System and method for preeclampsia prevention and early management

Abstract

The invention provides a system and a method for preeclampsia prevention and early management, which take digital intelligent management of gestational hypertension and preeclampsia as an entry point, focus medical and defense intelligent fusion, and depend on means such as digital and intelligent auxiliary diagnosis and treatment and information acquisition. The accuracy of early prevention and management of gestational hypertension and preeclampsia is effectively improved, pregnant women with bad outcome risks are identified, living habits of the pregnant women are continuously improved, and adverse effects caused by diseases are reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of disease risk prediction, and in particular relates to a system for the prevention and early management of preeclampsia. Background Art

[0002] Preeclampsia (PE) is a common pregnancy-specific disease and a major cause of maternal and fetal morbidity and mortality. The core of the pathogenesis of preeclampsia is the production of antiangiogenic and inflammatory factors by the hypoxic placenta, which leads to downstream manifestations of the disease, including hypertension and end-organ damage. Currently, effective treatments for PE are limited and are limited to prolonged pregnancy and anticonvulsant treatment to prevent seizures, as delivery of the placenta and fetus is the only way to cure the disease. However, as the incidence of PE continues to rise worldwide, considerable efforts have been made to develop treatments that are safe for both mothers and infants. PE is a variable disease, and the disease can progress from non-severe PE progressively to severe PE or even eclampsia, or it can develop very rapidly and in leaps and bounds, with eclamptic convulsions occurring in the absence of obvious prodromal manifestations or symptoms. Although termination of pregnancy is the only treatment for PE, it may be difficult to decide the timing of termination of pregnancy for PE, especially early-onset PE. Prevention and early warning systems are needed to guide patients to actively carry out early management, control blood pressure levels, improve pregnancy outcomes, and help identify pregnant women at risk of adverse outcomes in order to balance the risks to mothers and infants brought about by termination of pregnancy and expectant management. Summary of the invention

[0003] The purpose of the present invention is to provide a system for the prevention and early management of preeclampsia on the basis of the existing technology, taking the digital intelligent management of gestational hypertension and preeclampsia as the entry point, focusing on the intelligent integration of medical and prevention, relying on digital and intelligent auxiliary diagnosis and treatment and information collection and other means, to effectively improve the accuracy of early prevention and management of gestational hypertension and preeclampsia, identify pregnant women with risks of adverse outcomes, continuously improve the living habits of pregnant women, and reduce the adverse effects of the disease.

[0004] Another object of the present invention is to provide the above method for the prevention and early management of preeclampsia.

[0005] The technical solution of the present invention is as follows:

[0006] Pregnant women: enter the health data of pregnant women and create a database;

[0007] Risk grading and early warning: input the critical value range of indicators and the criteria for determining the risk level of hypertension to establish an early warning model;

[0008] Doctor management end: Doctors process warning information and feedback to pregnant women;

[0009] Integrated prenatal and postnatal recording terminal: input the follow-up results of pregnant women, and intelligently screen and classify the data of healthy and unhealthy pregnant women;

[0010] AI intelligent analysis terminal: Statistical analysis of the differences between healthy and unhealthy pregnant women, and use of big data to verify the effectiveness of early management interventions.

[0011] In a preferred solution, on the pregnant woman's side, a Chatgpt-like model is used to intelligently inquire about the pregnant woman's health data, including entering test reports, previous medical history, basic health information, psychological stress survey results, physiological stress assessment results and discomfort symptoms, and intelligent information collection is performed on the entered health data to create a database.

[0012] Among them, the test reports include: blood routine, urine routine, BNP, cTnI, blood sugar and B-ultrasound test reports. Past medical history includes history of hypertension during pregnancy, history of chronic hypertension, history of preeclampsia, reproductive history and history of diabetes. Basic health includes: age during pregnancy, height, weight and blood pressure. Discomfort symptoms include: dizziness, headache, lower limb edema and blurred vision.

[0013] In a preferred embodiment, in the risk grading and early warning end, the critical value range of indicators and the criteria for determining the risk level of hypertension are input, and an artificial neural network learning model is used for deep learning. The risk grading and early warning are determined based on whether abnormal indicators are identified. If abnormal indicators are not identified, the risk is low, and if abnormal indicators are identified, the risk is medium or high. An early warning model is established. When the risk grading is low risk, close follow-up is performed; when the risk grading is medium, high or low risk, a warning message is issued and transmitted to the doctor management end.

