Method and system for individually accelerating multi-decision multi-objective optimization of clinical path of rehabilitation surgery

Abstract

The invention belongs to the technical field of medical treatment, and discloses an auxiliary decision-making method for accelerating rehabilitation surgery clinical path multi-decision multi-objective optimization. The method comprises the following steps: constructing a data set of preoperative states and medical decisions of a patient; unsupervised clustering is adopted, the clustering method is beneficial to dividing the patients into different phenotypes and risk levels according to unique features of the patients, and the obtained clusters provide valuable insights for heterogeneity of patient groups. And then, according to the decision data of each clustering phenotype, constructing a Bayesian network model, and adding some regulations in the construction process to trim the network, so that the constructed Bayesian network model better accords with clinic. And finally, inferring the state posterior probability of each node by using the Bayesian network, and inferring a decision scheme of each risk group by using the support degree so as to accelerate the postoperative rehabilitation process of the patient, reduce the hospitalization time of the patient and reduce the hospitalization cost.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and its medical applications, and particularly relates to a method and system for multi-decision optimization of personalized enhanced recovery after surgery (ERAS) clinical pathways. Background Art

[0002] The selection of clinical interventions for enhanced recovery after surgery (ERAS) is guided by existing guidelines and relevant clinical studies. However, the high-quality evidence of existing guideline clinical interventions all comes from randomized controlled trials (RCTs), but RCTs have certain limitations and can only verify the impact of a single variable on the outcome each time. During the entire ERAS process, the interaction between clinical interventions is ignored. The decision-making recommendations of the guidelines mainly target a group of patients with a certain disease, lacking personalization; and not all decisions come from high-quality clinical studies, but also from clinical experience, lacking evidence. With the continuous development of artificial intelligence technology, using artificial intelligence technology to mine medical knowledge from a large amount of medical data has begun to be applied to medical decision-making, but such medical decision-making lacks personalization and only targets one type or one intervention measure. Therefore, there is an urgent need for a method for personalized multi-decision and multi-objective optimization. Summary of the Invention

[0003] In view of the problems existing in the prior art, the present invention provides a method for multi-decision and multi-objective optimization of personalized enhanced recovery after surgery (ERAS) clinical pathways.

[0004] The present invention is implemented as follows. A method for multi-objective optimization and auxiliary decision-making of personalized enhanced recovery after surgery (ERAS) clinical pathways includes:

[0005] Step 1, constructing a dataset of the preoperative status of ERAS patients and medical decision-making;

[0006] Step 2, performing data preprocessing on the indicators related to the preoperative status of patients, including deleting outlier data or abnormal data, filling missing data, etc.;

[0007] Step 3, using the method of unsupervised clustering analysis to perform risk stratification on the preoperative status data of patients and determining the risk degree of the patient clustering phenotype.

[0008] Step 4, dividing the intervention degree of the medical decision-making data of patients;

[0009] Step 5, establishing a rule base for constructing a Bayesian network model according to expert experience or medical knowledge;

[0010] Step 6, grouping according to the clustering situation in Step 3, and constructing a Bayesian network model for each risk group by combining the rule base in Step 5;

[0011] Step 7: Using the Bayesian network model constructed in Step 6, determine the probability of the occurrence of the target outcome node and infer the posterior probability of each other decision node;

[0012] Step 8: Calculate the support degree of each state of each decision node by dividing the posterior probability by the prior probability, select each node, and use the state with the largest support degree as the recommended state for this decision.

[0013] Step 9: Integrate the recommended states of each decision node in different risk groups, and form a set of decision recommendation plans for each risk group.

[0014] Step 10: For newly admitted patients, input their preoperative status data into the unsupervised clustering model in Step 3 to determine the risk level of a single patient and give a decision recommendation plan for the corresponding risk level.

[0015] Furthermore, in step 1, the patient's preoperative status data includes but is not limited to demographic information (age, gender, BMI), smoking history, drinking history, allergy history, family history, disease history, surgical history, underlying diseases (hypertension, heart disease, diabetes, pulmonary disease COPD, asthma, cerebrovascular disease, peripheral vascular disease, liver and kidney diseases, tumors, etc.), ASA classification, type of surgery, preoperative vital signs (respiration, heart rate, blood pressure, oxygen saturation), preoperative blood routine, preoperative liver and kidney function electrolytes, preoperative blood gas, preoperative coagulation, preoperative inflammatory indicators, preoperative infection indicators, preoperative myocardial injury markers, preoperative examinations (preoperative electrocardiogram, preoperative cardiac ultrasound, preoperative CT, preoperative MRI, etc.). The medical decision-making data is divided into preoperative decision-making data, intraoperative decision-making data, and postoperative decision-making data according to the operation time. The preoperative decision-making data includes but is not limited to preoperative fasting time, preoperative carbohydrate load, preoperative use of regional anesthesia or analgesia, preoperative use of opioids, preoperative use of sedatives, preoperative use of tramadol, preoperative use of acetaminophen, preoperative use of non-steroidal anti-inflammatory drugs, preoperative use of gabapentinoids, preoperative use of antiemetics, preoperative use of antibiotics, preoperative use of antiplatelet drugs, preoperative use of anticoagulants, preoperative use of hormones, etc. The intraoperative decision-making data includes but is not limited to anesthesia method (general anesthesia, intraspinal anesthesia, and peripheral nerve block, etc.), intraoperative opioids, intraoperative intravenous injection of lidocaine, intraoperative sedatives, intraoperative non-steroidal anti-inflammatory drugs, intraoperative antiemesis, intraoperative use of hormones, intraoperative use of antibiotics, intraoperative ventilator mode and parameters, intraoperative body temperature maintenance, intraoperative local infiltration analgesia, intraoperative fluid therapy, intraoperative use of tranexamic acid, intraoperative placement of drainage tubes, etc. The postoperative decision-making data includes but is not limited to postoperative use of opioids, postoperative use of acetaminophen, postoperative use of acetaminophen, postoperative use of antiplatelet drugs, postoperative use of anticoagulants, postoperative use of antiemetics, postoperative use of sedatives, postoperative use of antibiotics, postoperative peripheral nerve block, postoperative tramadol, postoperative nutrition, postoperative early mobilization, etc. The outcome data includes but is not limited to postoperative hospital stay, hospitalization cost, survival status, complication status, etc.

