Method and system for constructing cardiac resuscitation visualization process

Abstract

The invention discloses a method for constructing a cardiac resuscitation visualization process. The method comprises the following steps: collecting multi-source data; constructing an arrhythmia automatic identification model based on deep learning; judging the vital sign state of the patient; analyzing by using a machine learning algorithm to evaluate the severity of the illness state; through the Internet of Things technology, real-time positioning and state monitoring are carried out on first-aid resources in a hospital; an optimal first-aid resource allocation scheme is automatically calculated; analyzing the relationship between each factor and the cured patient to obtain key factors and potential problem points of the treatment success rate; a personalized cardiac resuscitation advanced life support process is generated based on an artificial intelligence algorithm and is visually displayed, and first-aid operation steps and time node guidance are provided for first-aid personnel. According to the method, the whole process of cardiac resuscitation is realized by fully utilizing an artificial intelligence technology, and intelligence and visualization of each link are realized, so that the treatment timeliness and success rate of sudden cardiac death patients are remarkably improved.

Description

Technical Field

[0001] The present invention generally relates to the field of artificial intelligence technology. More specifically, the present invention relates to a method and system for constructing a visualized cardiac resuscitation process. Background Art

[0002] In the modern medical field, cardiac arrest is a life-threatening emergency, and timely and effective resuscitation is crucial. With the continuous progress of medical technology and the gradual improvement of people's first aid awareness, the cardiac resuscitation process has received high attention in both clinical practice and first aid training.

[0003] Although a relatively standardized operation process for cardiac resuscitation has been formed at present, these processes are summarized based on a large amount of medical research and clinical experience, aiming to maximize the survival rate and rehabilitation quality of patients. However, there are still some problems to be solved urgently in the actual application of the existing cardiac resuscitation processes. On the one hand, despite the standard processes, during the execution process, due to the complexity of the on-site situation and individual differences among first aid personnel, it is difficult to ensure the standardization and consistency of operations. For example, during external chest compressions, the depth, frequency, and position of compressions may vary due to the skill level and experience of first aid personnel, and these subtle differences may have a significant impact on the resuscitation effect. On the other hand, the traditional cardiac resuscitation process is relatively fixed and lacks the ability to dynamically adjust according to the patient's real-time status and on-site conditions. The factors such as the age, physical condition, and underlying diseases of different patients are different, and a single standard process cannot fully meet the personalized first aid needs. Moreover, the data collection and analysis during the cardiac resuscitation process are not sufficient at present, making it difficult to deeply understand the precise relationship between each operation link and the final treatment effect of the patient, and thus unable to optimize the process targeted. These problems lead to a large room for improvement in the success rate of cardiac resuscitation, and there is an urgent need for an innovative technical solution that can monitor the cardiac resuscitation process in real time, accurately evaluate it, and adaptively optimize it according to the actual situation to improve the treatment effect and survival rate of patients with cardiac arrest.

[0004] In view of this, there is an urgent need to provide a method for constructing a visualized cardiac resuscitation process to improve the treatment effect and survival rate of patients with cardiac arrest. Summary of the Invention

[0005] In order to solve at least one or more of the above-mentioned technical problems, the present invention proposes a method for constructing a visualized cardiac resuscitation process in multiple aspects.

[0006] In a first aspect, the present invention provides a method for constructing a visualized cardiac resuscitation process, including: constructing a big data platform by collecting multi-source data; constructing an automatic arrhythmia recognition model based on deep learning to perform real-time analysis on electrocardiogram data to identify characteristic waveforms of cardiac arrest; analyzing the physical characteristics of a patient using image recognition technology according to on-site video monitoring data to determine the vital sign status of the patient; analyzing using a machine learning algorithm according to the basic information of the patient, the vital sign status, and real-time data collected on-site to evaluate the severity of the condition;

[0007] According to the visualized large screen of resource allocation in the hospital command center, the first aid resources in the hospital are subjected to real-time positioning and status monitoring through Internet of Things technology; according to the location and condition of the patient, as well as the location and status of the first aid resources, an optimal first aid resource allocation plan is automatically calculated through an artificial intelligence algorithm;

[0008] By collecting and sorting out the detailed data of each cardiac resuscitation event, a first aid effect evaluation model is established using a machine learning algorithm to analyze the relationship between various factors and the patient after cure, so as to obtain the key factors and potential problem points of the rescue success rate; generating a personalized advanced life support process for cardiac resuscitation based on an artificial intelligence algorithm and performing visualized display to provide the first aid personnel with operation steps and time node guidelines for first aid; according to the evaluation results of the first aid effect evaluation model, automatically generating process optimization suggestions using an artificial intelligence algorithm to achieve dynamic optimization and adaptive adjustment of the advanced life support process for cardiac resuscitation. Through an artificial intelligence algorithm, the on-site operation data collected is compared and analyzed with the operation parameters of the advanced life support process for cardiac resuscitation to adjust the cardiac resuscitation operation process in real time.

[0009] In a second aspect, the present invention provides a method and system for constructing a visualized cardiac resuscitation process, including: a data acquisition module: used to construct a big data platform by collecting multi-source data; an arrhythmia automatic recognition model construction module: used to construct an arrhythmia automatic recognition model based on deep learning to perform real-time analysis on electrocardiogram data to identify characteristic waveforms of cardiac arrest; a vital sign analysis module: used to analyze the physical characteristics of a patient using image recognition technology based on on-site video surveillance data to determine the patient's vital sign status; a condition assessment module: used to analyze using machine learning algorithms based on the patient's basic information, the vital sign status, and real-time data collected on-site to assess the severity of the condition; an emergency resource management module: used to perform real-time positioning and status monitoring of in-hospital emergency resources through Internet of Things technology according to the visualized large screen of resource allocation in the hospital command center; an emergency resource scheduling module: used to automatically calculate the best emergency resource allocation plan through artificial intelligence algorithms based on the patient's location and condition, as well as the location and status of emergency resources; an emergency treatment effect evaluation model construction module: used to establish an emergency treatment effect evaluation model using machine learning algorithms by collecting and collating detailed data of each cardiac resuscitation event, and analyze the relationship between various factors and the patient after recovery to obtain key factors for the success rate of treatment and potential problem points. And

[0010] a cardiac resuscitation process generation module: used to generate a personalized advanced life support process for cardiac resuscitation based on artificial intelligence algorithms and perform visualized display, providing operation steps and time node guidelines for first aid personnel; a cardiac resuscitation process optimization module: used to automatically generate cardiac resuscitation process optimization suggestions using artificial intelligence algorithms according to the evaluation results of the emergency treatment effect evaluation model to achieve dynamic optimization and adaptive adjustment of the advanced life support process for cardiac resuscitation; a cardiac resuscitation process adjustment module: used to perform comparative analysis on the on-site operation data collected and the operation parameters of the advanced life support process for cardiac resuscitation through artificial intelligence algorithms to adjust the cardiac resuscitation operation process in real time; a visualized display module: used to perform visualized display of each module of the system for the visualized cardiac resuscitation process through a visualized screen.

[0011] Through a method for constructing a visual process of cardiopulmonary resuscitation provided as above, embodiments of the present invention obtain and organize cardiopulmonary resuscitation-related data from multiple data sources to form a comprehensive, accurate, and uniformly formatted data set. Further, in some embodiments, machine learning algorithms are used to determine key factors, which helps to focus on the core elements affecting the first aid effect, making subsequent process optimization more targeted, avoiding wasting resources on irrelevant factors, and improving the optimization efficiency and effect. Furthermore, in some embodiments, by constructing a process optimization model based on reinforcement learning, the cardiopulmonary resuscitation process parameters and links are reasonably defined as the state space, and the optimizable action space is clarified, enabling the model to accurately perceive the state of the first aid process and make decisions within a limited and practically meaningful action range, enhancing the operability of the model and the rationality of decision-making. By reasonably setting the weight coefficients to balance the importance of various factors, the model fully considers various actual situations in the process of pursuing the maximization of the first aid effect, avoiding one-sided pursuit of a certain index while ignoring other important factors, and improving the clinical practicability and safety of the optimization strategy. Furthermore, in some embodiments, the data collected after each first aid is fed back to the model, enriching the training data of the model and enabling it to learn more first aid experiences and optimization strategies in different situations.

[0012] In summary, by fully utilizing artificial intelligence technology throughout the entire process of cardiopulmonary resuscitation, from data collection and diagnosis, resource allocation, operation guidance to process evaluation and improvement, the intelligentization and visualization of each link are realized, thereby significantly improving the timeliness and success rate of treating patients with sudden cardiac death and providing a more reliable guarantee for saving patients' lives. At the same time, with the continuous accumulation of data and the continuous optimization of algorithms, the performance of the system and the treatment effect will continue to improve, promoting the development and progress of cardiopulmonary resuscitation technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features, and advantages of the exemplary embodiments of the present invention will become readily understood. In the drawings, several embodiments of the present invention are shown in an exemplary rather than restrictive manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0014] Figure 1 An exemplary flowchart of a method for constructing a visual process of cardiopulmonary resuscitation according to some embodiments of the present invention is shown;

[0015] Figure 2 An exemplary structural block diagram of a system for constructing a visual process of cardiopulmonary resuscitation according to an embodiment of the present invention is shown. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present invention.

[0017] The following will specifically describe the embodiments of the present invention in conjunction with the accompanying drawings.

[0018] Figure 1 FIG. shows an exemplary flowchart of a method 100 for constructing a visual process of cardiopulmonary resuscitation according to some embodiments of the present invention;

[0019] As Figure 1 shown, in method 100, the present invention provides a method for constructing a visual process of cardiopulmonary resuscitation, including the following steps: Step S101: Construct a big data platform by collecting multi-source data.

[0020] When constructing a visual process of cardiopulmonary resuscitation, first, a multi-source data access system needs to be established. By collecting multi-source data, a big data platform integrating multi-source data is constructed.

[0021] Specifically, when constructing an all-round data collection network for data collection, various medical device data interfaces in the hospital can be integrated, including but not limited to electrocardiogram monitors, blood pressure monitors, blood oxygen saturation detectors, etc. At the same time, it is docked with the hospital information system (HIS), emergency information system (EMIS), electrocardiogram monitoring system, etc., to obtain the patient's basic information (such as name, age, past medical history, etc.), electronic medical record data, vital sign data (electrocardiogram, blood pressure, blood oxygen saturation, etc.), and the location of the event in real time.

[0022] Access the mobile medical device data of the emergency vehicle, including the patient monitoring data during transportation, to ensure the continuity of data throughout the process from the scene of the disease to the hospital.