[0014] In a preferred solution, on the doctor management side, the doctor processes the warning information and feeds it back to the pregnant woman. Intelligent monitoring reminds the pregnant woman to take medication (to control blood pressure, blood sugar and blood lipids), and pushes healthy exercises, diet recipes, popular science and music, thereby achieving the effect of relieving the pregnant mother's stress.

[0015] Integrated prenatal and postnatal recording terminal: record the follow-up results of pregnant women, such as pregnancy and delivery outcomes, incidence of gestational hypertension, incidence of preeclampsia, incidence of severe preeclampsia, incidence of eclampsia, time from onset of gestational hypertension to hospitalization for delivery, time from onset of preeclampsia to hospitalization for delivery, incidence of pregnancy complications, and incidence of postpartum hemorrhage, and intelligently filter and classify data of healthy and unhealthy pregnant women.

[0016] The present invention also provides a method for the prevention and early management of preeclampsia, comprising:

[0017] Obtain the health data of pregnant women through the pregnant women’s terminal and create a database;

[0018] The critical value range of indicators and the criteria for determining the risk level of hypertension are input through the risk grading and early warning terminal, and the artificial neural network learning model is used for deep learning. The risk grading and early warning are determined based on whether abnormal indicators are identified. If abnormal indicators are not identified, the risk is low, and if abnormal indicators are identified, the risk is medium or high. When the risk grading is low, close follow-up is performed; when the risk grading is medium or high, an early warning message is sent to the doctor management terminal.

[0019] The doctor management end processes the warning information and feeds back the processing results to the pregnant woman end, which can intelligently monitor and remind the pregnant woman to take medicine, and push healthy exercise, diet recipes, popular science and music;

[0020] The integrated prenatal and postnatal recording terminal records the follow-up results of pregnant women, and intelligently screens and classifies the data of healthy and unhealthy pregnant women;

[0021] Through AI intelligent analysis, statistical analysis is performed on the differences between healthy and unhealthy pregnant women, and big data is used to verify the effectiveness of early management interventions.

[0022] For the present invention, the health data of pregnant women is obtained through intelligent inquiry of a Chatgpt-like model, including obtaining test reports, previous medical history, basic health information, psychological stress survey results, physiological stress assessment results and discomfort symptoms, and intelligent information collection is performed on the obtained health data to create a database.

[0023] Among them, the test reports include: blood routine, urine routine, BNP, cTnI, blood sugar and B-ultrasound test reports. Past medical history includes history of hypertension during pregnancy, history of chronic hypertension, history of preeclampsia, reproductive history and history of diabetes. Basic health includes: age during pregnancy, height, weight and blood pressure. Discomfort symptoms include: dizziness, headache, lower limb edema and blurred vision.

[0024] Adopt the technical scheme of the present invention, the advantages are as follows:

[0025] The present invention provides a system and method for the prevention and early management of preeclampsia. It takes the digital intelligent management of gestational hypertension and preeclampsia as the entry point, focuses on the intelligent integration of medical and preventive care, relies on digital and intelligent auxiliary diagnosis and treatment and information collection, and effectively improves the accuracy of early prevention and management of gestational hypertension and preeclampsia, identifies pregnant women at risk of adverse outcomes, continuously improves the living habits of pregnant women, and reduces the adverse effects of the disease. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a flow chart of the system for the prevention and early management of preeclampsia provided by the present invention. DETAILED DESCRIPTION

[0027] The present invention can be better understood according to the following examples. However, it is easy for those skilled in the art to understand that the contents described in the examples are only used to illustrate the present invention, and should not and will not limit the present invention described in detail in the claims.

[0028] like Figure 1 As shown, the system for the prevention and early management of preeclampsia provided by the present invention comprises:

[0029] Pregnant women: The medical question-and-answer model based on deep learning is used to input the health data of pregnant women and create a database.

[0030] Through a Chatgpt-like model, the health data of pregnant women is intelligently queried, including entering test reports, previous medical history, basic health information, psychological stress survey results, physiological stress assessment results and discomfort symptoms, and the entered health data is intelligently processed to create a database.

[0031] Among them, the test reports include: blood routine, urine routine, BNP, cTnI, blood sugar and B-ultrasound test reports. Past medical history includes history of hypertension during pregnancy, history of chronic hypertension, history of preeclampsia, reproductive history and history of diabetes. Basic health includes: age during pregnancy, height, weight and blood pressure. Discomfort symptoms include: dizziness, headache, lower limb edema and blurred vision.

