Medical big data intelligent analysis method and system based on deep learning

Abstract

The invention relates to the technical field of medical big data intelligent analysis, and discloses a medical big data intelligent analysis method and system based on deep learning, and the method comprises the steps: obtaining patient image data, case text data and wearable equipment monitoring data, carrying out the preprocessing, and carrying out the standardized coding of each modal data; and calculating the dynamic weight of each modal data, and synthesizing a multi-level fusion weight. And in combination with the fusion weight and the clinical state parameters of the patient, constructing a modal value-clinical urgency degree two-factor resource decision model to obtain a priority score. According to the priority scores, the patients are divided into different emergency treatment levels, and different treatment strategies are provided. The model provided by the invention can maintain efficient prediction capability in various clinical scenes, improves the accuracy of clinical decisions, and enables the resource allocation of hospitals to be more scientific and reasonable. In an emergency treatment environment, the real-time priority evaluation of the patient can significantly improve the efficiency of emergency treatment, reduce the waiting time, and improve the survival rate and the treatment effect of the patient.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent analysis of medical big data, and specifically to a method and system for intelligent analysis of medical big data based on deep learning. Background Art

[0002] In recent years, with the popularization of medical imaging, electronic medical record (EMR) systems and wearable devices, the application of medical big data has developed significantly. The medical industry is gradually moving towards an era of data-driven intelligent diagnosis and treatment. Multimodal data sources such as imaging data, case text data and wearable device monitoring data provide clinicians with richer patient information. In particular, the introduction of deep learning technology has greatly improved the analysis of medical data. Deep learning can automatically extract effective features from large-scale medical data for efficient prediction and diagnosis, and has shown significant advantages in disease prediction, image processing, natural language processing and other fields. In addition, with the continuous advancement of cloud computing and big data technology, the storage, processing and analysis of medical data are no longer bottlenecks, and the deployment and reasoning capabilities of deep learning models are becoming more mature, which has greatly promoted the development of medical big data intelligent analysis systems.

[0003] Although the current intelligent analysis methods of medical big data have made significant progress, the existing technologies still have significant deficiencies in processing multimodal data fusion, personalized diagnosis and treatment, and clinical urgency assessment. Most existing methods often rely on single-modal data when processing medical data, such as diagnosing diseases based solely on image data or inferring symptoms based solely on case texts, lacking comprehensive analysis of multimodal data. The use of this single data source has led to limitations in diagnostic accuracy, especially in the diagnosis of some complex diseases, where single-modal information cannot fully reflect the patient's health status. In addition, although deep learning technology has been widely used in medical image analysis and text processing, how to effectively fuse multi-source information such as images, texts, and wearable device monitoring data, and generate comprehensive diagnostic opinions through deep learning models, is still an urgent problem to be solved. Existing fusion methods often lack an effective weighting mechanism, resulting in poor results in multimodal data fusion, and even information loss or data conflicts, which affects the accuracy and reliability of the model.

[0004] In addition, although deep learning models perform well in medical diagnosis, due to the diversity and complexity of training data, existing models often have difficulty in accurately assessing individual patient conditions. In clinical decision-making such as emergency treatment, personalized treatment plan recommendations, and patient priority sorting, existing deep learning models lack sufficient flexibility and accuracy and cannot be dynamically optimized based on patients' real-time data and changes in their condition. This makes it impossible for existing systems to provide comprehensive, real-time, and personalized medical support, and there is insufficient assessment of clinical urgency, resulting in uneven resource allocation or delays in the best treatment time for patients.

[0005] Our invention uses deep learning technology to efficiently fuse and analyze multimodal data in real time, and uses Bayesian optimization and weighted least squares to dynamically adjust weights, which can achieve accurate priority assessment of patients. Compared with existing technologies, our invention not only improves the fusion effect of multimodal data, but also can dynamically adjust treatment strategies according to changes in the patient's condition in real time, significantly improving the accuracy and timeliness of clinical decision-making. Summary of the invention

[0006] In view of the above-mentioned problems, the present invention is proposed.

[0007] Therefore, the technical problems solved by the present invention are: the existing medical data analysis methods have poor multimodal data fusion effect, lack of accuracy in personalized diagnosis and treatment and urgency assessment, and the inability to dynamically adjust resource allocation and treatment strategies according to patients' real-time data; as well as how to improve the accuracy of multimodal data fusion, optimize treatment plans in real time, and perform accurate clinical priority assessment.

[0008] To solve the above technical problems, the present invention provides the following technical solutions: a medical big data intelligent analysis method based on deep learning, comprising: obtaining patient imaging data, case text data and wearable device monitoring data, performing preprocessing, and standardizing the encoding of each modality data.

[0009] Calculate the dynamic weight of each modality data and synthesize multi-level fusion weights.