[0023] At the emergency scene, wearable medical devices and mobile data collection terminals are used to collect on-site information when the patient gets sick (such as environmental data, body posture when getting sick, etc.) and data such as the preliminary diagnosis and first aid measures taken by the emergency personnel on-site, and transmit them back to the hospital data center in real time through the 5G network.

[0024] Furthermore, the big data platform is constructed to store and process data. After collecting multi-source data, data cleaning algorithms are used to remove noise, outliers, and duplicate data in the collected data to ensure the accuracy and integrity of the data. And the data from different data sources are standardized and normalized to make them have a unified data format and dimension, which is convenient for subsequent data analysis and utilization.

[0025] Further, the process proceeds to step S102: constructing an automatic arrhythmia recognition model based on deep learning to perform real-time analysis on electrocardiogram data to identify characteristic waveforms of cardiac arrest.

[0026] In one embodiment, the constructing an automatic arrhythmia recognition model based on deep learning to perform real-time analysis on electrocardiogram data to identify characteristic waveforms of cardiac arrest includes: using a convolutional neural network architecture as the basic framework of the automatic arrhythmia recognition model, combined with an attention mechanism to enable the model to focus on key feature regions related to arrhythmia in the electrocardiogram signal; and combining a long short-term memory network to model the time series information of the electrocardiogram signal to obtain the dependency relationship and dynamic change law of the electrocardiogram signal in the time dimension.

[0027] First, collect a large amount of electrocardiogram (ECG) data from electrocardiogram monitoring devices in multiple departments of the hospital (such as the cardiology department, emergency department, intensive care unit, etc.), covering cases of various age groups, genders, underlying diseases, and different arrhythmia types. At the same time, incorporate external publicly available authoritative electrocardiogram databases, such as the MIT-BIH Arrhythmia Database, etc., to enrich the diversity and representativeness of the data.

[0028] For the collected electrocardiogram data, strict quality control and screening are carried out. Remove records with excessive signal noise, severe baseline drift, and obvious data missing to ensure the accuracy and usability of the data.

[0029] Organize a professional cardiovascular doctor team and use a standardized arrhythmia diagnosis standard and classification system to perform detailed annotation on the screened electrocardiogram data. The annotation content includes normal sinus rhythm, various arrhythmia types (such as ventricular fibrillation, pulseless ventricular tachycardia, atrial fibrillation, supraventricular tachycardia, etc.), and the start and end positions and categories of characteristic waveforms related to cardiac arrest. The annotation process uses a multi-person cross-check method to improve the accuracy and consistency of the annotation.

[0030] To increase the quantity and diversity of training data, various data augmentation techniques are used. For example, perform random time stretching, frequency offset, amplitude scaling, etc. on the electrocardiogram signal to simulate signal changes that may occur in actual clinical situations; enhance the model's robustness to noise by adding Gaussian noise, baseline drift noise, etc. at different levels; and can also perform segment interception and splicing on the electrocardiogram signal to generate new training samples to improve the model's recognition ability for different rhythm patterns.

[0031] Specifically, during the construction of the automatic arrhythmia recognition model, a convolutional neural network (CNN) architecture is adopted as the basic framework of the automatic arrhythmia recognition model. Through the combination of convolutional layers, pooling layers, and fully connected layers, the CNN architecture can automatically learn the feature patterns in the electrocardiogram (ECG) signals.

[0032] Multiple convolutional layers are designed, and each convolutional layer uses convolutional kernels of different sizes (such as 3x3, 5x5, 7x7, etc.) to capture feature information at different scales in the ECG signals. The convolutional kernels slide over the signals to perform convolutional operations, extract local features, and introduce non-linear characteristics through activation functions (such as ReLU) to enhance the expressive power of the model.

[0033] To enable the model to pay more attention to the key feature regions related to arrhythmias in the ECG signals, an attention mechanism is introduced. For example, a dual attention module based on channel attention and spatial attention is adopted. This module can adaptively learn the importance weights of each channel and each spatial position, thereby enhancing the important features in the ECG signals, suppressing the interference of irrelevant information, and improving the feature extraction ability and recognition accuracy of the model.

[0034] Furthermore, considering the temporal characteristics of the ECG signals, an LSTM layer is added on the basis of the CNN architecture. LSTM can model the time series information of the ECG signals, capture the dependency relationships and dynamic change laws in the time dimension of the signals, and contribute to better recognition of the occurrence and development processes of arrhythmias. Especially for some arrhythmia types with complex rhythm changes, such as atrial fibrillation, it can provide more accurate recognition results.

[0035] The labeled dataset is divided into a training set, a validation set, and a test set according to a certain ratio (such as 70%, 15%, 15%). The training set is used to train the model. During the training process, a mini-batch stochastic gradient descent (SGD) algorithm or its variants (such as Adam, Adagrad, etc.) is adopted to optimize the parameters of the model to minimize the cross-entropy loss function between the prediction results and the true labels.

[0036] To prevent the model from overfitting, various regularization techniques are adopted, such as L1 and L2 regularization, Dropout layers, etc. The Dropout layer randomly discards the outputs of some neurons during the training process, enabling the model to have different network structures in different training batches, thereby enhancing the generalization ability of the model; L1 and L2 regularization prevent the parameters from being too large and reduce the complexity of the model by imposing constraints on the model parameters.

[0037] Tune multiple hyperparameters of the model, such as the number of convolutional kernels, the depth of convolutional layers, the number of LSTM units, the learning rate, the Dropout probability, etc., to find the optimal model configuration. Adopt hyperparameter search algorithms, such as grid search, random search, or methods based on Bayesian optimization, to search and evaluate within the given range of hyperparameter values, and select the hyperparameter combination with the best performance on the validation set as the final model configuration.

[0038] During the training process, regularly evaluate the performance of the model using the validation set. Evaluation metrics include accuracy, recall, F1 score, sensitivity, specificity, etc., and comprehensively consider the model's ability to identify different types of arrhythmias and its overall performance. According to the evaluation results on the validation set, save the model checkpoint with the best performance, and after the training ends, select the model with the best performance on the validation set as the final arrhythmia recognition model.

[0039] Through the above steps of the embodiments, an accurate and efficient intelligent arrhythmia recognition model can be constructed, providing strong support for early diagnosis and rapid treatment in the cardiac resuscitation process, and significantly improving the success rate and timeliness of treating patients with sudden cardiac death.

[0040] Next, specifically describe the specific formula calculation process through an embodiment:

[0041] Assume that the input electrocardiogram signal data is X, and its shape is (batch s size, sequence l length, num c channels), where batch s size is the number of samples processed in one training, sequence l length is the time series length of the electrocardiogram signal, and num c channels is usually 1 (single-lead electrocardiogram signal).

[0042] Before inputting into the model, first perform normalization processing on the data, map the values of the electrocardiogram signal to a specific range, such as [-1, 1], and the formula is as follows: where μ is the mean value of the electrocardiogram signal in the training dataset, and σ is the standard deviation. Through this normalization operation, the training speed of the model can be accelerated and its stability can be improved.

[0043] Furthermore, the formula for the convolutional neural network (CNN) layer is:

[0044] Z convl = Conv1D(X norm , W conv1 , b conv1, stride = 1, padding ='same')

[0045] A conv1 = ReLU(Z conv1 )

[0046] Z pool1 = MaxPool1D(A conv1 , pool s size = 2, stride = 2)

[0047] Z conv2 = Conv1D(Z pool1 , W conv2 , b conv2 , stride = 1, padding ='same')

[0048] A conv2 = ReLU(Z conv2 )

[0049] Z pool2 = MaxPoollD(A conv2 , pool s size = 2, stride = 2)

[0050] For the convolutional kernel parameters (W conv1 , W conv2 and b conv1 , b conv2 ): In the first - layer convolution (Conv1D), set the number of convolutional kernels to n conv1 = 32, the size of the convolutional kernel to k conv1 = 5. These convolutional kernels slide over the input electrocardiogram signal and, through convolution operations with local regions of the signal, extract different feature maps. The weight matrix W conv1 has a shape of (k conv1 , num c channels, n conv1 ), and the bias vector b conv1 has a shape of (n conv1 ,). For example, for an electrocardiogram signal with a length of sequence l ength, after passing through the first convolutional layer, the shape of the output feature map is (batch s size, sequence l ength, n conv1 ), where each feature map represents the feature information of the electrocardiogram signal extracted under different convolutional kernels, such as the morphological features of the QRS complex, the preliminary features of the P wave and T wave, etc.

[0051] In the second - layer convolution, the number of convolutional kernels is increased to n conv2 = 64, the size of the convolutional kernel is k conv2 = 3, and the shape of the weight matrix W conv2 is (k conv2 , n conv1 , n conv2 ). The shape of the bias vector b conv2 is (n conv2 ). The second - layer convolution can capture the subtle differences in waveforms and the correlation relationships between features in the electrocardiogram signal based on the features extracted in the first layer, such as in different arrhythmia situations.

[0052] For the activation function (ReLU): The ReLU function (A = ReLU(Z)=max(0, Z)) is used to introduce non - linear characteristics and enhance the expressive power of the model. By converting the linear feature map output by the convolutional layer into a non - linear feature map, the model can learn more complex electrocardiogram signal feature patterns, rather than just simple linear combinations.

[0053] Pooling layer (MaxPooI1D) parameters: After each layer of convolution, the max - pooling layer (MaxPool1D) is used for downsampling operations to reduce the size of the feature map, reduce the computational amount, and extract the main features. The pooling window size of the first - layer pooling layer is pool s ize = 2, and the stride is 2. This means that it slides a window of size 2 on the feature map with a stride of 2 and takes the maximum value within the window as the output. After the first - layer pooling, the length of the feature map becomes half of the original, that is, the shape becomes With the same settings for the second - layer pooling layer, the length of the feature map is further reduced to The output shape is The pooling operation helps the model to extract the coarse - grained features of the electrocardiogram signal, while maintaining sensitivity to important features and reducing the over - fitting problem caused by small signal variations.

