Dynamic electrocardiogram early warning system based on artificial intelligence

Abstract

The invention discloses a dynamic electrocardiogram early warning system based on artificial intelligence, which relates to the technical field of medical health monitoring, and comprises a model building module, a threshold setting module, an early warning threshold setting module based on the baseline model, a threshold adjusting module, and a dynamic electrocardiogram early warning module based on the real-time activity state of the user. And dynamically adjusting the early warning threshold. By dynamically adjusting the early warning threshold value, daily activity changes of the user can be better adapted, unnecessary false alarm and missing alarm are reduced, the early warning reliability and the user experience are improved, in addition, a dynamic threshold value adjusting mechanism can also capture the influence of user emotion fluctuation on physiological indexes, and the user experience is improved for high-intensity motion and static states. The early warning threshold can be relaxed and tightened respectively, the sensitivity and accuracy of early warning are ensured, and the intelligent level of early warning monitoring is further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical health monitoring, and in particular to a dynamic electrocardiogram early warning system based on artificial intelligence. Background Art

[0002] In recent years, with the increasing incidence of cardiovascular diseases, the application of ECG early warning monitoring in the medical and health field has become increasingly widespread. Traditional static ECG monitoring equipment mainly relies on professional instruments in hospitals or clinics, and patients need to perform short-term ECG recordings in specific environments, which makes the early detection and continuous monitoring of heart abnormalities face many challenges.

[0003] Existing ECG early warning monitoring still has some shortcomings. Traditional ECG monitoring equipment usually uses a fixed threshold to determine whether to trigger an early warning. This static threshold cannot adapt to the individual differences of different users. Secondly, the user feedback mechanism is relatively weak, and the early warning strategy cannot be dynamically adjusted according to the user's actual health status, resulting in a low level of intelligence in early warning monitoring. Summary of the invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a dynamic electrocardiogram early warning system based on artificial intelligence to solve the problem that the inherent threshold cannot adapt to individual differences of users and the feedback mechanism is weak.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] The present invention provides a dynamic electrocardiogram early warning system based on artificial intelligence, which comprises:

[0008] Data acquisition module: users wear wearable devices to collect the user's electrocardiogram physiological signals and environmental data and extract multimodal features;

[0009] The data fusion module selects the deep learning model and inputs the multimodal features to form a multimodal feature vector;

[0010] Model building module, which builds a baseline model for users based on multimodal feature vectors;

[0011] The threshold setting module sets the warning threshold based on the baseline model;

[0012] Threshold adjustment module, dynamically adjusts the warning threshold according to the user's real-time activity status;

[0013] The anomaly detection module analyzes the multimodal feature vectors to obtain anomaly probability scores;

[0014] The model optimization module compares the anomaly probability score with the adjusted warning threshold to obtain the warning result, collects user feedback and optimizes the baseline model.

[0015] As a preferred solution of the dynamic electrocardiogram early warning system based on artificial intelligence described in the present invention, wherein:

[0016] The user's electrocardiogram physiological indicators include heart rate, blood pressure, blood oxygen saturation, body temperature and respiratory rate;

[0017] As a preferred solution of the artificial intelligence-based dynamic electrocardiogram early warning system described in the present invention, the environmental data includes movement status, location information and weather conditions.

[0018] As a preferred solution of the artificial intelligence-based dynamic electrocardiogram early warning system described in the present invention, the multimodal features include the R wave peak, PQ interval, QRS complex wave morphology, systolic pressure, diastolic pressure, SpO2 value of blood oxygen saturation, average change rate of body temperature and respiratory cycle length of respiratory frequency.

[0019] As a preferred solution of the dynamic electrocardiogram early warning system based on artificial intelligence described in the present invention, wherein: selecting a deep learning model and inputting multimodal features to form a multimodal feature vector includes the following steps:

[0020] Select Transformer as the core architecture of the model to build a deep learning model;

[0021] Integrate weight calculation units in deep learning models;

[0022] Use the weighted summation method to fuse multimodal features of the weight calculation unit and normalize them;

[0023] PCA is used to reduce the dimension of the fused multimodal features to obtain a multimodal feature vector.

[0024] As a preferred solution of the dynamic electrocardiogram early warning system based on artificial intelligence of the present invention, wherein: based on the multimodal feature vector, a baseline model is established for the user, including the following steps:

[0025] Select the long short-term memory network as the core structure of the baseline model;

[0026] Use the PyTorch framework to build a long short-term memory network model;

[0027] The multimodal feature vector is input into the long short-term memory network model, and the hidden layer is stacked and trained using the Adam optimizer to obtain the trained long short-term memory network model;

[0028] Based on the trained long short-term memory network model, the user's electrocardiogram physiological indicators are divided into time windows of set lengths through sliding window technology to generate time series segments;

[0029] Add the corresponding indicator label to each indicator label, obtain the time series segment with the indicator label and convert it into tensor form;

[0030] The converted time series segments of the indicator labels and the long short-term memory network model are assembled to obtain a complete baseline model.