[0010] Combining the fusion weights and the patient's clinical status parameters, a modality value-clinical urgency dual-factor resource decision-making model was constructed to obtain the priority score.

[0011] According to the priority score, patients are divided into different emergency treatment levels and different treatment strategies are proposed.

[0012] As a preferred solution of the medical big data intelligent analysis method based on deep learning described in the present invention, the preprocessing includes collecting multimodal data of patients from hospital imaging systems, electronic medical records, and wearable devices, including imaging data, case text data, and wearable device data.

[0013] De-noise and normalize the image data. Segment and vectorize the case text data, and use the word embedding model to convert the text data into vector representation. Detect and correct outliers on wearable device data, and perform standardization.

[0014] As a preferred solution of the medical big data intelligent analysis method based on deep learning described in the present invention, the standardized encoding of each modality data includes using a deep convolutional network to extract features from the image data and obtaining high-level feature vectors of the image through a pre-trained model.

[0015] The text data is encoded through a natural language processing model to obtain a vector representation of the text.

[0016] The monitoring data of wearable devices are modeled through time series through deep neural networks to obtain the time series feature vector of the device monitoring data.

[0017] The obtained vector representations of each modal data are integrated to form a unified multimodal feature vector, which is expressed as: ; in, represents the final multimodal feature vector, Represents the encoding characteristics of image data, Represents the encoding characteristics of text data, Indicates the data encoding characteristics of wearable devices.

[0018] As a preferred solution of the medical big data intelligent analysis method based on deep learning described in the present invention, wherein: the calculation of the dynamic weight of each modality data includes initializing the weight of each modality , and it is iteratively updated using the Bayesian optimization algorithm, expressed as: ; in, Indicates Mode in time The dynamic weight of represents the final multimodal feature vector, Indicates The initial parameters of the modes, Indicates the current time.

[0019] In order to further improve the accuracy of the dynamic weight, the weighted least squares method is introduced to fine-tune the dynamic weight according to the quality of each modal data. The optimized weight is expressed as: ; in, represents the number of modes, Indicates The standard deviation of a mode indicates the relative contribution of the mode to the final decision. Indicates Patients at time The multimodal feature vector of Indicates the prediction model based on Mode in time The estimated feature vector of .

[0020] As a preferred solution of the medical big data intelligent analysis method based on deep learning described in the present invention, the synthesized multi-level fusion weight includes normalizing the dynamic weights of each modality after obtaining them, and using the weighted average method to fuse the features of different modalities, which is expressed as: ; in, Indicates The normalized weights of the modes.

[0021] Based on the normalized weights, the features of each modality are weighted and fused to obtain the fused multimodal feature vector. The process of multimodal fusion is expressed as: ; in, represents the fused multimodal feature vector, Indicates Mode in time The original vector.

[0022] As a preferred solution of the medical big data intelligent analysis method based on deep learning described in the present invention, the patient's clinical status parameters are screened from the collected patient imaging data, case text data and wearable device monitoring data, including vital signs data, disease scores and clinical observation data.

[0023] Vital sign data include heart rate, respiratory rate, blood oxygen saturation and blood pressure, the condition score is the acute condition score, and clinical observation data include body temperature, weight and clinical symptom description.

[0024] As a preferred solution of the medical big data intelligent analysis method based on deep learning described in the present invention, the construction of the modal value-clinical urgency dual-factor resource decision model includes merging the fused multimodal feature vector and the patient's clinical status parameters into a unified input vector. Assuming that there are m clinical status parameters, then Will contain fusion weights and patient clinical status parameters The splicing result is expressed as: ; A multilayer perceptron network model is used to learn the relationship between the fused features and clinical status parameters, including input layer, hidden layer and output layer.

[0025] The input layer receives input consisting of multimodal features and clinical status , the input vector The dimension is , which is the sum of the dimensions of the fusion features and the dimensions of the clinical status parameters.

[0026] The hidden layer contains multiple neurons, each of which receives the weighted sum input from the input layer and is processed by a nonlinear activation function. Hidden layer neurons The output is represented as: ; in, express Activation function, Represents the connection input layer nodes and hidden layers The weight of the node, Represents the input layer The value of the element, Represents the hidden layer The bias term of the node.

[0027] The output layer weights the output of the hidden layer, and then maps the output of the network to the probability distribution of the three priority categories through the Softmax function, which is expressed as: ; in, Represents the connection of hidden layer The weights of the nodes and the output layer, Represents the hidden layer The output of a neuron, represents the bias term of the output layer, represents the hidden layer dimension, express Function, expressed as: ; in, represents the weighted sum of each output neuron, Corresponding to low, medium and high priority, Represents the exponential sum of all categories, used to normalize the output.

[0028] The probability values ​​of the three categories of the patient at time t are obtained: the probability values ​​of low priority, medium priority and high priority, and the category with the largest probability value is selected as the priority category of the patient.

