AI large model-based drug sensitivity test drug recommendation system, method and device and medium

Abstract

The invention discloses a drug sensitivity test drug recommendation system, method and device based on an AI large model and a medium, relates to the technical field of drug sensitivity tests, is applied to a computer device, and comprises a medical object information collection module used for collecting medical feature information from initial medical information of a medical object by using a feature information extraction AI large model; the medical feature information comprises basic information of a medical object, current and past drug allergy information and disease information; the drug recommendation module is used for determining a preliminary drug recommendation result according to a drug sensitive test result of the medical object and a preset expert database constructed based on a drug use industry standard; the medication recommendation module is used for inputting the medical feature information, the preliminary medication recommendation result and a medication database into a medication recommendation AI large model and outputting a target medication recommendation conclusion; the target drug use recommendation conclusion at least comprises recommended drugs, drug use risk information of the recommended drugs and drug use details. And more accurate and practical drug recommendation is realized.

Description

Technical Field

[0001] This invention relates to the field of drug susceptibility testing technology, and in particular to a drug susceptibility testing drug recommendation system, method, apparatus and medium based on AI large model. Background Technology

[0002] With the development of medical information technology, drug sensitivity testing has become an important diagnostic and treatment tool in clinical practice for inferring the effectiveness of antibacterial drugs based on test results. However, the current drug sensitivity testing drug recommendation system has insufficient accuracy, high drug risk, and high dispersion and poor readability of the recommendation conclusions. It cannot provide clear and intuitive reference for medical staff, nor can it dynamically adjust the medication plan according to the actual drug effect of the patient. At the same time, the drug recommendation information is not detailed enough, and medical staff still need to supplement and infer key information such as drug dosage based on experience, which increases the workload of medical staff.

[0003] In summary, how to achieve more accurate and practical medication recommendations is a problem that needs to be solved in this field. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a drug recommendation system, method, apparatus, and medium for drug sensitivity testing based on an AI large-scale model, to achieve more accurate and realistic drug recommendations. The specific solution is as follows:

[0005] In a first aspect, this application discloses a drug recommendation system for drug sensitivity testing based on an AI large model, applied to a computer device, comprising:

[0006] The medical patient information collection module is used to collect medical feature information from the initial medical information of the medical patient using a feature information extraction AI model; wherein, the medical feature information includes at least one of the following: basic information of the medical patient and current and past drug allergy information and disease information;

[0007] The drug recommendation module is used to determine preliminary drug recommendation results based on the drug sensitivity test results of the medical subject and a preset expert database; wherein, the preset expert database is constructed based on industry standards for drug use.

[0008] The medication recommendation module is used to input the medical feature information, the preliminary drug recommendation results, and the drug information already available in the drug database into the medication recommendation AI model, and output the target medication recommendation conclusion; wherein, the target medication recommendation conclusion includes at least the recommended drug, the medication risk information of the recommended drug, and medication details.

[0009] Optionally, the recommended conclusion for the target medication may also include contraindications.

[0010] Optionally, the AI-based large-scale model-based drug sensitivity testing recommendation system further includes:

[0011] The model function application optimization module is used to obtain the comprehensive evaluation results of the target drug recommendation conclusions in the preset initial stage of online deployment, and to determine whether the feature information extraction AI model and the drug recommendation AI model meet the preset compliance conditions based on the comprehensive evaluation results. If the preset compliance conditions are not met, a non-compliance conclusion is generated. If the main item in the non-compliance conclusion is a contraindicated drug, the feature information extraction AI model is determined as a model to be optimized. If the main item in the non-compliance conclusion is not a contraindicated drug, the drug recommendation AI model is determined as a model to be optimized, and the model to be optimized is fine-tuned.

[0012] Optionally, the model function application optimization module includes:

[0013] The item evaluation unit is used to evaluate each drug recommendation item in the target drug recommendation conclusions within the preset initial stage of online launch, so as to obtain the evaluation result; wherein, the evaluation result includes credible, neutral and unreliable;

[0014] The scoring unit for a single medication recommendation conclusion is used to quantify and average the evaluation results of each medication recommendation item in a single target medication recommendation conclusion to obtain the score of the single medication recommendation conclusion.

[0015] The comprehensive evaluation result generation unit is used to calculate the average score of each of the single medication recommendation conclusions, and to determine the average score as the credibility of the medication recommendation conclusion. The robustness is determined based on the median absolute deviation of the scores of each of the single medication recommendation conclusions.

[0016] Optionally, the model function application optimization module includes:

[0017] The model incremental fine-tuning unit is used to obtain the judgment reasons and correct conclusions for the drug recommendation items whose judgment results are unreliable, and to perform incremental fine-tuning on the model to be optimized using the judgment reasons and correct conclusions.

[0018] Optionally, the AI-based large-scale model-based drug sensitivity testing recommendation system further includes:

[0019] The medication feedback module is used to obtain the medication effects reported by each medical subject within a preset stable phase based on the target medication recommendation conclusion. If there is adverse reaction information in the medication effects, the target medication recommendation conclusion is adjusted using the adverse reaction information to obtain an optimized target medication recommendation conclusion.

[0020] Optionally, the medication recommendation module is further configured to store the target medication recommendation conclusion, evaluation result, evaluation reason, correct conclusion, and medication effect in the medication recommendation conclusion storage database.

[0021] Optionally, the medical subject information collection module includes:

[0022] The multimodal medical record digitization processing unit is used to parse and extract text information from the multimodal raw medical information of medical subjects to obtain initial medical information; wherein, the multimodal raw medical information is any one or more of the following: images, handwritten text, tables, and voice information;

[0023] A feature information storage unit is used to store the medical feature information in a feature information database; wherein, the feature information database is a relational database.

[0024] Optionally, the feature information extraction AI big model is obtained by incremental fine-tuning and instruction fine-tuning of the pre-trained big language model for medical information extraction.

[0025] Optionally, the medication recommendation AI model is obtained by incremental fine-tuning and instruction fine-tuning of a pre-trained large language model for medical question answering.

[0026] Secondly, this application discloses a drug recommendation method for drug sensitivity testing based on an AI large model, including:

[0027] The AI ​​model for feature extraction collects medical feature information from the initial medical information of medical subjects; wherein, the medical feature information includes basic information of the medical subject and at least one of current and past drug allergy information and disease information;

[0028] Preliminary drug recommendations are determined based on the drug sensitivity test results of the medical subjects and a pre-set expert database; wherein, the pre-set expert database is constructed based on industry standards for drug use.