[0031] As a preferred solution of the artificial intelligence-based dynamic electrocardiogram early warning system of the present invention, wherein: based on the baseline model, the early warning threshold is set, including the following steps:

[0032] Use TensorFlow to forward propagate the time steps of the baseline model, extract the implicit state of each time step and combine them into an implicit state vector;

[0033] Use clustering algorithms to perform cluster analysis on latent state vectors and identify typical activity patterns of users;

[0034] Correlate the user's typical activity patterns with the user's physiological indicators to form a data set of user behavior;

[0035] Based on the user behavior data set, all time steps are grouped according to the corresponding clusters, and the mean and standard deviation of the user's physiological indicators are calculated;

[0036] According to the mean and standard deviation of the user's physiological indicators, the initial warning threshold is set using the fixed multiple standard deviation method.

[0037] As a preferred solution of the dynamic electrocardiogram early warning system based on artificial intelligence of the present invention, wherein: dynamically adjusting the early warning threshold according to the real-time activity status of the user includes the following steps:

[0038] Use the constructed deep learning model to infer the multimodal feature vectors and identify the user's real-time activity status;

[0039] Select heart rate variability as an indicator of user emotion fluctuations, and calculate the standard deviation and mean of heart rate variability;

[0040] Combine the user's current real-time activity status and the standard deviation and mean of heart rate variability to calculate the user's comprehensive activity status score, expressed as:

[0041]

[0042] Where D is the standard deviation of heart rate variability, N is the number of heart beat intervals, and R is irepresents the i-th heartbeat interval, represents the average of all heartbeat intervals, S represents the user's comprehensive activity status score, p(y=j) represents the probability that the user is in the jth motion state, y represents the user's motion state variable, F represents the mean of heart rate variability, w1 represents the weight of the motion state, and w2 represents the weight of the user's emotional fluctuation;

[0043] Dynamically adjust the warning threshold of physiological indicators based on the user's comprehensive activity status score.

[0044] As a preferred solution of the dynamic electrocardiogram early warning system based on artificial intelligence described in the present invention, wherein: analyzing the multimodal feature vector to obtain the abnormal probability score includes the following steps:

[0045] Introduce graph neural network and list multimodal feature vectors as nodes in the graph neural network to establish a multimodal graph structure;

[0046] Use graph attention network to propagate features of multimodal graph structure and obtain new multimodal feature vectors;

[0047] Based on the new multimodal feature vector, dynamic prior distribution probability and marginal probability are defined for each user;

[0048] According to the dynamic prior probability distribution and marginal probability, the abnormal probability score of the user is calculated, and the expression is:

[0049]

[0050] in, represents the abnormal probability score of the user, e represents the health status of the user, R(z|e=1) represents the conditional probability that the user is in an abnormal health state when a new multimodal feature vector z is observed, O(e=1) represents the dynamic prior distribution probability that the user is in an abnormal health state, and Q(z) represents the marginal probability of the new multimodal feature vector.

[0051] As a preferred solution of the artificial intelligence-based dynamic electrocardiogram early warning system of the present invention, wherein: comparing the abnormal probability score with the adjusted early warning threshold to obtain an early warning result, collecting user feedback and optimizing the baseline model, the following steps are included:

[0052] When the value of the abnormal probability score is greater than the adjusted warning threshold, the risk warning mechanism is triggered and the risk warning level is divided;

[0053] Provide users with corresponding processing suggestions based on different risk warning levels;

[0054] The corresponding processing suggestion is a warning result;

[0055] The AR interface was designed using the Unity3D design platform, and the Blender modeling tool was used to create a 3D model of the user's body with animation effects;

[0056] Synchronize the warning results and the user's body 3D model to the AR interface;

[0057] The user views the warning results on the AR interface and interacts with the 3D model of his or her body through gestures and voice commands to obtain interactive results;

[0058] Provide users with a feedback time window, use mobile APP and voice interaction devices to send feedback requests to users, and collect user feedback information based on the interaction results;

[0059] The feedback information includes the user's health status, action status, medical follow-up and early warning evaluation;

[0060] Optimize the baseline model based on user feedback.

[0061] The beneficial effects of the present invention are: using the baseline model to set the warning threshold, and dynamically adjusting the warning threshold according to the user's real-time activity status. By dynamically adjusting the warning threshold, it can better adapt to the user's daily activity changes, reduce unnecessary false alarms and missed alarms, and improve the reliability of the warning and user experience. In addition, the dynamic threshold adjustment mechanism can also capture the impact of user emotional fluctuations on physiological indicators. For high-intensity exercise and static state, the warning threshold can be relaxed and tightened respectively to ensure the sensitivity and accuracy of the warning, further improving the intelligent level of warning monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0063] Figure 1 This is a flow chart of the dynamic electrocardiogram early warning system based on artificial intelligence in Example 1.

