Dynamic critical disease treatment path construction method and system based on historical medical record analysis

Abstract

The invention relates to the technical field of dynamic critical disease treatment path construction, in particular to a dynamic critical disease treatment path construction method and system based on historical medical record analysis, and the method comprises the steps: obtaining historical medical record data and real-time clinical data of a patient; analyzing the historical medical record data of the patient based on a natural language processing algorithm, and performing personalized risk assessment on the illness state of the patient in combination with a machine learning model; and correspondingly selecting a critical disease first-aid guidance model based on the patient personalized risk assessment result to generate an initial treatment path, and dynamically adjusting the initial treatment path according to the disease change in the real-time clinical data to obtain a final treatment path. According to the method, through multi-modal data fusion and dynamic path optimization, the coordination of personalized risk assessment and real-time treatment decision is realized, and the critical disease treatment efficiency and the clinical resource utilization rate are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of dynamic critical care treatment pathway construction, and in particular to a method and system for constructing a dynamic critical care treatment pathway based on historical medical record analysis. Background Art

[0002] The treatment process of critically ill patients involves multiple links and complex decisions, requiring the medical team to quickly assess the patient's condition and develop a suitable treatment plan within a limited time. With the development of medical informatization, a large amount of clinical data is systematically recorded and stored. At present, the existing critical care treatment plans often rely on standardized clinical guidelines, which makes it difficult to fully consider the individual differences of patients and the characteristics of disease evolution. In practical applications, doctors need to spend a lot of time reading and understanding patients' historical medical records, and manual summary and analysis are inefficient and easy to miss key information. At the same time, traditional treatment plans are relatively fixed once they are formulated, and it is difficult to make timely adjustments according to the dynamic changes in the patient's condition. This lag may affect the treatment effect. In addition, due to the lack of systematic analysis and utilization of historical data, it is difficult to extract valuable experience from past cases to guide current treatment decisions. Summary of the invention

[0003] The purpose of the present invention is to provide a method and system for constructing a dynamic critical care treatment pathway based on historical medical record analysis, which realizes the coordination of personalized risk assessment and real-time treatment decision-making through multimodal data fusion and dynamic pathway optimization, and significantly improves the efficiency of critical care and clinical resource utilization.

[0004] The present invention is achieved through the following technical solutions: A method for constructing a dynamic critical care treatment pathway based on historical medical record analysis, the method comprising the following steps: Obtain patient historical medical records and real-time clinical data; Analyze the patient's historical medical records based on natural language processing algorithms, and use machine learning models to conduct personalized risk assessments of the patient's condition; Based on the patient's personalized risk assessment results, the critical care emergency guidance model is selected to generate an initial treatment pathway. At the same time, the initial treatment pathway is dynamically adjusted according to changes in the condition in real-time clinical data to obtain the final treatment pathway.

[0005] Optionally, after obtaining the patient's historical medical record data and real-time clinical data, the patient's historical medical record data and real-time clinical data are cleansed and standardized in turn.

[0006] Optionally, the patient's historical medical record data is parsed based on a natural language processing algorithm, which is specifically: Identify medical entities in patients’ historical medical records through entity recognition models, and annotate the medical entity recognition results and their boundaries; The entity relationship extraction model is used to extract the relationship of medical entity recognition results, and the relationship type and corresponding probability score of each entity pair are obtained. The relationship type of each entity pair includes: First relationship: determine whether the patient's main symptoms are caused by a specific disease or cause; Second relationship: Identify allergic drugs and their corresponding symptoms in the patient's historical medical record data; The third relationship: identifying the association between surgical methods and complications in the patient's historical medical record data; The fourth relationship: distinguish whether the patient is in the early or late stages of the disease, mild or severe stages; The preset a priori medical rule library is called to verify the consistency of the relationship type and corresponding probability score of each entity pair, and the medical record parsing structure is output.