[0029] The medical feature information, the preliminary drug recommendation results, and the drug information already available in the drug database are input into the drug recommendation AI model, and the target drug recommendation conclusion is output. The target drug recommendation conclusion includes at least the recommended drug, the drug use risk information, and the drug use details.

[0030] Thirdly, this application discloses a computer device, comprising:

[0031] Memory, used to store computer programs;

[0032] A processor is used to execute the computer program to implement the steps of the aforementioned disclosed method for recommending drug use in drug sensitivity testing based on a large AI model.

[0033] Fourthly, this application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the aforementioned disclosed method for recommending drug use for drug sensitivity testing based on an AI large model.

[0034] The beneficial effects of this application are as follows: Leveraging the advantages of AI models for feature information extraction, which have been pre-trained on massive amounts of medical data, and combining the high semantic generalization and contextual understanding capabilities of these AI models, basic information, drug allergy information, and symptom information of medical subjects can be accurately extracted from their initial medical information, laying a reliable data foundation for subsequent medication recommendations. The medication recommendation module determines preliminary medication recommendations based on drug sensitivity test results and an explicit, structured, and interpretable expert database built upon industry standards for drug use, ensuring the traceability of the reasoning basis. The medication recommendation module will accurately extract medical... Feature information, traceable preliminary drug recommendation results, and drug database information are input into the drug recommendation AI model. Leveraging the AI ​​model's high semantic generalization and contextual understanding capabilities, it can accurately output target drug recommendation conclusions that include recommended drugs, drug risk information, and drug details, based on the individual characteristics of the medical patient. This significantly improves the accuracy and reliability of drug recommendation conclusions. At the same time, through the cascaded reasoning mode of the expert database, the feature information extraction AI model, and the drug recommendation AI model, the reasoning process of the target drug recommendation conclusions can be decomposed and the basis can be traced, greatly improving the interpretability of drug recommendation conclusions. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0036] Figure 1 This is a schematic diagram of a drug recommendation system for drug sensitivity testing based on an AI large model disclosed in this application;

[0037] Figure 2 This is a schematic diagram of a specific AI-based large model-based drug susceptibility testing drug recommendation system disclosed in this application;

[0038] Figure 3 This application discloses a flowchart of a drug recommendation method for drug sensitivity testing based on an AI large model.

[0039] Figure 4 This is a schematic diagram illustrating a specific AI-based large model-based drug recommendation for drug sensitivity testing disclosed in this application;

[0040] Figure 5 This is a schematic diagram illustrating a specific AI large-scale model optimization disclosed in this application;

[0041] Figure 6 This is a schematic diagram illustrating a specific AI large-scale model construction disclosed in this application;

[0042] Figure 7 This is a structural diagram of a computer device disclosed in this application. Detailed Implementation

[0043] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0044] With the development of medical information technology, drug sensitivity testing has become an important diagnostic and treatment tool in clinical practice for inferring the effectiveness of antibacterial drugs based on test results. However, the current drug sensitivity testing drug recommendation system has insufficient accuracy, high drug risk, and high dispersion and poor readability of the recommendation conclusions. It cannot provide clear and intuitive reference for medical staff, nor can it dynamically adjust the medication plan according to the actual drug effect of the patient. At the same time, the drug recommendation information is not detailed enough, and medical staff still need to supplement and infer key information such as drug dosage based on experience, which increases the workload of medical staff.

[0045] Therefore, this application provides a drug recommendation scheme for drug sensitivity testing based on an AI large model, so as to achieve more accurate and realistic drug recommendations.

[0046] See Figure 1 As shown in the figure, this application discloses a drug recommendation system for drug sensitivity testing based on an AI large model, applied to a computer device, including:

[0047] The medical subject information collection module 11 is used to collect medical feature information from the initial medical information of the medical subject using a feature information extraction AI model; wherein, the medical feature information includes at least one of the following: basic information of the medical subject and current and past drug allergy information and disease information;

[0048] The drug recommendation module 12 is used to determine preliminary drug recommendation results based on the drug sensitivity test results of the medical subject and a preset expert database; wherein, the preset expert database is constructed based on industry standards for drug use.

[0049] The medication recommendation module 13 is used to input the medical feature information, the preliminary drug recommendation results, and the drug information already existing in the drug database into the medication recommendation AI big model, and output the target medication recommendation conclusion; wherein, the target medication recommendation conclusion includes at least the recommended drug, the medication risk information of the recommended drug, and medication details.

[0050] The following is an explanation of the collection of information on medical subjects.

[0051] In this embodiment, the medical object information acquisition module includes: a multimodal medical record digitization processing unit, used to parse and extract text information from the multimodal raw medical information of the medical object to obtain initial medical information; wherein, the multimodal raw medical information is any one or more of the following: images, handwritten text, tables, and voice information; and a feature information storage unit, used to store the medical feature information in a feature information database; wherein, the feature information database is a relational database.

[0052] As a core component of the patient information collection module, the multimodal medical record digitization processing unit performs text information parsing and extraction on any one or more types of multimodal raw medical information, such as images, handwritten text, tables, and voice information. Specifically, it can parse and extract information from scanned copies of paper medical records, images of physical examination reports, handwritten medical records, physical examination data tables, video consultation recordings, and electronic consultation records. If the original medical information is paper data, it can be converted into electronic version data using OCR (Optical Character Recognition) and handwritten font recognition methods to complete the parsing and extraction, and finally convert it into standard format electronic initial medical information.

[0053] By leveraging a large-scale AI model for feature extraction—a core component of the patient information collection module—medical feature information is collected from the initial medical information obtained through multimodal medical record digitization. This medical feature information includes basic information about the patient (such as age, weight, body mass index, liver and kidney function, pregnancy status, smoking history, and alcohol consumption history, or any one or more of these), as well as at least one of current and past drug allergies and current and past symptoms. The collected medical feature information is categorized, processed, and stored in a feature information database. It is evident that by utilizing the medical semantic understanding and information extraction capabilities of the large-scale AI model, key medical feature information can be accurately extracted, avoiding omissions and errors inherent in traditional manual and algorithmic extraction. This ensures the comprehensiveness and accuracy of the medical feature information, providing precise input data for the medication recommendation AI model and enabling feature information traceability through database storage. This helps improve the accuracy and interpretability of medication recommendations and reduces medication risks.