[0064] Figure 2 This is a flow chart of obtaining the early warning result in Example 1. DETAILED DESCRIPTION

[0065] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0066] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0067] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0068] Example 1, reference Figure 1 and Figure 2 , which is the first embodiment of the present invention, provides a dynamic electrocardiogram early warning system based on artificial intelligence, comprising the following steps:

[0069] Data acquisition module, which collects the user's electrocardiogram physiological signals and environmental data and extracts multimodal features;

[0070] The user's electrocardiogram physiological signals include heart rate, blood pressure, blood oxygen saturation, body temperature, and respiratory rate;

[0071] This step can comprehensively evaluate the user's heart health by collecting multiple physiological signals, improving the accuracy of abnormality detection;

[0072] Environmental data includes movement status, location information (such as whether the user is near a medical institution), and weather conditions (temperature, humidity, and air pressure);

[0073] Multimodal features include the peak R wave of the electrocardiogram (the process of ventricular depolarization), the PQ interval (the time of conduction from the atrium to the ventricle), the QRS complex wave morphology, the systolic pressure of blood pressure (the highest pressure in the artery when the heart contracts), the diastolic pressure (the lowest pressure in the artery when the heart relaxes), the SpO2 value of blood oxygen saturation, the average rate of change of body temperature, and the length of the respiratory cycle of respiratory rate;

[0074] This step can extract multimodal features to conduct a more detailed analysis of the user's heart health status and capture more abnormal signals. This refined feature extraction can detect potential heart problems in users at an early stage and improve the sensitivity and specificity of abnormality detection.

[0075] The data fusion module selects the deep learning model and inputs the multimodal features to form a multimodal feature vector.

[0076] The following steps are included:

[0077] Select Transformer as the core architecture of the model to build a deep learning model;

[0078] Transformer is a learning model based on the self-attention mechanism. It can effectively process sequence data and capture long-distance dependencies through parallel computing and self-attention mechanism. In addition, Transformer has strong generalization ability and can maintain stable performance under different users and scenarios, reducing the risk of overfitting.

[0079] Integrate weight calculation units in deep learning models;

[0080] In this step, the weight calculation unit can automatically adjust the weight according to the data characteristics to adapt to different application scenarios. By allocating the weight model, it can better handle the imbalance problem between different modal features.

[0081] Use the weighted summation method to fuse multimodal features of the weight calculation unit and normalize them;

[0082] For each modal feature, the corresponding weight value is first generated by the weight calculation unit, and then each modal feature is multiplied by its corresponding weight and summed to obtain the fused feature representation;

[0083] The purpose of normalization processing (Min-Max) is that since the features of different modalities have different numerical ranges, direct fusion may cause some features to dominate, thus affecting the performance of the model;

[0084] Use PCA (principal component analysis) to reduce the dimension of the fused multimodal features to obtain a multimodal feature vector;

[0085] When selecting the output dimension and the target dimension, the appropriate number of principal components can usually be selected through the cumulative contribution rate. For example, the principal components with a cumulative contribution rate of more than 95% can be selected, and the fused multimodal features can be input into the PCA algorithm. After covariance matrix calculation and eigenvalue decomposition, a set of eigenvectors are obtained. These eigenvectors are the multimodal feature vectors after dimensionality reduction.

[0086] The model building module builds a baseline model for the user based on the multimodal feature vector.

[0087] The following steps are included:

[0088] The long short-term memory network (LSTM) is selected as the core structure of the baseline model;

[0089] The reason for the selection is that LSTM can effectively handle the long-term dependency problem in sequence data. For time series data (such as electrocardiogram physiological indicators), LSTM can capture the temporal relationship in the data, thereby improving the accuracy of prediction and classification;

[0090] Use the PyTorch framework to build a long short-term memory network model;

[0091] The PyTorch framework is an open source machine learning library that provides a concise and powerful API for quickly building and training complex neural network models. Its modular design makes the code structure clear, easy to understand and maintain.

[0092] The multimodal feature vector is input into the long short-term memory network model, and the hidden layer is stacked and trained using the Adam optimizer to obtain the trained long short-term memory network model;

[0093] Select the hidden layer dimension (size of the LSTM internal state) and number of layers (depth of the LSTM), input the multimodal feature vector into the LSTM, process the data at each time step, generate the corresponding output and extract the output of the last time step, pass it to the fully connected layer to obtain the final prediction result;

[0094] Select an appropriate loss function according to the type of training task (classification or regression). For example, for classification tasks, cross entropy loss is usually used, and for regression tasks, mean square error loss can be used. Compare the model's prediction results with the true labels, calculate the LSTM loss value, and calculate the gradient of each layer layer by layer through the back propagation algorithm to obtain the LSTM parameters. Finally, use the Adam optimizer to update the LSTM parameters (adjust the update amplitude according to the current learning rate and momentum), and repeat all steps multiple times (iterative training) until the LSTM model completes convergence;

[0095] Based on the trained long short-term memory network model, the user's electrocardiogram physiological indicators are divided into time windows of set lengths through sliding window technology to generate time series segments;