[0054] The formula of the attention mechanism layer includes:

[0055] e t = Attention(Z pool2 , W e , b e ); α t = Softmax(e t );

[0056] Among them, the attention mechanism parameters (W e and b e ): A simple attention mechanism based on a fully - connected layer is adopted. The weight matrix W eis of shape (n conv2 , 1), and the bias vector b e is of shape (1,). For the pooled feature sequence Z pool2 , its shape is (batch s size, T, n conv2 )(where ), and the attention scores e t are calculated through the fully connected layer, and its shape is (batch s size, T, 1). Then, the attention weights α t are obtained through the Softmax function, with a shape of (batch s size, T, 1). Finally, the attention-weighted feature representation S t is obtained by performing weighted summation on the feature sequence, with a shape of (batch s size, n conv2 ). The attention mechanism enables the model to automatically learn the importance weights of different time steps in the feature sequence, giving higher weights to the key feature regions related to arrhythmia in the electrocardiogram signal, thereby highlighting these important features, suppressing the interference of irrelevant information, and improving the model's ability to extract and recognize arrhythmia features.

[0057] In the long short-term memory network (LSTM) layer: Z lstm = LSTM(S t , h0, W lstm , b lstm );

[0058] Among them, regarding the LSTM cell parameters (W lstm and b lstm ): The LSTM layer is used to perform time series modeling on the features weighted by attention, learning the dependency relationship and dynamic change law of the electrocardiogram signal in the time dimension. The weight matrix W lstm contains the input weight, forget gate weight, output gate weight, and cell state weight, and its shape is (4 * n lstm , n conv2 + n lstm ), where n lstm is the number of hidden units of the LSTM cell, set to 128. The bias vector b lstm is of shape (4 * n lstm,). h0 is the initial hidden state, initialized as a vector of all zeros. At each time step, the LSTM cell determines whether to retain or forget the information in the cell state based on the input features and the hidden state at the previous moment through a gating mechanism, and generates the output and hidden state at the current moment. For example, for the features processed by the attention mechanism, the LSTM layer can capture the rhythm change patterns of the electrocardiogram signal over a period of time, such as the irregular fluctuations of atrial electrical activity during atrial fibrillation and the chaotic state during ventricular fibrillation. These time series features are crucial for accurately identifying the types of arrhythmias.

[0059] Fully connected layer (FC) and output layer: Z fc = Linear(Z lsim , W fc , b fc ); y = Softmax(Z fc );

[0060] Among them, the fully connected layer parameters (W fc and b fc ): The fully connected layer maps the output of the LSTM layer to the final probability distribution of arrhythmia categories. The weight matrix W fc has a shape of (n lstm , n classes ), where n classes is the number of arrhythmia categories (for example, including normal rhythm, ventricular fibrillation, pulseless ventricular tachycardia, atrial fibrillation, etc., set as n classes = 5), and the bias vector b fc has a shape of (n classes ). Through the linear transformation of the fully connected layer, the feature vector output by the LSTM is converted into scores corresponding to each category, and then these scores are converted into a probability distribution through the Softmax function.

[0061] Softmax function: Used to convert the scores output by the fully connected layer into probability values for each arrhythmia category, such that the sum of the probabilities of all categories is 1. The formula is:

[0062] The final output y is a tensor with a shape of (batch s size, n classes ), where each element y ij represents the probability that the i-th sample belongs to the j-th arrhythmia category.

[0063] During the training process of the embodiment, the model will, according to the input electrocardiogram data, in accordance with the above formulas and parameter settings, sequentially perform feature extraction and classification prediction through the convolutional layer, attention mechanism layer, LSTM layer, and fully connected layer, and update the model's parameters according to the gradient information of the loss function through the backpropagation algorithm. After multiple iterative trainings, the recognition accuracy of the characteristic waveforms of cardiac arrest and other types of arrhythmias will be gradually improved, and finally, the real-time and accurate analysis of electrocardiogram data will be achieved, providing reliable technical support for the cardiac resuscitation process.

[0064] Further, the process proceeds to step S103: According to the on-site video surveillance data, use image recognition technology to analyze the patient's physical characteristics to determine the patient's vital sign status.

[0065] In one embodiment, it may include: First, extract the image frame sequence from the on-site video surveillance data to obtain an image sequence dataset. Based on the high-resolution and stable frame rate video data obtained by the surveillance cameras from multiple angles on-site, ensure that the whole body of the patient is covered, including key parts such as the face, limbs, and torso. Perform time synchronization processing on the collected video data and mark accurate timestamps for subsequent fusion analysis with other data.

[0066] Extract image frames from the video at fixed time intervals (such as 5 frames per second) to form an image sequence dataset. Adopt a screening algorithm based on image quality assessment to remove blurred, unevenly illuminated, or occluded image frames, and retain clear, complete, and representative image frames for subsequent analysis.

[0067] Then, construct a feature extraction network for multi-scale feature fusion, and use the image sequence dataset to obtain the patient's body part localization and local feature extraction.

[0068] Construct a CNN architecture for multi-scale feature fusion, the formula is as follows:

[0069] F l =Conv l (F l-1 , W l , b l )+BN(F l-1 )+ReLU(F l-1 );

[0070] F s =Upsample(Conv s (F l , W s , b s )+BN(F l )+ReLU(F l ));

[0071] F m = Concat(F l ,F s ).

[0072] Among them, F l-1 is the feature map of the upper layer, and F l is the feature map of the current convolutional layer. W l and b l are the weights and biases of the convolutional layer. BN is the batch normalization operation, ReLU is the activation function, Conv s is a convolutional layer with a small convolutional kernel (such as 3×3) used to generate high-resolution feature maps. Upsample is the upsampling operation, F s is the feature map after upsampling, and F m is the fused feature map.

[0073] Convolutional layer parameter settings: The starting layer of the network uses a larger convolutional kernel (such as 7×7) with a stride of 2 to quickly extract the general features of the image and reduce the resolution, thereby reducing the computational amount. As the network depth increases, the convolutional kernel size is gradually reduced, and the number of feature maps is increased to extract more detailed and abstract features. For example, the middle layer can be set with a 5×5 convolutional kernel, and the last few layers use a 3×3 convolutional kernel. The number of feature maps in each layer gradually increases from 32 to 256 or more, and the specific number is adjusted according to the complexity of the dataset and the performance requirements of the model.

[0074] Batch normalization layer (BN) parameter settings: The momentum (β) is set to $0.9$, which is used to calculate the moving average of the mean and variance of the feature map during training, accelerating the model convergence and improving the stability and generalization ability of the model.

[0075] The activation function (ReLU) is used to introduce non-linearity and enhance the expression ability of the model, enabling the model to learn complex feature patterns.

[0076] In the above embodiments, by fusing feature maps of different scales, the model can simultaneously capture the global structural information and local detailed features of body parts. For example, it can not only identify the general expression and skin color changes of the entire face but also accurately detect details such as the color of the lips and the state of the eyes, thereby more comprehensively and accurately analyzing the association between the patient's body features and vital signs.

[0077] Then, further, an attention mechanism is introduced into the feature extraction network to assign different attention weights to the features of key regions.

[0078] An attention mechanism is introduced into the feature extraction network, and the formula is as follows:

[0079] A = Softmax(Conv(Fm , W a , b a )); F a = F m ⊙ A;

[0080] Among them, F m is the fused feature map, W a and b a are the weights and biases of the attention convolution layer, A is the attention weight map obtained through the Softmax function, ⊙ represents the element-wise multiplication operation, and F a is the feature map after attention weighting. For example, for the chest area related to breathing and the facial area related to blood circulation, the model can automatically assign higher attention weights, highlight the features of these key areas, suppress the interference of the background and irrelevant areas, thereby improving the effectiveness and pertinence of feature extraction.

[0081] Finally, the image frame feature sequence after feature extraction and attention weighting is input into the long short-term memory network to obtain the vital sign status of the patient.

[0082] The image frame feature sequence after feature extraction and attention weighting is input into the long short-term memory network (LSTM), and the formula is as follows: P = Softmax(W p h T + b p );

[0083] Among them, is the attention-weighted feature map of the t-th image frame, h t is the hidden state of the LSTM at time t, W h and b h are the weights and biases of the LSTM unit, T is the length of the image frame sequence, W p and b p are the weights and biases of the fully connected layer, used to map the final hidden state of the LSTM to the probability distribution P of the vital sign status.

[0084] LSTM unit parameter settings: The number of hidden units (n h ) is set to 128 or 256, and adjusted according to the size and complexity of the dataset. More hidden units can learn more complex time series features, but it will also increase the computational cost and the training difficulty of the model.

[0085] Fully connected layer and Softmax function: The fully connected layer maps the output features of the LSTM to specific vital sign status categories, such as normal, shortness of breath, abnormal heart rate, shock, etc. (assuming a total of C categories). The Softmax function converts the output of the fully connected layer into probability values for each category, such that the sum of the probabilities of all categories is 1, thereby obtaining the final vital sign status classification result P, where P i represents the probability of the i-th category.

[0086] Collect a large amount of on-site video data of patients with different vital sign statuses, and invite professional emergency doctors and medical experts to accurately label the vital sign status of the patients in the videos to form a training data set and a validation data set. The annotation information should include detailed vital sign categories, descriptions of relevant physical characteristics, and corresponding timestamp information to ensure the accuracy and consistency of the annotation. And use the cross-entropy loss function as the training loss function of the model, and select the Adam optimizer for parameter update.

[0087] After each training epoch, use the validation data set to evaluate the model, calculate evaluation metrics such as accuracy, recall, F1 value, etc. to monitor the performance changes of the model. Adjust the hyperparameters of the model (such as the number of convolutional layers, the number of feature maps, the number of LSTM hidden units, etc.) and the training strategy (such as the learning rate, the number of training epochs, etc.) according to the evaluation results, and continuously optimize the model until satisfactory performance metrics are achieved on the validation data set.

[0088] The above embodiments can effectively extract the physical feature information of patients from on-site video surveillance data and accurately judge their vital sign status through improved convolutional neural network (CNN) feature extraction, attention mechanism, and recurrent neural network (RNN) time series analysis. Compared with traditional image recognition methods, this method can more comprehensively and deeply mine the information in the image, considering the multi-scale, time series changes of physical characteristics, and the importance of key features, thereby improving the accuracy and reliability of vital sign status judgment, providing more timely and accurate information support for the first aid of cardiac arrest patients, and helping to improve the success rate and timeliness of first aid.

[0089] Further, in step S104: According to the basic information of the patient, the vital sign status, and the real-time data collected on-site, use machine learning algorithms for analysis to evaluate the severity of the condition.

[0090] In one embodiment, it includes: performing data preprocessing on the basic information of the patient, the vital sign status, and the real-time data collected on-site.