[0032] Medical question answering model based on deep learning:

[0033] In terms of building an intelligent inquiry model, MedicalGPT was selected as the basic inquiry framework. MedicalGPT is a pre-trained language model based on OpenAI's GPT-3 architecture, which is customized for the medical field and aims to provide intelligent auxiliary diagnosis, consultation and knowledge retrieval services by understanding and generating medical-related texts. Whether doctors, researchers or medical industry practitioners, they can all benefit from its powerful natural language processing capabilities.

[0034] The core of MedicalGPT is the GPT-3 model, a large-scale Transformer architecture that can understand and generate texts of various lengths and complexities. In medical applications, it has been pre-trained on a large amount of medical literature, clinical reports and other data, and has the ability to understand professional terms and concepts. The pre-training phase of the project is carried out on a large amount of medical data, which gives the model a broad knowledge base. In order to better adapt to specific scenarios, data can also be fine-tuned according to actual needs to improve the performance of the model on specific tasks. MedicalGPT supports multi-round dialogues, can understand context, conduct continuous questions and answers, and provide users with a smooth conversation experience. This is very helpful for simulating communication between doctors and patients, or helping doctors quickly obtain medical records information.

[0035] The model training process is divided into the following four stages:

[0036] Phase 1: PT (Continue PreTraining) incremental pre-training, which pre-trains the GPT model twice on massive domain document data to inject domain knowledge.

[0037] The second stage: SFT (Supervised Fine-tuning) supervised fine-tuning, constructing an instruction fine-tuning dataset, and performing instruction fine-tuning based on the pre-trained model to align instruction intent.

[0038] The third stage: RM (Reward Model) reward model modeling, constructing a human preference ranking dataset, and training the reward model to align human preferences.

[0039] The fourth stage: RL (Reinforcement Learning) is based on reinforcement learning with human feedback (RLHF). The reward model is used to train the SFT model, and the generative model uses rewards or penalties to update its strategy in order to generate higher quality text that is more in line with human preferences.

[0040] In the first phase of incremental training, encyclopedia document datasets are used for incremental pre-training or secondary pre-training on domain datasets, hoping to inject domain knowledge into the model. In the present invention, it is hoped that incremental pre-training can enable the model to understand the current symptoms, causes, therapeutic drugs, treatment methods, drug efficacy, etc. of pregnant women, so that subsequent SFT supervised fine-tuning can activate these intrinsic knowledge.

[0041] In the second stage of supervised fine-tuning, supervised fine-tuning was performed using medical question-answering data. In order to further improve the overall performance of the model, enable it to fully understand human intentions and reduce unsafe outputs, human-feedback training (HFT) was performed based on the model fine-tuned by instructions. In training, we used a method based on human feedback reinforcement learning (RM, PPO) and combined with a variety of other means for joint training, including human feedback fine-tuning (HFFT), chain-of-hindsight fine-tuning (COHFT), AI feedback (AI Feedback) and rule-based reward system (RBRS), etc., to make up for the shortcomings of the PPO method and accelerate training. In the process of fine-tuning, it is first necessary to establish a special medical corpus. These Chinese data come from cleaned diagnostic reports. After deduplication, model scoring, data bucketing, rule filtering, sensitive topic filtering and data evaluation of the original data, we finally obtained valid data of about 5,000 pregnant patients. Then, you need to inject professional knowledge in the medical field. For example: "I am a 25-year-old pregnant woman who is five months pregnant. My current symptom is dizziness." "My blood pressure is usually around 135 / 85mmHg, but today it suddenly rose to 150 / 90mmHg." "I started to feel swollen 2 weeks ago."

[0042] In the third phase of reward model building, a reward model RM is trained with the human annotation set before entering the RL loop. The purpose of the reward model is to simulate human scoring of text. The best practice for building a reward model is to predict the order of results, that is, for each prompt (input text) corresponding to two results, the model predicts which score is higher. The RM model is trained by manually annotating the scoring results of the SFT model. The purpose is to replace manual scoring. It is essentially a regression model used to align human preferences, mainly the "HHH" principle, specifically "helpful, honest, harmless".

[0043] In the fourth stage of reinforcement learning based on human feedback, the purpose of the RL (Reinforcement Learning) model is to maximize the output of the reward model. Based on the above steps, we have a fine-tuned language model and reward model and can start executing the RL cycle. This process is roughly divided into three steps:

[0044] (1) Input prompt, and the model generates a response;

[0045] (2) using a reward model to score responses;

[0046] (3) Based on the scores, a round of policy optimization reinforcement learning (PPO) is performed.