[0054] It is important to note that the aforementioned feature information extraction AI model is obtained by incremental and instruction-based fine-tuning of a pre-trained large language model for medical information extraction. The incremental fine-tuning stage uses publicly available medical datasets and hospital-owned medical records and physical examination information annotations as training data, specifically enhancing the model's information extraction capabilities in the medical vertical domain, enabling it to more accurately identify key entities and relationships in medical information. The instruction-tuning stage employs an instruction-tuning format to construct data with various expressions, allowing the model to output medical feature information in a standard format, avoiding inconsistent output formats. Based on the massive medical corpus already learned by the pre-trained large model, after these two fine-tuning stages, the model possesses stronger medical semantic generalization and contextual understanding, as well as ER (Entity-Relation) information extraction capabilities in the medical field. This allows for efficient and accurate extraction of medical feature information, providing precise data support for subsequent medication recommendations, reducing errors, and improving the overall accuracy and efficiency of the system's medication recommendations.

[0055] The preliminary drug recommendation process is explained below.

[0056] First, obtain the drug susceptibility test results for the medical subjects. Drug susceptibility testing is performed on the medical subjects to obtain the results, which include information on the bacterial strain causing the infection, the antimicrobial drug, and the corresponding minimum inhibitory concentration (MIC). This may involve data on one or more bacterial strains, or multiple antimicrobial drugs corresponding to one strain. It is understood that the bacterial strain causing the infection is a specific genus / species (e.g., Staphylococcus aureus, Escherichia coli, Streptococcus pneumoniae) isolated and identified from pathogenic microorganism samples collected from the infection site (e.g., sputum, blood, urine, wound pus). In a specific case, the drug susceptibility test result might be: the bacterial strain causing the infection is Staphylococcus aureus, the antimicrobial drug is cefuroxime, and the MIC value is ≤1.

[0057] Secondly, the drug sensitivity test results of medical subjects are combined with a pre-set expert database to determine preliminary drug recommendations. This pre-set expert database is maintained by professionals and built based on industry standards for drug use. Specifically, it is based on domestic and international drug use industry standards such as CLSI (Clinical and Laboratory Standards Institute) and UCAST, and includes explicit, structured, and interpretable reasoning knowledge such as breakpoints, expert rules, and ECV epidemiological threshold databases. This database is updated as industry standards are updated. First, the antibiotic susceptibility interpretation of a strain to a specific antimicrobial drug (S) or resistance (R) is derived from the inflection points in the pre-defined expert database. Then, using this interpretation as a prerequisite, and combining additional experiments, sample type, and applicable symptoms, expert rules with higher priority than the inflection points are applied (e.g., if strain A is judged to be sensitive to a certain antimicrobial drug B according to the inflection point, but the expert rule clearly states that strain A is naturally resistant to a certain antimicrobial drug B, the expert rule conclusion prevails). Finally, a preliminary drug recommendation result is determined, including the strain's sensitivity / resistance conclusion to each antimicrobial drug, the expert rule interpretation, and the strain's resistance trend. This result is also stored in the drug recommendation conclusion storage database. The expert database, built based on industry standards, provides authoritative and interpretable reasoning basis for the preliminary drug recommendation, ensuring the professionalism and compliance of the preliminary recommendation result. Furthermore, the combination of inflection points and expert rules (with expert rules taking priority) avoids judgment bias that may be caused by a single basis. At the same time, the result storage facilitates the traceability of subsequent drug recommendation basis and expert evaluation reference, laying a reliable foundation for the accurate reasoning of the subsequent drug recommendation AI model and improving the overall accuracy and credibility of drug recommendations.

[0058] The following section explains the process of generating the target drug recommendation conclusion.

[0059] Medical characteristic information, preliminary drug recommendation results, and existing drug information from the drug database are all input into the drug recommendation AI model. The drug database includes two parts: a comprehensive drug database based on sources such as the NMPA (National Medical Products Administration) drug database and a local (hospital-level) drug database. The drug database is accessed via a RAG middleware that receives requests from the drug recommendation AI model, queries the database, assembles the results, and sends them back to the model. Therefore, in outputting the target drug recommendation conclusion, the model first considers the medical characteristic information, preliminary drug recommendation results, and the optimal drug recommendation principle, starting with the drug... The entire database is matched to obtain initially matched drugs. If the same drug also exists in the local (hospital-wide) drug database, it is used directly. If the drug does not exist in the local (hospital-wide) drug database, the optimal alternative drug to the initially matched drug is searched. If no alternative drug is found, a suggestion is made in the medication recommendation conclusion. If an alternative drug exists, it is used to replace the initially matched drug to obtain the final recommended drug. This generates a target medication recommendation conclusion that includes at least the recommended drug, its medication risk information, and medication details. The target medication recommendation conclusion is in a standardized format. In this way, the workload of medical staff in medication judgment is reduced, the local drug database connection improves medication accessibility, and the conclusion is traceable, which helps to improve the accuracy of medication recommendations and reduce medication risks.

[0060] Furthermore, in this embodiment, the recommended conclusion for the target medication also includes contraindications.

[0061] The generated target medication recommendation conclusion is presented in a standardized format, covering four parts: recommended drugs, medication risk information, medication details, and contraindications. The recommended drug section combines drug sensitivity test results and medical characteristic information to clearly define the recommended drugs and the basis for the recommendation. If the local drug database lacks the corresponding drug, a shortage warning or a similar alternative drug will be provided. The medication risk information is based on domestic and international adverse drug reaction databases and medical characteristic information, describing potential adverse reactions or precautions. The medication details strictly follow the doctor's prescription format, based on data from national standards (such as the National Pharmacopoeia Commission database and the Administrative Measures for the Clinical Application of Antibacterial Drugs), clearly indicating the drug name, dosage, method of administration, and frequency of administration. The data on medication contraindications comes from two sources: firstly, drug allergy history and contraindication information extracted from medical characteristic information by a feature information extraction AI model; and secondly, the medication recommendation AI model infers from the patient's basic information, current / past illnesses, and adverse reactions (not the first time the medication was recommended and existing). In one scenario, the target medication recommendation conclusion is as follows:

[0062] Recommended drugs: [1. Meropenem (sensitive to strain A, with good efficacy)]

[0063] Meropenem: Insufficient stock; similar alternatives are recommended, such as imipenem-cilastatin.

[0064] Contraindicated medications: [1. Aspirin (contraindicated in patients with asthma);]

[0065] Medication risk information: [1. Amoxicillin may cause skin rash;]

[0066] 2. Renal function needs to be monitored after taking cefuroxime;

[0067] 3. Cefuroxime may cause hearing loss;

[0068] Medication details: [1. Imipenem-cilastatin 0.5g intravenous drip, once every 6 hours;

[0069] 2. Cefuroxime 0.25g intravenously twice daily;

[0070] It is evident that the standardized conclusion format covers key information throughout the entire medication process, enabling medical staff to quickly and clearly obtain the core content of medication, significantly reducing the workload of judgment, and avoiding potential medication risks in advance by clearly identifying contraindications and risks. It also improves the feasibility of medication by combining recommendations with local drug conditions, thus comprehensively ensuring the accuracy, safety and convenience of medication.