[0029] As a preferred solution of the medical big data intelligent analysis method based on deep learning described in the present invention, wherein: the construction of the modal value-clinical urgency dual-factor resource decision model also includes optimizing the network using a multi-class cross entropy loss function, training the deep learning model by minimizing the loss, and the objective function is expressed as: ; in, Indicates patient In time The true category label of the patient Belongs to category ,but , otherwise 0. Indicates patient In time Corresponding category The predicted probability of . represents the number of samples, Corresponding to low, medium, and high priority categories.

[0030] As a preferred solution of the medical big data intelligent analysis method based on deep learning described in the present invention, the patients are divided into different emergency treatment levels according to the priority score, and different treatment strategies are proposed, including that for patients with the highest probability value belonging to the high priority category, the hospital must immediately carry out emergency treatment, and all relevant medical staff must receive the patient as soon as possible. The patient is assigned to a dedicated intensive care unit, and all relevant medical resources are immediately activated and provided with priority.

[0031] Patients will be monitored in real time 24 hours a day during the emergency process, and medical staff will check key physiological parameters every hour to ensure a rapid response to any changes in their condition.

[0032] If necessary, emergency treatment and surgery will be carried out immediately to ensure that the patient's condition is stabilized as soon as possible.

[0033] For patients whose probability values ​​belong to the medium priority category and whose conditions need to be treated in the short term, the hospital will arrange diagnosis and treatment within 48 hours and ensure that the patients are admitted to the ward as soon as possible.

[0034] Assess the patient's condition every 24 hours to check for worsening conditions and other emergencies. If the patient's condition suddenly becomes serious, transfer them to a high-priority treatment process.

[0035] Based on the results of clinical evaluation, patients are provided with necessary drug treatment to ensure that their condition remains stable.

[0036] For patients whose probability values ​​belong to the low priority category, the hospital will arrange for them to be diagnosed and treated through regular outpatient clinics.

[0037] For low-priority patients, the hospital will prioritize the use of routine resources, including general wards and routine examination equipment.

[0038] The patient will be scheduled for a follow-up visit within an appropriate time after the visit to ensure that the condition has not deteriorated further. If there is a change, reconsider whether it needs to be upgraded based on the latest assessment of the condition. Focus on the patient's quality of life and provide symptom relief and rehabilitation treatment according to the patient's condition.

[0039] A medical big data intelligent analysis system based on deep learning, characterized by: including: The preprocessing module obtains patient image data, case text data and wearable device monitoring data, performs preprocessing, and standardizes the encoding of each modality data.

[0040] The calculation module calculates the dynamic weight of each modality data and synthesizes the multi-level fusion weight. Combining the fusion weight and the patient's clinical status parameters, a dual-factor resource decision model of modality value-clinical urgency is constructed to obtain the priority score.

[0041] The classification module divides patients into different emergency treatment levels according to the priority scores and proposes different treatment strategies.

[0042] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0043] A computer-readable storage medium stores a computer program, which implements the steps of the method described above when executed by a processor.

[0044] The beneficial effects of the present invention are as follows: by collecting, preprocessing and standardizing image data, case text data and wearable device monitoring data, combined with deep learning technology, feature extraction and fusion of data of different modalities are performed. The implementation of this step enables the system to integrate data from different sources, improving the comprehensiveness and accuracy of data analysis. Compared with traditional single data analysis methods, multimodal data fusion can overcome the limitations of each single data source and improve the predictive accuracy and reliability of the analysis model. This method reduces the probability of misdiagnosis and missed diagnosis by integrating information, effectively supporting more accurate personalized diagnosis and treatment.

[0045] During the optimization process of the model, the Bayesian optimization algorithm was introduced to tune the model hyperparameters, and the weighted least squares method was combined to adjust the weighted data of the model. Through this dynamic optimization process, the system can more accurately deal with the heterogeneity of the data and improve the prediction effect. The introduction of this optimization step enables the model to not only achieve efficient training with less sample data, but also dynamically adjust according to the individual differences of patients, so that the model can maintain efficient prediction capabilities in a variety of clinical scenarios and improve the accuracy of clinical decision-making.

[0046] The fused data is processed by a multi-layer perceptron (MLP) network model, and the patient's health status is combined with clinical urgency parameters (such as vital signs data, condition score, etc.) to generate a priority score for each patient. This score is classified into high priority, medium priority, and low priority based on the output value of the Softmax function, ensuring that the emergency treatment level of each patient can be accurately judged. Unlike the traditional priority evaluation method based on deep learning, the present invention dynamically evaluates the patient's urgency through a deep learning model, avoiding the subjectivity and errors of human judgment, making the hospital's resource allocation more scientific and reasonable. Especially in an emergency environment, real-time priority assessment of patients can significantly improve the efficiency of emergency treatment, reduce waiting time, and improve patients' survival rate and treatment effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them: Figure 1 An overall flow chart of a medical big data intelligent analysis method and system based on deep learning provided for the first embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0049] Example 1, reference Figure 1 , is an embodiment of the present invention, and provides a medical big data intelligent analysis method based on deep learning, comprising: S1: Obtain patient imaging data, case text data, and wearable device monitoring data, perform preprocessing, and standardize the encoding of each modality data.