[0047] After the question-answering model is trained, the trained model parameters are loaded. Pregnant women provide corresponding inquiries. After receiving prompts from the question-answering model, pregnant women need to upload the corresponding diagnosis report. The model needs to use the OCR detection model to extract the report text for professional analysis.

[0048] OCR text report recognition based on deep learning:

[0049] Various medical reports are transmitted to the cloud in the form of images, and the optical character recognition (OCR) detection model recognizes the report content one by one and converts it into text. It is a technology that converts printed or handwritten text into editable text. OCR technology plays an important role in the field of image processing and artificial intelligence. It can help convert text information in paper documents, photos or scans into digital data, thereby realizing automated processing and analysis.

[0050] The basic principle of OCR technology is to analyze images or scans, identify the characters and text information in them, and then convert them into computer-readable text format. Currently, OCR recognition based on deep learning includes the following four basic steps:

[0051] (1) Image acquisition and preprocessing: First, it is necessary to acquire images or scans containing text information. Then, the images are preprocessed, including removing noise, adjusting the brightness and contrast of the images, etc., to improve the subsequent recognition accuracy.

[0052] (2) Text localization and segmentation: The OCR system needs to be able to accurately locate the text area in the image and segment it into individual characters or words. This step usually involves image processing and computer vision techniques, such as edge detection and connected component analysis.

[0053] (3) Character Recognition: Once the text is located and segmented, the OCR system will recognize each character. This step usually uses machine learning algorithms, such as convolutional neural networks (CNN) or recurrent neural networks (RNN), to classify and recognize characters.

[0054] (4) Post-processing of results: After character recognition is completed, the recognition results may need to be post-processed, including correcting recognition errors, merging segmented characters, correcting recognized text, etc., to improve the final recognition accuracy.

[0055] Considering the complexity of various medical diagnosis reports, we used the most popular visual language modeling network (VisionLAN) as the basic recognition framework for the text recognition model. This framework treats visual and language information as a whole and directly gives the visual model language capabilities. In the training phase, text recognition based on character occlusion feature maps was introduced. When the visual clues are confused (occlusion, noise, etc.), the visual model uses the visual texture of the characters and the language information of the visual context for recognition. Since the language information is obtained together with the visual features and no additional language model is required, VisionLAN has higher recognition efficiency and is more in line with the actual needs of medical scenarios. VisionLAN is an end-to-end trainable framework consisting of three parts: the backbone network, the masked language perception module (MLM), and the visual reasoning module (VRM).

[0056] In the training phase, given an input image, the two-dimensional features V are first extracted from the backbone network. Then, MLM takes the extracted features V and character index P as input and generates position-aware character mask mapping Mask through weakly supervised complementary learning. c .Mask c It is used to occlude the character-by-character visual information in V to simulate the lack of character-by-character visual semantics. Afterwards, VRM converts the occluded feature map v m As input, predictions are made under full word-level supervision. During the test phase, MLM is removed and only VRM is used for prediction.

[0057] In terms of text, to mask character visual cues to guide language learning, VisionLAN has implemented a Masked Language Perception Module (MLM) to automatically generate mask maps containing only the original word-level annotations. MLM takes visual features V and character index p as input. The character index p∈[1,N w ] represents the index of the masked character, for each length N w The input word image is randomly obtained to obtain the index of the occluded character. Then the Transformer unit is used to improve the feature representation capability. Finally, the character mask mapping Mask is obtained by combining the character index information through the sigmoid layer. c , used to generate the occlusion feature map V m To guide Mask c In the learning process, two parallel branches are designed based on weakly supervised complementary learning (WCL). WCL aims to guide Mask c Cover more areas of the blocked characters, thereby complementing the 1-Mask cThe region containing more other characters. In the first branch, the visual feature v and Mask c The element-wise product between them is used to generate a feature map v containing the visual semantics of the occluded characters. mas In contrast, the visual features v and 1-Mask in the second branch c The element-wise product between them is used to generate a feature map v containing the visual semantics of other characters. rem Through such operations, the complementary learning process guides Mask c Only the characters at the corresponding position are covered without overlapping with other characters, which is more suitable for the actual situation where medical reports have dense text. We share the weights of the transformer unit and the prediction layer between the two parallel branches for feature representation enhancement and semantic guidance. in ∈R hw×c is the feature map, Att∈R hw×N is the attention map, where c = 512 is the number of channels, N = 25 is the maximum time step, and h and w are the height and width of the map respectively. c is the positional encoding of the character sequence [2]. w1, w2, w3 are trainable weights, and t is the time step.