[0071] It is important to note that the aforementioned medication recommendation AI model is obtained through incremental and instruction-based fine-tuning of a pre-trained medical question-and-answer language model. Based on this pre-trained model, which excels in conversational healthcare, the model first undergoes incremental fine-tuning using labeled domestic and international adverse drug reaction (ADR) / allergy-related databases to enhance its medication risk reasoning ability. Simultaneously, it fine-tunes its dosage reasoning ability by incorporating national standard data such as the Chinese Guidelines for Clinical Application of Antimicrobial Drugs, the National Pharmacopoeia Commission database, and the Administrative Measures for Clinical Application of Antimicrobial Drugs. Then, it employs instruction-tuning to construct data in various expression formats for instruction-based fine-tuning, and adds prompt constraints to ensure standardized output format. The final result is an AI model capable of recommending medications in the field of drug sensitivity testing. This dual fine-tuning allows the model to retain the dialogue understanding and reasoning foundation of the pre-trained medical question-and-answer model while accurately adapting to drug sensitivity testing scenarios, improving the accuracy of medication risk assessment and dosage calculation. Furthermore, the standardized output format facilitates efficient acquisition of key information by medical personnel, reducing medication judgment errors and workload.

[0072] Furthermore, the drug recommendation system for drug sensitivity testing based on the AI ​​big model also includes: a model function application optimization module, used to obtain the comprehensive evaluation results of the target drug recommendation conclusions in the preset initial stage of online deployment, and to determine whether the feature information extraction AI big model and the drug recommendation AI big model meet the preset compliance conditions based on the comprehensive evaluation results. If the preset compliance conditions are not met, a non-compliance conclusion is generated. If the main item in the non-compliance conclusion is a contraindicated drug, the feature information extraction AI big model is determined as a model to be optimized. If the main item in the non-compliance conclusion is not a contraindicated drug, the drug recommendation AI big model is determined as a model to be optimized, and the model to be optimized is fine-tuned.

[0073] In the initial phase of the pre-set launch, doctors or experts with certain professional qualifications first evaluate the credibility of each recommendation item (i.e., recommended drugs, contraindicated drugs, drug risk information, and drug details) in each target drug recommendation conclusion within this phase, thereby obtaining a comprehensive evaluation result for each target drug recommendation conclusion. Based on the pre-set compliance conditions, it is determined whether the feature information extraction AI model and the drug recommendation AI model meet the standards. If they do not meet the standards, a non-compliance conclusion is generated. If the main item in the non-compliance conclusion is a contraindicated drug (since the contraindicated drug data mainly comes from the feature information extraction AI model), then the feature information extraction AI model is identified as a model to be optimized. If the main item is not a contraindicated drug, then the drug recommendation AI model is identified as a model to be optimized. Subsequently, the model to be optimized is fine-tuned using the labeled data (such as correct conclusions from expert feedback, relevant incremental data, and domestic and international standard data) in the drug recommendation conclusion storage database. By accurately identifying problems in the model through conclusions that do not meet the standards, we can ensure a clear direction for model optimization, continuously improve model performance, and thus guarantee the accuracy and reliability of subsequent medication recommendations, thereby reducing medication risks.

[0074] In this embodiment, the model function application optimization module includes: an item evaluation unit, used to evaluate each medication recommendation item in each target medication recommendation conclusion within the preset initial stage of online deployment, to obtain an evaluation result; wherein, the evaluation result includes credible, neutral, and unreliable; a single medication recommendation conclusion score calculation unit, used to quantify and average the evaluation results of each medication recommendation item in a single target medication recommendation conclusion, to obtain a score for a single medication recommendation conclusion; and a comprehensive evaluation result generation unit, used to calculate the average score of each single medication recommendation conclusion, and determine the average value as the credibility of the medication recommendation conclusion, and determine robustness based on the median absolute deviation of the scores of each single medication recommendation conclusion.

[0075] The item evaluation unit is performed by doctors or experts with certain professional qualifications. They evaluate the credibility of each medication recommendation item in the target medication recommendation conclusions in the initial stage of the pre-set online launch. Based on professional judgment, they determine the evaluation results as credible (the item is judged to be professionally credible and reliable), neutral (the credibility of the item is uncertain), or unreliable (the item is judged to be clearly wrong). If an item is judged to be unreliable, they also need to provide feedback on deletion or correction and store it in the medication recommendation conclusion storage database.

[0076] The scoring unit for a single medication recommendation conclusion first quantifies and averages the evaluation results of each item in the single target medication recommendation conclusion. Specifically, it assigns a value of Smax to credible, kSmax to neutral (where k is the quantification coefficient for a neutral evaluation, preferably 0.7-0.9), and 0 to unreliable. Then, it calculates the score of the single medication recommendation conclusion using an averaging formula, which is:

[0077] ;

[0078] in, This represents the score for a single item, where n is the total number of recommendations in a single instance. An example of how to calculate the score for a single medication recommendation conclusion is shown below:

[0079] Recommended drugs: [1. Meropenem (sensitive to strain A, with good efficacy)]

[0080] Meropenem: Insufficient stock, recommend similar alternatives such as imipenem-cilastatin; reliable.

[0081] 2. Cefuroxime (high safety profile), reliable;

[0082] Contraindicated medications: [1. Aspirin (contraindicated in patients with asthma), credible;]

[0083] Medication risk information: [1. Amoxicillin may cause skin rash reactions, which is credible;]

[0084] 2. Cefuroxime requires monitoring of renal function; neutral approach recommended.

[0085] 3. The claim that cefuroxime may cause hearing damage is unreliable.

[0086] Medication details: [1. Imipenem-cilastatin 0.5g intravenous drip, once every 6 hours, reliable;]

[0087] 2. Cefuroxime 0.25g intravenously twice daily is reliable;

[0088] Furthermore, when the confidence level is "confident," the corresponding score is 100; when the confidence level is "neutral," k is 0.8, and the corresponding score is 80; when the confidence level is "disagree," the corresponding score is 0, as detailed below:

[0089] .

