method for electrocardiogram (ECG) detection and diagnosis, comprising the following steps: S1. Construct a three-dimensional mapping database.

[0021] A 3D mapping database is constructed using the labeled dataset of samples. Each sample was associated with a thoracic morphology feature vector, a cardiac anatomy feature vector, an electrocardiogram (ECG) signal feature vector, and optimal lead position parameters. To accommodate different patient groups, the samples covered individuals of varying ages (3-90 years), body types (BMI 16-40 kg / m²), types of thoracic deformities (scoliosis, pectus carinatum, pectus excavatum), and underlying diseases (hypertension, diabetes, coronary heart disease). All samples were jointly annotated with optimal lead positions by professional physicians, using the criteria of maximum electrical signal intensity and minimum waveform distortion in the cardiac projection area.

[0022] S2. Positioning and multimodal data acquisition of the multimodal acquisition terminal.

[0023] This step includes: Thoracic contour feature extraction: The 3D point cloud data of the thoracic contour acquired by the infrared scanning unit (sampling density 100 points / cm²) was filtered by Gaussian (σ=0.8) to remove noise such as hair and clothing wrinkles. The equation of the thoracic contour was fitted by an algorithm to extract key feature points, including 12 anatomical landmarks such as the suprasternal notch, midpoint of the clavicle, xiphoid process, left and right nipple points, and midline projection point of the spine. The Euclidean distance, angle and surface curvature between each landmark were calculated to form a 36-dimensional thoracic contour feature vector.

[0024] Cardiac structural feature extraction: The short-axis and long-axis cross-sectional images of the heart acquired by the ultrasound probe (resolution 512×512 pixels) were segmented into the left ventricle, right ventricle, atrium and main coronary artery region using the U-Net semantic segmentation model, and 18-dimensional cardiac structural feature vectors such as the coordinates of the heart's center of gravity, ventricular wall thickness and cardiac chamber volume were extracted.

[0025] Initial ECG signal feature extraction: The initial ECG signal (sampling frequency 500Hz) acquired by 12-lead dry electrodes was bandpass filtered (0.5-100Hz) to remove electromyographic interference and power frequency noise, and 12-dimensional signal feature vectors (initial ECG signal feature vectors) such as QRS complex amplitude, RR interval standard deviation, and ST segment baseline drift were extracted.

[0026] 3D ECG Lead Localization: Initial ECG signals and chest wall feature data are input into the 3D ECG lead localization module. A 3D mapping database is called, and a similarity retrieval algorithm is used for matching. The similarity retrieval algorithm is an improved K-Nearest Neighbors (KNN) algorithm, which introduces a weighted distance calculation model. During retrieval, the top-20 similar samples are quickly matched, and their lead position parameters are extracted as a candidate set. The lead position parameters (x, y, z 3D coordinates) of the candidate set are calculated using a weighted average method to obtain initial predicted values. The weights are positively correlated with sample similarity (the higher the similarity, the greater the weight). Based on the initial predicted positions, the electrode positions (i.e., the 12-lead dry electrodes) are adjusted to achieve the first electrode position optimization.

[0027] ECG adaptive optimization positioning: Based on the initial predicted position, a 3-second ECG signal is acquired, and the signal-to-noise ratio (SNR) and QRS complex identification (using waveform correlation coefficient) are calculated. If the SNR < 30 dB or the identification < 0.85, iterative optimization of the electrode position is initiated: the electrodes are finely adjusted along the x / y / z axes by ±0.5 mm, ±1.0 mm, and ±0.3 mm (step size 0.1 mm) respectively using a micro-adjustment mechanism. A 1-second ECG signal is acquired after each adjustment, and the index values ​​are calculated until the optimal combination of SNR and QRS complex identification is found. The optimal lead position parameters are then output, and precise positioning is completed based on real-time signal feedback calibration, i.e., the second optimization positioning. Finally, the final ECG signal is acquired based on the precisely positioned 12-lead dry electrodes.

[0028] S3. Perform signal analysis and diagnosis using a fusion diagnostic model.