[0058]

[0059] Att=Softmax(G(V in ))

[0060] G(V in )=W1tanh(W2O c +W3V in )

[0061] Compared with BERT, although both methods mask information within a certain time step, the MLM proposed in the present invention masks the visual features of the two-dimensional spatial space instead of covering token-level information. In addition, since the STR dataset is weakly labeled, it is difficult to obtain accurate character-by-character pixel-level annotations. Therefore, it is impractical to directly implement BERT-based methods in the STR task. On this basis, MLM helps the model learn language information from a new perspective, which is irreplaceable by existing masking methods. The supervision of WCL is automatically obtained using the original word-level annotations and randomly generated character indices. Therefore, MLM automatically generates accurate character mask mappings without the need for additional annotations, making practical applications possible.

[0062] In terms of vision, the Visual Reasoning Module (VRM) differs from previous methods that use a two-step structure to capture visual and language information, and models both information simultaneously with a unified structure. As a purely vision-based structure, the VRM aims to leverage character information in the visual context to infer word-level predictions from occluded features. It consists of two parts: a Visual Semantic Reasoning (VSR) layer and a Parallel Prediction (PP) layer. The VSR layer consists of N Transformer units, which have been shown to be effective in modeling long-term dependencies in recent computer vision tasks. In particular, positional encoding is used to perceive the position information of pixels. The proposed VRM uses Transformer units for sequence modeling, which is not affected by word length. Then, a PP layer is designed to predict characters in parallel.

[0063] In order to implement the language modeling process y i =f(yN,...,y i+1 ,y i-1 ,...,y1), the i-th character y i The reasoning process of MLM needs to rely entirely on the information of other characters. Since MLM accurately masks character information during the training phase, it guides the VSR layer to predict the dependencies between character visual features, thereby inferring the semantics of the occluded characters. Therefore, under word-level supervision, the VSR layer learns to actively model language information in the visual context to assist recognition. In the test phase, the VSR layer is able to adaptively consider language information for visual feature enhancement in the case of current visual semantic confusion (such as occlusion, noise, etc.).

[0064] When training the OCR recognition model VisionLAN, the training objectives are defined as follows:

[0065] L=L rec +λ1L mas +λ2L rem

[0066]

[0067] L rec is the loss in VRM, L mas and L rem are the losses of predicting masked characters and other characters in MLM, respectively. λ1 and λ2 are used to balance the two loss functions. t and g t represent the predicted value and the true value respectively.

[0068] Risk grading and early warning: input the critical value range of indicators and the criteria for determining the risk level of hypertension to establish an early warning model.

[0069] In the risk grading and warning end, the critical value range of indicators and the criteria for determining the risk level of hypertension are input, and the artificial neural network learning model is used for deep learning. The risk grading and warning are determined based on whether abnormal indicators are identified. If abnormal indicators are not identified, the risk is low, and if abnormal indicators are identified, the risk is medium-high risk. An early warning model is established. When the risk grading is low risk, close follow-up is performed; when the risk grading is medium-high-low risk, a warning message is issued and transmitted to the doctor management end.

[0070] The present invention uses Bayesian networks and deep neural networks to evaluate patient risks in medical diagnosis, wherein Bayesian networks are used to evaluate some simple symptoms, and deep neural networks are used to evaluate complex symptoms.

[0071] A Bayesian network is a probabilistic graphical model that can describe the probabilistic dependencies between variables. A Bayesian network represents the dependencies between variables through a directed acyclic graph, where nodes represent variables and edges represent the dependencies between variables. A Bayesian network uses Bayes’ theorem for reasoning and learning, and can update the probability distribution of unknown variables based on existing evidence, thereby performing risk assessment and decision-making.

[0072] By analyzing the patient's medical records and clinical test results, Bayesian networks can help doctors assess the patient's risk and provide support for medical decision-making. The following will introduce the specific process of Bayesian networks in the application of patient risk assessment.

[0073] Identify variables and factors: First, it is necessary to identify variables and factors relevant to risk assessment. This may include patient indicators (such as age, gender, weight, etc.), disease risk factors (such as family history, smoking history, etc.), and other relevant factors.