[0007] Optionally, the method of combining a machine learning model to perform a personalized risk assessment of the patient's condition is specifically as follows: The medical record parsing structure is associated and fused with real-time clinical data, and the fused data is vectorized to form a comprehensive feature set, which is then divided into a training set and a test set; Taking the probability of complication occurrence as the prediction target, a personalized risk assessment model for patients is constructed, and the training set is calculated through the personalized risk assessment model for patients. During the training process, the personalized risk assessment model for patients is optimized through cross-validation based on the test set as a supervision signal until the training of the personalized risk assessment model for patients is completed; The real-time clinical data is evaluated by the trained patient personalized risk assessment model to obtain the patient's complication probability, and the patient's personalized risk assessment result is formed by combining the previous medical record analysis structure.

[0008] Optionally, the patient personalized risk assessment model specifically takes balancing prediction accuracy and potential clinical costs as an optimization goal, which is specifically: Based on the comprehensive feature set, a training set including feature vectors, true complication labels and potential clinical costs is constructed; Each training sample is input into the patient-specific risk assessment model for forward calculation to obtain the probability of complication occurrence; Solve the correctness score and cost penalty for each training sample: The parameters of the patient personalized risk assessment model are iteratively updated based on gradient descent. In each iteration, the patient personalized risk assessment model improves the accuracy of training samples according to the accumulated value of the correctness score, and limits the error frequency of training samples according to the cost penalty to balance the accuracy of the patient personalized risk assessment model and clinical cost control.

[0009] Optionally, the correctness score: maps the true complication label from a binary form to positive or negative to distinguish the direction of samples with complications and samples without complications in the objective function, and measures the accuracy of the match between the complication probability and the true complication label by the positive or negative difference between the complication probability output by the patient's personalized risk assessment model and the neutral threshold, and amplifies the degree of match between the complication probability and the true complication label in exponential form to obtain the correctness score of each training sample.

[0010] Optionally, the cost penalty is: for training samples that incorrectly predict the probability of complication occurrence, their potential clinical cost value is introduced as the error cost, and combined with the set cost amplification parameter and amplified exponentially to obtain the cost penalty for each training sample.

[0011] Optionally, the medical record parsing structure is associated and integrated with the real-time clinical data, which is specifically: The relationship types in the medical record parsing structure are standardized and vectorized, and the associated features of the medical record parsing structure are extracted based on the attention mechanism to represent them as text vectors; Perform consistency check on real-time clinical data and obtain numerical vectors in real-time clinical data through sliding windows; The text vector and the numerical vector are fused in the feature dimension and the dimension is reduced by an autoencoder to generate a preliminary feature vector of multimodal fusion. After outlier detection and correction are performed on the preliminary feature vector of multimodal fusion, the output is the comprehensive feature set.

[0012] Optionally, the critical care first aid guidance model is selected based on the patient's personalized risk assessment result to generate an initial treatment pathway, which is specifically: Obtain the probability of complications of patients, and classify patients into corresponding risk levels based on the probability of complications of patients and pre-set risk classification rules; Based on the risk level, selecting a treatment plan template matching the risk level from a preset critical care first aid guidance model library; The selected critical care emergency guidance model applies the patient's previous medical record analysis structure and the probability of patient complications, and customizes the treatment pathway nodes in the critical care emergency guidance model to generate an initial treatment pathway.

[0013] A dynamic critical care treatment pathway construction system based on historical medical record analysis, including: Data acquisition module, used to obtain patients' historical medical records and real-time clinical data; Personalized risk assessment module, which is used to analyze the patient's historical medical record data based on natural language processing algorithms and conduct personalized risk assessment of the patient's condition in combination with machine learning models; The treatment pathway processing module is used to select the critical care emergency guidance model based on the patient's personalized risk assessment results to generate an initial treatment pathway. At the same time, the initial treatment pathway is dynamically adjusted according to changes in the condition in real-time clinical data to obtain the final treatment pathway.