[0096] First, you need to determine the size and step size of each time window (the degree of overlap between adjacent time windows). For example, if you want each time window to contain 5 seconds of data (that is, the number of data points in each window depends on the sampling frequency), and the sampling frequency is 100 data points per second, then each window will contain 500 data points. Divide the continuous user ECG physiological index data into multiple time windows. Starting from the starting position of the ECG physiological index data, gradually move the window according to the set step size and extract a fixed-length time window each time until all data are processed. The data in each time window will form a time series segment;

[0097] Add the corresponding indicator label to each indicator label, obtain the time series segment with the indicator label and convert it into tensor form (PyTorch tensor);

[0098] The indicator label can be a binary label (0 for normal, 1 for abnormal) or a multi-classification label (such as normal, slightly abnormal, and seriously abnormal). The specific label type depends on the application scenario and requirements.

[0099] The purpose of converting to tensor form is to ensure the consistency and integrity of input and output data, reduce the workload of data preprocessing, and facilitate the establishment of subsequent baseline models;

[0100] The converted time series segments of the indicator labels and the long short-term memory network model are assembled to obtain a complete baseline model;

[0101] The baseline model obtained in this step can be directly used in practical applications. Whether it is used for real-time monitoring of ECG abnormalities or for long-term health data analysis of users, the model can provide strong support.

[0102] The threshold setting module sets the warning threshold based on the baseline model.

[0103] The following steps are included:

[0104] Use TensorFlow to forward propagate the time steps of the baseline model, extract the implicit state of each time step (the user's heart health baseline characteristics at that time step, i.e., the user's current health status) and combine them into an implicit state vector;

[0105] Use TensorFlow to load the baseline model, set the baseline model to inference mode and turn off the Dropout layer (with random operation function), and perform forward propagation calculations on all time steps of the baseline model (that is, generate an implicit state for each time step). After the calculation is completed, record the implicit state of each time step based on the calculation results and combine them into a fixed-dimensional implicit state vector (high-level expression in complex mode);

[0106] Use a clustering algorithm (DBSCAN) to perform cluster analysis on the latent state vector and identify the user's typical activity patterns (such as stillness, walking, and running);

[0107] First, the number of clusters is determined by the elbow method. Then, the DBSCAN algorithm iteratively clusters the implicit state vector until convergence. During each iteration, the algorithm updates the cluster center and redistributes the data points to the nearest cluster center (obtaining the corresponding cluster label).

[0108] After clustering, each cluster represents a typical activity mode of a user. For example, one cluster may correspond to a social mode, another cluster may correspond to a fitness mode, and another cluster may correspond to a running mode.

[0109] Correlate the user's typical activity patterns with the user's physiological indicators to form a data set of user behavior;

[0110] The user behavior data set generated in this step can be comprehensively analyzed to reveal the relationship between user activity patterns and physiological indicators. Through correlation analysis, physiological changes under different activity patterns can be more accurately identified, providing strong support for subsequent early warning and health management.

[0111] Based on the user behavior dataset, all time steps are grouped according to the corresponding clusters (according to the cluster label to which each implicit state vector belongs), and the mean and standard deviation of the user's physiological indicators are calculated (reflecting the average level and fluctuation of the user's physiological indicators in each activity mode);

[0112] This step quantitatively analyzes the user's health status in different activity modes by calculating the mean and standard deviation, which helps to identify potential health problems of users and take preventive measures in advance;

[0113] According to the mean and standard deviation of the user's physiological indicators, the initial warning threshold is set using the fixed multiple standard deviation method; for example, for the user's blood oxygen saturation indicator, the initial warning threshold is set as:

[0114] T = M + k × σ;

[0115] Wherein, T represents the initial warning threshold of the user's blood oxygen saturation, M represents the mean of the blood oxygen saturation, σ represents the standard deviation of the blood oxygen saturation, and k represents a constant (usually 2) that deviates from the mean and the standard deviation.

[0116] The initial warning thresholds of other physiological indicators (such as blood pressure and heart rate) can also be calculated based on their means and standard deviations.

[0117] The threshold adjustment module dynamically adjusts the warning threshold according to the user's real-time activity status.

[0118] The following steps are included:

[0119] Use the constructed deep learning model to infer the multimodal feature vectors and identify the user's real-time activity status;

[0120] First, define the user's real-time activity state (such as still, moving, resting);

[0121] Update the implicit state vector and classify it through the fully connected layer in the deep learning model to output a user's activity probability distribution. For example, assuming that the output activity probability distribution is [0.6, 0.3, 0.1], the probability that the user is currently in a static state is 60%, the probability of being in a moving state is 30%, and the probability of being in a resting state is 10%;

[0122] Heart rate variability (HRV, i.e., the variability of heartbeat intervals, obtained by extracting the peak value of the R wave of the electrocardiogram) is selected as an indicator of user emotional fluctuations, and the standard deviation and mean of heart rate variability are calculated;

[0123] The reason for choosing HRV is that HRV reflects the regulation of the user's autonomic nervous system on cardiac activity, especially the balance between the sympathetic and parasympathetic nerves. When a person is in an emotional state (such as tension, anxiety, anger), the sympathetic nervous system will be activated, resulting in a decrease in HRV.