[0090] The comprehensive evaluation result generation unit generates credibility and robustness based on the scores of each individual medication recommendation conclusion. Specifically, it calculates the average score of each individual medication recommendation conclusion and determines the average score as the credibility of the medication recommendation conclusion, using the following formula:

[0091] ;

[0092] Where N represents the number of scores for a single medication recommendation conclusion, and the credibility score of the medication recommendation conclusion directly reflects the degree of consistency between the model's recommendation conclusion and medical professional judgment. Robustness is determined based on the median absolute deviation of the scores for each single medication recommendation conclusion, as shown below:

[0093] ;

[0094] Where median() is the function that takes the median value, and If most scores are concentrated in 80 or 100, the MAD value will be small, R will be close to 1, and the robustness will be strong. When the credibility score of the target drug recommendation model is >95 and R>0.9, the credibility performance of the AI ​​large model drug recommendation is considered stable. In other words, this score can reflect the stability of the model's recommendation performance under different sample input scenarios. The credibility score and robustness score together constitute the quantitative evaluation result of the model performance, ensuring that the performance evaluation of the target drug recommendation model is both accurate and comprehensive, and providing a clear quantitative basis for whether the model meets the requirements of clinical application.

[0095] In this embodiment, the model function application optimization module includes: a model incremental fine-tuning unit, used to obtain the judgment reasons and correct conclusions of the drug recommendation items whose judgment results are unreliable, and to use the judgment reasons and correct conclusions to incrementally fine-tune the model to be optimized.

[0096] The model incremental fine-tuning unit first retrieves relevant data from the medication recommendation conclusion storage database for medication recommendation items that were judged as unreliable in the initial stage of online deployment. This data includes the reasons for the judgment marked by experts for the unreliable items (such as "inconsistent with existing clinical evidence or pharmacological mechanisms, there is no clinical support for cefuroxime to cause hearing loss") and the corresponding correct conclusions. At the same time, it combines the medical feature information stored in the database as incremental data. After organizing this data into a format that meets the requirements of model training, it performs incremental fine-tuning on the previously judged model to be optimized. Specifically, for the feature information extraction AI model, it will focus on strengthening its ability to extract key features such as drug allergy information and drug contraindication information. For the medication recommendation AI model, it will focus on optimizing its core functions such as medication risk reasoning and recommended drug matching. It is evident that this embodiment can accurately utilize the root cause data of unreliable entries and relevant authoritative data to perform targeted optimization of the model, directly compensate for the performance shortcomings of the model in specific scenarios, avoid the inefficiency caused by indiscriminate training, continuously improve the accuracy of the model in extracting medical features or generating medication recommendation conclusions, thereby ensuring the reliability of the output results of the subsequent medication recommendation system and reducing medication risks.

[0097] Furthermore, in the model functionality optimization module, for the large-scale medication recommendation AI model, to ensure the output format is in a standard format, a Prompt (prompt word constraint) is added to further tighten the input and output format. The Prompt design is as follows:

[0098] {

[0099] Your task is to generate a structured target medication recommendation based on the input of adverse reaction information, drug database data, preliminary drug recommendation results, and medical characteristic information.

[0100] The output must strictly follow the following format to provide medication recommendations; do not provide additional explanations:

[0101] [Recommendation for Targeted Medication]:

[0102] Recommended medications: [List of recommended medications and the rationale; if the local pharmacy database does not contain the corresponding medication, a shortage alert or alternative medication recommendation will be provided.]

[0103] Contraindicated medications: [List of medications that should not be used and the reasons]

[0104] Medication risk information: [List describing potential adverse reactions or precautions]

[0105] Medication details: [List, in doctor's order format, such as medication name + dosage + usage instructions]

[0106] [Input Data]

[0107] Medical characteristics information: {Enter the patient's current age, weight, and current / past medical conditions here}

[0108] Drug database data: {This section retrieves the current drug information database data from the local (hospital-level) database and the full drug database from the NMPA (National Medical Products Administration) database}

[0109] Preliminary drug recommendation results: {Enter the drug susceptibility test conclusion here, for example, "Strain A is sensitive to drug B; strain A is sensitive to drug C; strain A is sensitive to drug D and resistant to drug F"}

[0110] Adverse reactions: {Enter the adverse reaction here, such as "skin rash after medication"}

[0111] }

[0112] If this is a first-time medication recommendation or if the medication has shown good results after use based on the recommendation, the [Adverse Reaction Information] field in [Input Data] should be empty.

[0113] In addition, the drug recommendation system for drug sensitivity testing based on AI big data models also includes a drug feedback module, which is used to obtain the drug effect feedback of each medical subject based on the target drug recommendation conclusion within a preset stable phase. If there is adverse reaction information in the drug effect, the target drug recommendation conclusion is adjusted using the adverse reaction information to obtain an optimized target drug recommendation conclusion.

[0114] The medication feedback module is used to obtain medication effect feedback information from each medical subject after taking medication according to the target medication recommendation conclusion during the preset stable phase of the system. This feedback information is stored in the medication recommendation conclusion storage database. When adverse reaction information (such as skin rash, abnormal kidney function, etc.) is found in the feedback medication effect, the medication feedback module will use this adverse reaction information as key input data and pass it to the medication recommendation AI model in the medication recommendation module. The medication recommendation AI model combines the medical subject's original medical characteristics, drug sensitivity test results, drug database data, and the initial medication recommendation conclusion to perform secondary reasoning and adjustment on the target medication recommendation conclusion. For example, it may replace drugs that may cause adverse reactions, adjust the dosage or method of administration, etc., and finally obtain an optimized target medication recommendation conclusion for medical staff to use as a reference for adjusting medication. It can capture actual problems in the medication process in real time, and avoid the risks of continuous medication by dynamically adjusting the medication recommendation conclusions. At the same time, adverse reaction information can be used to incrementally fine-tune the medication recommendation AI model, further improving the model's ability to cope with complex medication scenarios, realizing continuous optimization of medication recommendations, and ensuring the safety and efficacy of medication for medical patients.

[0115] Furthermore, after administering medication to the patients based on the recommended medication, information on adverse reactions is collected, including the type, severity, duration, whether medication was discontinued, and intervention measures. This information is then quantified according to a pre-defined scoring standard to obtain an adverse reaction score. The quantification formula is:

[0116] ;

[0117] in, The number of times the medication is effective. This represents the number of adverse reactions after medication. Generally, an adverse reaction score > 95 indicates that the system performs well in terms of adverse reactions to the recommended medication; preferably, the adverse reaction score should be greater than 98. In summary, when reliability > 95, robustness > 0.9, and adverse reaction score > 98, the system's overall medication recommendation conclusions are considered reliable. Even when reliability > 95, robustness > 0.9, and adverse reaction score > 98 are simultaneously met, the system still needs to maintain regular monitoring of the above three parameters (reliability score, robustness score, and adverse reaction score) and continue fine-tuning training. Ideally, reliability score = 100, robustness score = 1, and adverse reaction score = 100.

[0118] In this embodiment, the medication recommendation module is also used to store the target medication recommendation conclusion, evaluation result, evaluation reason, correct conclusion, and medication effect in the medication recommendation conclusion storage database.