[0014] The technical solution of the present invention has at least the following advantages and beneficial effects: By integrating natural language processing and machine learning models, the present invention not only realizes the multi-dimensional integration of historical medical records and real-time clinical data, but also constructs a closed-loop optimization mechanism for personalized risk assessment and dynamic treatment pathways. On the one hand, the automated medical record parsing based on NLP combined with real-time data feature extraction significantly improves the accuracy and timeliness of risk assessment; on the other hand, through the intelligent matching and dynamic adjustment of the critical care first aid guidance model, the treatment pathway can be optimized in real time as the disease evolves. In addition, the system embeds clinical cost factors into the decision-making process to reduce resource waste while ensuring treatment effectiveness. The present invention not only shortens the decision-making response time of critically ill patients, but also improves the adaptability of pathway optimization through a continuous learning mechanism, providing clinicians with intelligent decision-making support that takes into account both efficiency and safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic diagram of a process for constructing a dynamic critical care treatment pathway based on historical medical record analysis provided by the present invention; Figure 2 A logical schematic diagram of the association fusion provided by the present invention; Figure 3 This is a schematic diagram of the construction of a dynamic critical care treatment pathway based on historical medical record analysis provided by the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be described clearly and completely below in conjunction with the drawings in the present invention. Obviously, what is described is only a part of the present invention, not all of it. Generally, the components of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0017] like Figure 1As shown, the present invention provides one embodiment: a method for constructing a dynamic critical care treatment pathway based on historical medical record analysis, the method comprising the steps of: Obtain patient historical medical records and real-time clinical data; Analyze the patient's historical medical records based on natural language processing algorithms, and use machine learning models to conduct personalized risk assessments of the patient's condition; Based on the patient's personalized risk assessment results, the critical care emergency guidance model is selected to generate an initial treatment pathway. At the same time, the initial treatment pathway is dynamically adjusted according to changes in the condition in real-time clinical data to obtain the final treatment pathway.

[0018] In the specific implementation of this embodiment, this embodiment is aimed at the diagnosis and treatment process of critically ill patients. In order to realize the automatic acquisition of patient historical medical records and real-time clinical data, natural language processing and machine learning technology are used to realize personalized risk assessment, and the appropriate critical emergency guidance model is selected based on the risk assessment results to generate an initial treatment path. At the same time, the treatment plan is dynamically adjusted in combination with real-time data, and the optimal treatment path is finally output to provide intelligent support for clinical decision-making. This embodiment obtains patient historical medical record data and real-time clinical data from the hospital information management system, electronic medical record system and real-time monitoring platform through an interface. Historical data includes text-based medical record descriptions, diagnostic information and previous treatment records; real-time data mainly covers physiological indicators, images and monitoring parameters. All data are cleaned, standardized and structured by the preprocessing module. Based on the natural language processing (NLP) algorithm, the structured and unstructured historical medical record data are semantically parsed to extract key information and feature vectors. The extracted feature data is then fused with real-time clinical data to form a multi-dimensional input. In the personalized risk assessment stage, this embodiment uses a multi-layer perceptron model to analyze the condition data of each patient and output the predicted probability of complications such as critical illness. This risk assessment result not only reflects the patient's historical condition characteristics, but also takes into account real-time changes in the condition, achieving more accurate risk prediction. After obtaining the personalized risk assessment results, the appropriate critical care emergency guidance model is selected based on the risk classification and the corresponding clinical cost, the patient's current risk is mapped, and the initial treatment path is generated. Considering that the patient's condition is in a dynamic change process, this embodiment again uses the real-time clinical data obtained continuously to monitor and analyze the initial treatment path in real time. Using the objective function for the multi-layer perceptron model, the actual effect of the current treatment path in risk control and clinical resource utilization is judged. If a change in the condition or a mismatch in the path is detected, the preset dynamic adjustment strategy is automatically adopted to optimize the treatment path, and the updated final treatment path is output again to form a closed-loop feedback mechanism to ensure that the treatment strategy is always adapted to the patient's actual condition, which not only improves the accuracy of clinical decision-making, but also reduces the waste of resources and risk losses caused by misjudgment, and significantly optimizes the medical treatment process.