[0074] Constructing a Bayesian network: Based on the determined variables and factors, a Bayesian network model is constructed. A Bayesian network consists of nodes and edges, where nodes represent variables and edges represent dependencies between variables. The probability distribution of nodes can be estimated based on domain knowledge or data.

[0075] Collect data: In order to build and train the Bayesian network model, relevant data needs to be collected. This data can include patient indicator data, disease occurrence, etc. The quality and quantity of the data are very important for the accuracy and reliability of the model.

[0076] Parameter estimation: Use the collected data to estimate the parameters in the Bayesian network model. Parameter estimation methods can include maximum likelihood estimation, Bayesian estimation, etc.

[0077] Inference and risk assessment: Once the Bayesian network model is built and its parameters estimated, it can be used for inference and risk assessment. Given a patient’s indicator value, the probability that the patient faces a specific risk can be calculated.

[0078] By collecting a large amount of medical record data and using Bayesian networks for parameter learning, the system can predict the risks of pregnant women based on the input variables. For example, for a pregnant patient over 40 years old, if her body mass index is high and her blood sugar level is elevated, then she may have a higher probability of diabetes and need regular checkups.

[0079] For scenarios with complex indicators, this paper conducts risk assessment based on deep learning, taking into account the contextual associations between different indicators:

[0080] On the basis of the LSTM network model, its external structure is further improved to obtain enhanced sequence modeling capabilities and provide better explanations. The bidirectional structure allows the model to more fully examine contextual information, which is of great significance in risk prediction. On the one hand, the forward and reverse LSTMs may obtain better and more comprehensive information and discover more potential patterns in the sequence. On the other hand, more accurate predictions are also of great significance. The attention model is of equal significance. It can determine which part of the input needs to be focused on. The patient's future treatment needs to focus on historical treatment fragments that are highly relevant to the current condition. At the same time, when the model accurately predicts and builds a connection with all historical information, the relationship between all visits at different times can be determined.

[0081] First, various indicator data are processed into two sets of temporal data and one set of static data, which are fed into bidirectional heterogeneous LSTMs. Heterogeneous LSTMs can effectively process multimodal data, thereby learning the actual hidden state of the patient at each moment. The attention mechanism can learn the current state at each moment, thereby selecting state information that is useful for treatment. The essence of the model is a multi-label classifier, so the multimodal sequence of each medical record is used to train the multi-label classification model end-to-end, and the risk level of the current patient is predicted by learning the knowledge hidden in large-scale data.

[0082] Given the patient's indicator information, it is input into a bidirectional heterogeneous LSTM, which outputs the forward real hidden state and the backward real hidden state as the patient representation. A set of real hidden states can calculate the relative importance of each treatment stage. Subsequently, the forward and backward context states are calculated based on the relative importance and hidden states. This process is the attention model. Finally, the predicted risk score is obtained through the fully connected layer.

[0083] The forward LSTM predicts the future risk factor based on the multimodal time series information of the previous moment. The reverse LSTM not only needs to make predictions based on the information of the past time, but also can predict the risk factor based on the multimodal time series information of the future moment.

[0084] Doctor management side: Doctors process warning information and provide feedback to pregnant women.

[0085] On the doctor management side, the doctor processes the warning information and feeds it back to the pregnant woman. Intelligent monitoring reminds pregnant women to take medication (to control blood pressure, blood sugar and blood lipids), and pushes healthy exercises, diet recipes, popular science and music, so as to relieve the stress of pregnant mothers.

[0086] Integrated prenatal and postnatal recording terminal: input the follow-up results of pregnant women, and intelligently filter and classify the data of healthy and unhealthy pregnant women.

[0087] The follow-up results of pregnant women are recorded, such as pregnancy and delivery outcomes, incidence of gestational hypertension, incidence of preeclampsia, incidence of severe preeclampsia, incidence of eclampsia, time from onset of gestational hypertension to hospitalization for delivery, time from onset of preeclampsia to hospitalization for delivery, incidence of pregnancy complications, and incidence of postpartum hemorrhage, and data of healthy and unhealthy pregnant women are intelligently screened and classified.

[0088] AI intelligent analysis end: Statistical analysis of the differences between healthy and unhealthy pregnant women, and use big data to verify the effectiveness of early management intervention. Pregnant women with complete follow-up results and test results included in the system are included and randomly divided into training and validation sets at a ratio of 7:3.