[0119] The system-generated target medication recommendations, the evaluation results of each medication recommendation item by doctors or experts with certain professional titles, the reasons for the evaluation when an item is judged as unreliable by experts, the correct conclusions marked by experts for unreliable items, and the medication effects reported by medical patients based on the target medication recommendations are all uniformly stored in the medication recommendation conclusion storage database. The database manages this data by dimension, realizing the traceability of key data in the entire medication recommendation process. This provides complete data support for subsequent judgment of modules that need to be optimized in the system, incremental fine-tuning of the feature information extraction AI model and the medication recommendation AI model, and at the same time, it helps to continuously iterate and optimize system performance, ensure the accuracy and reliability of medication recommendation conclusions, and reduce medication risks.

[0120] For example Figure 2As shown, the drug recommendation system for drug sensitivity testing based on an AI large model includes a medical subject information collection module, a drug recommendation module, a medication recommendation module, a model function application optimization module, and a medication feedback module. Among them, the medical subject information collection module includes a multimodal medical record digitization processing unit and a feature information storage unit. The model function application optimization module includes an item evaluation unit, a score calculation unit for word medication recommendation conclusions, a comprehensive evaluation result generation unit, and a model incremental fine-tuning unit.

[0121] The beneficial effects of this application are as follows: Leveraging the advantages of AI models for feature information extraction, which have been pre-trained on massive amounts of medical data, and combining the high semantic generalization and information extraction capabilities of these AI models, basic information, drug allergy information, and symptom information of medical subjects can be accurately extracted from their initial medical information, laying a reliable data foundation for subsequent medication recommendations. The medication recommendation module determines preliminary medication recommendations based on drug sensitivity test results and an explicit, structured, and interpretable expert database built upon industry standards for drug use, ensuring the traceability of the reasoning basis. The medication recommendation module will accurately extract... Medical characteristic information, traceable preliminary drug recommendation results, and drug database information are input into the drug recommendation AI model. Leveraging the AI ​​model's high semantic generalization, contextual understanding, and question-answering capabilities, it can accurately output target drug recommendation conclusions containing recommended drugs, medication risk information, and medication details, significantly improving the accuracy and reliability of the drug recommendation conclusions. Simultaneously, through the cascaded reasoning mode of the expert database, the feature information extraction AI model, and the drug recommendation AI model, the reasoning process of the target drug recommendation conclusions is deconstructible and its basis is traceable, greatly improving the interpretability of the drug recommendation conclusions. Access to the local (hospital-wide) drug database enhances the implementability of the target drug recommendation conclusions.

[0122] Understandably, before using a large-scale AI model for feature extraction to collect medical feature information from the initial medical information of medical subjects, it is necessary to construct a large-scale AI model for feature extraction. In the construction phase of the large-scale AI model for feature extraction, a pre-trained large language model for medical information extraction is selected as the pre-training model for the large-scale AI model for feature extraction. For example, the pre-trained large language model for medical information extraction is a fine-tuned version of Qwen-Med IE. The fine-tuned version of Qwen-Med IE is trained on tens of millions of Chinese electronic medical records and has rich medical corpus and semantics. Then, incremental fine-tuning and instruction fine-tuning are performed on the model using data from a certain number of domestic and foreign standard medical databases, datasets in specific formats, and data from drug recommendation conclusion databases. In this way, the large language model for medical information extraction is obtained, and the large language model for medical information extraction is determined as the large-scale AI model for feature extraction. The incremental fine-tuning primarily utilizes publicly available datasets (such as CMeEE (Chinese Medical Entity Recognition) and CMeIE (Medical Information Extraction), while also employing labeled data based on hospital-owned medical records and physical examination information, especially various clinical descriptions of rare diseases and uncommon drug allergies. This data is used to fine-tune the feature information extraction AI model using NER / RE (Named Entity Recognition / Relation Extraction Fine-tuning). In instruction fine-tuning, data with multiple expression formats is constructed using the instruction-tuning format. The input to the instruction fine-tuning training samples is: context (structured input) + instruction (optional instructions), and the output is: response (fixed JSON). The initial training samples used for instruction fine-tuning are in the tens of thousands. Further instruction fine-tuning will be determined based on the model's output data format and content performance.

[0123] Similarly, before inputting medical feature information, preliminary drug recommendation results, and existing drug information from the drug database into the large-scale drug recommendation AI model, it is necessary to construct the large-scale drug recommendation AI model. During the construction phase, incremental and instruction-based fine-tuning is performed on the pre-trained medical question-and-answer language model to obtain the medical question-and-answer language model, which is then designated as the large-scale drug recommendation AI model. For example, the pre-trained medical question-and-answer language model could be Med-GLM. The training samples for incremental fine-tuning of the large-scale drug recommendation AI model prioritize data from domestic national standard databases related to adverse drug reactions (ADRs) and allergic reactions, such as the labeled National Adverse Drug Reaction Monitoring System, the National Pharmacopoeia Commission database, the Administrative Measures for Clinical Application of Antibacterial Drugs, the Pharmacopoeia / Instructions Database, and the Hospital ADR Reporting System. Secondly, labeled international databases related to adverse drug reactions (ADRs) and allergic reactions, such as SIDER and MedDRA, are used to continuously perform incremental fine-tuning on the large-scale drug recommendation AI model. In instruction fine-tuning, similar to instruction fine-tuning of large AI models for feature information extraction, data with multiple representations is constructed using the Instruction-tuning format. The input of the instruction fine-tuning training samples is: context (structured input) + instruction (optional instructions), and the output is: response (fixed JSON). The number of training samples used for instruction fine-tuning for the first time is around ten thousand. Subsequently, it is necessary to determine whether to continue instruction fine-tuning of the model based on the performance of the model's data output format and content.

[0124] The trial uses models (including the feature extraction AI model and the medication recommendation AI model) that have undergone instruction-based fine-tuning and incremental fine-tuning. During the trial phase, experts manually evaluate and quantify the medication recommendation conclusions output by the models over nearly N months / N thousand times. Then, the model's performance is comprehensively judged based on the expert evaluation quantification values. If it meets the standards, the model is deployed online. If it does not meet the standards, the main types of non-compliant conclusions are identified. If the main type of non-compliant conclusions is contraindicated medication, the feature extraction AI model is fine-tuned and optimized. If the main type of non-compliant conclusions is non-contraindicated medication, the medication recommendation AI model is fine-tuned and optimized.

