A refined antibacterial drug prescription review method

Abstract

This invention belongs to the technical field of pharmaceutical assessment methods, and relates to a refined method for reviewing antimicrobial drug prescriptions. Based on input information related to antimicrobial drug prescriptions, this invention intelligently identifies, processes, and learns the patterns in antimicrobial drug prescription review, thereby achieving intelligent review of medical prescriptions. This invention supports multiple antimicrobial drugs and has better robustness. Without the need for pre-setting relevant rules, after inputting data, this invention can automatically identify and process information and learn the potential patterns in antimicrobial drug prescription review. It can output multi-label review results, which helps hospitals to refine the clinical application of antimicrobial drugs. The output results of this invention are standardized, helping hospitals to standardize management decisions.

Description

Technical Field

[0001] This invention belongs to the technical field of pharmaceutical evaluation methods, and relates to a refined method for evaluating antibacterial drug prescriptions. Background Technology

[0002] Antimicrobial resistance is a major challenge in global public health. Infections caused by drug-resistant bacteria have higher mortality rates and are more expensive to treat, and the increasing resistance could lead to a situation where no effective treatments are available. The information management system for the clinical use of antimicrobial drugs in medical institutions is still imperfect, and the rationality of antimicrobial drug use needs to be improved. Existing methods or software cannot adequately assess preoperative patients, such as whether they have underlying diseases, immunodeficiency such as diabetes, or are elderly; nor can they provide comprehensive information about intraoperative conditions, such as the level of surgery, duration, and whether it was an emergency surgery. Current technologies largely focus on evaluating the rationality of prescriptions, lacking more precise evaluation; furthermore, the lack of standardized prescription processing procedures leads to management chaos.

[0003] Existing literature related to this invention includes:

[0004] Beaudoin,M.,Kabanza,F.,Nault,V.,&Valiquette,L,2016.Evaluation of amachine learning capability for a clinical decision support system to enhance antimicrobial stewardship programs.Artificial intelligence in medicine 68,29-36.

[0005] This study uses a knowledge base and a supervised learning module to discover expert rules for a clinical decision support system. This module extracts classification rules for inappropriate antimicrobial prescriptions based on past recommendations for dose and dosing frequency adjustments, discontinuation of treatment, early conversion from intravenous to oral therapy, and redundant antimicrobial spectrum. The advantage of this approach is its ability to extract clinically relevant rules for multiple types of antimicrobial alerts, and the learned rules have been shown to extend the baseline system's knowledge base by identifying pharmacist interventions missed by the baseline system.

[0006] However, this method only applies to piperacillin sodium and tazobactam sodium, and its effectiveness against other antibacterial drugs is currently unknown. Furthermore, this method only performs a binary classification of whether a drug prescription is appropriate, lacking the ability to extract relevant rules for multiple types of early warning systems.

[0007] Shi,ZY,Hon,JS,Cheng,CY,Chiang,HT,&Huang,HM,2022. ApplyingMachine Learning Techniques to the Audit of Antimicrobial Prophylaxis.Applied Sciences12.5,2586.

[0008] This study applied supervised machine learning classifiers (Auto-WEKA, multilayer perceptron, decision tree, SimpleLogistic, Bagging, and AdaBoost) to an antimicrobial prophylaxis dataset to determine the level of adherence to the antimicrobial practice guidelines based on their recommendations, and validated the results in more than 600 instances.

[0009] The main limitation of this study is that it categorizes adherence to practice guidelines for prophylactic antibiotic use into five levels and only makes predictions for these five levels. Therefore, the prediction results cannot be well adapted to clinical antibiotic management practices.

[0010] Wang Guifeng, Liu Ruifeng, Li Xueqin, Li Yunjing. Evaluation of drug utilization of piperacillin sodium and tazobactam sodium for injection based on weighted TOPSIS method [J]. Chinese Journal of Hospital Pharmacy, 2021, 41(01):52-56. DOI:10.13286 / j.1001-5213.2021.01.10.