[0089] The present invention uses Bayesian networks and deep neural networks to evaluate the risks of pregnant women with hypertension / preeclampsia / eclampsia / complications, etc. Among them, Bayesian networks are used to evaluate some simple symptoms, and deep neural networks are used to evaluate complex symptoms.

[0090] A Bayesian network is a probabilistic graphical model that can describe the probabilistic dependencies between variables. A Bayesian network represents the dependencies between variables through a directed acyclic graph, where nodes represent variables and edges represent the dependencies between variables. A Bayesian network uses Bayes’ theorem for reasoning and learning, and can update the probability distribution of unknown variables based on existing evidence, thereby performing risk assessment and decision-making.

[0091] By analyzing the patient's medical records and clinical test results, Bayesian networks can help doctors assess the patient's risk and provide support for medical decision-making. The following will introduce the specific process of Bayesian networks in the application of patient risk assessment.

[0092] Identify variables and factors: First, it is necessary to identify variables and factors relevant to risk assessment. This may include patient indicators (such as age, gender, weight, etc.), disease risk factors (such as family history, smoking history, etc.), and other relevant factors.

[0093] Constructing a Bayesian network: Based on the determined variables and factors, a Bayesian network model is constructed. A Bayesian network consists of nodes and edges, where nodes represent variables and edges represent dependencies between variables. The probability distribution of nodes can be estimated based on domain knowledge or data.

[0094] Collect data: In order to build and train the Bayesian network model, relevant data needs to be collected. This data can include patient indicator data, disease occurrence, etc. The quality and quantity of the data are very important for the accuracy and reliability of the model.

[0095] Parameter estimation: Use the collected data to estimate the parameters in the Bayesian network model. Parameter estimation methods can include maximum likelihood estimation, Bayesian estimation, etc.

[0096] Inference and risk assessment: Once the Bayesian network model is built and its parameters estimated, it can be used for inference and risk assessment. Given a patient’s indicator value, the probability that the patient faces a specific risk can be calculated.

[0097] By collecting a large amount of medical record data and using Bayesian networks for parameter learning, the system can predict the risks of pregnant women based on the input variables. For example, for a pregnant patient over 40 years old, if her body mass index is high and her blood sugar level is elevated, then she may have a higher probability of diabetes and need regular checkups, etc.

[0098] The use process of the present invention is as follows, comprising the following steps:

[0099] (1) Obtain the health data of pregnant women through intelligent inquiries using a Chatgpt-like model, including test reports, medical history, basic health information, psychological stress survey results, physiological stress assessment results, and discomfort symptoms, and intelligently process the acquired health data to create a database.

[0100] Among them, the test reports include: blood routine, urine routine, BNP, cTnI, blood sugar and B-ultrasound test reports. Past medical history includes history of hypertension during pregnancy, history of chronic hypertension, history of preeclampsia, reproductive history and history of diabetes. Basic health includes: age during pregnancy, height, weight and blood pressure. Discomfort symptoms include: dizziness, headache, lower limb edema and blurred vision.

[0101] (2) The critical value range of indicators and the criteria for determining the risk level of hypertension are input through the risk grading and early warning terminal, and the artificial neural network learning model is used for deep learning. The risk grading and early warning are determined based on whether abnormal indicators are identified. If abnormal indicators are not identified, the risk is low, and if abnormal indicators are identified, the risk is medium or high. When the risk grading is low, close follow-up is performed; when the risk grading is medium or high, an early warning message is issued and transmitted to the doctor management terminal;

[0102] (3) On the doctor management side, the doctor processes the warning information and feeds it back to the pregnant woman. Intelligent monitoring reminds the pregnant woman to take medication (to control blood pressure, blood sugar and blood lipids), and pushes healthy exercises, diet recipes, popular science and music, thereby achieving the effect of relieving the pregnant mother's stress.

[0103] (4) Record the follow-up results of pregnant women, such as pregnancy and delivery outcomes, incidence of gestational hypertension, incidence of preeclampsia, incidence of severe preeclampsia, incidence of eclampsia, time from onset of gestational hypertension to hospitalization for delivery, time from onset of preeclampsia to hospitalization for delivery, incidence of pregnancy complications, and incidence of postpartum hemorrhage, and intelligently filter and classify the data of healthy and unhealthy pregnant women.