[0124] This step can effectively quantify the user's emotional fluctuations through HRV calculation, improving the accuracy of the comprehensive activity status score;

[0125] Combine the user's current real-time activity status and the standard deviation and mean of heart rate variability to calculate the user's comprehensive activity status score, expressed as:

[0126]

[0127] Where D represents the standard deviation of heart rate variability, which is used to measure the discreteness of heartbeat intervals, N represents the number of heartbeat intervals, and R i Indicates the i-th heartbeat interval, usually in milliseconds (ms). represents the average of all heartbeat intervals, S represents the user's comprehensive activity status score, p(y=j) represents the probability that the user is in the jth motion state, y represents the user's motion state variable, F represents the mean of heart rate variability, when D is close to F, the ratio is close to 0, indicating that the mood fluctuation is small, when D is greater than F, the ratio is large, indicating that the mood fluctuation is large, w1 represents the weight of the motion state, and w2 represents the weight of the user's mood fluctuation;

[0128] Dynamically adjust the warning threshold of physiological indicators based on the user's comprehensive activity status score;

[0129] By introducing a function based on the comprehensive activity status score, the warning threshold of the user's new physiological index can be expressed as:

[0130] T = M + k × σ × f (S);

[0131]

[0132] Among them, f(S) represents a function based on the comprehensive activity status score S, S1 represents the threshold point of a low comprehensive activity status score, S2 represents the threshold point of a high comprehensive activity status score, a1 represents the threshold adjustment coefficient for tightening the warning threshold, a2 represents the threshold adjustment coefficient for keeping the warning threshold unchanged, and a3 represents the threshold adjustment coefficient for relaxing the warning threshold (the specific values of these three adjustment coefficients are set according to the specific needs of users in the actual application scenario);

[0133] When S < S1, it means that the user is in a static or emotionally stable state, and the warning threshold can be tightened (i.e., a1 is smaller, for example, 0.9). For example, assume that the average value of the user's heart rate is 70 beats per minute, the standard deviation is 5 beats per minute, and the constant k is 2. Then the original warning threshold is 70 + 2×5 = 80 beats per minute. Using a1 = 0.9, the new warning threshold is 70 + 2×5×0.9 = 79 beats per minute, tightening the warning threshold;

[0134] When S1 ≤ S < S2, it means that the user is in a general activity state or has minor emotional fluctuations, and the warning threshold can be kept unchanged (i.e., a2 is close to 1, for example, 1.0);

[0135] When S ≥ S2, it means that the user is in a state of high-intensity exercise or has large emotional fluctuations, and the warning threshold can be appropriately relaxed (i.e., a3 is larger, for example, 1.1).

[0136] The anomaly detection module analyzes the multi-modal feature vector to obtain the anomaly probability score.

[0137] It includes the following steps

[0138] Introduce a graph neural network and list the multi-modal feature vector as a node in the graph neural network to establish a multi-modal graph structure;

[0139] According to the similarity of nodes (entity objects), use cosine similarity to define the edges in the nodes (the relationships between nodes), connect the edges with high similarity to form an edge set and initialize it (for example, if the health status feature vectors of two users are very similar, then these two nodes can be connected by an edge in the graph), forming a multi-modal graph structure;

[0140] Use the graph attention network to perform feature propagation on the multi-modal graph structure to obtain a new multi-modal feature vector;

[0141] In this step, use the graph attention network to automatically learn the dependency relationships of the multi-modal graph structure and perform feature propagation, which has strong dynamic adaptability and is suitable for different application scenarios;

[0142] Based on the new multi-modal feature vector, define the dynamic prior distribution probability and marginal probability for each user;

[0143] The marginal probability of a user can be expressed as:

[0144] Q(z)=R(z|e=0)O(e=0)+R(z|e=1)O(e=1);

[0145] Wherein, P(z|e=0) represents the conditional probability that the user is in a normal health state when the multimodal feature vector z is observed, and O(e=0) represents the dynamic prior distribution probability that the user is in a normal health state;

[0146] According to the dynamic prior probability distribution and marginal probability, the abnormal probability score of the user is calculated, and the expression is:

[0147]

[0148] in, represents the abnormal probability score of the user (i.e., the possibility that the user is in an abnormal health state), e represents the user's health state, which takes a value of 0 (normal) or 1 (abnormal), R(z|e=1) represents the conditional probability that the user is in an abnormal health state when the multimodal feature vector z is observed, O(e=1) represents the dynamic prior distribution probability that the user is in an abnormal health state, and Q(z) represents the marginal probability of the new multimodal feature vector (taking into account all possible health states, that is, the total probability of observing the multimodal feature vector z).