[0050] Collect patients' multimodal data from hospital imaging systems, electronic medical records, and wearable devices, including imaging data, case text data, and wearable device data.

[0051] It should be noted that medical image acquisition equipment uses CT scanners, MRI (magnetic resonance imaging) equipment, X-ray equipment, and ultrasound scanners. The acquisition process is as follows: In the hospital imaging department, radiologists select appropriate imaging equipment for examination based on the patient's symptoms and clinical needs. The scan results generated by the imaging equipment will be automatically recorded in the hospital imaging system, and all imaging data (stored in DICOM format and transmitted to the hospital data management system) will be automatically recorded in the hospital imaging system.

[0052] The process of collecting case text data is as follows: When a patient visits a hospital, the doctor will record the patient's medical history, chief complaints, physical signs, diagnosis, and treatment plan, and this information will be entered into the hospital's electronic medical record system.

[0053] The electronic medical record system automatically generates and archives the patient's medical record files. These text data contain the patient's detailed medical history information, diagnosis description, treatment records, etc.

[0054] Wearable device data collection devices include smart bracelets, smart watches, wearable physiological monitoring devices (such as heart rate monitors, blood oxygen meters, sleep monitoring devices, etc.). The collection process is as follows: After the patient wears the wearable device, the device will monitor and record physiological data (such as heart rate, body temperature, respiratory rate, blood oxygen saturation, etc.) through sensors.

[0055] The data is transmitted to the hospital data platform and synchronized to the hospital data management system in real time.

[0056] De-noise and normalize the image data. Segment and vectorize the case text data, and use the word embedding model to convert the text data into vector representation. Detect and correct outliers on wearable device data, and perform standardization.

[0057] It should be noted that denoising uses the U-Net model in the convolutional neural network (CNN) for image denoising. U-Net is widely used in denoising and segmentation tasks in medical images and has powerful feature extraction capabilities. Normalization processing normalizes the pixel values ​​of image data so that the pixel values ​​of the image are in the [0,1] interval to facilitate subsequent model training.

[0058] Furthermore, case text data processing includes the following: Word segmentation: Use spaCy (for English text segmentation) and Jieba (for Chinese text segmentation) to segment case text data. The segmented data can be more easily processed by deep learning models.

[0059] Word embedding model: Use the BERT model to vectorize text data. The BERT model is particularly suitable for processing medical case texts and can obtain contextual semantic information for each word.

[0060] Wearable device data processing includes the following: Outlier detection and correction: Use the isolation forest algorithm for outlier detection. These algorithms can effectively identify abnormal fluctuations in monitoring data and correct data points that do not conform to the rules.

[0061] Standardization: Use Z-score standardization to standardize the device data so that they have the same scale to facilitate subsequent model processing.

[0062] Use deep convolutional networks to extract features from image data and obtain high-level feature vectors of images through pre-trained models.

[0063] The text data is encoded through a natural language processing model to obtain a vector representation of the text.

[0064] The monitoring data of wearable devices are modeled through time series through deep neural networks to obtain the time series feature vector of the device monitoring data.

[0065] It should be noted that image feature extraction uses ResNet50 in the deep convolutional neural network (CNN) for image feature extraction. Fine-tune the pre-trained model to extract high-level feature vectors from the image. Text data feature extraction uses the BERT model to encode case text data and convert it into a fixed-dimensional text vector representation. Wearable device data feature extraction: Use LSTM for time series modeling to extract the time series feature vector of device monitoring data.

[0066] The obtained vector representations of each modal data are integrated to form a unified multimodal feature vector, which is expressed as: ; in, represents the final multimodal feature vector, Represents the encoding characteristics of image data, Represents the encoding characteristics of text data, Indicates the data encoding characteristics of wearable devices.

[0067] S2: Calculate the dynamic weight of each modality data and synthesize multi-level fusion weights.

[0068] Initialize the weights of each modality , and it is iteratively updated using the Bayesian optimization algorithm, expressed as: ; in, Indicates Mode in time The dynamic weight of represents the final multimodal feature vector, Indicates The initial parameters of the modes, Indicates the current time.

[0069] In order to further improve the accuracy of the dynamic weight, the weighted least squares method is introduced to fine-tune the dynamic weight according to the quality of each modal data. The optimized weight is expressed as: ; Where n represents the number of modes, Indicates The standard deviation of a mode indicates the relative contribution of the mode to the final decision. Indicates Patients at time The multimodal feature vector of Indicates the prediction model based on Mode in time The estimated feature vector of .