[0011] This study used piperacillin sodium tazobactam sodium for injection as the research object. Based on the relevant drug instructions and the "Guidelines for Clinical Application of Antibacterial Drugs" (2015 edition), a detailed evaluation rule for the utilization of piperacillin sodium tazobactam sodium based on the weighted superiority-inferiority distance method (TOPSIS) was constructed to evaluate the use of piperacillin sodium tazobactam sodium in hospitals.

[0012] The main limitation of this study is that it only focuses on the injectable piperacillin sodium / tazobactam sodium, lacking general applicability to the evaluation of other drugs. Furthermore, the use of an attribute hierarchy model (AHM) to establish the weights of the evaluation indicators is somewhat subjective.

[0013] Patents related to existing technologies include:

[0014] CN202111193587.9 A prescription review method based on RPA. This invention provides a prescription review method based on RPA, which uses RPA to obtain the current prescription data table to be reviewed, and performs logical judgment between the patient review information and the corresponding drug review rule database according to the drug review rule database, thereby realizing standardized prescription review.

[0015] CN202110411958.X Regional Prescription Review Management System. This invention relates to a regional prescription review management system that processes prescription information, filters it according to preset sample extraction conditions, and sends prescription review information that meets the sample extraction conditions to the management terminal; after analysis and processing by the management terminal, rule adjustment information and quality control detailed prescription report information are obtained.

[0016] CN201811462365.0 Method, apparatus, device and readable storage medium for judging abnormal medical prescriptions. This invention discloses a method, apparatus, device and readable storage medium for judging abnormal medical prescriptions, which compares a reference baseline generated based on historical medical information from big data and the current medical information of the patient to determine the abnormality of the patient's current medical prescription.

[0017] Differences: The three patents mentioned above primarily use association rule-based methods to evaluate medical prescriptions, requiring pre-setting of these rules. The data used is mostly structured, lacking the ability to mine unstructured data. While the methods can determine whether medical prescriptions conform to relevant rules, the evaluation results are coarse-grained and lack the ability to identify specific reasons for anomalies in antibiotic prescriptions.

[0018] Our system, designed for various antimicrobial prescriptions and tailored to the specific circumstances of hospitals, intelligently identifies internal patterns in both structured and unstructured data. This eliminates the need for pre-set rules and subjective biases. For abnormal antimicrobial prescriptions, our system can specifically identify the anomalies and provide concrete improvement suggestions.

[0019] Detailed review of hospital antibiotic prescriptions is a prerequisite for curbing drug resistance in patients and improving hospital antibiotic clinical application management strategies. Current technologies have the following problems:

[0020] 1. The applicability of the method is limited. Existing studies mostly focus on prescription data for single antimicrobial drugs, and the relevant rules and conclusions obtained lack universality in the management of other antimicrobial drugs.

[0021] 2. It has poor practicality, requiring pre-setting of relevant rules, which increases the workload of relevant medical staff.

[0022] 3. Existing methods only determine whether a prescription is reasonable or abnormal, making it difficult to provide detailed comments.

[0023] 4. Lack of standardization leads to chaotic management. Summary of the Invention

[0024] To address the limitations of traditional rule-based methods for reviewing antimicrobial prescriptions, the inability of current methods to comprehensively consider relevant information and provide refined reviews, and the lack of standardization in the prescription review process, this invention proposes a refined method for reviewing antimicrobial prescriptions. This invention first extensively collects prescription information, patient-related information, surgical information, and drug information to establish a unified and integrated hospital antimicrobial prescription review database. The data is then processed using data preprocessing techniques and methods such as term frequency-inverse document frequency (TF-IDF). Next, the AdaBoost machine learning algorithm is used to learn relevant prescription review knowledge, establishing a refined medical prescription review model, which is then applied to the prescriptions to be reviewed.

[0025] The technical solution adopted in this invention is as follows:

[0026] A refined method for evaluating antimicrobial drug prescriptions, comprising the following steps:

[0027] Step 1: Establish a hospital antimicrobial drug prescription review database

[0028] Based on the Hospital Information Management System (HIS), relevant hospital regulations and procedures, and supplementary knowledge of physicians, this system collects antibiotic prescriptions generated by the hospital, along with related patient, surgical, and drug information. Patient information is linked based on patient hospital number and admission date; surgical information is linked based on patient hospital number, surgical date, and prescription date; and drug information is linked based on antibiotic name. Based on all linked information, the cost and dosage of each antibiotic prescription are calculated. All information is then integrated to establish a hospital antibiotic prescription review database.