[0149] The model optimization module compares the anomaly probability score with the adjusted warning threshold to obtain the warning result, collects user feedback and optimizes the baseline model.

[0150] The following steps are included:

[0151] When the value of the abnormal probability score is greater than the adjusted warning threshold, the risk warning mechanism (0.5-0.9) is triggered and the risk warning level is divided;

[0152] The risk warning levels are low risk, medium risk and high risk;

[0153] Provide users with corresponding processing suggestions based on different risk warning levels;

[0154] ^

[0155] When e is less than or equal to 0.7 and greater than 0.5, the user's current health status is low risk. It is recommended that the user record key physiological indicators (such as pulse and blood pressure) at least once a day, maintain a balanced diet (reduce the intake of high-salt and high-sugar foods, avoid processed foods and sugary drinks), and choose appropriate exercise according to personal physical fitness and interests (such as walking, jogging, yoga, swimming);

[0156] ^

[0157] When e is less than 0.9 and greater than or equal to 0.7, the user's current health status is medium risk. If the user is currently doing high-intensity exercise, it is recommended to switch to moderate-intensity exercise and arrange a short lunch break (20 to 30 minutes) at noon every day. If the user finds new symptoms (such as dizziness, chest tightness, fatigue), it is recommended to consult a doctor as soon as possible and conduct a comprehensive examination of heart function and blood indicators;

[0158] ^

[0159] When e is greater than 0.9, the user's current health status is high risk. Stop all current exercise and physical labor, immediately enter a rest state, prepare the latest physical examination report, medication record, allergy history and other medical records to go to the hospital for treatment. When contacting the doctor, describe your symptoms in detail, including the time of occurrence, duration, severity, etc. After arriving at the hospital, take appropriate measures (such as taking medicine) according to the doctor's instructions;

[0160] The corresponding processing suggestions are warning results;

[0161] The AR interface was designed using the Unity3D design platform, and the Blender modeling tool was used to create a 3D model of the user's body with animation effects (such as heartbeat and breathing) (with a focus on the heart area);

[0162] The AR interface includes a UI area (used to display warning results), a user interaction area, a text description of the warning, and a warning color code (defined according to the R wave peak, PQ interval, and QRS complex wave morphology of the user's electrocardiogram, and the color of the heart area can change dynamically);

[0163] In this step, AR technology is introduced to design an AR interface and a 3D model of the user's body. Through the 3D model and its animation effects, users can intuitively see their physical condition in a virtual environment, improving the effectiveness of their health management and self-monitoring.

[0164] Synchronize the warning results and the user's body 3D model to the AR interface;

[0165] The warning results are converted into XML format, and the communication interface between the server and the AR interface is established using the WebSocket real-time communication protocol. At the same time, according to different warning results, corresponding state mapping rules are defined for the user's body 3D model. For example, when the warning result is low risk, the heart area is green and the heartbeat animation is stable; when the risk is medium, the heart area is yellow and the heartbeat animation is accelerated; when the risk is high, the heart area is red and the heartbeat animation is violent;

[0166] Write scripts in the AR interface to dynamically update the status of the 3D model based on the warning results. For example, by adjusting the parameters in the animation controller, you can control the changes in animation effects such as heartbeat and breathing. At the same time, add visual feedback elements (such as flashing lights, vibration frequency, etc.) to this 3D model to further enhance the user experience.

[0167] Integrate the converted warning results and the updated 3D model of the user's body and synchronize them to the AR interface;

[0168] The user views the warning results on the AR interface and interacts with the 3D model of his or her body through gestures (such as pinching to zoom in, horizontal movement) and voice commands (such as zooming in on the heart area, flipping the model) to obtain interactive results;

[0169] Use gesture recognition plugin (Vuforia) to realize user gesture recognition function, for example, users can select specific body parts through gestures to view detailed health data;

[0170] Integrate the speech recognition API (Google Speech API) to enable users to query health information through voice commands. For example, users can say "check my heart rate" or "give me some treatment suggestions". At the same time, provide users with instant voice prompts. For example, when users click on a certain body part, a voice prompt is played to explain the current health status of that part.

[0171] Assume that the user selects the heart area through voice commands. The following is the interaction process and results:

[0172] The 3D model zooms in on the heart area (displaying the current ECG physiological indicators). Users can further choose to view their own ECG physiological indicators. The 3D model will display a detailed ECG waveform and provide a short video tutorial to explain the abnormalities in the current ECG and their causes.