[0029] The fusion diagnostic model includes a multimodal two-level fusion decision model, which consists of a first-level fusion component and a second-level fusion component.

[0030] The first-level fusion component is feature-layer fusion, which concatenates the thoracic morphology feature vector, cardiac structure feature vector, final electrocardiogram (ECG) signal feature vector, and clinical information feature vector to form a multimodal fusion feature vector (dimension = 36 + 18 + 12 + 8 = 74 dimensions). This vector serves as input to the AI ​​diagnostic model, allowing the model to learn the correlations between different modalities of data at the feature level. The multimodal fusion feature vector contains various features that significantly influence diagnostic results, such as the "electrode position offset" of the thoracic contour (monitored in real time by a positioning module), "cardiac structural abnormality markers" from ultrasound images (such as abnormal ventricular wall motion), and "pathological feature vectors" from ECG signals.

[0031] The second-level fusion component is the decision-level fusion, which constructs an ECG signal diagnosis module and a multimodal diagnosis module. The ECG signal diagnosis module analyzes and diagnoses the final acquired ECG signals to obtain ECG signal diagnosis results. The multimodal diagnosis module, based on the multimodal fusion feature vector output from the first-level fusion component, performs correlation anomaly diagnosis to obtain anomaly diagnosis results. Bayesian fusion rules are used to weight and fuse the ECG signal diagnosis results and anomaly diagnosis results to output the final diagnosis result. Bayesian fusion rules are used to calculate and adjust the confidence level, setting the fusion weight of the diagnosis results based on the confidence level. For example, if the ECG signal indicates an abnormality, but the multimodal fusion model indicates "electrode displacement" (e.g., chest features show electrode displacement >1mm, ultrasound shows normal heart structure), the confidence level of the abnormal diagnosis is reduced to avoid false positives; if both indicate an abnormality, the confidence level is increased to ensure diagnostic reliability.

[0032] The ECG signal diagnosis module includes a first model and a second model, as detailed below: The first model is formed by introducing an attention mechanism (SE) based on the DS-CNN (Depthly Separable Convolutional) model. The attention mechanism focuses on extracting key frequency bands (such as 0.5-10Hz corresponding to the ST segment) and local features (such as Q wave and T wave morphology) of the final ECG signal, avoiding interference from irrelevant features. The DS-CNN model adopts a lightweight structure combining grouped convolution and pointwise convolution. Simultaneously, its convolutional modules use multi-scale convolutional kernels (3×1, 5×1, 7×1), corresponding to short-cycle, medium-cycle, and long-cycle signal features respectively. Compared to single-scale convolutional kernels, this approach has a stronger ability to capture pathological features of different durations (such as transient ST-segment elevation and persistent atrial fibrillation waveforms). The first model outputs local pathological features.

[0033] The second model employs a bidirectional LSTM (Long Short-Term Memory) model, simultaneously capturing both the positive temporal dependence (e.g., compensatory pauses after premature beats) and the negative correlation (e.g., waveform precursors before ST segment changes) of the final ECG signal, overcoming the limitation of traditional LSTM which can only extract temporal features unidirectionally. The final ECG signal is divided into several cardiac cycle segments based on duration (e.g., a 1-minute signal) according to the RR interval. Each segment serves as the input unit for the LSTM, and the temporal correlation coefficients between segments are combined to improve the sensitivity for identifying rhythm abnormalities (e.g., paroxysmal atrial fibrillation). The second model outputs long-term rhythm features.

[0034] The ECG signal diagnosis module analyzes and diagnoses by fusing local pathological features and long-term rhythm features in a fusion layer. A dynamic weight adjustment mechanism is also introduced to obtain the ECG signal diagnosis results. The dynamic weight adjustment mechanism constructs a clinical information feature vector (containing 8 dimensions including age, gender, past medical history, and underlying diseases). This clinical information feature vector is mapped to weight coefficients (0.1-0.9) through a fully connected layer, dynamically adjusting the weight ratio of local pathological features and long-term rhythm features in the fusion layer. For example, for elderly patients (age ≥ 60 years), the weight of long-term rhythm features is increased by 0.2-0.3, as rhythm abnormalities are more common in elderly patients; for diabetic patients, the weight of ST segment features (part of the local pathological features) is increased by 0.15-0.25, as diabetes easily leads to myocardial micro-damage, and ST segment changes are more diagnostically valuable.