[0091] According to the extracted basic information of the patient, including age, gender, past medical history (such as history of heart disease, hypertension, diabetes, etc., represented by binary coding for presence or absence, for example, 1 means yes, 0 means no), allergy history, etc. Obtain the vital sign status data from on-site monitoring devices, such as heart rate (HR), respiratory rate (RR), blood pressure (systolic blood pressure SBP, diastolic blood pressure DBP), blood oxygen saturation (SpO2), etc., and record and organize them at fixed time intervals (for example, every 5 seconds). And the real-time data collected on-site includes environmental temperature (Temp), humidity (Humidity), implementation of on-site first aid measures (such as whether cardiopulmonary resuscitation was performed, number of defibrillations, etc., represented by integer coding), etc. For the vital sign data, statistical-based methods are used to identify and process outliers. For example, set the reasonable range of heart rate to 40 - 200 beats per minute, and data outside this range is regarded as an outlier, and linear interpolation or correction is performed according to the trend of adjacent data.

[0092] Then, construct combined features for the features after data preprocessing to obtain the processed feature vectors.

[0093] Construct combined features: For example, calculate the shock index (SI = HR / SBP), which combines heart rate and blood pressure information and is of great significance for evaluating the function of the circulatory system and the severity of the condition; Oxygenation index (OI = SpO2 / FiO2, assuming that FiO2 on-site is a known fixed value, such as the air oxygen concentration of 21%) can also be calculated to reflect the respiratory function status.

[0094] Encode categorical features: For categorical variables such as gender (Male: 1, Female: 0), past medical history, etc., use one-hot encoding to convert them into numerical vectors so that machine learning algorithms can process them. For example, if there are 3 common past medical histories (heart disease, hypertension, diabetes), then the medical history information of each patient is encoded as a three-dimensional vector, such as [1, 0, 1] indicating having heart disease and diabetes but no history of hypertension.

[0095] Finally, construct a gradient boosting decision tree model to predict the severity of the condition.

[0096] Gradient Boosting Decision Tree (GBDT) is an ensemble learning algorithm based on the Boosting idea. By gradually constructing multiple weak decision trees and combining them into a strong prediction model, it can effectively handle non-linear relationships and high-dimensional data in fields such as complex medical data mining and condition assessment.

[0097] Construct the GBDT model formula: Among them, is the predicted value of the severity of the condition for the i-th sample by the model, xi is the feature vector of the i-th sample (including basic information, vital sign status, and on-site real-time data features), K is the number of decision trees, and f k (x i ) is the predicted value (residual) of the k-th decision tree for the sample x i .

[0098] Decision tree parameter settings: The maximum depth (max d epth) of each decision tree is set to 5 - 8 layers. A shallower depth can prevent overfitting while maintaining the interpretability and computational efficiency of the model; a deeper depth can learn more complex feature interaction relationships but may lead to overfitting, and it needs to be adjusted according to the complexity of the data and the number of samples.

[0099] The minimum number of samples in a leaf node (min s amples l eaf) is set to 10 - 20. This parameter limits the minimum number of samples in each leaf node, avoids generating overly complex decision tree structures, helps improve the generalization ability and stability of the model, and prevents overfitting to noisy data.

[0100] The learning rate (learning r ate) is set to 0.05 - 0.1. The learning rate controls the contribution degree of each decision tree in the model iteration process. A smaller learning rate can make the model converge more robustly during training, but it requires a larger number of decision trees and training time; a larger learning rate may cause the model to skip the optimal solution during training but can speed up the training. In practical applications, the optimal learning rate value usually needs to be determined through multiple trials.

[0101] For new patient data, first process it according to the above data preprocessing and feature construction process, and then input the processed feature vector into the trained GBDT model to obtain the predicted value of the disease severity. The disease is divided into different grades (such as mild, moderate, severe) according to the size of the predicted value, providing intuitive and quantitative disease assessment results for first aid personnel so that they can quickly and accurately formulate corresponding treatment plans and resource allocation strategies, improving the treatment efficiency and success rate of patients with cardiac arrest.

[0102] The above-described embodiment for evaluating the condition provided by the present invention, compared with the traditional condition evaluation method, takes into account more complex relationships between factors and data characteristics. Through the automatic learning and optimization of the machine learning model, it can provide more accurate and objective condition evaluation results, provide strong support for first-aid decision-making, and help improve the prognosis of patients with cardiac arrest. In practical applications, it is also necessary to continuously collect more data to further optimize and improve the model to meet the needs of different clinical scenarios and patient groups.

[0103] Further, the process proceeds to step S105: According to the visual large screen of the resource allocation in the hospital command center, through the Internet of Things technology, the first-aid resources in the hospital are positioned and monitored in real time.

[0104] For first-aid equipment (such as defibrillators, cardiopulmonary resuscitation machines, first-aid medicine boxes, etc.), select appropriate active RFID (Radio Frequency Identification) tags as sensors. These tags have unique identification codes and can send the location information and basic status information of the equipment (such as battery power, whether it is in a working state, etc.) in real time.

[0105] For first-aid personnel, equip intelligent bracelets or name tags with positioning functions, which are built-in with GPS (Global Positioning System) chips and Bluetooth modules to obtain the location information of personnel in real time. At the same time, the bracelet or name tag can also integrate some simple sensors, such as heart rate monitors, to obtain the physical status information of first-aid personnel (this part of the information can be used as auxiliary data to determine whether the personnel are in an emergency or fatigued state, so as to reasonably allocate human resources).

[0106] Deploy RFID readers and Bluetooth gateways in key areas of the hospital (such as wards, corridors, stairwells, elevator entrances, emergency rooms, etc.). These devices are responsible for receiving signals from sensors and transmitting the data to the background server. After receiving the data, the server first parses and preprocesses the data, including removing invalid data, correcting data format errors, performing coordinate conversion on location data (unifying location data in different coordinate systems into the standard coordinate system of the hospital map), etc., for subsequent data storage and analysis and processing.

[0107] On the visual large screen of the resource allocation, draw a two-dimensional or three-dimensional map of the hospital. The map should be drawn based on the actual building layout and structure of the hospital, including the exact location and shape information of key areas such as each floor, room, corridor, stairwell, elevator, etc.

[0108] Design unique icons for different types of first aid resources so that their locations and status can be visually displayed on the map. For example, a defibrillator can be represented by a square icon with a lightning symbol, a CPR machine by a heart-shaped icon, a first aid medicine box by a medicine box icon, and first aid personnel by circular icons of different colors (color-coded according to the position or task status of the personnel, such as doctors in blue, nurses in green, stretcher bearers in orange, etc.).

[0109] Through the implementation steps of the above embodiments, by using Internet of Things technology in combination with database management and visualization display technology, it is possible to achieve real-time positioning and status monitoring of first aid resources in the hospital, and visually and dynamically display them to the command center personnel through a visualization large screen, providing strong technical support for the reasonable allocation and efficient utilization of first aid resources, helping to improve the hospital's ability and efficiency in dealing with emergencies such as cardiac arrest, shortening the first aid response time, and enhancing the success rate of patient treatment.

[0110] Further, the process proceeds to step S106: According to the location and condition of the patient, as well as the location and status of the first aid resources, an optimal first aid resource allocation plan is automatically calculated through artificial intelligence algorithms.

[0111] In one embodiment, automatically calculating the optimal first aid resource allocation plan includes: integrating patient data and first aid resource data to generate a unified data format and coordinate system; standardizing different types of data to obtain the same dimension and numerical range; constructing a hospital environment model based on graph theory, using path planning algorithms, combining heuristic information and actual path costs to obtain the optimal path from the location of the first aid resource to the location of the patient; using the linear weighted method to transform multiple objective functions into a single objective function, establishing a multi-objective optimization model, and using the genetic algorithm to decode and solve the multi-objective optimization model to obtain the optimal first aid resource allocation plan.

[0112] The following is a detailed description through an embodiment of obtaining a first aid resource allocation plan.

[0113] First, model the hospital environment. Create a hospital environment model based on graph theory, abstracting each area of the hospital (wards, corridors, stairwells, elevators, emergency rooms, etc.) as nodes of the graph, and abstracting the passages connecting these areas (doors, corridors, stairs, elevator passages, etc.) as edges of the graph. Assign a weight to each edge, representing the time or cost required to pass through this passage, and this weight can be comprehensively calculated based on factors such as the length, width, pedestrian flow, and passage speed limit of the passage.

[0114] For example, for a corridor with a length of L meters, an average pedestrian flow of P people per minute, and a passage speed limit of V meters per minute, its weight W can be calculated by the following formula: where P max is the maximum carrying capacity of people flow in this passageway. In this way, it can more accurately reflect the passing efficiency and cost of different passageways under actual conditions, providing more reasonable basic data for subsequent path planning.

[0115] Then, select and improve the path planning algorithm. An improved algorithm based on the A*(A-Star) algorithm or its variants is used as the basic path planning algorithm. This algorithm takes into account special situations in the hospital environment, such as the waiting time of elevators, the passing difficulty of stairs (for equipment such as stretchers), the temporary blockade or restricted passage of certain areas, etc. For example, when calculating the heuristic function, not only the straight-line distance is considered, but also the estimation of elevator waiting time (predicted based on historical elevator usage data and the current operating status of elevators) and the stair passing difficulty coefficient (determined according to factors such as the slope of the stairs and the height of the steps) are added to make the path planning more in line with the actual situation.

[0116] The formula is as follows: f(n) = g(n) + h(n), where f(n) is the comprehensive evaluation function of node n, g(n) is the actual path cost from the starting node to node n, and h(n) is the heuristic estimated cost from node n to the target node (including factors such as the improved straight-line distance, elevator waiting time estimation, and stair passing difficulty coefficient). By continuously expanding the node with the minimum f(n) value until the target node is found, the best path from the first-aid resources to the patient can be obtained.

[0117] Secondly, construct and optimize the resource allocation model. Establish a resource allocation model based on multi-objective optimization, considering multiple objectives of first-aid resource allocation, such as minimizing the time for resources to reach the patient, maximizing the availability and applicability of resources, minimizing the cost of resource allocation, etc., and establish a multi-objective optimization model.

[0118] The linear weighted method can be used to transform multiple objective functions into a single objective function. The formula is as follows: Z = ω1T + ω2(1 - A) + ω3C.

[0119] Among them, Z is the value of the comprehensive objective function, T is the total time for resources to reach the patient (the sum of the times for each resource to reach the patient calculated by the above path planning algorithm), A is the overall availability of resources (the ratio of the number of available resources to the total number of resources), C is the total cost of resource allocation (including equipment handling costs, personnel movement costs, etc., which can be quantitatively estimated according to the actual situation), ω1, ω2, ω3 are the weight coefficients of each objective, which are set according to actual needs and priorities, and satisfy ω1 + ω2 + ω3 = 1. By adjusting these weight coefficients, trade-offs and optimizations can be made between different objectives to meet the first-aid resource allocation requirements in different scenarios.