[0070] It should be noted that Bayesian optimization is a global optimization algorithm that uses a proxy model (such as a Gaussian process) to optimize the objective function. It can find the optimal hyperparameter configuration with a small number of evaluations, so it is suitable for deep learning models that require hyperparameter tuning. Through Bayesian optimization, model parameters can be dynamically adjusted during the training process, thereby improving model performance.

[0071] Furthermore, the weighted least squares method is introduced to optimize by assigning weights to each data point, which is very effective for the imbalance problem in the data. For medical data, the conditions of some patients may be more important, and giving these samples higher weights can help the model better learn the characteristics of these patients. The use of weighted least squares can improve the prediction accuracy of the model on important categories, especially in the medical field, which is crucial for the detection and classification of key diseases.

[0072] After obtaining the dynamic weights of each mode, normalization is performed and the weighted average method is used to fuse the features of different modes, which is expressed as: ; in, Indicates The normalized weights of the modes.

[0073] Based on the normalized weights, the features of each modality are weighted and fused to obtain the fused multimodal feature vector. The process of multimodal fusion is expressed as: ; in, represents the fused multimodal feature vector, Indicates Mode in time The original vector.

[0074] S3: Combining the fusion weights and the patient's clinical status parameters, a modality value-clinical urgency dual-factor resource decision model is constructed to obtain a priority score.

[0075] The patient's clinical status parameters are screened from the collected patient imaging data, case text data and wearable device monitoring data, including vital signs data, disease scores and clinical observation data.

[0076] Vital sign data include heart rate, respiratory rate, blood oxygen saturation and blood pressure, the condition score is the acute condition score, and clinical observation data include body temperature, weight and clinical symptom description.

[0077] Furthermore, the acute condition score is scored by doctors based on the patient's condition, combining clinical diagnosis, pathological indicators, vital signs and other dimensions to assess the severity of the patient's condition. This score is manually given by doctors based on clinical experience and the patient's specific condition through a scoring table.

[0078] The fused multimodal feature vector and the patient's clinical status parameters are combined into a unified input vector. Assuming there are m clinical status parameters, then Will contain fusion weights and patient clinical status parameters The splicing result is expressed as: ; A multilayer perceptron network model is used to learn the relationship between the fused features and clinical status parameters, including input layer, hidden layer and output layer.

[0079] The input layer receives input consisting of multimodal features and clinical status , the input vector The dimension is , which is the sum of the dimensions of the fusion features and the dimensions of the clinical status parameters.

[0080] The hidden layer contains multiple neurons, each of which receives the weighted sum input from the input layer and is processed by a nonlinear activation function. Hidden layer neurons The output is represented as: ; in, express Activation function, Represents the connection input layer nodes and hidden layers The weight of the node, Represents the input layer The value of the element, Represents the hidden layer The bias term of the node.

[0081] The output layer weights the output of the hidden layer, and then maps the output of the network to the probability distribution of the three priority categories through the Softmax function, which is expressed as: ; in, Represents the connection of hidden layer The weights of the nodes and the output layer, Represents the hidden layer The output of a neuron, represents the bias term of the output layer, represents the hidden layer dimension, express Function, expressed as: ; in, represents the weighted sum of each output neuron, Corresponding to low, medium and high priority, Represents the exponential sum of all categories, used to normalize the output.

[0082] The probability values ​​of the three categories of the patient at time t are obtained: the probability values ​​of low priority, medium priority and high priority, and the category with the largest probability value is selected as the priority category of the patient.

[0083] The multi-class cross entropy loss function is used to optimize the network and train the deep learning model by minimizing the loss. The objective function is expressed as: ; in, Indicates patient In time The true category label of the patient Belongs to category ,but , otherwise 0. Indicates patient In time Corresponding category The predicted probability of . N represents the number of samples, Corresponding to low, medium, and high priority categories.

[0084] It should be noted that by minimizing the multi-class cross entropy loss function, the classification performance of the model can be optimized, so that the gap between the probability distribution of the model output and the actual label is minimized. The optimization of the loss function can improve the model's ability to identify different priority categories, ensure that the model can accurately assign priorities to patients, and optimize the allocation of medical resources.

[0085] S4: According to the priority score, patients are divided into different emergency treatment levels and different treatment strategies are proposed.

[0086] For patients with the highest probability value and belonging to the high priority category, the hospital must immediately provide emergency treatment, and all relevant medical staff must receive the patient as soon as possible. The patient will be assigned to a dedicated intensive care unit, and all relevant medical resources will be immediately activated and provided with priority.

[0087] Patients will be monitored in real time 24 hours a day during the emergency process, and medical staff will check key physiological parameters every hour to ensure a rapid response to any changes in their condition.

[0088] If necessary, emergency treatment and surgery will be carried out immediately to ensure that the patient's condition is stabilized as soon as possible.