[0019] In the application of this embodiment, the patient's historical medical record data is parsed based on the natural language processing algorithm, which is specifically: Identify medical entities in patients’ historical medical records through entity recognition models, and annotate the medical entity recognition results and their boundaries; The entity relationship extraction model is used to extract the relationship of medical entity recognition results, and the relationship type and corresponding probability score of each entity pair are obtained. The relationship type of each entity pair includes: First relationship: determine whether the patient's main symptoms are caused by a specific disease or cause; Second relationship: Identify allergic drugs and their corresponding symptoms in the patient's historical medical record data; The third relationship: identifying the association between surgical methods and complications in the patient's historical medical record data; The fourth relationship: distinguish whether the patient is in the early or late stages of the disease, mild or severe stages; The preset a priori medical rule library is called to verify the consistency of the relationship type and corresponding probability score of each entity pair, and the medical record parsing structure is output.

[0020] In this embodiment, the entity recognition model is based on a deep learning algorithm and is used to analyze the patient's historical medical record text data. It can automatically scan the medical terms in the text, identify the text entities, including disease names, symptoms, drug names, surgical methods, and allergy records, and mark their boundaries in the original text. After completing the entity recognition, this embodiment calls the entity relationship extraction model, which is also based on a deep neural network to identify whether there is a relationship between entity pairs, and can also output a probability score for each relationship established in this embodiment. The results are verified for consistency with the preset prior medical rule library to ensure that the relationship type obtained is consistent with clinical professional knowledge, thereby constructing a reliable medical record parsing structure. It can be understood that in this embodiment, regarding the judgment of whether the patient's main symptoms are caused by a specific disease or cause, the specific disease or cause is understood as: if the patient's main symptoms are angina pectoris, myocardial infarction, etc., then the specific disease or cause is a cardiovascular disease; if the patient's main symptoms are sepsis, influenza, etc., then the specific disease or cause is an infectious disease.

[0021] In the specific implementation of this embodiment, the personalized risk assessment of the patient's condition is performed in combination with the machine learning model, which is specifically: The medical record parsing structure is associated and fused with real-time clinical data, and the fused data is vectorized to form a comprehensive feature set, which is then divided into a training set and a test set; Taking the probability of complication occurrence as the prediction target, a personalized risk assessment model for patients is constructed, and the training set is calculated through the personalized risk assessment model for patients. During the training process, the personalized risk assessment model for patients is optimized through cross-validation based on the test set as a supervision signal until the training of the personalized risk assessment model for patients is completed; The real-time clinical data is evaluated by the trained patient personalized risk assessment model to obtain the patient's complication probability, and the patient's personalized risk assessment result is formed by combining the previous medical record analysis structure.

[0022] Specifically, this embodiment associates and fuses the medical record parsing structure with the real-time clinical data through the data docking interface, uniformly vectorizes the two types of data, and generates a comprehensive feature set. The comprehensive feature set covers the patient's previous condition, current state, and semantic information implicitly extracted from multi-source data. This embodiment takes the probability of complication occurrence as the prediction target to construct a personalized risk assessment model for patients. The personalized risk assessment model for patients is specifically built based on a multi-layer perceptron to model and analyze the comprehensive feature set. During the training process, the training set is input into the personalized risk assessment model for patients, and the correctness score and cost penalty of each sample are calculated through the objective function, and then the parameters of the personalized risk assessment model for patients are gradient updated. At the same time, the test set is used as a supervisory signal, and the personalized risk assessment model for patients is tuned by cross-validation technology until the evaluation index meets the preset requirements, and the training of the personalized risk assessment model for patients is completed. After the model training is completed, the trained personalized risk assessment model for patients is used to evaluate the acquired real-time clinical data. The probability of complication occurrence for the corresponding patient is output, and the risk assessment result is integrated with the historical information previously obtained through the medical record parsing structure to form the final patient personalized risk assessment result, which not only reflects the patient's current clinical status, but also comprehensively considers the evolution of his or her previous condition, providing comprehensive and timely data support for clinical physicians to formulate targeted and dynamically adjusted treatment plans.