[0120] Set a series of constraint conditions according to the actual situation. For example: Resource quantity constraint: The allocated quantity of each first aid resource cannot exceed its actual available quantity, that is, x i ≤X i where x i is the allocated quantity of the i-th resource, and X i is the actual available quantity of this resource. Personnel qualification constraint: For certain specific first aid tasks, personnel with corresponding qualifications and skills must be equipped. For example, for the first aid of patients with cardiac arrest, at least one doctor with advanced cardiovascular life support (ACLS) qualification must be present, that is where y ij indicates whether the j-th person participates in the i-th first aid task and this person has ACLS qualification. Time constraint: The time for the resource to reach the patient must be within a certain time limit. For example, T≤T max where T max is the maximum allowable waiting time set according to the severity of the patient's condition. For critically ill patients, this time limit is shorter to ensure timely treatment.

[0121] After the plan is determined, plan evaluation and dynamic adjustment should be carried out. First, establish a comprehensive evaluation index system for the first aid resource allocation plan, including but not limited to the following indicators:

[0122] Resource arrival time: Calculate the total time required for all key first aid resources (such as defibrillators, first aid personnel, etc.) to reach the patient's location from the issuance of the allocation instruction, and compare it with the preset time target to evaluate the performance of the plan in terms of timeliness.

[0123] Resource utilization rate: Statistically analyze the actual usage of the allocated first aid resources during the entire first aid process, calculate the utilization rate of the resources (the ratio of the actual usage time or frequency to the total available time or frequency), and evaluate whether the resource configuration is reasonable to avoid resource idleness and waste.

[0124] Patient treatment effect: Indirectly evaluate the impact of the resource allocation plan on the patient's treatment effect through the subsequent development of the patient's condition (such as whether successful resuscitation, length of hospital stay, mortality rate, etc.). This is the ultimate key indicator to measure the quality of the plan.

[0125] Before implementing the resource allocation plan, use historical data and the hospital environment model for simulation evaluation. By simulating different first aid scenarios and resource allocation plans multiple times, collect and analyze the data of the above evaluation indicators, pre-evaluate and optimize the plan, and identify potential problems and improvement spaces. For example, if it is found that the actual resource arrival time is longer than the simulated prediction time, analyze the reason that some channels may be temporarily congested or equipment failures and other factors are not fully considered in the model, and then make corresponding adjustments and improvements to the model to improve the accuracy and reliability of the model.

[0126] In some embodiment scenarios, a dynamic adjustment mechanism is further included. Since the hospital environment and the patient's condition may change at any time, a dynamic adjustment mechanism is established so that the resource allocation plan can be optimized and adjusted in a timely manner according to the real-time situation. For example, during the first aid process, if it is found that the patient's condition suddenly deteriorates and more first aid resources or different types of professionals are needed, the system can obtain the latest patient information and resource status in real time, re-run the resource allocation model, generate a new allocation plan, and notify the relevant personnel in time for adjustment.

[0127] Meanwhile, using real-time positioning technology and communication systems, track and monitor the allocated resources, and timely grasp the moving speed, position changes, and possible abnormal situations (such as personnel getting lost, equipment failures, etc.) of the resources, so as to take corresponding measures for remedy and adjustment at the first time, ensure that the resources can reach the patient accurately and quickly, and provide strong guarantee for the treatment of the patient.

[0128] Through the above implementation steps, combined with advanced artificial intelligence algorithms and multi-source data fusion technologies, it is possible to automatically calculate the optimal first aid resource allocation plan according to the patient's location and condition, as well as the location and status of the first aid resources, and continuously optimize the implementation effect of the plan through continuous evaluation and dynamic adjustment mechanisms, improve the first aid efficiency and quality of the hospital in dealing with emergencies such as cardiac arrest, and maximize the protection of the patient's life safety.

[0129] Further, the process proceeds to step S107: By collecting and sorting out the detailed data of each cardiac resuscitation event, use machine learning algorithms to establish a first aid effect evaluation model, analyze the relationship between various factors and the patient after recovery, so as to obtain the key factors for the success rate of treatment and potential problem points.

[0130] In one embodiment, the use of machine learning algorithms to establish a first aid effect evaluation model, analyze the relationship between various factors and the patient after recovery, so as to obtain the key factors for the success rate of treatment and potential problem points, includes: extracting and constructing specific features from the detailed data of each cardiac resuscitation event to obtain a cardiac resuscitation event feature dataset; the specific features are features that reflect the relationship between each feature and the first aid effect, including time-related features, combined features, and trend features; using the cardiac resuscitation event feature dataset, based on the ensemble learning algorithm of decision trees, construct a first aid effect evaluation model; wherein the first aid effect evaluation model includes a feature importance evaluation function; using the feature importance evaluation function of the first aid effect evaluation model, analyze the contribution weights of each feature to the first aid effect; by analyzing the decision tree structure and feature splitting situation in the first aid effect evaluation model, explore the potential relationships and laws between each feature and the first aid effect, so as to obtain the key factors for the success rate of treatment and potential problem points.

[0131] Obtain detailed data for each cardiac resuscitation event, including basic patient information: age, gender, height, weight, past medical history (such as cardiovascular diseases, respiratory diseases, diabetes, etc., using structured disease coding and severity grading), allergy history, lifestyle habits (smoking, drinking, exercise frequency, etc., quantitatively represented).

[0132] Situation at the onset site: onset location (indoor / outdoor, specific places such as home, public places, etc.), environmental factors (temperature, humidity, altitude, etc., precisely recorded), activity status at the time of onset (rest, exercise, work, etc., classified records), bystander first aid measures and start time (whether cardiopulmonary resuscitation, defibrillation and other operations were performed, as well as the duration and quality indicators of the operations, such as compression depth, frequency, etc., obtained through on-site records or device data).

[0133] Vital sign data: continuous electrocardiogram monitoring data (heart rate, rhythm, electrocardiogram waveform, etc., collected at high frequency), blood pressure (systolic pressure, diastolic pressure, continuously monitored), blood oxygen saturation (recorded in real time), respiratory rate and depth (obtainable through respiratory monitoring devices), body temperature (measured at regular intervals) from the onset to a period of time after resuscitation.

[0134] Treatment process data: composition of the first aid team (qualifications and experience levels of doctors, nurses, first aid personnel, quantitatively scored), arrival time at the scene, start time of cardiopulmonary resuscitation, number of defibrillations and energy settings, tracheal intubation time and operation conditions, types of drugs used, dosage and administration time (precisely recorded), implementation details of advanced life support measures (such as mechanical ventilation parameters, use of extracorporeal membrane oxygenation (ECMO), etc.), maintenance of vital signs during transportation and transportation time.

[0135] Post-resuscitation indicators: time to return of spontaneous circulation (ROSC), time to restoration of sinus rhythm, time to recovery of consciousness, neurological function assessment (using the Glasgow Coma Scale (GCS) or other professional assessment scales, regularly evaluated), organ function indicators (such as liver and kidney functions, myocardial enzyme spectrum, etc., laboratory test data), length of hospital stay, physical condition at discharge (cured, improved, disabled, dead, etc., clearly classified).

[0136] By collecting and collating the detailed data of each cardiac resuscitation event, use machine learning algorithms to establish an emergency treatment effect evaluation model. The training process is as follows:

[0137] First, considering the complexity of the first aid effect evaluation problem and the characteristics of the data (including numerical, categorical data, and the non-linear relationships between various factors), the machine learning algorithm can choose the Random Forest algorithm as the basic model. Random Forest is an ensemble learning algorithm based on decision trees. By constructing multiple decision trees and voting or averaging their results, it can effectively handle high-dimensional data, the interaction between features, and avoid overfitting problems. It has high prediction accuracy and stability and is suitable for discovering the complex relationships and potential laws between various factors and the first aid effect.

[0138] The data of cardiac resuscitation events after preprocessing and feature engineering is divided into a training set (70%), a validation set (15%), and a test set (15%). The training set is used to train the Random Forest model. The goal is to minimize the difference between the predicted first aid effect (such as the physical condition classification at discharge) and the actual result. The cross-entropy loss function is used as the objective function for training (for classification problems), and the formula is as follows:

[0139] where N is the number of training samples, C is the number of categories of the first aid effect (such as cured, improved, disabled, dead, etc., a total of C categories), and y ij is the true class label of the i-th sample (in one-hot encoding form, that is, if it belongs to the j-th class, then y ij = 1, otherwise y ij = 0), is the probability value that the model predicts the i-th sample belongs to the j-th class.

[0140] During the training process, the Random Forest model constructs multiple decision trees by sampling the training set with replacement (Bootstrap sampling). When constructing each decision tree, a part of the features (feature subset) is randomly selected from all the features for node splitting to increase the diversity and generalization ability of the model. For example, for a data set with M features, when constructing each decision tree, randomly select features (which can be adjusted according to the actual situation) as candidate features for splitting decisions. By continuously recursively splitting nodes until the stopping conditions are met (such as the number of samples in the node is less than a certain threshold, the depth of the tree reaches the preset maximum value, etc.), a complete decision tree is constructed.

[0141] The parameters of the Random Forest model also include the number of decision trees (n trees ), which is generally set between 100 - 500. More decision trees can improve the stability and generalization ability of the model, but it will also increase the computational cost and training time. Through multiple experiments and tuning on the validation set, the number of decision trees that performs best on the validation set is selected as the final model parameter.

[0142] During the training process, the performance of the model is regularly evaluated using a validation set, and evaluation metrics such as accuracy, recall, F1-score, and the area under the receiver operating characteristic curve (ROC curve) (AUC) are calculated to monitor the training effect of the model and prevent overfitting.

[0143] During the process of model analysis, it is necessary to analyze the feature importance. Using the feature importance evaluation function of the random forest model, the contribution degree of each factor (feature) to the first aid effect is analyzed. The random forest determines the importance score of a feature by calculating the average information gain (or other impurity reduction metrics, such as the Gini index) of each feature for node splitting in all decision trees. The greater the information gain of a feature, the more crucial the role of this feature in the classification decision, and the greater the impact on the first aid effect. For example, for a dataset with n features, the importance score Ik of the k-th feature k can be calculated by the following formula:

[0144] where, is the information gain of the k-th feature in the t-th decision tree, and n trees is the number of decision trees. By sorting the feature importance scores, key factors that have a greater impact on the first aid effect can be identified, such as the patient's age, past medical history, start time of cardiopulmonary resuscitation, number of defibrillations, etc. These key factors will provide important references for subsequent in-depth analysis and clinical decisions.