[0104] (5) Through AI intelligent analysis, statistical analysis is performed on the differences between healthy and unhealthy pregnant women, and big data is used to verify the effectiveness of early management interventions.

[0105] 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 aforementioned embodiments, a person skilled in the art should understand that it is still possible to modify the technical solutions described in the aforementioned embodiments, or to perform equivalent replacements on some of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A system for the prevention and early management of preeclampsia, characterized in that: include: Pregnant women: enter the health data of pregnant women and create a database; Risk grading and early warning: input the critical value range of indicators and the criteria for determining the risk level of hypertension to establish an early warning model; Doctor management end: Doctors process warning information and feedback the processing results to pregnant women; Integrated prenatal and postnatal recording terminal: input the follow-up results of pregnant women, and intelligently screen and classify the data of healthy and unhealthy pregnant women; AI intelligent analysis terminal: Statistical analysis of the differences between healthy and unhealthy pregnant women, and use of big data to verify the effectiveness of early management interventions.

2. The system for the prevention and early management of preeclampsia according to claim 1, characterized in that: On the pregnant woman's side, a Chatgpt-like model is used to intelligently inquire about the pregnant woman's health data, including entering test reports, previous medical history, basic health information, psychological stress survey results, physiological stress assessment results and discomfort symptoms, and intelligently collect information on the entered health data to create a database.

3. The system for the prevention and early management of preeclampsia according to claim 2, characterized in that: The test report includes: blood routine, urine routine, BNP, cTnI, blood sugar and B-ultrasound test reports; the past medical history includes the history of hypertension during pregnancy, chronic hypertension, preeclampsia, reproductive history and diabetes history; the basic health includes: age, height, weight and blood pressure during pregnancy; the discomfort symptoms include: dizziness, headache, lower limb edema and blurred vision.

4. The system for the prevention and early management of preeclampsia according to claim 1, characterized in that: In the risk grading and warning end, the critical value range of indicators and the criteria for determining the risk level of hypertension are input, and the artificial neural network learning model is used for deep learning. The risk grading and warning are determined based on whether abnormal indicators are identified. If abnormal indicators are not identified, the risk is low, and if abnormal indicators are identified, the risk is medium or high. An early warning model is established. When the risk grading is low risk, close follow-up is performed; when the risk grading is medium or high risk, a warning message is issued and transmitted to the doctor management end.

5. The system for the prevention and early management of preeclampsia according to claim 1, characterized in that: On the doctor management side, the doctor processes the warning information and feeds it back to the pregnant woman. Intelligent monitoring reminds pregnant women to take medication and pushes healthy exercises, diet recipes, popular science and music.

6. A method for the prevention and early management of preeclampsia, characterized in that: include: Obtain the health data of pregnant women through the pregnant women’s terminal and create a database; Through the risk grading and early warning input indicators critical value range and hypertension risk level determination criteria, the artificial neural network learning model is used for deep learning, and the risk grading and early warning are determined based on whether abnormal indicators are identified. If abnormal indicators are not identified, it is low risk, and if abnormal indicators are identified, it is medium-high risk. When the risk is classified as low risk, follow up closely; When the risk is classified as medium or high risk, an early warning message is sent to the doctor management end; The doctor management end processes the warning information and feeds back the processing results to the pregnant woman end, which can intelligently monitor and remind the pregnant woman to take medicine, and push healthy exercise, diet recipes, popular science and music; The integrated prenatal and postnatal recording terminal records the follow-up results of pregnant women, and intelligently screens and classifies the data of healthy and unhealthy pregnant women; Through AI intelligent analysis, statistical analysis is performed on the differences between healthy and unhealthy pregnant women, and big data is used to verify the effectiveness of early management interventions.

7. The method for the prevention and early management of preeclampsia according to claim 6, characterized in that: The health data of pregnant women are obtained through intelligent inquiries using a Chatgpt-like model, including test reports, past medical history, basic health information, psychological stress survey results, physiological stress assessment results, and discomfort symptoms. The obtained health data is then intelligently collected to create a database.

8. The method for the prevention and early management of preeclampsia according to claim 7, characterized in that: The test report includes: blood routine, urine routine, BNP, cTnI, blood sugar and B-ultrasound test reports; the past medical history includes the history of hypertension during pregnancy, chronic hypertension, preeclampsia, reproductive history and diabetes history; the basic health includes: age, height, weight and blood pressure during pregnancy; the discomfort symptoms include: dizziness, headache, lower limb edema and blurred vision.

Images