[0016] Furthermore, the content of the preoperative data preprocessing in step 2 includes but is not limited to checking for data outliers or abnormal values, deleting outlier or abnormal data, and checking for missing data. Then, imputation is performed on the null value data, and the imputation methods include but are not limited to multiple imputation, KNN imputation, median imputation, mode imputation, mean imputation, constructing a machine learning model for imputation, etc. If the multiple imputation method is adopted, indicators with a vacancy greater than 50% need to be deleted, otherwise imputation cannot be performed. To ensure the accuracy of the data, individual patients with a missing rate greater than 30% can also be deleted. Statistical methods can also be selected to test whether there is a statistical significance between the data before and after imputation.

[0017] Further, for the preoperative status data used for unsupervised clustering in step 3, either the data closest to the surgery can be used, or the time-series data for a period of time before the surgery can be used. Before formal clustering, the Hopkins statistic can be used to test whether the status data has clusterability. Methods such as the silhouette coefficient and the RI coefficient can also be selected to determine the data for the optimal clustering. The unsupervised clustering methods used include but are not limited to methods such as K-Means clustering, hierarchical clustering, and DBSCAN. After clustering, statistical analysis and other methods can be used to analyze whether there is a statistical significance in the outcome indicators between the clustered phenotype groups, and the risk level of the phenotype groups can be clarified based on the outcome indicators.

[0018] Further, the decision data degree division in step 4 is mainly divided into two types: one is operation data, which is mainly divided into whether there is such an operation; the other is medication data, and according to clinical experience, medication instructions, relevant guidelines and literature, the dosage of the medication can be divided into not used, low dose, medium dose, and high dose.

[0019] Further, the rule base of expert experience and clinical knowledge constructed in step 5 is mainly used to clarify the relationship between nodes when constructing the Bayesian network model. This rule mainly includes two aspects: one is the prohibited relationship rule, which refers to the relationship between decision-making measures that are determined not to exist, such as the relationship where postoperative decisions affect preoperative decisions, and this relationship is incorrect, so this relationship needs to be excluded when constructing the Bayesian network; the other is the existing relationship, which refers to the relationship that is determined to exist, and this relationship is determined to exist, so this relationship needs to be retained when constructing the Bayesian network.

[0020] Further, when constructing the Bayesian network in step 6, it is necessary to construct it separately for each clustering risk group. During the Bayesian network structure learning process, the knowledge base constructed in step 5 needs to be used to delete the prohibited relationships in the model rules and add the existing relationships in the rules to the model. Then, parameter learning is performed using the data of various clustering risk groups to obtain the conditional probability table, and finally the Bayesian network model is formed.

[0021] Further, the outcome node in step 7 can be one outcome node or multiple outcome nodes, so multi-decision and multi-objective optimization can be completed. By using the fact that the good outcome has occurred at the outcome node, the posterior probability of the decision node is inferred using the Bayesian network model.

[0022] After the Bayesian network models of each clustering risk group are constructed, the following Bayesian formula is used to infer the posterior probability of each decision node:

[0023]

[0024] P(A 1 ),P(A 2 ),...,P(An ) is called the prior probability, also known as the probability before the experiment, that is, the probability learned through parameter learning, P(A k |B) (k = 1, 2,... n) refers to the posterior probability, P(B|A i ) refers to the probability that B occurs under the result of A i occurring.

[0025] Furthermore, in step 8, the decision support degree calculation formula using the prior probability and the posterior probability is as follows, which is used to measure the influence of the decision on the outcome:

[0026]

[0027] Calculate the decision support degree for each state of each node respectively, and select the state with the largest decision support degree and greater than 1 for each decision node as the recommended state of this decision point, so as to form the decision-making plan for each clustering risk group.

[0028] Furthermore, step 10 is mainly the clinical application of the whole method. For newly admitted surgical patients, the state data of the patients are collected before surgery and input into the unsupervised clustering model determined in step 3, and a clustering risk result will be obtained. Select the corresponding recommended decision-making plan for treatment according to the clustering result.

[0029] Another object of the present invention is to provide a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of the multi-objective optimization assisted decision-making method for the enhanced recovery after surgery clinical pathway.

[0030] Combined with the above technical solutions and the solved technical problems, the advantages and positive effects of the technical solution to be protected by the present invention are:

[0031] The present invention proposes a brand-new decision optimization method, which can perform multi-objective optimization for complex decision-making problems in the medical field. This method comprehensively considers various treatment factors, such as surgical plans, postoperative rehabilitation strategies, drug selection, etc., and realizes the whole-process optimization. Compared with the traditional decision-making mode, the present invention not only focuses on the optimization of a single objective (such as treatment effect), but also can find the best balance among multiple objectives, avoid over-treatment or resource waste, and thus improve the overall medical efficiency.

[0032] Another major technical advantage of the present invention is to provide a precise treatment plan based on the individual needs of patients. By comprehensively analyzing the patient's medical history, physical constitution, genetic information, and clinical indicators, this method can dynamically adjust the treatment strategy to make the treatment plan more in line with the individual characteristics of the patient. Different from the traditional standardized treatment mode, this personalized method can improve the treatment effectiveness and reduce unnecessary side effects, and is particularly suitable for medical scenarios that require refined regulation, such as perimenopausal management, postoperative rehabilitation, and chronic disease management.

[0033] Since the present invention adopts an optimized decision-making method, it can accurately match the postoperative care needs of patients, accelerate the patient's rehabilitation process, and shorten the hospital stay. The personalized rehabilitation plan not only helps to improve the patient's recovery speed, but also can reduce the risks such as infection during hospitalization. For medical institutions, the shortening of the hospital stay can also improve the bed turnover rate, making the limited medical resources more efficiently utilized.