[0145] By analyzing the decision tree structure and feature splitting of the random forest model, the potential relationships and patterns between various factors and the first aid effect are further explored. For example, observe which feature combinations are frequently used for node splitting in different decision trees, and the corresponding relationships between these feature combinations and different first aid effect categories. If it is found that for cured patients, the feature combination of "early start time of cardiopulmonary resuscitation and few defibrillations" appears in multiple decision trees, this may imply that timely cardiopulmonary resuscitation and effective defibrillation operations are of great significance for improving the success rate of treatment, while too many defibrillations may be related to a poor prognosis, which provides clues for discovering potential problem points.

[0146] At the same time, combined with actual clinical knowledge and experience, the relationships and patterns mined by the model are explained and verified to judge their rationality and reliability. For some results that are inconsistent with traditional medical cognition, it is necessary to further analyze the data and the model in depth to find possible reasons, such as data quality problems, unreasonable feature construction, model overfitting, etc., to ensure that the discovered relationships and problem points have practical clinical value and guiding significance.

[0147] Over time and with the accumulation of new data on cardiac resuscitation events, the model is regularly optimized and updated to adapt to changing clinical situations and improve its performance, ensuring that the model can always accurately evaluate the emergency treatment effect, discover new key factors and potential problem points, and provide continuous technical support and guarantee for improving the success rate of treating patients with cardiac arrest.

[0148] By analyzing and modeling the detailed data of cardiac resuscitation events using the above-mentioned random forest machine learning algorithm, an emergency treatment effect evaluation model can be effectively established, the relationship between various factors and the patient's prognosis can be explored, the key factors for the success rate of treatment and potential problem points can be identified, which helps to continuously improve the treatment level and quality of patients with cardiac arrest, reduce the mortality and disability rate, and promote the development and progress of emergency medicine.

[0149] Further, the process proceeds to step S108: generating a personalized advanced life support process for cardiac resuscitation based on an artificial intelligence algorithm and visualizing it to provide emergency personnel with guidelines on the operation steps and time nodes for emergency treatment.

[0150] In one embodiment, generating a personalized advanced life support process for cardiac resuscitation based on an artificial intelligence algorithm and visualizing it to provide emergency personnel with guidelines on the operation steps and time nodes for emergency treatment includes: constructing a medical knowledge and case database; wherein the medical knowledge includes: authoritative medical literature, guidelines, and expert experience; the case database includes treatment cases of patients with cardiac arrest, and the treatment cases include successful cases and failed cases.

[0151] Constructing a hybrid neural network model based on deep learning; dividing the preprocessed and feature-extracted patient data, as well as the corresponding treatment process and outcome data in the case database, into a training set, a validation set, and a test set; training and optimizing the hybrid neural network model through the training set, the validation set, and the test set to obtain a trained hybrid neural network model.

[0152] According to the patient data collected in real time on-site, using the trained hybrid neural network model, a personalized advanced life support process for cardiac resuscitation for this patient is obtained; wherein the personalized advanced life support process for cardiac resuscitation includes the sequence of each operation step, the specific content, and the recommended execution time node for each operation step.

[0153] During the construction of the medical knowledge and case database, by collecting a large number of treatment cases of patients with cardiac arrest, including successful cases and failed cases, the patient characteristics, treatment measures, operation processes, and final treatment outcomes (such as whether resuscitation is successful, survival time, neurological function recovery, etc.) in each case are detailedly recorded.

[0154] Organize and input authoritative medical literature, guidelines, and expert experience, covering all aspects of cardiac resuscitation, such as the treatment methods for different types of arrhythmias, the indications and dosage ranges of drug use, the standard operation techniques of cardiopulmonary resuscitation, etc., to form a comprehensive medical knowledge database for subsequent algorithm training and decision support.

[0155] Adopt a neural network model based on deep learning, such as a hybrid architecture combining convolutional neural network (CNN) and recurrent neural network (RNN), to fully utilize the advantages of CNN in feature extraction and the ability of RNN in processing time series data. Among them, the CNN part is used to automatically extract spatial features in patient data, such as learning key feature patterns from electrocardiogram waveforms and the image representation of vital sign data (data over a period of time can be plotted in image form); the RNN part (such as long short-term memory network LSTM or gated recurrent unit GRU) is used to model the patient's time series data, capture the dynamic change trends of vital signs and the time-dependent relationship between different treatment measures and patient responses, so as to dynamically adjust the treatment strategy and operation process according to the patient's real-time status.

[0156] Divide the preprocessed and feature-extracted patient data, as well as the corresponding treatment process and outcome data, into a training set, a validation set, and a test set, and the ratio can be set as 70%, 15%, 15%.

[0157] Use the training set to train the neural network model. The goal is to minimize the difference between the predicted treatment process and the actual treatment process. The cross-entropy loss function is used as the objective function for training, and the formula is as follows:

[0158] Among them, N is the number of training samples, M is the number of operation steps or decision nodes in the treatment process, and y ij is the true label of the i-th sample at the j-th operation step or decision node (for example, one-hot encoding is used to represent whether to perform a certain operation or select a certain treatment plan), is the probability value predicted by the model.

[0159] During the training process, use stochastic gradient descent (SGD) or its variant optimization algorithms (such as Adam, Adagrad, etc.) to update the model's parameters to adjust the weights and biases of the model, so that the loss function gradually decreases. For example, the parameter settings of the Adam optimization algorithm are as follows: the initial value of the learning rate (Ir) is set to 0.001. As the training progresses, a learning rate decay strategy can be adopted. For example, every 10 training epochs, the learning rate is multiplied by 0.9 to quickly converge in the initial stage of training and finely adjust the model parameters in the later stage, avoid skipping the optimal solution area, and improve the convergence accuracy and stability of the model.

[0160] The β1 and β2 parameters are set to 0.9 and 0.999 respectively. These two parameters are used to calculate the first-order and second-order moment estimates of the gradient, control the step size and direction when the Adam optimizer updates the parameters, and help adaptively adjust the learning rate at different training stages, accelerating the convergence speed of the model.

[0161] After each training epoch, the performance of the model is evaluated using the validation set, and metrics such as accuracy, recall, and F1-score are calculated to monitor whether the model is overfitting or underfitting.

[0162] In the generation of the personalized process and determination of time nodes, when a new cardiac arrest patient appears, the patient data collected in real time at the scene (including basic information, vital sign data, on-site situation of the onset, etc.) is preprocessed and feature-extracted, and then input into the trained neural network model. Then, based on the input data, the model combines the knowledge and patterns learned during the training process to predict the personalized advanced life support process for cardiac resuscitation for this patient, including the order of each operation step, specific content (such as the depth and frequency of chest compressions, energy settings for defibrillation, types and dosages of drugs, etc.), and the recommended execution time nodes for each operation step.

[0163] Considering various factors and uncertainties in the actual first-aid environment, the operation steps and time nodes predicted by the model are further optimized and adjusted. For example, in combination with the number and skill levels of on-site first-aid personnel, tasks and operation sequences are reasonably allocated to avoid situations of personnel idleness or operation conflicts; according to the actual distribution and availability of first-aid equipment in the hospital, the time nodes and execution sequences of equipment-related operations are adjusted to ensure that the equipment can be in place in time and used correctly; at the same time, considering the real-time changes in the patient's vital signs, dynamic time windows and trigger conditions are set. When significant changes occur in the patient's vital signs (such as a sudden drop in heart rate, ventricular fibrillation, etc.), the treatment process can be adjusted in time, advancing or delaying the execution time of certain operation steps to maximize the success rate of resuscitation.

[0164] The generated personalized advanced life support process for cardiac resuscitation is visually displayed on a tablet computer, smart glasses, or other wearable devices at the first-aid scene, providing intuitive and clear operation guidelines for first-aid personnel. For example, for chest compression operations, the correct compression posture, depth, and frequency are demonstrated through animations, and the real-time compression count and remaining time are displayed beside; for defibrillation operations, the operation process of the defibrillator, the placement position of the electrode pads, and the recommended energy settings are shown, and a reminder signal is automatically emitted when the countdown ends to ensure that first-aid personnel can accurately and timely execute each operation step, improving the quality and efficiency of cardiac resuscitation.

[0165] By using the artificial intelligence algorithm in combination with multi-source data and the continuous learning mechanism as described above, it is possible to generate a personalized advanced life support process for cardiac arrest patients and display it visually, and provide accurate and real-time operation steps and time node guidelines for first aid personnel, which helps to improve the success rate and quality of cardiac resuscitation and provide stronger protection for the patient's life safety.

[0166] Further, in step S109: According to the evaluation result of the first aid effect evaluation model, use the artificial intelligence algorithm to automatically generate process optimization suggestions to achieve dynamic optimization and adaptive adjustment of the advanced life support process for cardiac resuscitation.

[0167] In one embodiment, according to the evaluation result of the first aid effect evaluation model, use the artificial intelligence algorithm to automatically generate cardiac resuscitation process optimization suggestions to achieve dynamic optimization and adaptive adjustment of the advanced life support process for cardiac resuscitation, including: obtaining the key data and other associated data of the first aid effect evaluation according to the first aid effect evaluation model, and the key data of the first aid effect evaluation includes patient basic information, on-site situation of the onset, vital sign data, treatment process data, and first aid effect indicators.

[0168] Construct a process optimization model, including: defining each link and related parameters in the advanced life support process for cardiac resuscitation as the state space of the reinforcement learning model; determining the action space of the reinforcement learning model to obtain the taken process optimization measures; using the reward function related to the first aid effect to evaluate the positive or negative impact of the actions taken by the reinforcement learning model on the first aid process, so that the reinforcement learning model can learn the optimal process optimization strategy.

[0169] Define each link and related parameters in the advanced life support process for cardiac resuscitation as the state space of the reinforcement learning model. For example, the state can include the current compression depth, frequency, ventilation parameters, preparation status of the defibrillation device, drug use situation, vital sign status of the patient (discretely represented, such as the heart rate is divided into several intervals such as normal, high, and low), etc.

[0170] Determine the action space of the model, that is, the process optimization measures that can be taken. These actions include adjusting the increment or decrement of the compression depth and frequency (such as increasing the compression depth by 0.5 cm, reducing the compression frequency by 5 times per minute, etc.), changing the ventilation volume and respiratory frequency, selecting different drugs or adjusting the drug dose, deciding whether to perform defibrillation and adjusting the energy setting of defibrillation, adjusting the personnel configuration and division of labor of the first aid team, etc.