[0023] Furthermore, the patient personalized risk assessment model specifically takes balancing prediction accuracy and potential clinical costs as the optimization goal, which is specifically: Based on the comprehensive feature set, a training set including feature vectors, true complication labels and potential clinical costs is constructed; Each training sample is input into the patient-specific risk assessment model for forward calculation to obtain the probability of complication occurrence; Solve the correctness score and cost penalty for each training sample: The parameters of the patient personalized risk assessment model are iteratively updated based on gradient descent. In each iteration, the patient personalized risk assessment model improves the accuracy of training samples according to the accumulated value of the correctness score, and limits the error frequency of training samples according to the cost penalty to balance the accuracy of the patient personalized risk assessment model and clinical cost control.

[0024] The calculation formula of the objective function is:

[0025] in, is the objective function, N is the total number of training samples, i is the sample index, are the hyperparameters, A hyperparameter that controls the discrimination of the exponential part of the correctness score. The larger its value, the stronger the discrimination between the prediction and the true label matching results. Adjust the hyperparameters of the cost amplification degree in the cost penalty to determine the weight amplification effect when the combined clinical cost enters the objective function. The hyperparameter that balances the weight between prediction accuracy and clinical cost in the overall objective function, the larger the represents the amplification of the impact of cost penalty, is the true complication label of the i-th sample, 1 indicates a complication occurs, 0 indicates no complication occurs, is the output probability of the patient personalized risk assessment model for sample i, is the error cost measurement function, which is defined as , is the potential clinical cost, used to measure the waste of medical resources caused by misdiagnosis. Based on the threshold of 0.5 Discrete prediction labels obtained by binary classification.

[0026] Specifically, the correctness score is as follows: the true complication label is mapped from a binary form to positive or negative to distinguish the direction of samples with complications and samples without complications in the objective function, and the positive or negative difference between the complication probability output by the patient's personalized risk assessment model and the neutral threshold is used to measure the accuracy of the match between the complication probability and the true complication label, and the matching degree of the complication probability to the true complication label is amplified in exponential form to obtain the correctness score of each training sample.

[0027] It can be understood that the neutral threshold is used to determine whether the probability of complication output by the patient's personalized risk assessment model is high enough to determine the boundary value for classifying the sample as "complication occurred" or "no complication occurred". In this embodiment, the neutral threshold is set to 0.5: when the probability of complication occurrence ≥ 0.5, the sample is considered to be "complication occurred". When the probability of complication occurrence < 0.5, the sample is considered to be "no complication occurred".

[0028] The cost penalty is as follows: for training samples that incorrectly predict the probability of complication occurrence, their potential clinical cost value is introduced as the error cost, and combined with the set cost amplification parameter and amplified exponentially to obtain the cost penalty for each training sample.

[0029] like Figure 2 As shown, in a further application of this embodiment, the medical record parsing structure is associated and integrated with the real-time clinical data, which is specifically: The relationship types in the medical record parsing structure are standardized and vectorized, and the associated features of the medical record parsing structure are extracted based on the attention mechanism to represent them as text vectors; Perform consistency check on real-time clinical data and obtain numerical vectors in real-time clinical data through sliding windows; The text vector and the numerical vector are fused in the feature dimension and the dimension is reduced by an autoencoder to generate a preliminary feature vector of multimodal fusion. After outlier detection and correction are performed on the preliminary feature vector of multimodal fusion, the output is the comprehensive feature set.

[0030] In a specific implementation, the critical care first aid guidance model is selected based on the patient's personalized risk assessment results to generate an initial treatment path, which is specifically: Obtain the probability of complications of patients, and classify patients into corresponding risk levels based on the probability of complications of patients and pre-set risk classification rules; Based on the risk level, selecting a treatment plan template matching the risk level from a preset critical care first aid guidance model library; The selected critical care emergency guidance model applies the patient's previous medical record analysis structure and the probability of patient complications, and customizes the treatment pathway nodes in the critical care emergency guidance model to generate an initial treatment pathway.