[0034] Through multi-objective optimization decision-making, the present invention can effectively avoid the complications that occur after surgery during the formulation of the treatment plan. For example, in the postoperative care plan, the present invention can intelligently adjust the anti-infection, anticoagulation, and nutritional support strategies to reduce the incidence of postoperative infection, thrombosis, and bleeding. The patient recovers in a safer environment, which not only improves the success rate of treatment, but also reduces the need for secondary treatment or long-term medical care.

[0035] Benefiting from the optimized decision-making process, the present invention can reduce unnecessary hospital stays and medical items, thereby reducing the overall medical expenses of patients. The precise treatment strategy can reduce repeated examinations, ineffective treatments, and over-interventions, making each medical expenditure of the patient play the greatest value. This optimized plan not only reduces the economic pressure on patients, but also improves the resource utilization rate of the hospital, making the overall operation of the medical system more efficient.

[0036] The present invention provides an intelligent decision support tool that can combine the patient's clinical data, medical knowledge, and treatment goals to provide doctors with suggestions for optimized treatment plans. Doctors can use this tool to quickly analyze the advantages and disadvantages of different treatment strategies and make more precise decisions in combination with their own experience. This not only improves the scientific nature of medical decision-making, but also reduces misjudgments caused by lack of experience or information asymmetry, thereby enhancing the patient's treatment experience and overall medical quality. Description of the Drawings

[0037] Figure 1 is a flowchart of a multi-objective optimization assisted decision-making method for an enhanced recovery after surgery clinical pathway provided by an embodiment of the present invention.

[0038] Figure 2 is a flowchart of a method for establishing a Bayesian network model provided by an embodiment of the present invention.

[0039] Figure 3 It is the structural block diagram of the multi - objective optimization assisted decision - making system for the enhanced recovery after surgery (ERAS) clinical pathway provided by the embodiments of the present invention.

[0040] Figure 4 It is the development data clustering result graph provided by the embodiments of the present invention.

[0041] Figure 5 It is the structural diagram of the Bayesian network model for each risk phenotype provided by the embodiments of the present invention.

[0042] Figure 6 It is the relationship graph between the number of decision recommendations that meet each risk phenotype and the average length of hospital stay provided by the embodiments of the present invention.

[0043] Figure 7 It is the detailed flowchart of the multi - objective optimization assisted decision - making method for the enhanced recovery after surgery (ERAS) clinical pathway provided by the embodiments of the present invention. Specific embodiments

[0044] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0045] As Figure 1 shown, a multi - objective optimization assisted decision - making method for the enhanced recovery after surgery (ERAS) clinical pathway provided by the embodiments of the present invention includes the following steps:

[0046] S101, construct a dataset of the preoperative status and medical decisions of ERAS patients;

[0047] The preoperative status data of patients includes, but is not limited to, demographic data, comorbidity data, disease history and surgical history, test and examination data, etc. Medical decision data includes, but is not limited to, various operations and medications, etc.

[0048] S102, perform data pre - processing on the indicators related to the preoperative status of patients, including deleting outlier data or abnormal data, filling in missing data, etc.

[0049] The content of preoperative data pre - processing includes, but is not limited to, checking for outlier values or abnormal values in the data, deleting outlier or abnormal data, and checking for missing data.

[0050] S103, perform risk stratification on the preoperative status data of patients using the method of unsupervised clustering analysis, and determine the risk level of the patient's clustering phenotype.

[0051] S104, divide the intervention degree of the medical decision data of patients.

[0052] There are mainly two types: one is operation data, which is mainly divided into whether there is such an operation; the other is medication data, and according to clinical experience, medication instructions, relevant guidelines and literature, the dosage of medication can be divided into no use, low dose, medium dose and high dose. Both operation data and medication data can be divided into preoperative, intraoperative and postoperative use.

[0053] S105, establish a rule base for constructing a Bayesian network model according to expert experience or medical knowledge.

[0054] This rule base uses expert knowledge and clinical experience to prune the Bayesian network structure, making the Bayesian network model closer to clinical practice.

[0055] S106, group according to the clustering situation in S103, and construct a Bayesian network model for each risk group by combining the rule base of S105 respectively.

[0056] S107, use the Bayesian network model constructed in S106 to determine the probability of the occurrence of the target outcome node and infer the posterior probability of each other decision node.

[0057] After the Bayesian network models of each clustering risk group are constructed, use the following Bayesian formula to complete the inference of the posterior probability of each decision node:

[0058]

[0059] P(A 1 ), P(A 2 ),..., P(A n ) are called prior probabilities, also known as probabilities before the experiment, that is, probabilities obtained through parameter learning. P(A k |B)(k = 1, 2,... n) refers to the posterior probability, and P(B|A i ) refers to the probability of B occurring under the result of the occurrence of A i .

[0060] S108, calculate the support degree of each state of each decision node by dividing the posterior probability by the prior probability, select each node, and use the state with the largest support degree as the recommended state of this decision.

[0061] The calculation formula for decision support using prior probability and posterior probability is as follows, which is used to measure the influence of this decision on the outcome:

[0062]

[0063] Calculate the decision support degree for each state of each node respectively, select the state with the largest decision support degree and greater than 1 for each decision node as the recommended state of this decision point, so as to form the decision-making plan for each clustering risk group.

[0064] S109. Integrate the recommended status of each decision node for different risk groups, and each risk group forms a set of decision recommendation schemes.

[0065] S110. For newly admitted patients, use their preoperative status data to input into the unsupervised clustering model in S103 to determine the risk level of a single patient and give a decision recommendation scheme corresponding to the risk level.

[0066] As Figure 7 shown, a multi-objective optimization assisted decision-making method for the enhanced recovery after surgery clinical pathway provided by an embodiment of the present invention includes the following steps:

[0067] Dataset construction;

[0068] The patient data in this example is sourced from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database, which is an open-access intensive care database. This latest version contains extensive clinical information of patients admitted to Beth Israel Deaconess Medical Center from 2008 to 2019.