[0089] For patients whose probability values ​​belong to the medium priority category and whose conditions need to be treated in the short term, the hospital will arrange diagnosis and treatment within 48 hours and ensure that the patients are admitted to the ward as soon as possible.

[0090] Assess the patient's condition every 24 hours to check for worsening conditions and other emergencies. If the patient's condition suddenly becomes serious, transfer them to a high-priority treatment process.

[0091] Based on the results of clinical evaluation, patients are provided with necessary drug treatment to ensure that their condition remains stable.

[0092] For patients whose probability values ​​belong to the low priority category, the hospital will arrange for them to be diagnosed and treated through regular outpatient clinics.

[0093] For low-priority patients, the hospital will prioritize the use of routine resources, including general wards and routine examination equipment.

[0094] The patient will be scheduled for a follow-up visit within an appropriate time after the visit to ensure that the condition has not deteriorated further. If there is a change, reconsider whether it needs to be upgraded based on the latest assessment of the condition. Focus on the patient's quality of life and provide symptom relief and rehabilitation treatment according to the patient's condition.

[0095] The above embodiments also include a medical big data intelligent analysis system based on deep learning, specifically: The preprocessing module obtains patient image data, case text data and wearable device monitoring data, performs preprocessing, and standardizes the encoding of each modality data.

[0096] The calculation module calculates the dynamic weight of each modality data and synthesizes the multi-level fusion weight. Combining the fusion weight and the patient's clinical status parameters, a dual-factor resource decision model of modality value-clinical urgency is constructed to obtain the priority score.

[0097] The classification module divides patients into different emergency treatment levels according to the priority scores and proposes different treatment strategies.

[0098] The computer device may be a server. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data cluster data of the power monitoring system. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a medical big data intelligent analysis method based on deep learning is implemented.

[0099] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0100] Example 2 is an embodiment of the present invention, which provides a medical big data intelligent analysis method and system based on deep learning. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through simulation experiments.

[0101] The goal of this example is to verify the effectiveness of the medical big data intelligent analysis method and system based on deep learning, especially the application effect in multimodal data fusion, clinical priority assessment and resource allocation. The main object of the experiment is 100 patient data from the emergency department of a large hospital, covering imaging data, case text data and wearable device monitoring data. The specific implementation process is as follows: Image data: CT image data of 100 patients with conditions ranging from mild respiratory diseases to acute myocardial infarction were collected through the hospital's imaging department. The image data were stored in DICOM format and uploaded to the hospital's data management system.

[0102] Case text data: The patient's medical record text is collected through the hospital's electronic medical record system (EMR), including medical history, physical signs, diagnosis information, etc. The medical record text is processed by the natural language processing (NLP) model to extract key information such as symptoms and past medical history.

[0103] Wearable device monitoring data: Use wearable devices such as smart watches and smart heart rate monitors to collect patients' real-time vital signs data, including heart rate, blood oxygen saturation, body temperature, etc. All wearable device data is uploaded to the cloud platform through the hospital's smart terminal and standardized.

[0104] Data preprocessing and feature extraction: Image data preprocessing: Use the U-Net model to denoise the image data and normalize the image data to the [0,1] interval to facilitate the input of the subsequent deep learning model. Use the ResNet50 model to extract high-level feature vectors from the image data.

[0105] Case text data processing: The case text data is converted into vector representation through word segmentation and word embedding model (BERT). The BERT model can capture the contextual semantic relationship in the text and generate accurate text vector representation.

[0106] Wearable device data processing: Use the LSTM model to model the time series of the monitoring data of wearable devices and extract the time series feature vector. Data standardization ensures that data from different devices can be analyzed at the same scale.

[0107] Training and optimization of deep learning models: The Bayesian optimization algorithm is used to optimize the model hyperparameters, and the weighted least squares method is combined to adjust the model's weights for each modality data. The preprocessed image, text, and wearable device data are fused and input into the deep learning model. The model is based on a multi-layer perceptron (MLP) network structure and outputs the priority score of each patient through the Softmax function.

[0108] The model is trained through a multi-class cross entropy loss function to ensure that the final model achieves optimal performance on different priority classifications.

[0109] During the training process, the model parameters are optimized by comparing the model prediction results with the priority labels evaluated by experts to improve the accuracy and stability of the model in clinical applications.

[0110] Evaluation criteria: Accuracy: Use the test set to evaluate the accuracy of the model in classifying patient priorities.

[0111] Resource allocation efficiency: Patients are assigned to different treatment priorities based on the priority scores predicted by the model to evaluate the effectiveness of the model in actual clinical resource allocation.

[0112] Patient survival rate and treatment time: The effectiveness of the deep learning model was verified by comparing the survival rate and treatment time of emergency patients after priority was assigned using the deep learning model.

[0113] The experimental results are shown in Table 1.