[0171] Design a reward function related to the first aid effect to evaluate the positive or negative impact of the actions taken by the model on the first aid process, so as to guide the model to learn the optimal process optimization strategy. The reward function comprehensively considers multiple factors. For example, if the actions taken make the patient's vital signs (such as heart rate, blood pressure, blood oxygen saturation, etc.) improve towards the normal range, a positive reward is given, and the reward value is proportional to the degree of improvement of the vital signs. For example, when the blood oxygen saturation increases by a certain percentage (such as 5%), a reward R1 = 10 is given; when the heart rate returns to the normal range, a reward R2 = 20 is given, etc.

[0172] For shortening the time of key treatment links (such as the start time of cardiopulmonary resuscitation, defibrillation time, etc.), an additional reward is given. For example, for every 1 minute earlier of effective cardiopulmonary resuscitation, a reward R3 = 5 is given; for a 30-second earlier defibrillation time, a reward R4 = 8 is given, etc., to encourage the model to optimize the process and improve the timeliness of first aid.

[0173] If the actions taken lead to the deterioration of the patient's vital signs or the occurrence of complications (such as rib fractures, barotrauma, etc.), a negative reward is given. For example, when a rib fracture occurs (judged by the pressing pressure, acceleration data, and clinical symptoms), a penalty R5 = -15 is given to prevent the model from taking optimization measures that may cause harm to the patient.

[0174] The comprehensive reward function R can be expressed as:

[0175] where R i are the above various specific reward or penalty items, and w i is the corresponding weight coefficient, which is set and adjusted according to the importance of each factor to the first aid effect (for example, the weight value is determined through the analysis of expert experience and historical data to ensure that the weight of the improvement of vital signs is relatively high, and the penalty weight for complications is also large enough to balance the influence of different factors on the reward function).

[0176] During cardiopulmonary resuscitation, obtain the patient's vital sign data and first aid operation data to obtain the patient's current status information. Using the patient's current status, according to the trained process optimization model, predict the best optimization actions that should be taken in the current state, and generate corresponding process optimization suggestions. The first aid personnel timely adjust the cardiopulmonary resuscitation operation process and related parameters according to the process optimization suggestions generated by the model.

[0177] Through the embodiments provided above, by using advanced artificial intelligence algorithms and data-driven methods, it is possible to automatically generate targeted optimization suggestions for the advanced life support process of cardiopulmonary resuscitation according to the results of the first aid effect evaluation model, and achieve dynamic optimization and adaptive adjustment of the process through real-time monitoring and feedback, providing an innovative and efficient technical means for improving the first aid effect of patients with cardiac arrest, contributing to the development and progress of first aid medicine, reducing the mortality and disability rate of patients, and having important value and significance in actual clinical applications.

[0178] Further, in step S110: The on-site operation data collected is compared and analyzed with the operation parameters of the advanced life support process of cardiopulmonary resuscitation through an artificial intelligence algorithm to adjust the cardiopulmonary resuscitation operation process in real time.

[0179] In one embodiment, the operation parameters of the advanced life support process of cardiopulmonary resuscitation are chest compressions, airway management, and defibrillation. According to the latest international cardiopulmonary resuscitation guidelines and professional medical literature, combined with the actual experience of the hospital and expert opinions, the detailed standard operation parameters of the advanced life support process of cardiopulmonary resuscitation are determined.

[0180] The on-site operation data collected is compared and analyzed with the standard operation parameters of the advanced life support process of cardiopulmonary resuscitation through an artificial intelligence algorithm to adjust the cardiopulmonary resuscitation operation process in real time, including:

[0181] Using a hybrid architecture that combines long short-term memory networks with convolutional neural networks to construct a comparative analysis model. Specifically, a comparative analysis model can be constructed using a hybrid architecture that combines long short-term memory networks (LSTM) with convolutional neural networks (CNN): Among them, the CNN part: is used to automatically extract spatial features in the operation data. For example, for the two-dimensional image representation of the pressing pressure data (arranging the pressure values within a period of time into an image matrix in a time series), the CNN can pass through the convolutional layer (setting multiple convolutional kernels, such as 32 convolutional kernels with a size of 3×3, a stride of 1, and a padding method of "same" to extract local features of different scales), the pooling layer (using max pooling, with a pooling window of 2×2 and a stride of 2 to reduce the data dimension and highlight the main features), and the fully connected layer (mapping the features to a low-dimensional vector space, and setting the number of neurons to 64 or 128, etc. according to the feature complexity), to extract the key feature patterns of the pressing operation, such as the uniformity of the pressure distribution, the trend of pressure change, etc. These features are of great significance for judging the quality and standardization of the pressing operation. The LSTM part: is used to process the time series characteristics of the operation data. Taking the pressing depth and frequency data as an example, the LSTM unit (setting the number of hidden units to 128 and the input dimension to the feature dimension after CNN processing plus the time stamp information dimension of the original data) can learn the dynamic change rules of the pressing operation in time, such as the gradual change trend of the pressing depth, the stability of the frequency, and the relationship with the changes in the patient's vital signs (if real-time vital sign data is input into the model).

[0182] By performing sequence modeling on the operation data over a period of time, the LSTM can predict the trend of future operations and determine whether adjustments are needed based on the comparison with the standard parameters.

[0183] Then, using the historical cardiopulmonary resuscitation operation data and the corresponding standard operation parameters as the training set, the comparative analysis model is trained. Using a large amount of historical cardiopulmonary resuscitation operation data (including operation cases that meet and do not meet the standards, classified by expert annotation) and the corresponding standard operation parameters as the training set, the constructed hybrid model is trained. Using the mean squared error (MSE) loss function as the training objective, the formula is as follows:

[0184] Among them, N is the number of training samples, y i is the standard operation parameter value of the i-th sample (such as the standard pressing depth, frequency, etc., represented in vector form), is the operation parameter value predicted by the model. By continuously adjusting the weights and biases of the model through the backpropagation algorithm, the error between the model prediction value and the standard value is minimized, so as to learn the complex mapping relationship and feature patterns between the operation data and the standard parameters.

[0185] Then, using the comparative analysis model, predict the deviation between the on-site operation data and the standard operation parameters.

[0186] At the scene of cardiopulmonary resuscitation, input the operation data collected in real time and preprocessed into the trained comparative analysis model at fixed time windows (such as one window every 5 seconds). Based on the input data, the model combines the knowledge and patterns it has learned to predict the deviation between the current operation and the standard operation parameters, and outputs the predicted value of each operation index and the degree of deviation from the standard value (expressed as a percentage or a specific value). For example, it is predicted that the current compression depth is 4.5 cm, and compared with the standard depth of 5 - 6 cm, the degree of deviation is -10% (indicating shallower); the ventilation frequency is predicted to be 15 times per minute, and compared with the adult standard of 10 - 12 times per minute, the degree of deviation is +25% (indicating faster), etc.

[0187] According to the deviation situation predicted by the model, adjust the on-site cardiopulmonary resuscitation operation process in real time. For example, if the compression depth is too shallow, the model recommends increasing the compression force, but ensure that the pressure does not exceed the safety limit (500N). By calculating the difference between the current pressure and the pressure required to reach the standard depth, gradually increase the compression force according to a certain proportion (such as 10% - 20%), and at the same time monitor the change of the compression depth to avoid complications such as rib fractures caused by excessive force.

[0188] For the situation where the compression frequency is too fast or too slow, the model emits a prompt sound or displays adjustment information on a visualization device (such as a field tablet or smart glasses display), asking the first aid personnel to slow down or speed up the compression speed accordingly. The adjustment step size is set to 5 - 10 times per minute according to the actual situation to gradually approach the standard frequency range.

[0189] Through the above embodiments, using advanced sensor technology to collect on-site operation data, combined with deep learning models for comparative analysis, and through visual and voice feedback to guide first aid personnel to adjust the operation process in real time, while continuously optimizing the model, can significantly improve the quality and effect of cardiopulmonary resuscitation operations, provide more powerful technical support and guarantee for the treatment of cardiac arrest patients, reduce mortality and disability rates, and promote the development and progress of first aid medicine.

[0190] In summary, through the method embodiments of constructing a visual cardiopulmonary resuscitation process provided by the present invention, through the close cooperation and synergy of each step, a complete closed loop from data collection to model construction, optimization suggestion generation and implementation, and model continuous learning and update is achieved, aiming to continuously improve the quality and effect of the advanced life support process of cardiopulmonary resuscitation, provide strong technical guarantee for the treatment of cardiac arrest patients, and have significant clinical application value and positive technical promotion effect.

[0191] Figure 2The exemplary structural block diagram of the system 200 for constructing a visual cardiac resuscitation process according to an embodiment of the present invention is shown.

[0192] As Figure 2 shown, in the system 200, a system for constructing a visual cardiac resuscitation process provided by the present invention includes: a data acquisition module 201: used to construct a big data platform by collecting multi-source data. An arrhythmia automatic recognition model construction module 202: used to construct an arrhythmia automatic recognition model based on deep learning, and perform real-time analysis on electrocardiogram data to identify the characteristic waveforms of cardiac arrest. A vital sign analysis module 203: used to analyze the physical characteristics of the patient according to the on-site video surveillance data by using image recognition technology to judge the vital sign status of the patient. A condition assessment module 204: used to analyze according to the basic information of the patient, the vital sign status and the real-time data collected on-site by using machine learning algorithms to evaluate the severity of the condition.

[0193] An emergency resource management module 205: used to perform real-time positioning and status monitoring of the emergency resources in the hospital through the Internet of Things technology according to the resource allocation visual big screen of the hospital command center. An emergency resource scheduling module 206: used to automatically calculate the best emergency resource allocation plan according to the location and condition of the patient, as well as the location and status of the emergency resources through artificial intelligence algorithms.

[0194] An emergency treatment effect evaluation model construction module 207: used to establish an emergency treatment effect evaluation model by collecting and sorting out the detailed data of each cardiac resuscitation event, and analyze the relationship between various factors and the patient after cure to obtain the key factors and potential problem points of the rescue success rate. A cardiac resuscitation process generation module 208: used to generate a personalized advanced life support process for cardiac resuscitation based on artificial intelligence algorithms and perform visual display, providing the operation steps and time node guidance for first aid personnel. A cardiac resuscitation process optimization module 209: used to automatically generate cardiac resuscitation process optimization suggestions according to the evaluation results of the emergency treatment effect evaluation model by using artificial intelligence algorithms to realize the dynamic optimization and adaptive adjustment of the advanced life support process for cardiac resuscitation.