[0069] The development dataset is extracted from the MIMIC-IV database. The raw data is obtained using Structured Query Language and Navicat Premium software (version 15.0.12). The criteria for selecting patients are as follows: (1) having a record of total hip arthroplasty; (2) age > 18 years old; (3) the first surgery during a single hospitalization.

[0070] Then, perioperative data is extracted, and the included indicators and decisions are as follows:

[0071] (1) Demographic information;

[0072] (2) Comorbidities;

[0073] (3) Laboratory indicators: blood routine, liver function, coagulation, renal function, electrolytes, glucose, C-reactive protein;

[0074] (4) Medication interventions: antiemetics, antiplatelet drugs, anticoagulants, novel oral anticoagulants (NOACs), aminocaproic acid, painkillers, non-steroidal anti-inflammatory drugs, opioids, ketamine, sedatives, glucocorticoids, antibiotics, nerve blocks, drainage, infusions, anesthesia methods. (See supplementary materials for details)

[0075] Regarding clinical interventions, the choices were guided by existing enhanced recovery after surgery (ERAS) total hip arthroplasty and related clinical studies. However, for patients undergoing surgery in the MIMIC-IV database, certain indicators showed a lack of these specific items or complete absence of data. This included preoperative carbohydrate treatment, surgical methods, artificial joint materials, epidural analgesia, periarticular injection analgesia, and the lack of specific drugs in certain drug classes, such as tranexamic acid and etoricoxib. In such cases, these intervention indicators were systematically excluded from the analysis. The postoperative length of stay (PLOS) was used as the outcome measure, and the status was divided depending on the mean PLOS in the development dataset.

[0076] The external validation dataset was from patients who underwent elective total hip arthroplasty at the First Affiliated Hospital of Third Military Medical University from January 1, 2018, to December 31, 2018. This study was approved by the Ethics Committee of Southwest Hospital, Third Military Medical University (number KY201936). The inclusion and exclusion criteria were the same as above, and patients who underwent multiple surgeries during one hospitalization were excluded.

[0077] Data preprocessing;

[0078] Data preprocessing was completed using R software v4.3.1. Missing values in the status data of the MIMIC-IV database and the external validation dataset were imputed multiple times using the "MICE" package version 3.16.021 by chained equations.

[0079] K-means clustering;

[0080] To utilize the characteristics of individual patients, including demographic information, comorbidities, and laboratory indicators, as shown in Table 1, K-means clustering was adopted. This clustering method helps to divide patients into different phenotypes based on their unique characteristics. Before clustering, Hopkins statistics were used to evaluate the clustering tendency of the data, and the analysis was performed using the cluster package (version 2.1.6) of the R language. The K-means algorithm systematically divides patients into K clusters, optimizing the grouping according to feature similarity, and the resulting clusters provide valuable insights into the heterogeneity of the patient population. Subsequently, by associating each identified phenotype with a specific risk pattern, an attempt was made to enhance the ability to predict and manage hospitalization outcomes. By analyzing these phenotypes, the respective risks of phenotypes related to the length of stay can be elucidated.

[0081] Division of the degree of intervention for the patients' medical decision-making data;

[0082] For drug interventions, different classification methods were adopted. Some drugs were classified according to dosage, and some were classified according to usage status, and they were converted into categorical variables. For anticoagulants, opioids, glucocorticoids, gabapentin, acetaminophen, and ketamine, the dosages were calculated and divided into different statuses using quartiles. Heparin, enoxaparin sodium, and fondaparinux sodium were calculated based on the equivalent dosages of low-molecular-weight heparin for combination. The combination of opioids was calculated based on the dose equivalent of morphine equivalents. At the same time, for non-steroidal anti-inflammatory drugs, sedatives, NOACs, antiplatelet drugs, antiemetics, aminocaproic acid, and antibiotics, the status was determined according to the usage situation. For drugs for which the dosage could be extracted, the status was divided by quartiles. Those less than the 25th percentile were classified as low dosage, those between the 25th and 75th percentiles were classified as medium dosage, those greater than the 75th percentile were classified as high dosage, and those not used. For drugs for which the dosage could not be determined or extracted, as well as medical decision-making measures, they were classified as used and not used. Specifically as shown in Table 2.

[0083] Construct the rule base of the Bayesian network

[0084] In the rule base of this example, only the prohibited rules were filled. The postoperative-related decisions and medications were prohibited from pointing to the preoperative-related decisions and medications, that is, the postoperative indicators could not affect the preoperative indicators.

[0085] Model construction based on the Bayesian network;

[0086] After determining different phenotypes through K-means clustering, a BN model was constructed separately for each phenotype to facilitate variable analysis and decision support for different patient types. The BN model was constructed using the bnlearn 4.9.3 version of the R 4.3.1 software package. The clustering results of the development data are as Figure 4 shown.

[0087] As Figure 2 shown, the steps for establishing the Bayesian network model provided by the embodiment of the present invention are as follows:

[0088] S201, variable selection: Select the variables of medical intervention. Divide the data according to the phenotypes, create different subsets, and use the decisions inherent in each phenotype for model construction.

[0089] S202, structure learning: Use the tabu search algorithm to learn the Bayesian network structure from the data. Use the tabu search algorithm to learn the Bayesian network structure from the data. This algorithm can iteratively identify the optimal relationships between variables, identify their causal relationships, and integrate expert knowledge to construct the topological structure of the Bayesian network.

[0090] S203, Parameter learning: After the structure of the Bayesian network model is completed, the conditional probability of each variable needs to be obtained. Using the maximum likelihood estimation (MLE) algorithm based on decision data, parameter learning is carried out to obtain the conditional probability table of each variable, thereby constructing the Bayesian network model. The Bayesian network model structures of each risk phenotype are as shown in Figure 5 .

[0091] S204, Bayesian decision-making: After constructing the Bayesian network model, the Netica 32Bit V5.18 software is used to visualize the prior probability of each decision point. According to Bayes' formula, the posterior probability under the condition that PLOS = 0, that is, the postoperative length of stay is not extended, is calculated, and the support degree is introduced to reflect the influence degree of the decision point on the result. A value greater than 1 indicates support for the decision, while a value less than 1 indicates non-support for the decision. The larger the value, the stronger the support for the decision. The support degrees of each node state in this embodiment are shown in Table 3 in detail.