[0035] Training of the fusion diagnostic model: Sample augmentation strategies are introduced during model training. Addressing the issue of insufficient samples for rare diseases, a Generative Adversarial Network (GAN) is used to generate simulated samples. Based on the feature distribution of real samples, synthetic samples are generated at a 1:5 ratio to alleviate model overfitting and improve the ability to identify rare diseases. On the other hand, signal perturbation enhancement is performed during training. Reasonable noise (such as EMG interference, power line interference), baseline drift, amplitude scaling, and other perturbations are added to the raw ECG signal to simulate signal states under different clinical acquisition environments, improving the model's robustness.

[0036] Quantization compression and inference optimization of the integrated diagnostic model: The trained model is compressed using INT8 quantization technology, reducing the model size by 75%. At the same time, the TensorRT inference engine is deployed on the edge computing chip (NVIDIA Jetson AGX Orin). Through layer fusion, accuracy optimization and other means, the diagnostic latency is controlled to ≤5s, which is significantly more efficient than traditional cloud diagnosis (latency 30s-5min) and local single model diagnosis (latency 10-15s).

[0037] Through 5,000 clinical tests (including 800 simulated electrode displacement samples and 4,200 real pathological samples), the false positive rate of this multimodal fusion scheme was 7.2%, which is more than 60% lower than the traditional single ECG signal diagnosis (false positive rate of 18%), and the true positive rate was not reduced (maintained at more than 95%), effectively reducing clinical misdiagnosis and unnecessary examinations.

[0038] After obtaining the fused diagnostic results, hierarchical early warning can be performed based on the entire structure. Specifically: Red Alert (Emergency): Alert information is pushed out via SMS, APP, departmental large screen, etc., and standardized reports of diagnostic results can be generated.

[0039] Yellow alert (suspicious): The diagnosis result is pushed to the corresponding doctor's client, and the doctor will provide suggestions for follow-up examination.

[0040] Green (Safe): In non-alert situations, the results are pushed directly to the patient's mobile device, where relevant diagnostic results and health guidance are displayed.

[0041] Ultimately, the above diagnostic results and corresponding warning information are stored and can be backed up in the cloud.

[0042] The present invention also provides an AI-based integrated intelligent auxiliary analysis system for the entire process of electrocardiogram (ECG) detection and diagnosis, which is used to implement the above-mentioned AI-based integrated intelligent auxiliary analysis method for the entire process of ECG detection and diagnosis.

[0043] The system includes a multimodal data acquisition module, a three-dimensional electrocardiogram lead localization module, and a fusion diagnostic module, as detailed below: Multimodal data acquisition module: Composed of a 12-lead ECG electrode array, a miniature ultrasound probe, and an infrared chest morphology scanning unit. The ECG electrodes adopt a dry electrode design (no conductive gel required), the ultrasound probe is used to acquire images of the heart structure, and the infrared scanning unit realizes rapid scanning of the three-dimensional contour of the chest (scanning time ≤ 2s). The sampling frequency of each acquisition unit can be dynamically adjusted by the central controller (ECG signal sampling frequency range 250-1000Hz, image data sampling frame rate 15-30fps). The integrated design of the multimodal acquisition terminal adopts modular packaging technology, integrating a 12-lead dry electrode array, a miniature ultrasound probe, and an infrared scanning unit into a single housing (20cm×15cm×3cm). The electrode array features a retractable design, connected via a flexible circuit board, allowing it to adaptively conform to the shape of the chest. The emission windows of the ultrasound probe and the infrared scanning unit are treated with anti-fog and anti-sweat coatings to ensure acquisition stability. Each acquisition unit employs a time-sharing power supply strategy, with the central controller dynamically allocating power according to the workflow (initiating infrared scanning and initial electrode acquisition during the positioning phase, and synchronous acquisition of ultrasound and ECG signals during the diagnostic phase), reducing system power consumption and extending the built-in lithium battery's battery life to ≥6 hours (in emergency mode).