[0029] Step 2: Data Preprocessing

[0030] Based on data from the hospital's antimicrobial drug prescription review database, structured data and unstructured data (text data) were distinguished. Then, the structured and unstructured data were preprocessed separately, as follows:

[0031] First, rows containing missing values ​​are removed from the structured data. Then, for numerical data in the structured data, including patient age, height, weight, body mass index (BMI), admission date, discharge date, prescription date, discontinuation date, single dose, total medication dosage, surgery time, and surgery duration, no special processing is performed. For categorical data in the structured data, including patient gender, whether the patient has diabetes, department, medication name, medication frequency, surgical level, surgical procedure grade, and medication grade, dummy variables are introduced for modeling.

[0032] Then, for unstructured data, including text data such as surgery and procedure names and diagnostic information, TF-IDF is used for modeling.

[0033] Specifically: 1) It is necessary to calculate the frequency of a term in the text, that is, to calculate the term frequency (TF). The formula for TF is as follows:

[0034]

[0035] Where n i,j It indicates a certain word t i In document d j The number of times it appears, ∑ k n k,j Document d represents document d j The total number of words in TF. i,j The word t indicates i In document d j The frequency of occurrence.

[0036] 2) It is necessary to calculate the Inverse Document Frequency (IDF) value. IDF represents the prevalence of a particular word, and its formula is as follows:

[0037]

[0038] Where |D| represents the total number of documents, |j:t i ∈d j | indicates that it contains a certain word t i The total number of documents, plus 1, is to prevent the inclusion of a certain word t. i The total number of documents is 0.

[0039] 3) The formula for TF-IDF is as follows:

[0040] TF-IDF = TF·IDF#(3)

[0041] TF-IDF indicates that a word has a high frequency in a specific document but a low frequency across all documents, thus carrying a higher weight. Therefore, TF-IDF can help filter commonly used words in medical diagnosis while retaining key terms, aiding in prescription review. TF-IDF modeling transforms text data such as surgical and procedural names and diagnostic information into vectors.

[0042] Step 3: Multi-tag processing

[0043] For manually reviewed data, since the review results are not limited to a single aspect, for example, a prescription may contain two problems at the same time: the reason for changing medication is not explained and the course of treatment is not explained. Therefore, the review results need to be converted into a multi-label classification format.

[0044] The evaluation aspects involved in this method include whether the drug prescription is reasonable, and the evaluation result index of reasonable prescriptions is set to 0.

[0045] If the prescription is deemed unreasonable, specific aspects include: ① unreasonable drug selection; ② skin test performed despite no history of cephalosporin allergy; ③ unjustified change of drug type; ④ unreasonable timing of medication (not administered 30 minutes to 1 hour before surgery); ⑤ unreasonable single dose; ⑥ unreasonable frequency of medication; ⑦ unreasonable course of treatment; ⑧ medication used without indication. The indices are 1 to 8 respectively. Therefore, when a prescription simultaneously contains issues of not explaining the reason for changing medication and not specifying the course of treatment, the multi-tag classification format for its review is [0,0,0,1,0,0,0,1,0].

[0046] The integrated Label Powerset method is adopted, which builds a separate classifier for each label combination and supports the mutual conversion between multi-label and single-label by creating a mapping relationship.

[0047] The specific processing method is as follows:

[0048] Let y0, ..., y8 represent the nine categories mentioned above, X be the input vector for the predicted labels, and x1, x2, x3, x4 be the review labels for the four prescription data points. Then, according to the multi-label classification format described above, as shown in Table 1:

[0049] Table 1. Classification Format for Multi-Label Categories

[0050] X <![CDATA[y0]]> <![CDATA[y1]]> <![CDATA[y2]]> <![CDATA[y3]]> <![CDATA[y4]]> <![CDATA[y5]]> <![CDATA[y6]]> <![CDATA[y7]]> <![CDATA[y8]]> <![CDATA[x1]]> 1 0 0 0 0 0 0 0 0 <![CDATA[x2]]> 0 0 0 1 0 0 0 1 0 <![CDATA[x3]]> 0 1 0 0 1 0 0 0 0 <![CDATA[x4]]> 0 0 0 1 0 0 0 1 0

[0051] The x1 label indicates that the medication is appropriate; the x2 and x4 labels indicate that the reason for changing the medication and the course of treatment are not explained; the x3 label indicates that the medication selection is inappropriate and the timing of medication is inappropriate.