[0092] Bayesian decision-making: The topological structure of the Bayesian network and the conditional probability between nodes are obtained through structure learning and parameter learning. In the constructed Bayesian network, except for the observed node, the postoperative length of stay (PLOS), other nodes are decision nodes. Therefore, Bayes' formula can be used to calculate the posterior probability of the decision under the condition of reduced PLOS. Bayes' formula is as follows:

[0093]

[0094] is usually called the prior probability, also known as the probability before the experiment (here it is the probability learned from the data), and is called the posterior probability, which refers to the probability of occurring under the condition that occurs.

[0095] Then, calculate the decision support degree:

[0096]

[0097] It represents the degree of support for a decision and also reflects the influence of the decision on the observed index (result). A value greater than 1 indicates support for the decision, while a value less than 1 indicates non-support for the decision. The larger the value, the stronger the support for the decision.

[0098] Form the decision recommendation scheme of this example.

[0099] Internal validation and external validation.

[0100] Use the development dataset for internal validation and the Southwest Hospital dataset for external validation. Respectively, count how many medical decisions of each patient in the two datasets conform to the decisions recommended by the Bayesian network, and calculate the average length of stay of patients who conform to 1 - 3 decisions, 4 - 6 decisions, 7 - 9 decisions, and 10 - 12 decisions. The results are shown in Table 4 and asFigure 6 as shown

[0101] such as Figure 3 As shown, a multi-objective optimization assisted decision-making system for an enhanced recovery after surgery clinical pathway provided by an embodiment of the present invention includes:

[0102] A data collection module, configured to collect historical data from a database of a hospital information system;

[0103] A data extraction module, configured to extract relevant patient status indicators and medical decision-making measures from the database;

[0104] A data preprocessing module, configured to enable a doctor to process relevant indicators, such as imputing missing status data and classifying the status of medical decision-making measures;

[0105] A clustering module, configured to perform unsupervised clustering on patient status data, including demographic information, comorbidities, and laboratory indicators, etc., classify patients into several phenotypes, and then divide risk levels according to outcome indicators. Commonly used clustering algorithms can be selected.

[0106] A rule base module, configured to store relevant expert knowledge and empirical knowledge and modify the structure of a Bayesian network model.

[0107] A Bayesian network module, after determining different phenotypes and risk levels, constructs a BN model separately for each phenotype to facilitate variable analysis and decision support for different patient types; The BN model is constructed using the bnlearn 4.9.3 version of the R 4.3.1 software package.

[0108] A decision recommendation module, configured to perform relevant decision recommendations using the Bayesian network model and form a decision recommendation plan.

[0109] For newly admitted patients, the system collects the preoperative status data of the patient, including preoperative information such as patient demographic data, personal history, medical history, comorbidities, and tests and examinations. The decision plan generation module will recommend a decision plan according to the clustering phenotype of the patient based on the patient's preoperative data.

[0110] The following are two specific embodiments of the present invention:

[0111] Embodiment 1: Personalized enhanced recovery after surgery treatment for high-risk patients undergoing hip replacement surgery

[0112] A 61-year-old female patient was admitted to the hospital due to "right hip pain for more than 11 months after a fall". The patient is an elderly female with a history of surgical-related trauma surgery and a long course of disease. The patient has underlying diseases such as diabetes and mild hypertension. The traditional postoperative rehabilitation plan failed to fully consider the individual characteristics of the patient, resulting in general postoperative recovery and a relatively long postoperative hospital stay.

[0113] Implementation process of the present invention:

[0114] 1. Personalized data collection: First, through the data collection module, preoperative data such as the patient's personal information, medical history, weight, and examination results are collected. The unsupervised clustering algorithm is used to identify the risk level where the patient's phenotype lies, and the patient is identified as a high-risk patient.

[0115] 2. Decision optimization: In the decision-making plan generation module, according to the patient's preoperative status data, including demographic data, comorbidity conditions, preoperative examination data, etc., the patient's clustering phenotype and risk level are identified, and the corresponding enhanced recovery after surgery treatment plan is recommended for the patient.

[0116] 3. Implement treatment and adjustment: According to the optimized plan, a specific rehabilitation treatment plan is formulated. A total of 10 recommended treatment plans were implemented for this patient.

[0117] 4. Treatment effect: By implementing this optimized plan, the patient's postoperative rehabilitation progress has been significantly accelerated. The actual postoperative hospital stay of this patient is 3.66 days. Compared with the average postoperative hospital stay of 4.19 days for this cluster of patients, the postoperative hospital stay of this patient is significantly shortened.

[0118] Advantages:

[0119] The personalized treatment plan provides precise treatment suggestions according to the patient's special medical history and recovery situation.

[0120] The optimized drug treatment and nursing plan reduce postoperative complications and shorten the hospital stay.

[0121] Save costs for patients and medical insurance funds. Since the hospital stay is shortened, the hospitalization expenses are also less.

[0122] Assist doctors in decision-making. The optimized method considers more comprehensive patient information and assists doctors in making decisions for which there are no relevant guidelines and consensus yet.

[0123] Example 2: Personalized enhanced recovery after surgery for low-risk patients undergoing hip replacement surgery

[0124] A patient, male, 37 years old, was admitted to the hospital due to "right hip pain for more than 1 year". The patient is a middle-aged male with a long course of disease, a smoking history, and denies a history of alcohol use, personal diseases, and surgery. Denied diabetes and hypertension. The auxiliary examinations showed no obvious abnormalities. The general condition of this patient is relatively good. If a personalized optimized treatment plan for enhanced recovery after surgery is adopted, can the postoperative hospital stay be further shortened?