[0114] Table 1 Experimental data of the model of the present invention in the priority assessment of emergency patients

[0115] It can be seen from the experimental data table that the intelligent analysis method of medical big data based on deep learning has significant advantages in the priority assessment of emergency patients. In the high-priority patient prediction probability column, the prediction probability processed by the deep learning model is significantly higher than the traditional manual evaluation method, and has a high consistency with the priority labels annotated by clinical experts. For example, the high priority prediction probabilities of patients 1, 3, and 5 are 75.2%, 80.1%, and 85.5%, respectively. These data show that the model can maintain a high accuracy in the prediction of high-priority patients. Moreover, the priority score predicted by this method can effectively guide the allocation of hospital resources, so that high-priority patients in the emergency room can be treated in a timely manner.

[0116] In addition, by comparing the data in the resource allocation efficiency (processing time, minutes), it is found that the efficiency of resource allocation of this method has been significantly improved. For example, the treatment time of patient 1 is 30 minutes, while traditional methods usually take longer to make similar diagnoses and allocations. This shows that the application of deep learning models in the priority assessment of emergency patients can significantly reduce the waiting time of patients and improve the efficiency of the use of emergency resources in hospitals.

[0117] From the case text data processing time and wearable device data processing time columns, we can see that the entire data processing process is completed within a reasonable time, ensuring the real-time nature of the data and the rapid response capability of the model. In particular, in the processing of wearable device data, the use of deep neural networks for time series modeling effectively captures the dynamic changes of patients and further improves the accuracy of predictions.

[0118] Through the above data analysis, it can be concluded that the deep learning model of the present invention can accurately and quickly evaluate the priority of emergency patients, optimize the resource allocation process, and improve the efficiency and treatment effect of emergency treatment in hospitals. Compared with the existing technology, the present invention is outstanding in its innovation and novelty in multimodal data fusion and priority assessment, and can provide more intelligent and accurate decision support in emergency medical environments, with significant clinical application prospects.

[0119] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A medical big data intelligent analysis method based on deep learning, characterized in that: include: Obtain patient imaging data, case text data, and wearable device monitoring data, perform preprocessing, and perform standardized coding on each modality data; Calculate the dynamic weight of each modality data and synthesize multi-level fusion weights; Combining the fusion weights and the patient's clinical status parameters, a dual-factor resource decision-making model of modality value and clinical urgency was constructed to obtain a priority score. According to the priority score, patients are divided into different emergency treatment levels and different treatment strategies are proposed.

2. The medical big data intelligent analysis method based on deep learning as claimed in claim 1, characterized in that: The preprocessing includes collecting multimodal data of patients from hospital imaging systems, electronic medical records, and wearable devices, including imaging data, case text data, and wearable device data; The image data is denoised and normalized. The case text data is segmented and vectorized, and the word embedding model is used to convert the text data into a vector representation. The wearable device data is detected and corrected for outliers and standardized.

3. The medical big data intelligent analysis method based on deep learning as claimed in claim 2, characterized in that: The standardized encoding of each modality data includes extracting features from the image data using a deep convolutional network and obtaining a high-level feature vector of the image through a pre-trained model; Encode the text data through a natural language processing model to obtain a vector representation of the text; The monitoring data of the wearable device is modeled through a deep neural network to obtain the time series feature vector of the device monitoring data; The obtained vector representations of each modal data are integrated to form a unified multimodal feature vector, which is expressed as: ； in, represents the final multimodal feature vector, Represents the encoding characteristics of image data, Represents the encoding characteristics of text data, Indicates the data encoding characteristics of wearable devices.

4. The medical big data intelligent analysis method based on deep learning as claimed in claim 3, characterized in that: The calculation of the dynamic weight of each modal data includes initializing the weight of each modal , and it is iteratively updated using the Bayesian optimization algorithm, expressed as: ； in, Indicates Mode in time The dynamic weight of represents the final multimodal feature vector, Indicates The initial parameters of the modes, Indicates the current time; In order to further improve the accuracy of the dynamic weight, the weighted least squares method is introduced to fine-tune the dynamic weight according to the quality of each modal data. The optimized weight is expressed as: ； in, represents the number of modes, Indicates The standard deviation of each mode indicates the relative contribution of the mode to the final decision; Indicates Patients at time The multimodal feature vector of Indicates the prediction model based on Mode in time The estimated feature vector of .

5. The medical big data intelligent analysis method based on deep learning as claimed in claim 4, characterized in that: The synthesis of multi-level fusion weights includes normalizing the dynamic weights of each modality after obtaining them, and using a weighted average method to fuse the features of different modalities, which is expressed as: ； in, Indicates Normalized weights of each mode; Based on the normalized weights, the features of each modality are weighted and fused to obtain the fused multimodal feature vector. The process of multimodal fusion is expressed as: ； in, represents the fused multimodal feature vector, Indicates Mode in time The original vector.