[0173] When the user selects a solution, the 3D model will display the previously generated warning results (for example, it is recommended that the user reduce caffeine intake and increase exercise frequency). The user uses voice commands to simulate increasing or decreasing the amount of exercise according to his or her physical condition (such as should I increase the amount of exercise). At this time, the 3D model will simulate and display the user's heart rate change trend and blood pressure after increasing or decreasing the amount of exercise, so as to formulate a specific action plan and timetable;

[0174] Provide users with a feedback time window (48 hours), use mobile APP and voice interaction devices to send feedback requests to users, and collect user feedback information based on the interaction results;

[0175] Feedback information includes the user's health status (whether there are any discomfort symptoms, whether they have taken any medication), action status (whether they have reduced their exercise intensity, whether they have contacted a doctor), medical follow-up (whether they have undergone further medical examinations and what the examination results are), and warning evaluation (whether they think the warning is accurate and whether they are satisfied with the handling suggestions);

[0176] This step ensures that users submit feedback information in a timely manner by setting a feedback time window and sending feedback requests, so as to better understand the actual situation of users and continuously optimize early warning mechanisms and handling suggestions.

[0177] Optimize the baseline model based on user feedback;

[0178] Using the statistical method of mutual information, blood pressure, blood oxygen saturation, exercise intensity (such as daily steps and calories consumed) and sleep quality are selected from user feedback information as the most important quantitative indicators. The reason is that a sudden increase or decrease in blood pressure may be an early signal of potential health problems, such as acute heart failure and renal insufficiency.

[0179] Blood oxygen saturation (less than 90%) may indicate respiratory problems, such as chronic obstructive pulmonary disease (COPD), asthma, etc. In acute situations (such as infection, dyspnea), a decrease in blood oxygen saturation can serve as an early warning sign of worsening condition;

[0180] Moderate physical activity helps maintain cardiovascular health, enhance immunity and improve mental health, but excessive exercise or lack of exercise may lead to cardiovascular health problems, and long-term lack of sleep or poor sleep quality is associated with a variety of health problems, such as diabetes and depression;

[0181] Use XGBoost (a machine learning algorithm for gradient boosting decision trees) to generate a sample classification data set and a real-time data set and extract the training set. Change a certain quantitative indicator into a data format dedicated to XGBoost and form a set of split nodes in the decision tree. At the same time, use its built-in scoring function to calculate the average gain of a certain quantitative indicator used for splitting nodes and obtain the weight of the quantitative indicator.

[0182] The weight of adjusting the quantitative index (taking sleep quality as an example) is expressed as:

[0183]

[0184] Among them, w(H) represents the weight of the quantitative index of the user's sleep quality, Y represents the number of decision trees, S u represents the set of split nodes of the u-th tree, 1(G∈S u) represents the indicator function of the quantitative index of sleep quality appearing in the split node of the u-th tree (1 if it appears, otherwise 0);

[0185] Use the learning mechanism of the incremental database to add the adjusted weights of quantitative indicators and user feedback information to the real-time dataset;

[0186] When data flows in, real-time quantitative indicators are extracted and converted to generate new indicators that can be used to optimize the baseline model, such as entropy features (to detect acute disease attacks). At the same time, recursive feature elimination technology is applied to gradually remove unimportant indicators and retain only the indicators that contribute most to the optimization model to complete the optimization of the baseline model.

[0187] In summary, the present invention uses a baseline model to set a warning threshold, and dynamically adjusts the warning threshold according to the user's real-time activity status. By dynamically adjusting the warning threshold, it can better adapt to the user's daily activity changes, reduce unnecessary false alarms and missed alarms, and improve the reliability of the warning and user experience. In addition, the dynamic threshold adjustment mechanism can also capture the impact of user emotional fluctuations on physiological indicators. For high-intensity exercise and static state, the warning threshold can be relaxed and tightened respectively to ensure the sensitivity and accuracy of the warning, further improving the intelligent level of warning monitoring.

[0188] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A dynamic electrocardiogram early warning system based on artificial intelligence, characterized by: include, Data collection module, which collects the user's electrocardiogram physiological indicators and environmental data and extracts multimodal features; The data fusion module selects the deep learning model and inputs the multimodal features to form a multimodal feature vector; Model building module, which builds a baseline model for users based on multimodal feature vectors; The threshold setting module sets the warning threshold based on the baseline model; Threshold adjustment module, dynamically adjusts the warning threshold according to the user's real-time activity status; The anomaly detection module analyzes the multimodal feature vectors to obtain anomaly probability scores; The model optimization module compares the anomaly probability score with the adjusted warning threshold to obtain the warning result, collects user feedback and optimizes the baseline model.

2. The artificial intelligence-based dynamic electrocardiogram early warning system according to claim 1, characterized in that: The user's electrocardiogram physiological indicators include heart rate, blood pressure, blood oxygen saturation, body temperature and respiratory rate.

3. The artificial intelligence-based dynamic electrocardiogram early warning system according to claim 2, characterized in that: The environmental data includes motion status, location information and weather conditions.

4. The artificial intelligence-based dynamic electrocardiogram early warning system according to claim 3, characterized in that: The multimodal features include the R wave peak, PQ interval, QRS complex wave morphology of the electrocardiogram, systolic pressure, diastolic pressure of blood pressure, SpO2 value of blood oxygen saturation, average change rate of body temperature and respiratory cycle length of respiratory frequency.