[0195] A cardiac resuscitation process adjustment module 210: used to compare and analyze the on-site operation data collected with the operation parameters of the advanced life support process for cardiac resuscitation through artificial intelligence algorithms to adjust the cardiac resuscitation operation process in real time.

[0196] A visual display module 211: used to perform visual display on each module of the system for the visual cardiac resuscitation process through a visual screen.

[0197] As can be seen from the above embodiments, the system of the embodiment provided by the present invention realizes the above Figure 1A specific implementation form of the method shown, because the description of the embodiments in the above combination Figure 1 The description of the embodiments in the method is also applicable here. The present invention will not be elaborated herein.

[0198] Although multiple embodiments of the present invention have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art can envision many variations, changes, and alternative ways without departing from the spirit and scope of the present invention. It should be understood that various alternatives to the embodiments of the present invention described herein may be employed in practicing the present invention. The appended claims are intended to define the scope of the present invention and thus cover equivalents or alternatives within the scope of these claims.

Claims

1. A method for constructing a visual process of cardiac resuscitation, characterized in that: include: Build a big data platform by collecting multi-source data; Using the deep learning-based automatic arrhythmia recognition model, the patient's ECG data is analyzed in real time to identify the characteristic waveform of cardiac arrest; Based on the on-site video surveillance data, image recognition technology is used to analyze the patient's physical characteristics to determine the patient's vital signs; Using machine learning algorithms to analyze the patient's basic information, the state of the vital signs, and real-time data collected on site to assess the severity of the disease; Based on the resource configuration visualization screen of the hospital command center, the real-time positioning and status monitoring of emergency resources in the hospital are carried out through the Internet of Things technology; Based on the patient's location and condition, as well as the location and status of emergency resources, the best emergency resource allocation plan is automatically calculated through artificial intelligence algorithms; Generate personalized advanced cardiac resuscitation life support processes based on artificial intelligence algorithms and display them visually, providing emergency personnel with emergency operation steps and time node guidance; According to the evaluation results of the pre-built emergency effect evaluation model, an artificial intelligence algorithm is used to automatically generate process optimization suggestions to achieve dynamic optimization and adaptive adjustment of the cardiac resuscitation advanced life support process; The collected on-site operation data is compared and analyzed with the operation parameters of the cardiac resuscitation advanced life support process through an artificial intelligence algorithm to adjust the cardiac resuscitation operation process in real time.

2. The method according to claim 1, characterized in that The deep learning-based automatic arrhythmia recognition model performs real-time analysis on the patient's ECG data to identify characteristic waveforms of cardiac arrest, including: The convolutional neural network architecture is used as the basic framework of the automatic arrhythmia recognition model, combined with the attention mechanism, so that the model can focus on the key feature areas related to arrhythmia in the ECG signal; and combined with the long short-term memory network, the time series information of the ECG signal is modeled to obtain the dependency and dynamic change law of the ECG signal in the time dimension.

3. The method according to claim 1, characterized in that The method of analyzing the patient's physical characteristics using image recognition technology based on on-site video surveillance data to determine the patient's vital signs includes: Extracting image frame sequences from live video surveillance data to obtain an image sequence dataset; Construct a feature extraction network with multi-scale feature fusion and use the image sequence dataset to obtain the patient's body part positioning and local feature extraction; Introducing an attention mechanism into the feature extraction network to assign different attention weights to features in key areas; The feature sequence of image frames after feature extraction and attention weighting is input into the long short-term memory network to obtain the patient's vital signs status.

4. The method according to claim 1, characterized in that: The method uses a machine learning algorithm to analyze the patient's basic information, the vital signs and real-time data collected on site to assess the severity of the disease, including: Performing data preprocessing on the patient's basic information, the vital signs status, and real-time data collected on site; Construct combined features for the features after data preprocessing to obtain the processed feature vector; A gradient boosting decision tree model was constructed to predict the severity of the disease.

5. The method according to claim 1, characterized in that According to the patient's location and condition, as well as the location and status of emergency resources, the optimal emergency resource allocation plan is automatically calculated through an artificial intelligence algorithm, including: Integrate patient data and emergency resource data to generate a unified data format and coordinate system; Standardize different types of data to obtain the same dimensions and numerical range; Construct a hospital environment model based on graph theory, adopt a path planning algorithm, combine heuristic information and actual path cost, and obtain the optimal path from the location of emergency resources to the location of patients; The linear weighting method is used to transform multiple objective functions into a single objective function and establish a multi-objective optimization model; The multi-objective optimization model is decoded and solved using a legacy algorithm to obtain the best emergency resource allocation plan.

6. The method according to claim 1, characterized in that The pre-built first aid effectiveness evaluation model includes: Extracting and constructing specific features from the detailed data of each cardiac resuscitation event to obtain a cardiac resuscitation event feature data set; the specific features are features that reflect the relationship between each feature and the first aid effect, including time-related features, combination features, and trend features; Using the cardiac resuscitation event feature data set and an integrated learning algorithm based on a decision tree, a first aid effect evaluation model is constructed; wherein the first aid effect evaluation model includes a feature importance evaluation function; Utilizing the feature importance evaluation function of the first aid effect evaluation model, the contribution weight of each feature to the first aid effect is analyzed; By analyzing the decision tree structure and feature splitting in the first aid effect evaluation model, the potential relationship and rules between each feature and the first aid effect are explored to obtain the key factors and potential problem points of the treatment success rate.

7. The method according to claim 1, characterized in that The personalized advanced cardiac resuscitation life support process is generated based on the artificial intelligence algorithm and visualized, providing emergency personnel with emergency operation steps and time node guidance, including: Constructing a medical knowledge and case library; wherein the medical knowledge includes: authoritative medical literature, guidelines and expert experience; the case library includes cases of treatment of cardiac arrest patients, and the treatment cases include successful cases and failed cases; Build a hybrid neural network model based on deep learning; The preprocessed and feature-extracted patient data, as well as the corresponding treatment process and result data in the case library are divided into a training set, a validation set, and a test set; Training and optimizing the hybrid neural network model through a training set, a validation set, and a test set to obtain a trained hybrid neural network model; According to the patient data collected in real time on site, the trained hybrid neural network model is used to obtain a personalized cardiac resuscitation advanced life support process for the patient; wherein the personalized cardiac resuscitation advanced life support process includes the sequence of each operation step, the specific content and the recommended execution time node of each operation step.

8. The method according to claim 1, characterized in that The method of automatically generating a cardiac resuscitation process optimization suggestion based on the evaluation result of the pre-built emergency response effect evaluation model by using an artificial intelligence algorithm to achieve dynamic optimization and adaptive adjustment of the cardiac resuscitation advanced life support process includes: According to the pre-built first aid effect evaluation model, key data and related data for first aid effect evaluation are obtained to build a process optimization model. The key data for first aid effect evaluation include basic information of patients, on-site conditions, vital signs data, treatment process data, and first aid effect indicators; Constructing a process optimization model includes: constructing a reinforcement learning model, wherein the reinforcement learning model includes defining each link in the cardiac resuscitation advanced life support process and parameters related to each link as the state space of the reinforcement learning model; determining the action space of the reinforcement learning model to obtain the process optimization measures taken; using a reward function related to the first aid effect to evaluate the positive or negative impact of the actions taken by the reinforcement learning model on the first aid process, so that the reinforcement learning model learns the optimal process optimization strategy; During the cardiac resuscitation process, obtain the patient's vital signs data and emergency operation data to obtain the patient's current status information; Using the patient's current status information and based on the trained process optimization model, predict the best optimization action that should be taken under the current status, and generate corresponding process optimization suggestions; First aid personnel adjust the cardiopulmonary resuscitation operation process and relevant parameters of each link based on the process optimization suggestions generated by the model.

9. The method according to claim 1, characterized in that: The operating parameters of the cardiac resuscitation advanced life support process are chest compression, airway management, and electric defibrillation; the collected on-site operation data is compared and analyzed with the operating parameters of the cardiac resuscitation advanced life support process by the artificial intelligence algorithm to adjust the cardiac resuscitation operation process in real time, including: A hybrid architecture of long short-term memory network combined with convolutional neural network is used to build a comparative analysis model; Using historical cardiac resuscitation operation data and corresponding standard operation parameters as a training set to train the comparative analysis model; Using the comparative analysis model, predicting the deviation between the field operation data and the standard operation parameters; According to the deviations predicted by the model, the on-site cardiopulmonary resuscitation operation process is adjusted in real time.

10. A system for constructing a visual process of cardiac resuscitation, characterized in that: include: Data collection module: used to build a big data platform by collecting multi-source data; Arrhythmia automatic recognition model building module: used to build an arrhythmia automatic recognition model based on deep learning, and perform real-time analysis of ECG data to identify the characteristic waveform of cardiac arrest; Vital signs analysis module: used to analyze the patient's physical characteristics based on on-site video surveillance data using image recognition technology to determine the patient's vital signs status; Condition assessment module: used to analyze the patient's basic information, the vital signs and real-time data collected on site using a machine learning algorithm to assess the severity of the condition; Emergency Resource Management Module: It is used to configure a large visualization screen based on the resources of the hospital command center, and to conduct real-time positioning and status monitoring of emergency resources in the hospital through the Internet of Things technology; Emergency resource dispatch module: used to automatically calculate the best emergency resource deployment plan based on the patient's location and condition, as well as the location and status of emergency resources through artificial intelligence algorithms; First aid effect evaluation model building module: It is used to collect and organize detailed data of each cardiac resuscitation event, use machine learning algorithms to build a first aid effect evaluation model, analyze the relationship between various factors and the patient's recovery, and obtain the key factors and potential problems of the success rate of treatment; Cardiac resuscitation process generation module: used to generate personalized cardiac resuscitation advanced life support processes based on artificial intelligence algorithms and display them visually, providing emergency personnel with emergency operation steps and time node guidance; Cardiac resuscitation process optimization module: used to automatically generate cardiac resuscitation process optimization suggestions based on the evaluation results of the first aid effect evaluation model using an artificial intelligence algorithm to achieve dynamic optimization and adaptive adjustment of the cardiac resuscitation advanced life support process; Cardiac resuscitation process adjustment module: used to compare and analyze the collected on-site operation data with the operation parameters of the cardiac resuscitation advanced life support process through an artificial intelligence algorithm, so as to adjust the cardiac resuscitation operation process in real time; Visual display module: used for visually displaying each module of the system of the cardiac resuscitation visualization process through a visualization screen.

Images