[0125] Implementation process of the present invention:

[0126] 1. Personalized data collection: First, through the data collection module, preoperative data such as the patient's personal information, medical history, weight, and test results are collected. The unsupervised clustering algorithm is used to identify that the patient phenotype is at a low-risk level, classifying the patient as a low-risk patient.

[0127] 2. Decision optimization: In the decision-making plan generation module, based on the patient's preoperative status data, including demographic data, comorbidity conditions, preoperative test data, etc., the patient's clustered phenotype and risk level are identified, and the corresponding enhanced recovery after surgery treatment plan is recommended for the patient.

[0128] 3. Treatment implementation: According to the optimized plan, a specific rehabilitation treatment plan is formulated. A total of 6 recommended treatment plans were implemented for this patient.

[0129] 4. Treatment effect: By implementing this optimized plan, the patient's postoperative rehabilitation progress has been significantly accelerated. The actual postoperative hospital stay of this patient is 3.53 days. Compared with the average postoperative hospital stay of 3.98 days for this clustered group of patients, the patient's postoperative hospital stay has been shortened.

[0130] Advantages:

[0131] The personalized treatment plan provides precise treatment recommendations based on the patient's specific medical history and recovery situation.

[0132] The optimized drug treatment and nursing plan reduce postoperative complications and shorten the hospital stay.

[0133] It saves costs for patients and medical insurance funds. Since the hospital stay is shortened, the hospitalization expenses are also less.

[0134] It assists doctors in making decisions. This optimization method considers more comprehensive patient information and helps doctors complete decisions for which there are no relevant guidelines and consensus yet.

[0135] Total hip arthroplasty is a surgical method that replaces the diseased femoral head and acetabulum with an artificial joint prosthesis to restore its normal function, effectively treating various hip joint diseases such as hip osteoarthritis and avascular necrosis of the femoral head, and improving the quality of life. With the increase in population aging, the demand for total hip arthroplasty is expected to grow exponentially. Therefore, reducing postoperative complications and minimizing the hospital stay have become the focus. This embodiment aims to classify different risks of patients based on personalized indicators of patients, and then use the Bayesian network algorithm to determine the perioperative measures that affect the prognosis of total hip arthroplasty patients, establish a relationship network between indicators based on the mutual connections between indicators, formulate an optimal decision-making combination for total hip arthroplasty patients, guide clinical decisions, thereby accelerating the patient's recovery and achieving the goal of reducing the hospital stay.

[0136] The development dataset was extracted from the MIMIC-IV database. Hospitalized patients were selected according to the following criteria: (1) surgical records of total hip arthroplasty; (2) age > 18 years; (3) complete decision data; (4) the first surgery during a single hospitalization.

[0137] Step 1, the personalized status data of the extracted patients, including demographic information, comorbidities, and laboratory indicators, are shown in Table 1 for specific indicators and missing status.

[0138] Table 1. Personalized status indicators of the included patients

[0139]

[0140]

[0141] Step 2, the KNN data imputation method was used to impute the missing data, and the K-means method was further used to cluster the patient status data. The results are as Figure 4 shown.

[0142] Step 3, extract the perioperative decision data of the patients and divide the degree of intervention for the patients' medical decision data;

[0143] For drug interventions, different classification methods were adopted. Some drugs were classified according to the dose, and some were classified according to the usage status, and they were converted into categorical variables. For anticoagulants, opioids, glucocorticoids, gabapentin, acetaminophen, and ketamine, the dose was calculated and divided into different statuses using quartiles. Heparin, enoxaparin sodium, and fondaparinux sodium were calculated based on the equivalent dose of low-molecular-weight heparin for merging. The combination of opioids was calculated based on the dose equivalent of morphine equivalents. At the same time, for non-steroidal anti-inflammatory drugs, sedatives, NOACs, antiplatelet drugs, antiemetics, aminocaproic acid, and antibiotics, the status was determined according to the usage status. Drugs for which the dose could be extracted were divided into statuses using quartiles, with less than the 25th percentile classified as low dose, 25%-75% percentile classified as medium dose, greater than the 75th percentile classified as high dose, and not used. Drugs for which the dose could not be determined or extracted, as well as medical decision measures, were divided into used and not used. See Table 2 for details.

[0144] Table 2. The included decision data and the corresponding status classification

[0145]

[0146]

[0147] Step 4: Construct a rule base for the Bayesian network. There are no specific rules in this particular embodiment, but it is stipulated that postoperative indicators are prohibited from pointing to preoperative indicators, that is, postoperative decisions do not affect preoperative decisions.

[0148] Step 5: Construct a Bayesian network model for each patient phenotype, specifically as Figure 5 , and use Bayes' formula to calculate the posterior probability of the decision under the condition of reduced postoperative hospital stay. Bayes' formula is as follows:

[0149]

[0150] is usually called the prior probability, also known as the probability before the experiment (here it is the probability learned from data), and is called the posterior probability, which refers to the probability of occurring under the condition that has occurred.

[0151] Then, use the following decision support formula to calculate the support degree:

[0152]

[0153] The support degrees of the node states of each risk phenotype are specifically shown in Table 3.

[0154] Table 3. Support degrees of node states of each risk phenotype

[0155]

[0156]

[0157]

[0158] Step 6: If the specific support degree is greater than 1, it indicates support for the decision, while if it is less than 1, it indicates non - support for the decision. The larger the value, the stronger the support for the decision, thus forming the decision recommendation scheme for this instance.

[0159] Step 7: Statistically analyze the relationship between the number of medical decisions that conform in the development dataset and the average postoperative hospital stay, as shown in Table 4 and Figure 6 as follows.

[0160] Step 8: Use the clustering model constructed from the previous development dataset to cluster the patient status data of the validation data, and then, according to the decisions recommended by the patient phenotype, statistically analyze the relationship between the number of medical decisions that conform to the recommended decisions in the validation dataset and the average postoperative hospital stay, as shown in Table 4 and Figure 6 as follows.

[0161] From the results in Table 4 and Figure 6 , it can be concluded that the more the actual medical decisions of the patients conform to the medical decisions recommended by the model, the fewer the average postoperative hospital stay days. The decisions recommended by this model are due to reducing the average postoperative hospital stay days.