[0031] This embodiment uses a trained personalized risk assessment model to evaluate the newly collected real-time clinical data of patients and output the probability of complications for each patient. Based on the pre-set risk grading rules, patients are divided into different risk levels. The risk grading rules can be set according to the predicted probability range or the standards set by clinical experts. For example, patients with a probability of complications below a specific threshold are classified as low risk, and patients above a specified threshold are classified as high risk, or subdivided into low, medium, and high levels to meet the requirements of risk response in different clinical scenarios. According to the risk grading results, the system automatically matches the treatment plan template corresponding to the patient's risk level from the preset critical care first aid guidance model library. The guidance model library contains various templates constructed in accordance with clinical guidelines and historical treatment experience. Each template is designed for the corresponding risk level, covering key links such as emergency medication, intervention operations and life support, ensuring that corresponding and standardized medical response strategies can be provided at different risk levels. After selecting the first aid guidance model template that matches the patient's risk level, the system uses the patient's previous medical record analysis structure information and the probability of complications output by the model as customized inputs to personalize the various treatment path nodes in the template. This adjustment process fine-tunes the various intervention measures in the template by carefully matching the patient's historical condition, clinical data, and actual pathological characteristics, ensuring that the initial treatment pathway not only meets the standard requirements of the emergency guidelines, but also fully reflects the individual differences in the patient's condition. The resulting initial treatment pathway can provide clinicians with customized treatment strategies, allowing the decision-making process to take into account timeliness, accuracy, and clinical operability.

[0032] like Figure 3 As shown, the present invention provides another embodiment: a dynamic critical care treatment pathway construction system based on historical medical record analysis, comprising: Data acquisition module, used to obtain patients' historical medical records and real-time clinical data; Personalized risk assessment module, which is used to analyze the patient's historical medical record data based on natural language processing algorithms and conduct personalized risk assessment of the patient's condition in combination with machine learning models; The treatment pathway processing module is used to select the critical care emergency guidance model based on the patient's personalized risk assessment results to generate an initial treatment pathway. At the same time, the initial treatment pathway is dynamically adjusted according to changes in the condition in real-time clinical data to obtain the final treatment pathway.

[0033] The above is only the preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for constructing a dynamic critical care treatment pathway based on historical medical record analysis, characterized in that: The steps of the method include: Obtain patient historical medical records and real-time clinical data; Analyze the patient's historical medical records based on natural language processing algorithms, and use machine learning models to conduct personalized risk assessments of the patient's condition; Based on the patient's personalized risk assessment results, the critical care emergency guidance model is selected to generate an initial treatment pathway. At the same time, the initial treatment pathway is dynamically adjusted according to changes in the condition in real-time clinical data to obtain the final treatment pathway.

2. The method for constructing a dynamic critical care treatment pathway based on historical medical record analysis according to claim 1, characterized in that: After obtaining the patient's historical medical record data and real-time clinical data, the patient's historical medical record data and real-time clinical data are cleaned and standardized in turn.

3. The method for constructing a dynamic critical care treatment pathway based on historical medical record analysis according to claim 1, characterized in that: The analysis of the patient's historical medical record data based on the natural language processing algorithm is specifically as follows: Identify medical entities in patients’ historical medical records through entity recognition models, and annotate the medical entity recognition results and their boundaries; The entity relationship extraction model is used to extract the relationship of medical entity recognition results, and the relationship type and corresponding probability score of each entity pair are obtained. The relationship type of each entity pair includes: First relationship: determine whether the patient's main symptoms are caused by a specific disease or cause; Second relationship: Identify allergic drugs and their corresponding symptoms in the patient's historical medical record data; The third relationship: identifying the association between surgical methods and complications in the patient's historical medical record data; The fourth relationship: distinguish whether the patient is in the early or late stages of the disease, mild or severe stages; The preset a priori medical rule library is called to verify the consistency of the relationship type and corresponding probability score of each entity pair, and the medical record parsing structure is output.

4. The method for constructing a dynamic critical care treatment pathway based on historical medical record analysis according to claim 3, characterized in that: The personalized risk assessment of the patient's condition is performed by combining the machine learning model, which is specifically: The medical record parsing structure is associated and fused with real-time clinical data, and the fused data is vectorized to form a comprehensive feature set, which is then divided into a training set and a test set; Taking the probability of complication occurrence as the prediction target, a personalized risk assessment model for patients is constructed, and the training set is calculated through the personalized risk assessment model for patients. During the training process, the personalized risk assessment model for patients is optimized through cross-validation based on the test set as a supervision signal until the training of the personalized risk assessment model for patients is completed; The real-time clinical data is evaluated by the trained patient personalized risk assessment model to obtain the patient's complication probability, and the patient's personalized risk assessment result is formed by combining the previous medical record analysis structure.