[0125] See Figure 3 As shown in the embodiment of this application, a method for recommending drugs for drug sensitivity testing based on an AI large model is disclosed, characterized by comprising:

[0126] Step S11: Use the feature information extraction AI model to collect medical feature information from the initial medical information of the medical object; wherein, the medical feature information includes at least one of the following: basic information of the medical object and current and past drug allergy information and disease information.

[0127] In this embodiment, the process of collecting medical feature information from the initial medical information of a medical object using a feature information extraction AI model includes: parsing and extracting text information from the multimodal original medical information of the medical object to obtain the initial medical information; wherein, the multimodal original medical information is any one or more of the following: images, handwritten text, tables, and voice information; and storing the medical feature information in a feature information database; wherein, the feature information database is a relational database.

[0128] The system parses and extracts text information from one or more of the following multimodal raw medical information sources: images of medical subjects (such as scanned copies of paper medical records and images of physical examination reports), handwritten text (such as handwritten medical records), tables (such as physical examination data tables), and voice information (such as video consultation recordings and electronic dialogue consultation records). This information is then transformed into initial medical information. A feature information extraction AI model is then used to collect medical feature information (including basic information of the medical subject, current and past drug allergy information, and symptom information) from this initial medical information. The collected medical feature information is stored in a feature information database, which is a relational database (such as MSSQL or ORACLE). This database is used for the traceability of subsequent medication recommendation data to improve the interpretability of medication recommendations.

[0129] Information in the medical characteristic information that is not directly related to the target medication recommendation conclusion includes basic information about the patient and their condition, such as the patient's current / past conditions, current age, and weight, which serves as input data for the medication recommendation AI model. Characteristic information in the medical characteristic information that is directly related to the target medication recommendation conclusion includes drug allergy history and medication contraindications.

[0130] In this embodiment, the feature information extraction AI model is obtained by incrementally fine-tuning and instruction fine-tuning a pre-trained large language model for medical information extraction. The acquisition process of the feature information extraction AI model is based on the pre-trained large language model for medical information extraction, and incremental and instruction fine-tuning are carried out to obtain a feature information extraction AI model with the ability to accurately extract medical feature information.

[0131] Step S12: Determine preliminary drug recommendations based on the drug sensitivity test results of the medical subject and the preset expert database; wherein, the preset expert database is constructed based on industry standards for drug use.

[0132] The results of drug susceptibility testing on medical subjects include information on the strains that infect the medical subjects, the antimicrobial drugs and their corresponding minimum inhibitory concentrations, and may involve data on one or more strains or multiple antimicrobial drugs corresponding to one strain.

[0133] The pre-built expert database is maintained by professionals and built based on drug use industry standards. Specifically, it is built based on domestic and international drug use industry standards such as CLSI and UCAST, and includes explicit structured and interpretable reasoning knowledge such as inflection points, expert rules, and ECV epidemiological threshold databases under the standards. It will be updated as industry standards are updated.

[0134] The drug susceptibility test results of medical subjects are combined with a pre-set expert database. First, the drug susceptibility interpretation of the strain to a specific antimicrobial drug (S) or resistance (R) is obtained through the breakpoints in the pre-set expert database. Then, using the drug susceptibility interpretation as a prerequisite, combined with additional experiments, sample type, applicable symptoms, etc., expert rules with higher priority than the breakpoints are applied to finally determine the preliminary drug recommendation results, which include the susceptibility / resistance conclusions of the strain to each antimicrobial drug, the expert rule interpretations, and the strain's resistance trend. This result is also stored in the drug recommendation conclusion storage database.

[0135] Step S13: Input the medical feature information, the preliminary drug recommendation results, and the drug information already in the drug database into the drug recommendation AI model, and output the target drug recommendation conclusion; wherein, the target drug recommendation conclusion includes at least the recommended drug, the drug use risk information of the recommended drug, and the drug use details.

[0136] For example Figure 4 As shown, medical characteristic information, preliminary drug recommendation results, and drug information from both the NMPA drug database and the local (hospital) drug database are input into the drug recommendation AI model. The model first matches the drugs from the full database according to the principle of optimal drug use. If the same drug is available locally, it is used directly; otherwise, it searches for the best alternative drug. If no alternative drug is available, a suggestion is given; if one is available, the drug is replaced. Finally, a standardized target drug recommendation conclusion is generated, including recommended drugs, drug risk information, and detailed drug usage. This reduces the workload of medical staff in drug evaluation, improves drug accessibility, ensures the traceability of conclusions, and helps improve the accuracy of drug recommendations and reduce drug risks.

[0137] In this embodiment, the target medication recommendation conclusion also includes medication contraindications. The generated target medication recommendation conclusion is in a standardized format, including four parts: recommended drug, medication risk information, medication details, and medication contraindications. The recommended drug combines drug sensitivity test results and medical characteristic information to clearly identify the drug and the basis for the recommendation, and provides a prompt or similar alternative drug if the drug is unavailable locally. The medication risk information is based on domestic and international adverse drug reaction databases and medical characteristic information to explain potential adverse reactions or precautions. The medication details are marked with the drug name, dosage, method of administration, and frequency according to the doctor's order format. The medication contraindications are inferred from the history of drug allergies and medication contraindication information, and the reasons for the contraindications are also indicated.

[0138] For example Figure 5As shown in this embodiment, the medication recommendation AI model is obtained by incremental fine-tuning and instruction fine-tuning of a pre-trained medical question-and-answer language model.

[0139] The process of obtaining the large-scale AI model for medication recommendation is based on a pre-trained large language model for medical question answering. Incremental fine-tuning and instruction fine-tuning are carried out. The initial fine-tuning sample size is around 10,000. Subsequent optimization is determined based on the model's output format and content performance. Finally, a large-scale AI model for medication recommendation adapted to the medication recommendation scenario is obtained through dual fine-tuning.

[0140] This embodiment also includes: obtaining a comprehensive evaluation result of the target drug recommendation conclusions within the preset initial stage of online deployment, and judging whether the feature information extraction AI model and the drug recommendation AI model meet the preset compliance conditions based on the comprehensive evaluation result. If the preset compliance conditions are not met, a non-compliance conclusion is generated. If the main item in the non-compliance conclusion is a contraindicated drug, the feature information extraction AI model is determined as a model to be optimized. If the main item in the non-compliance conclusion is not a contraindicated drug, the drug recommendation AI model is determined as a model to be optimized, and the model to be optimized is fine-tuned.