[0044] This module integrates the ECG electrode array, ultrasound probe, and infrared scanning unit into a single monitoring terminal. Compared to traditional "separate acquisition devices" (such as a separate ECG machine and ultrasound machine), it is smaller and lighter, enabling rapid deployment in mobile scenarios such as bedside and ambulances. The electrode array adopts a modular design, supporting flexible switching according to monitoring needs (such as conventional 12-lead or dynamic 3-lead), resulting in stronger compatibility.

[0045] Lead adaptive adjustment structure: A miniature electric adjustment mechanism (accuracy ±0.1mm) is configured below the ECG electrode array. Driven by the position parameters output by the three-dimensional ECG lead positioning module, lead positioning can be completed without manual intervention. Compared with traditional "manual positioning" or "fixed position electrodes", the positioning time is shortened and the positioning error is reduced, significantly improving the standardization of signal acquisition.

[0046] The three-dimensional ECG lead localization module includes a three-dimensional mapping database, a three-dimensional ECG localization unit, and a data preprocessing unit.

[0047] Data preprocessing unit: performs noise reduction and feature extraction (such as key points of the thoracic cavity and coordinates of the heart) on the collected thoracic contour data and ultrasound images.

[0048] 3D mapping database: A labeled dataset that stores sample-associated thoracic morphological feature vectors, cardiac anatomical feature vectors, electrocardiogram signal feature vectors, and optimal lead position parameters.

[0049] The 3D ECG localization unit: It calls upon a 3D mapping database, obtains a candidate set of lead position parameters through a similarity retrieval algorithm, and calculates initial predicted values ​​using a weighted average method to achieve the first electrode position optimization. Subsequently, based on real-time signal feedback calibration, it achieves a second electrode position optimization, controlling the 12-lead dry electrode array to complete precise positioning.

[0050] Fusion Diagnosis Module: It is equipped with a fusion diagnosis model, which performs analysis and diagnosis based on step S3 of the AI-based full-process ECG detection and diagnosis integrated intelligent auxiliary analysis method.

[0051] This module is used to extract local pathological features (such as ST segment deviation and Q wave morphology) from electrocardiogram signals, while the LSTM module is used to capture long-term temporal dependencies of signals (such as rhythm changes in arrhythmias). The model has a built-in adaptive weight adjustment unit that can dynamically optimize feature extraction weights based on clinical information such as patient age and underlying diseases. At the same time, it integrates a real-time inference engine to ensure that the diagnostic analysis delay is ≤5s.

[0052] By integrating the fusion model and inference engine onto the system's local chip (using the NVIDIA Jetson AGX Orin edge computing chip), the diagnostic latency is significantly reduced compared to the traditional architecture of "data acquisition terminal + cloud diagnosis," while avoiding diagnostic stagnation caused by network interruptions, thus meeting the real-time requirements of emergency care.

[0053] The system also includes a real-time feedback module and a data storage module.

[0054] The real-time feedback module includes a touch screen, a voice broadcast unit, and a wireless communication interface. It can display the original ECG waveform, lead positioning status, and diagnostic results (including pathological feature annotations) in real time. It can also synchronize the data to the hospital information system (HIS) or the patient's mobile device via 5G / Wi-Fi. Abnormal results trigger tiered warnings (red: emergency, requiring immediate action; yellow: suspicious, re-examination recommended; green: normal).

[0055] The data storage module adopts a combination of local encrypted storage and cloud backup. Local storage supports ≥1000 complete monitoring data, while the cloud uses blockchain technology to achieve data traceability and privacy protection, meeting medical data compliance requirements.

[0056] This system integrates the entire process of "location-collection-analysis-diagnosis-feedback," significantly reducing the time from the start of patient monitoring to obtaining a diagnostic report. It is more efficient than the traditional process (average 2-4 hours) and can effectively support rapid decision-making in emergency care. It also has the following advantages: (1) High lead positioning accuracy: By establishing a mapping relationship between chest morphology and ECG signal through a three-dimensional ECG lead positioning database, it can adapt to special chest morphologies such as obesity and scoliosis, improve the signal-to-noise ratio, and reduce the invalid monitoring rate caused by lead displacement.