5. The method for constructing a dynamic critical care treatment pathway based on historical medical record analysis according to claim 4, characterized in that: The patient personalized risk assessment model specifically aims to balance prediction accuracy and potential clinical costs, which is specifically: Based on the comprehensive feature set, a training set including feature vectors, true complication labels and potential clinical costs is constructed; Each training sample is input into the patient-specific risk assessment model for forward calculation to obtain the probability of complication occurrence; Solve the correctness score and cost penalty for each training sample: The parameters of the patient personalized risk assessment model are iteratively updated based on gradient descent. In each iteration, the patient personalized risk assessment model improves the accuracy of training samples according to the accumulated value of the correctness score, and limits the error frequency of training samples according to the cost penalty to balance the accuracy of the patient personalized risk assessment model and clinical cost control.

6. The method for constructing a dynamic critical care treatment pathway based on historical medical record analysis according to claim 5, characterized in that: The correctness score is as follows: the true complication label is mapped from a binary form to positive or negative to distinguish the direction of samples with complications and samples without complications in the objective function, and the positive or negative difference between the complication probability output by the patient's personalized risk assessment model and the neutral threshold is used to measure the accuracy of the matching of the complication probability to the true complication label, and the matching degree of the complication probability to the true complication label is amplified in exponential form to obtain the correctness score of each training sample.

7. The method for constructing a dynamic critical care treatment pathway based on historical medical record analysis according to claim 6, characterized in that: The cost penalty is as follows: for training samples that incorrectly predict the probability of complication occurrence, their potential clinical cost value is introduced as the error cost, and combined with the set cost amplification parameter and amplified exponentially to obtain the cost penalty for each training sample.

8. The method for constructing a dynamic critical care treatment pathway based on historical medical record analysis according to claim 7, characterized in that: The medical record analysis structure is associated and integrated with the real-time clinical data, which is specifically: The relationship types in the medical record parsing structure are standardized and vectorized, and the associated features of the medical record parsing structure are extracted based on the attention mechanism to represent them as text vectors; Perform consistency check on real-time clinical data and obtain numerical vectors in real-time clinical data through sliding windows; The text vector and the numerical vector are fused in the feature dimension and the dimension is reduced by an autoencoder to generate a preliminary feature vector of multimodal fusion. After outlier detection and correction are performed on the preliminary feature vector of multimodal fusion, the output is the comprehensive feature set.

9. The method for constructing a dynamic critical care treatment pathway based on historical medical record analysis according to claim 8, characterized in that: The critical care first aid guidance model is selected based on the patient's personalized risk assessment results to generate an initial treatment path, which is specifically: Obtain the probability of complications of patients, and classify patients into corresponding risk levels based on the probability of complications of patients and pre-set risk classification rules; Based on the risk level, selecting a treatment plan template matching the risk level from a preset critical care first aid guidance model library; The selected critical care emergency guidance model applies the patient's previous medical record analysis structure and the probability of patient complications, and customizes the treatment pathway nodes in the critical care emergency guidance model to generate an initial treatment pathway.

10. A dynamic critical care treatment pathway construction system based on historical medical record analysis, characterized in that: include: Data acquisition module, used to obtain patients' historical medical records and real-time clinical data; Personalized risk assessment module, which is used to analyze the patient's historical medical record data based on natural language processing algorithms and conduct personalized risk assessment of the patient's condition in combination with machine learning models; The treatment pathway processing module is used to select the critical care emergency guidance model based on the patient's personalized risk assessment results to generate an initial treatment pathway. At the same time, the initial treatment pathway is dynamically adjusted according to changes in the condition in real-time clinical data to obtain the final treatment pathway.

Images