[0162] Table 4. Relationship between the number of decision recommendations met by each risk phenotype and the average length of hospital stay

[0163]

[0164] Figure 4 . Develop a data clustering result graph

[0165] Figure 5 . Structure diagram of the Bayesian network model for each risk phenotype

[0166] Figure 6 . Relationship between the number of decision recommendations met by each risk phenotype and the average length of hospital stay

[0167] Example 1: Optimization of postoperative analgesia regimen for enhanced recovery after surgery (ERAS)

[0168] A 65-year-old male patient with a BMI of 28.5 needed to undergo laparoscopic surgery for colorectal cancer. He had a history of hypertension and diabetes, and no coagulation disorder. Preoperative examinations showed mild anemia, and the ASA score was II.

[0169] 1) Construct a patient dataset

[0170] Enter the patient's preoperative data, including demographic information (age, BMI), underlying diseases (hypertension, diabetes), preoperative laboratory test results (hemoglobin, coagulation function), etc.

[0171] Enter medical decision data, such as preoperative analgesia regimen, intraoperative anesthesia method, postoperative analgesia regimen, etc.

[0172] 2) Data preprocessing

[0173] Process missing data, exclude extreme outliers, and ensure data integrity.

[0174] 3) Unsupervised clustering risk stratification

[0175] Use the K-Means clustering method to divide the patients into the "medium-risk" group (the risk level is between high and low, and moderate postoperative pain occurs) based on preoperative data.

[0176] 4) Optimization of postoperative analgesia decision

[0177] Combine postoperative analgesia-related decision data and classify the medication regimens at different stages of preoperative, intraoperative, and postoperative:

[0178] Preoperative: Low-dose opioids + nonsteroidal anti-inflammatory drugs (NSAIDs).

[0179] Intraoperative: Use general anesthesia + local infiltration analgesia (lidocaine).

[0180] Postoperative: Based on Bayesian network analysis, low-dose opioids combined with acetaminophen are recommended to reduce the use of opioids and postoperative complications (such as nausea and vomiting, respiratory depression).

[0181] 5) Bayesian network inference and decision support

[0182] Calculate the probability of postoperative pain control in patients, and determine the optimal postoperative analgesia plan based on the posterior probability.

[0183] The system recommends that the patient use acetaminophen + NSAIDs + low-dose opioids as an analgesia plan after surgery, and reduce the duration of opioid use to within 24 hours.

[0184] 6) Postoperative follow-up and efficacy evaluation

[0185] The pain scores were followed up at 24 hours and 48 hours after surgery, and it was found that the patient's pain was well controlled and there were no obvious opioid side effects.

[0186] The length of hospital stay was shortened by 1 day, and compared with the traditional opioid-dominated analgesia strategy, the hospital costs and opioid use-related risks were reduced.

[0187] Example 2: Preoperative optimization intervention strategy for reducing postoperative complications

[0188] A patient, 72 years old, female, BMI 30.0, is scheduled to undergo hip replacement surgery. She has a history of chronic obstructive pulmonary disease (COPD), hypertension, and osteoporosis. The preoperative hemoglobin level was low (Hb 105 g / L), and there were no obvious abnormalities in the preoperative cardiac function assessment.

[0189] 1) Construct a patient dataset

[0190] Enter preoperative status data, including underlying diseases (COPD, hypertension), preoperative laboratory tests (hemoglobin), preoperative pulmonary function tests, etc.

[0191] Preoperative medical decision-making data, including preoperative smoking cessation education, respiratory rehabilitation training, etc.

[0192] 2) Data preprocessing

[0193] Process abnormal preoperative hemoglobin values to ensure the accuracy of hematological data.

[0194] 3) Risk stratification and clustering analysis

[0195] Using the unsupervised clustering (DBSCAN) method, the patient was classified into the "high-risk" group, indicating a higher risk of postoperative pulmonary complications.

[0196] 4) Preoperative optimization decision

[0197] Combined with Bayesian network analysis, the system recommends the following preoperative optimization plan:

[0198] Conduct respiratory rehabilitation training 4 weeks before surgery (to enhance postoperative lung function recovery).

[0199] Supplement iron before surgery to increase the preoperative hemoglobin level.

[0200] Quit smoking for more than 2 weeks before surgery to reduce the risk of postoperative pulmonary infection.

[0201] 5) Intraoperative management optimization

[0202] Anesthesia method: Select spinal anesthesia + local nerve block to reduce the adverse effects of general anesthesia on lung function.

[0203] Intraoperative fluid management: Maintain appropriate fluid balance to avoid pulmonary edema caused by excessive fluid replacement.

[0204] 6) Postoperative optimization management

[0205] Early postoperative activity: Encourage the patient to get out of bed within 6 hours after surgery to reduce the risk of pulmonary complications.

[0206] Postoperative analgesia plan: Non-steroidal anti-inflammatory drugs (NSAIDs) + local analgesics to avoid the inhibitory effect of excessive use of opioid drugs on respiration.

[0207] Postoperative respiratory management: Conduct pulmonary rehabilitation training daily to improve lung compliance and reduce the incidence of atelectasis.

[0208] 7) Bayesian network inference and decision support

[0209] Calculate the probability of postoperative pulmonary complications. The results show that the use of preoperative optimization strategies can reduce the risk of postoperative complications by 35%.

[0210] The system recommends continuing to implement the preoperative optimization plan and strictly monitoring the recovery of lung function after surgery.

[0211] 8) Follow-up and efficacy evaluation

[0212] The patient had no hypoxemia within 3 days after surgery, did not develop atelectasis or pneumonia, and had a good postoperative recovery.

[0213] The length of hospital stay was shortened by 2 days. Compared with patients who did not adopt the optimization strategy, the incidence of postoperative complications decreased significantly, and the overall medical cost decreased by 15%.