[0057] (2) High diagnostic accuracy and adaptability: The DS-CNN and LSTM fusion model can extract local pathological features and long-term rhythm features at the same time, and has a high accuracy rate in the comprehensive identification of common cardiovascular diseases such as myocardial infarction, atrial fibrillation, and premature ventricular contractions.

[0058] (3) Multimodal fusion reduces false positives: By fusing chest images and electrocardiogram signals, electrode displacement and real pathological abnormalities can be accurately distinguished, reducing the false positive rate and minimizing clinical misdiagnosis and unnecessary subsequent examinations for patients.

[0059] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other occasions without modification, are all within the protection scope of the present invention.

Claims

1. An AI-based integrated intelligent auxiliary analysis method for the entire process of electrocardiogram (ECG) detection and diagnosis, characterized in that, Includes the following steps: S1. Construct a three-dimensional mapping database, which includes a labeled dataset of samples. Each sample is associated with a thoracic morphology feature vector, a cardiac anatomy feature vector, an electrocardiogram signal feature vector, and optimal lead position parameters. S2. The positioning and multimodal data acquisition of the multimodal acquisition terminal first extracts the features of the thoracic contour, the features of the heart structure, and the features of the initial electrocardiogram signal. Then, it calls the three-dimensional mapping database to perform three-dimensional electrocardiogram lead positioning, and then performs electrocardiogram adaptive optimization positioning to acquire the final electrocardiogram signal. S3. Signal analysis and diagnosis are performed using a fusion diagnostic model, which includes a two-level fusion decision model. The two-level fusion decision model includes: In the first-level fusion section, the thoracic morphology feature vector, cardiac structure feature vector, final electrocardiogram signal feature vector, and clinical information feature vector are concatenated to form a multimodal fusion feature vector; The secondary fusion section includes an ECG signal diagnosis module and a multimodal diagnosis module. The ECG signal diagnosis module analyzes and diagnoses the final acquired ECG signal to obtain the ECG signal diagnosis result. The multimodal diagnosis module performs correlation anomaly diagnosis based on the multimodal fusion feature vector output from the primary fusion section to obtain the anomaly diagnosis result. The ECG signal diagnosis result and the anomaly diagnosis result are then weighted and fused to output the final diagnosis result.

2. The AI-based integrated intelligent auxiliary analysis method for the entire process of electrocardiogram detection and diagnosis according to claim 1, characterized in that, In step S2, Thoracic contour feature extraction includes: 3D point cloud data of the thoracic cavity collected by the infrared scanning unit, fitting the equation of the thoracic cavity surface through an algorithm, extracting key feature points, calculating the Euclidean distance, angle and surface curvature between each marker point, and forming a thoracic cavity morphological feature vector. Cardiac structural feature extraction includes: short-axis and long-axis cross-sectional images of the heart acquired by an ultrasound probe, segmenting the left ventricle, right ventricle, atrium and main coronary artery region using the U-Net semantic segmentation model, and extracting cardiac structural feature vectors; The initial ECG signal feature extraction includes: acquiring the initial ECG signal using a 12-lead dry electrode, removing electromyographic interference and power frequency noise through bandpass filtering, and extracting the initial ECG signal feature vector.

3. The AI-based integrated intelligent auxiliary analysis method for the entire process of electrocardiogram detection and diagnosis according to claim 2, characterized in that, In step S2, the three-dimensional ECG lead localization includes: the initial ECG signal and chest feature data are input into the three-dimensional ECG lead localization module, the three-dimensional mapping database is called, and a similarity retrieval algorithm is used for matching retrieval; the similarity retrieval algorithm introduces a weighted distance calculation model based on the K-nearest neighbor algorithm, and quickly matches several most similar samples during retrieval, extracts the lead position parameters of the matched samples as a candidate set, calculates the initial predicted value of the lead position parameters of the candidate set using the weighted average method, and adjusts the electrode position based on the initial predicted position.