6. The medical big data intelligent analysis method based on deep learning as claimed in claim 5, characterized in that: The patient clinical status parameters are screened from the collected patient imaging data, case text data and wearable device monitoring data, including vital signs data, disease score and clinical observation data; Vital sign data include heart rate, respiratory rate, blood oxygen saturation and blood pressure, the condition score is the acute condition score, and clinical observation data include body temperature, weight and clinical symptom description.

7. The medical big data intelligent analysis method based on deep learning according to claim 6, characterized in that: The construction of the modality value-clinical urgency dual-factor resource decision model includes combining the fused multimodal feature vector and the patient's clinical status parameter into a unified input vector. Assuming there are m clinical status parameters, then Will contain fusion weights and patient clinical status parameters The splicing result is expressed as: ； A multilayer perceptron network model is used to learn the relationship between the fused features and clinical status parameters, including an input layer, a hidden layer, and an output layer; The input layer receives input consisting of multimodal features and clinical status , the input vector The dimension is , is the sum of the dimensions of the fusion features and the dimensions of the clinical status parameters; The hidden layer contains multiple neurons, each of which receives the weighted sum input from the input layer and is processed by a nonlinear activation function. Hidden layer neurons The output is represented as: ； in, express Activation function, Represents the connection input layer nodes and hidden layers The weight of the node, Represents the input layer The value of the element, Represents the hidden layer The bias term of each node; The output layer weights the output of the hidden layer, and then maps the output of the network to the probability distribution of the three priority categories through the Softmax function, which is expressed as: ； in, Represents the connection of hidden layer The weights of the nodes and the output layer, Represents the hidden layer The output of a neuron, represents the bias term of the output layer, represents the hidden layer dimension, express Function, expressed as: ； in, represents the weighted sum of each output neuron, Corresponding to low, medium and high priority, represents the exponential sum of all categories, used to normalize the output; The probability values ​​of the three categories of the patient at time t are obtained: the probability values ​​of low priority, medium priority and high priority, and the category with the largest probability value is selected as the priority category of the patient.

8. The medical big data intelligent analysis method based on deep learning as claimed in claim 7, characterized in that: The constructing of the modality value-clinical urgency dual-factor resource decision model also includes optimizing the network using a multi-class cross entropy loss function and training the deep learning model by minimizing the loss. The objective function is expressed as: ； in, Indicates patient In time The true category label of the patient Belongs to category ,but , otherwise 0; Indicates patient In time Corresponding category The predicted probability of represents the number of samples, Corresponding to low, medium, and high priority categories.

9. The medical big data intelligent analysis method based on deep learning as claimed in claim 8, characterized in that: According to the priority score, the patients are divided into different emergency treatment levels, and different treatment strategies are proposed, including that for patients with the highest probability value belonging to the high priority category, the hospital must immediately carry out emergency treatment, and all relevant medical staff must receive the patient as soon as possible; the patient is assigned to a dedicated intensive care unit, and all relevant medical resources are immediately activated and provided with priority; Patients will be monitored 24 hours a day in real time during the emergency treatment process, and medical staff will check key physiological parameters every hour to ensure a quick response to any changes in their condition; If necessary, emergency treatment and surgery will be carried out immediately to ensure that the patient's condition is stabilized as soon as possible; For patients with medium-priority probability values, whose conditions need to be treated in the short term, the hospital will arrange diagnosis and treatment within 48 hours and ensure that the patients are admitted to the ward as soon as possible; Assess the patient's condition every 24 hours to check for worsening conditions and other emergencies. If the patient's condition suddenly becomes serious, transfer them to a high-priority treatment process. Provide patients with necessary medication based on clinical assessment results to ensure their condition remains stable; For patients whose probability values ​​belong to the low priority category, the hospital will arrange for the patients to be diagnosed and treated through regular outpatient clinics; For low-priority patients, the hospital will prioritize the use of conventional resources, including general wards and routine examination equipment; The patient arranges a follow-up visit within an appropriate time after the visit to ensure that the condition has not deteriorated further. If there is a change, reconsider whether the patient's priority should be raised based on the latest assessment of the condition; Focus on the patient's quality of life and provide symptom relief and rehabilitation treatment according to the patient's condition.

10. A medical big data intelligent analysis system based on deep learning using the method according to any one of claims 1 to 7, characterized in that: The preprocessing module obtains patient imaging data, case text data, and wearable device monitoring data, performs preprocessing, and performs standardized coding on each modality data; The calculation module calculates the dynamic weight of each modality data and synthesizes the multi-level fusion weight; Combining the fusion weights and the patient's clinical status parameters, a dual-factor resource decision-making model of modality value and clinical urgency was constructed to obtain a priority score. The classification module divides patients into different emergency treatment levels according to the priority scores and proposes different treatment strategies.

Images