[0052] Next, the prediction problem is transformed into three different classification problems, with the multi-label classification as follows:

[0053]

[0054] Y' represents the converted single tag.

[0055] Since x2 and x4 have the same label, they are converted to the same natural number. The mapping is stored, and a reverse mapping is generated to ensure the availability of the review results.

[0056] Step 4: Learning and Reviewing Knowledge

[0057] Using the processed data, an AdaBoost machine learning model is built to learn parameters. The model iterates through multiple iterations, integrating multiple weak classifiers (decision trees) to construct a strong classifier. In each iteration, the weights of data points that performed poorly in the previous iteration are increased, while the weights of data points that performed well are decreased, thus improving the classifier. After the model learns the feedback knowledge, the best model is stored, and this model is used to provide machine feedback on the added prescription data. Finally, the feedback results are obtained and output through the inverse mapping relationship obtained in step three.

[0058] The beneficial effects of this invention are:

[0059] This invention proposes a refined method for reviewing antimicrobial drug prescriptions. This method intelligently identifies, processes, and learns review patterns based on input antimicrobial drug prescription information, thereby achieving intelligent review of medical prescriptions. Compared with existing technologies, this invention has the following advantages:

[0060] 1. This invention supports a variety of antibacterial drugs and has better robustness.

[0061] 2. Without the need to pre-set relevant rules, after inputting data, this invention can automatically identify and process information and learn the potential patterns of antimicrobial drug prescription review.

[0062] 3. It can output multi-labeled review results, which helps hospitals to manage the clinical application of antimicrobial drugs more precisely.

[0063] 4. The output results of this invention are standardized, which helps hospitals make standardized management decisions. Attached Figure Description

[0064] Figure 1 This is a framework diagram of the refined antibacterial drug prescription evaluation method of the present invention. Detailed Implementation

[0065] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0066] The following describes the specific implementation scheme of the present invention.

[0067] Example 1:

[0068] The steps for a refined method of reviewing antimicrobial drug prescriptions are as follows:

[0069] The first step is to establish a hospital antimicrobial drug prescription review database.

[0070] Antibiotic prescription data, along with related patient, surgical, and drug information, were collected from a hospital between November 2021 and January 2022. The hospital's pharmacy department provided relevant manual review information. After data processing and correction, the selected antibiotic prescription review variables are shown in Table 3.

[0071] Table 3. Selected Antimicrobial Drug Prescription Review Variables

[0072]

[0073]

[0074] The second step involves processing structured and unstructured data using dummy variables, TF-IDF, and other methods, based on the obtained hospital antimicrobial prescription review database.

[0075] The third step is to integrate multi-label processing methods and establish mapping relationships for the acquired manually labeled multi-label information.

[0076] The fourth step is to integrate machine learning methods, build a model, and learn and evaluate knowledge.

[0077] Example 2:

[0078] This invention proposes a refined method for reviewing antibacterial drug prescriptions. It processes structured and unstructured data from input medical prescription data and achieves refined prescription review through multi-label modeling, including the following steps:

[0079] Step 000: Collect hospital antimicrobial drug prescription data and related patient information, surgical information, drug information, manually labeled multi-label information, hospital rules and regulations, and supplementary knowledge from doctors.

[0080] Step 010: Integrate the collected multi-source data through relevant fields to establish a hospital antimicrobial drug prescription review database.

[0081] Step 020: Based on the established hospital antimicrobial drug prescription review database, identify structured and unstructured data and perform preprocessing.

[0082] Step 030: Based on the obtained manually labeled multi-label information, establish a mapping relationship to obtain the processed single-label information.