4. The AI-based integrated intelligent auxiliary analysis method for the entire process of electrocardiogram detection and diagnosis according to claim 3, characterized in that, In step S2, after three-dimensional ECG lead localization, ECG signals are acquired based on the initial predicted position, and judgment indicators are calculated. The judgment indicators include signal-to-noise ratio and QRS complex recognition. If any judgment indicator is lower than the set threshold, the electrode position is iteratively optimized. ECG signals are acquired after each adjustment of the electrode position, and calibration is performed based on real-time signal feedback until the judgment indicators reach the comprehensive optimum. Finally, the optimal lead position parameters are output. Accurate localization is completed based on real-time signal feedback calibration, and then the final ECG signal is acquired.

5. The AI-based integrated intelligent auxiliary analysis method for the entire process of electrocardiogram detection and diagnosis according to claim 1, characterized in that, In step S3, the electrocardiogram signal diagnosis module includes: The first model introduces an attention mechanism on the basis of the DS-CNN model. The attention mechanism focuses on extracting key frequency bands and local features of the final electrocardiogram signal and outputs local pathological features. The second model uses a bidirectional LSTM model to simultaneously capture the positive temporal dependence and negative correlation of the final electrocardiogram signal and output long-term rhythm features. The ECG signal diagnosis module introduces a dynamic weight adjustment mechanism to analyze and diagnose local pathological features and long-term rhythm features after fusing them in the fusion layer, thereby obtaining ECG signal diagnosis results.

6. The AI-based integrated intelligent auxiliary analysis method for the entire process of electrocardiogram detection and diagnosis according to claim 5, characterized in that, In step S3, the dynamic weight adjustment mechanism includes: constructing a clinical information feature vector, mapping the clinical information feature vector to weight coefficients through a fully connected layer, and dynamically adjusting the weight ratio of local pathological features and long-term rhythm features in the fusion layer.

7. The AI-based integrated intelligent auxiliary analysis method for the entire process of electrocardiogram detection and diagnosis according to claim 5, characterized in that, The first model adopts a lightweight structure that combines grouped convolution and pointwise convolution. Meanwhile, the convolution module uses multi-scale convolution kernels, which correspond to signal features with short, medium, and long periods, respectively.

8. The AI-based integrated intelligent auxiliary analysis method for the entire process of electrocardiogram detection and diagnosis according to claim 5, characterized in that, The final electrocardiogram signal is divided into several cardiac cycle segments based on the duration and the RR interval. Each segment serves as the input unit of the LSTM, and the temporal correlation coefficient between the segments is combined to identify rhythm abnormalities.

9. The AI-based integrated intelligent auxiliary analysis method for the entire process of electrocardiogram detection and diagnosis according to claim 1, characterized in that, In step S3, the Bayesian fusion rule in the secondary fusion part performs weighted fusion of the ECG signal diagnosis results and abnormal diagnosis results. The Bayesian fusion rule is used to calculate and adjust the confidence level, and the fusion weight of the diagnosis results is set based on the confidence level.

10. An AI-based integrated intelligent auxiliary analysis system for the entire process of electrocardiogram (ECG) detection and diagnosis, characterized in that: The method is used to implement the AI-based integrated intelligent auxiliary analysis method for the entire process of electrocardiogram detection and diagnosis as described in any one of claims 1-9, and includes: The multimodal data acquisition module integrates the ECG electrode array, ultrasound probe, and infrared scanning unit into the same monitoring terminal, and is equipped with a lead adaptive adjustment structure, which includes a miniature electric adjustment mechanism configured below the ECG electrode array. The three-dimensional ECG lead positioning module includes a three-dimensional mapping database, a three-dimensional ECG positioning unit, and a data preprocessing unit. The data preprocessing unit is used to preprocess the collected data. The three-dimensional mapping database stores the labeled dataset of the samples. The three-dimensional ECG positioning unit calls the three-dimensional mapping database to optimize the position of the motor and outputs position parameters to control the adaptive adjustment structure of the lead to complete the precise positioning of the ECG electrode array. Fusion Diagnosis Module: This module includes the fusion diagnosis model used for analysis and diagnosis.

Images