[0214] These two embodiments demonstrate the application of the present invention in personalized postoperative analgesia optimization and preoperative optimization interventions. Through the Bayesian network model combined with unsupervised clustering analysis, high-risk patients can be accurately identified, and the best preoperative, intraoperative, and postoperative management plans can be provided, thereby reducing postoperative complications, shortening the hospital stay, optimizing the utilization of medical resources, and improving the quality of patient recovery. It should be noted that the implementation mode of the present invention can be realized through hardware, software, or a combination of software and hardware. The hardware part can be realized using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those of ordinary skill in the art can understand that the above devices and methods can be realized using computer-executable instructions and / or included in processor control code, such as providing such code on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits of programmable hardware devices such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or can be implemented by software executed by various types of processors, or can be implemented by a combination of the above hardware circuits and software such as firmware.

[0215] As described above, it is only the specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention by those skilled in the art within the technical scope disclosed by the present invention shall be covered by the protection scope of the present invention.

Claims

1. A multi-decision and multi-objective optimization decision-making auxiliary method for the clinical pathway of enhanced recovery after surgery, characterized in that: The following steps are involved: Step 1: construct a data set containing preoperative patient status data and medical decision data, wherein the preoperative patient status data includes demographic information, comorbidity information, medical history, surgical history, and test data, and the medical decision data includes preoperative, intraoperative, and postoperative operation and medication data; Step 2: preprocess the preoperative patient status data, including outlier detection, outlier data removal, missing data filling, etc. Step 3: Use unsupervised clustering methods to perform cluster analysis on preoperative patient status data to form different risk groups and determine the risk level of each cluster phenotype; Step 4: classify the medical decision data by state, including the presence or absence of operation data and the dosage of medication data; Step 5: construct a Bayesian network rule base based on medical knowledge and expert experience, and use the rule base to prune the Bayesian network structure to optimize the model structure; Step 6: Based on the different risk groups formed in step 3 and combined with the rule base in step 5, a Bayesian network model corresponding to each risk group is constructed; Step 7, using the Bayesian network model constructed in step 6, calculate the probability of the target outcome node and infer the posterior probability of each decision node; Step 8: Calculate the decision support based on the posterior probability, and select the state with the largest support as the recommended state for each decision node to form a decision plan for each risk group; Step 9: Integrate the decision plans of each risk group to form a complete clinical pathway decision recommendation plan; Step 10: Input the preoperative status data of the newly admitted patients into the unsupervised clustering model to determine their risk groups and provide corresponding decision-making recommendations.

2. The method according to claim 1, characterized in that The data set includes preoperative patient status data and medical decision data. The preoperative patient status data includes demographic information, comorbidity information, surgical history, underlying diseases, ASA grade, preoperative vital signs, preoperative blood routine, preoperative coagulation, preoperative imaging examinations, etc. The medical decision data is divided into three stages of decision data: preoperative, intraoperative and postoperative according to the operation time.

3. The method according to claim 1, characterized in that The cluster analysis of the preoperative patient status data adopts unsupervised clustering methods, including but not limited to K-Means clustering, hierarchical clustering, and DBSCAN methods, and uses the Hopkins statistic or silhouette coefficient to determine the clusterability of the data before formal clustering, and analyzes the risk level of each cluster phenotype based on the outcome indicators.

4. The method according to claim 1, characterized in that The status division of the medical decision data includes operation data and medication data. The operation data is divided according to whether the operation is performed. The medication data is divided into four states: not used, low dose, medium dose and high dose according to clinical guidelines, drug instructions or statistical methods, and is divided into preoperative, intraoperative and postoperative medication data according to the operation time node.

5. The method according to claim 1, characterized in that The construction of the Bayesian network model uses the rule base established in step 5 to perform structural pruning, including deleting invalid relationships that violate medical knowledge, adding associations that conform to medical knowledge, and performing parameter learning based on the data of each risk group to form a conditional probability table.

6. The method according to claim 1, characterized in that The target outcome node of the Bayesian network model can be single or multiple, the starting node and the intermediate node correspond to the medical decision, and the posterior probability of each decision node is calculated using the Bayesian formula based on the occurrence of the target outcome.

7. The method according to claim 1, characterized in that The decision support calculation is based on the ratio of prior probability to posterior probability, and the support is calculated for different states of each node. The state with the largest support and greater than 1 is selected as the recommended decision state to form a decision plan for each risk group.

8. A multi-objective optimization decision-making support system for an accelerated recovery surgery clinical pathway that implements the multi-objective optimization decision-making support method for an accelerated recovery surgery clinical pathway as described in any one of claims 1 to 7, characterized in that: The auxiliary decision-making system for multi-decision and multi-objective optimization of the clinical pathway of enhanced recovery after surgery includes: A data collection module is used to collect historical data from the database of the hospital information system; A data extraction module is used to extract relevant patient status indicators and medical decision measures from the database; Data preprocessing module, used by doctors to process relevant indicators, such as interpolating missing status data and classifying the status of medical decision-making measures; The clustering module is used to perform unsupervised clustering of patients' status data, including demographic information, comorbidities, and laboratory indicators, to differentiate patients into several phenotypes. When dividing risk levels according to outcome indicators, commonly used clustering algorithms can be selected. The rule base module is used to store relevant expert knowledge and experience knowledge and to modify the structure of the Bayesian network model. The Bayesian network module, after determining different phenotypes and risk levels, constructs a Bayesian network model for each phenotype separately to facilitate variable analysis and decision support for different patient types; the decision recommendation module uses the Bayesian network model and support formula to make relevant decision recommendations and form a decision recommendation plan. For newly admitted patients, the system collects the patient's personal preoperative status data, including patient demographic data, personal history, medical history, comorbidities, test results and other preoperative information. The decision plan generation module will recommend a decision plan based on the patient's cluster phenotype based on the patient's preoperative data.

9. A computer device, characterized in that: The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the auxiliary decision-making method for multi-decision and multi-objective optimization of the accelerated recovery surgery clinical pathway as described in any one of claims 1-7.

10. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the auxiliary decision-making system for multi-decision and multi-objective optimization of the accelerated recovery surgery clinical pathway as described in claim 8.

Images