[0083] Step 040: Based on the processed data features and single-label information, integrate the machine learning method AdaBoost to build a model and learn and comment on knowledge.

[0084] 2.1 Step 000 includes the following steps:

[0085] Step 001: Collect data from the hospital information management system, hospital regulations and rules related to antimicrobial drugs, and supplementary information from doctors.

[0086] Step 002: Based on the antimicrobial drug prescription data, obtain the patient information, patient surgical information, and drug information involved in the antimicrobial drug prescription.

[0087] 2.2 Step 010 includes the following steps:

[0088] Step 011: Associate patient information, surgical information, and medication information with prescription information according to the relevant fields.

[0089] Step 012: Based on all related information, calculate the cost and dosage of each antibacterial drug prescription using information such as drug treatment course and frequency of use.

[0090] 2.3 Step 020 includes the following steps:

[0091] Step 021: Identify and separate structured data from unstructured data.

[0092] Step 022: Remove missing values ​​from the structured data. No special processing is performed on numerical data within the structured data. For categorical data within the structured data, dummy variables are introduced for modeling.

[0093] Step 023: For unstructured data, use TF-IDF for modeling.

[0094] Step 030 in section 2.4 includes the following steps:

[0095] Step 031: Construct a label vector space based on manually labeled multi-label information.

[0096] Step 032: Establish mapping and reverse mapping relationships based on the label vector space.

[0097] Step 033: Based on the established mapping relationship, convert manually labeled multi-label information into manually labeled single-label information.

[0098] Step 040 in section 2.5 includes the following steps:

[0099] Step 041: Based on the structured data, unstructured data, and manually labeled single-label information obtained above, integrate the machine learning method AdaBoost to build a model and learn the parameters.

[0100] Step 042: Use the obtained model to review the prescription to be reviewed, and output the review results according to the inverse mapping relationship.

Claims

1. A method of fine-tuning an antibiotic prescription review, characterized by, The steps are as follows: Step 1: Establish a hospital antibacterial drug prescription review database Based on the hospital information management system (HIS), hospital related rules and regulations and doctor's supplementary knowledge, collect the hospital generated antibacterial drug prescriptions and the related patient information, operation information and drug information involved in the prescriptions; according to the patient's hospitalization number and admission time, associate the patient information; according to the patient's hospitalization number, operation time and prescription time, associate the operation information; according to the name of antibacterial drugs, associate the drug information; according to all the associated information, calculate the cost and drug amount of each antibacterial drug prescription; integrate all the information to establish a hospital antibacterial drug prescription review database; Step 2: Data preprocessing According to the data in the hospital antibacterial drug prescription review database, distinguish between structured data and unstructured data; then, preprocess the structured data and unstructured data respectively, the steps are as follows: Firstly, delete the rows containing missing values in the structured data; then, for the numerical data in the structured data, including patient age, patient height, patient weight, patient body mass index (BMI), admission date, discharge date, prescription time, stop prescription time, drug dose, total drug amount, operation time, operation duration information, no special treatment is needed; for the classification data in the structured data, including patient gender, patient with diabetes, patient department, drug name, drug frequency, operation level, operation and grade, drug grade information, introduce virtual variables for modeling; Then, for the unstructured data, including operation and operation name, diagnosis information text data, use TF-IDF for modeling; Step 3: Multi-label processing For artificial review data, since the review results are not limited to a single aspect, the review results need to be converted to a multi-label classification format; The review aspects involved in this method include whether the drug prescription is reasonable, and the review results of reasonable prescriptions are indexed as 0; If not reasonable, the specific aspects of unreasonable include: ① unreasonable drug selection; ② no cephalosporin allergy still skin test; ③ no reason to change drug varieties; ④ unreasonable drug opportunity; ⑤ unreasonable single dose; ⑥ unreasonable drug frequency; ⑦ unreasonable drug course; ⑧ no indication of drug use; their indexes are 1 to 8 respectively; Therefore, when a prescription contains both the reasons for changing drugs and the course of treatment, the multi-label classification format of its review is [0, 0, 0, 1, 0, 0, 0, 1, 0]; Using the integrated Label Powerset method, a separate classifier is established for each label combination, and a mapping relationship is created to support the mutual conversion between multi-label and single label; Step 4: Learn review knowledge 4.1 According to the structured data, unstructured data and artificial single label information obtained in the foregoing, integrate the machine learning method AdaBoost to establish the model and learn the parameters; 4.2 Use the obtained model to review the prescription to be reviewed, and output the review results according to the reverse mapping relationship.