5. The artificial intelligence-based dynamic electrocardiogram early warning system according to claim 4, characterized in that: Select a deep learning model and input multimodal features to form a multimodal feature vector. The following steps are included: Select Transformer as the core architecture of the model to build a deep learning model; Integrate weight calculation units in deep learning models; Use the weighted summation method to fuse multimodal features of the weight calculation unit and normalize them; PCA is used to reduce the dimension of the fused multimodal features to obtain a multimodal feature vector.

6. The artificial intelligence-based dynamic electrocardiogram early warning system according to claim 5, characterized in that: Based on the multimodal feature vector, a baseline model is established for the user, including the following steps: Select the long short-term memory network as the core structure of the baseline model; Use the PyTorch framework to build a long short-term memory network model; The multimodal feature vector is input into the long short-term memory network model, and the hidden layer is stacked and trained using the Adam optimizer to obtain the trained long short-term memory network model; Based on the trained long short-term memory network model, the user's electrocardiogram physiological indicators are divided into time windows of set lengths through sliding window technology to generate time series segments; Add the corresponding indicator label to each indicator label, obtain the time series segment with the indicator label and convert it into tensor form; The converted time series segments of the indicator labels and the long short-term memory network model are assembled to obtain a complete baseline model.

7. The artificial intelligence-based dynamic electrocardiogram early warning system according to claim 6, characterized in that: Based on the baseline model, set the warning threshold, including the following steps: Use TensorFlow to forward propagate the time steps of the baseline model, extract the implicit state of each time step and combine them into an implicit state vector; Use clustering algorithms to perform cluster analysis on latent state vectors and identify typical activity patterns of users; Correlate the user's typical activity patterns with the user's physiological indicators to form a data set of user behavior; Based on the user behavior data set, all time steps are grouped according to the corresponding clusters, and the mean and standard deviation of the user's physiological indicators are calculated; According to the mean and standard deviation of the user's physiological indicators, the initial warning threshold is set using the fixed multiple standard deviation method.

8. The artificial intelligence-based dynamic electrocardiogram early warning system according to claim 7, characterized in that: Dynamically adjust the warning threshold according to the user's real-time activity status, including the following steps: Use the constructed deep learning model to infer the multimodal feature vectors and identify the user's real-time activity status; Select heart rate variability as an indicator of user emotion fluctuations, and calculate the standard deviation and mean of heart rate variability; Combine the user's current real-time activity status and the standard deviation and mean of heart rate variability to calculate the user's comprehensive activity status score, expressed as: Where D is the standard deviation of heart rate variability, N is the number of heart beat intervals, and R is i represents the i-th heartbeat interval, represents the average of all heartbeat intervals, S represents the user's comprehensive activity status score, p(y=j) represents the probability that the user is in the jth motion state, y represents the user's motion state variable, F represents the mean of heart rate variability, w1 represents the weight of the motion state, and w2 represents the weight of the user's emotional fluctuation; Dynamically adjust the warning threshold of physiological indicators based on the user's comprehensive activity status score.

9. The artificial intelligence-based dynamic electrocardiogram early warning system according to claim 8, characterized in that: Analyzing the multimodal feature vector to obtain the abnormal probability score includes the following steps: Introduce graph neural network and list multimodal feature vectors as nodes in the graph neural network to establish a multimodal graph structure; Use graph attention network to propagate features of multimodal graph structure and obtain new multimodal feature vectors; Based on the new multimodal feature vector, dynamic prior distribution probability and marginal probability are defined for each user; According to the dynamic prior probability distribution and marginal probability, the abnormal probability score of the user is calculated, and the expression is: in, represents the abnormal probability score of the user, e represents the health status of the user, R(z|e=1) represents the conditional probability that the user is in an abnormal health state when a new multimodal feature vector z is observed, O(e=1) represents the dynamic prior distribution probability that the user is in an abnormal health state, and Q(z) represents the marginal probability of the new multimodal feature vector.

10. The artificial intelligence-based dynamic electrocardiogram early warning system according to claim 9, characterized in that: Compare the abnormal probability score with the adjusted warning threshold to obtain the warning result, collect user feedback and optimize the baseline model, including the following steps: When the value of the abnormal probability score is greater than the adjusted warning threshold, the risk warning mechanism is triggered and the risk warning level is divided; Provide users with corresponding processing suggestions based on different risk warning levels; The corresponding processing suggestion is a warning result; The AR interface was designed using the Unity3D design platform, and the Blender modeling tool was used to create a 3D model of the user's body with animation effects; Synchronize the warning results and the user's body 3D model to the AR interface; The user views the warning results on the AR interface and interacts with the 3D model of his or her body through gestures and voice commands to obtain interactive results; Provide users with a feedback time window, use mobile APP and voice interaction devices to send feedback requests to users, and collect user feedback information based on the interaction results; The feedback information includes the user's health status, action status, medical follow-up and early warning evaluation; Optimize the baseline model based on user feedback.

Images