2. The method of claim 1, wherein the method is a refined antibacterial drug prescription review method, characterized in that, In the second step, the TF-IDF modeling is used, which is as follows: 1) The frequency of the word in the text needs to be calculated, that is, the word frequency (TF) is calculated, and the formula of TF is as follows: where n i,j represents the number of occurrences of a term t i in a document d j ;∑ k n k,j represents the total number of terms in a document d j ;TF i,j represents the frequency of a term t i in a document d j ; 2) The value of inverse document frequency (IDF) needs to be calculated; IDF represents the universality of a certain word, and its formula is as follows: where |D| represents the total number of documents, |j:t i ∈d j represents the total number of documents containing a term t i , and the plus one is to prevent the total number of documents containing a term t i from being zero. 3) The formula of TF-IDF is as follows: TF-IDF=TF·IDF#(3) Through TF-IDF modeling, the text data of operation and operation name, diagnosis information are converted into vectors.

3. The method of claim 1 or 2, wherein the method is a refined antibacterial drug prescription review method, characterized in that, In the third step, the integrated Label Powerset method is used to establish a separate classifier for each label combination, and a mapping relationship is created to support the mutual conversion between multi-label and single label, and the specific processing method is as follows: Let y0,…,y8 be the above nine categories, X be the input vector of the predicted label, x1,x2,x3,x4 be the four prescription data respectively. The x1 label represents reasonable drug use; the x2 and x4 labels represent no explanation for drug replacement and no explanation for the treatment course; the x3 label represents unreasonable drug selection and unreasonable drug use timing; Then, the prediction problem is converted into three different classification problems, that is: Y’ represents the converted single label; Since the labels of x2 and x4 are the same, they are converted to the same natural number, a mapping relationship is stored, and a reverse mapping relationship is generated to ensure the availability of the review results.

4. The method of claim 1 or 2, wherein the method is a refined antibacterial drug prescription review method, characterized in that, In the fourth step, the specific operation is as follows: Using the processed data, the integrated machine learning method AdaBoost is used to establish a model and learn parameters; the model will integrate multiple weak classifiers to build a strong classifier through multiple iterations; for each iteration, the weight of the data point with poor classification in the last iteration is increased, and the weight of the data point with good classification is reduced, so as to achieve the purpose of improvement; after learning the review knowledge by using the model, the best model is stored, and the model is used to machine review the added prescription data to be reviewed; Then, the review results are obtained through the reverse mapping relationship obtained in step three and output.

5. The method of claim 3, wherein the refined antibacterial drug prescription review method is characterized by, In the fourth step, the specific operation is as follows: Using the processed data, the integrated machine learning method AdaBoost is used to establish a model and learn parameters; the model will integrate multiple weak classifiers to build a strong classifier through multiple iterations; for each iteration, the weight of the data point with poor classification in the last iteration is increased, and the weight of the data point with good classification is reduced, so as to achieve the purpose of improvement; after learning the review knowledge by using the model, the best model is stored, and the model is used to machine review the added prescription data to be reviewed; Then, the review results are obtained through the reverse mapping relationship obtained in step three and output.

6. The method of claim 1 or 2 or 5, wherein the method is a refined antibacterial drug prescription review method, characterized in that, In the third step, the unreasonable drug use timing determination standard is: not given within 30min-1h before operation.

7. The method of claim 3, wherein the method is a refined antibacterial drug prescription review method, characterized by, In the third step, the unreasonable drug use timing determination standard is: not given within 30min-1h before operation.

8. The method of claim 4, wherein the refined antibacterial drug prescription review method is characterized by, In the third step, the unreasonable drug use timing determination standard is: not given within 30min-1h before operation.

Images