[0141] For example Figure 6 As shown, in the initial stage of the pre-set online deployment, doctors or experts with certain professional titles evaluate the credibility of recommended drugs, contraindicated drugs, and other items in the drug recommendation conclusions for each target. After obtaining a comprehensive evaluation result, it is determined whether the two models meet the preset standards. If they do not meet the standards, a non-compliance conclusion is generated. The model to be optimized is determined according to whether the main item is a contraindicated drug (the AI ​​large model for extracting feature information for contraindicated drug issues is optimized, and the AI ​​large model for drug recommendation is optimized for non-contraindicated drug issues). Then, the model to be optimized is fine-tuned using the labeled data in the drug recommendation conclusion storage database. In this way, the problem is accurately located, the optimization direction is clarified, the model performance is continuously improved, and the accuracy and reliability of subsequent drug recommendation conclusions are ensured, and the drug use risks are reduced.

[0142] In this embodiment, obtaining the comprehensive evaluation results of the target drug recommendation conclusions within the preset initial launch stage includes: evaluating each drug recommendation item in each target drug recommendation conclusion within the preset initial launch stage to obtain an evaluation result; wherein, the evaluation result includes credible, neutral, and unreliable; quantifying and averaging the evaluation results of each drug recommendation item in a single target drug recommendation conclusion to obtain a score for a single drug recommendation conclusion; calculating the average score of each single drug recommendation conclusion, and determining the average value as the credibility of the drug recommendation conclusion; and determining robustness based on the median absolute deviation of the scores of each single drug recommendation conclusion.

[0143] The process is conducted by qualified doctors or experts who assess the credibility of each medication recommendation in the initial phase of the pre-set online platform. Based on their professional judgment, they determine whether the recommendation is credible (the item is deemed professionally credible and reliable), neutral (the credibility of the item is uncertain), or unreliable (the item is deemed clearly erroneous). If an item is deemed unreliable, feedback on deletion or correction is required and the information is stored in the medication recommendation database.

[0144] First, the evaluation results of each item in the single target medication recommendation conclusion are quantified and averaged. Specifically, the credibility is assigned a value of Smax, the neutrality is assigned a value of kSmax (where k is the quantification coefficient for a neutral judgment, preferably 0.7-0.9), and the unreliability is assigned a value of 0. Then, the score of a single medication recommendation conclusion is calculated using the averaging formula:

[0145] ;

[0146] in, This represents the score for a single item, where n is the total number of recommendations in a single instance. An example of how to calculate the score for a single medication recommendation conclusion is shown below:

[0147] Recommended drugs: [1. Meropenem (sensitive to strain A, with good efficacy)]

[0148] Meropenem: Insufficient stock, recommend similar alternatives such as imipenem-cilastatin; reliable.

[0149] 2. Cefuroxime (high safety profile), reliable;

[0150] Contraindicated medications: [1. Aspirin (contraindicated in patients with asthma), credible;]

[0151] Medication risk information: [1. Amoxicillin may cause skin rash reactions, which is credible;]

[0152] 2. Cefuroxime requires monitoring of renal function; neutral approach recommended.

[0153] 3. The claim that cefuroxime may cause hearing damage is unreliable.

[0154] Medication details: [1. Imipenem-cilastatin 0.5g intravenous drip, once every 6 hours, reliable;]

[0155] 2. Cefuroxime 0.25g intravenously twice daily is reliable;

[0156] Furthermore, when the confidence level is credible, the corresponding score is 100; when the confidence level is neutral, k is 0.8, and the corresponding score is 80; when the confidence level is unreliable, the corresponding score is 0, as shown below:

[0157] .

[0158] Confidence and robustness are generated based on the scores of each individual medication recommendation. Specifically, the average score of each individual medication recommendation is calculated, and this average is used to determine the confidence level of the medication recommendation. The formula is as follows:

[0159] ;

[0160] Where N represents the number of scores for a single medication recommendation conclusion, and the credibility score of the medication recommendation conclusion directly reflects the degree of consistency between the model's recommendation conclusion and medical professional judgment. Robustness is determined based on the median absolute deviation of the scores for each single medication recommendation conclusion, as shown below:

[0161] ;

[0162] Where median() is the function that takes the median value, and If most scores are concentrated in 80 or 100, the MAD value will be small, R will be close to 1, and the robustness will be strong. When the credibility score of the target drug recommendation model is >95 and R>0.9, the credibility performance of the AI ​​large model drug recommendation is considered stable. In other words, this score can reflect the stability of the model's recommendation performance under different sample input scenarios. The credibility score and robustness score together constitute the quantitative evaluation result of the model performance, ensuring that the performance evaluation of the target drug recommendation model is both accurate and comprehensive, and providing a clear quantitative basis for whether the model meets the requirements of clinical application.

[0163] In this embodiment, the fine-tuning of the model to be optimized includes: obtaining the reasons for the judgment and the correct conclusions of the drug recommendation items whose judgment results are unreliable, and using the reasons for the judgment and the correct conclusions to perform incremental fine-tuning of the model to be optimized.

[0164] First, data on medication recommendations deemed unreliable during the initial online phase is retrieved from the medication recommendation conclusion storage database. This includes expert annotations explaining the reasons for the unreliability and the corresponding correct conclusions. Simultaneously, medical feature information stored in the database is used as incremental data. After organizing this data into a format suitable for model training, incremental fine-tuning is performed on the previously identified model to be optimized. Specifically, the AI ​​model for feature extraction will be strengthened to enhance its ability to extract key features such as drug allergy information and contraindications. The AI ​​model for medication recommendation will be optimized to improve its core functions such as medication risk reasoning and recommended drug matching. This embodiment can accurately utilize the root cause data of unreliable items and relevant authoritative data to specifically optimize the model, directly compensating for performance shortcomings in specific scenarios, avoiding inefficiency caused by indiscriminate training, continuously improving the accuracy of the model in extracting medical features or generating medication recommendation conclusions, thereby ensuring the reliability of the subsequent medication recommendation system output and reducing medication risks.

[0165] This embodiment also includes: obtaining the medication effect feedback of each medical subject based on the target medication recommendation conclusion within a preset stable phase; if there is adverse reaction information in the medication effect, the target medication recommendation conclusion is adjusted using the adverse reaction information to obtain an optimized target medication recommendation conclusion.

[0166] After the system enters the preset stable phase, it acquires feedback information on the medication effects of each medical subject after taking the medication according to the target medication recommendation conclusion. This feedback information is stored in the medication recommendation conclusion storage database. When adverse reaction information is found in the feedback medication effects (such as skin rash, abnormal kidney function, etc. after medication), the medication feedback module will use this adverse reaction information as key input data and transmit it to the medication recommendation AI model in the medication recommendation module. The medication recommendation AI model combines the medical subject's original medical characteristics, drug sensitivity test results, drug database data, and the initial medication recommendation conclusion to perform secondary reasoning and adjustment on the target medication recommendation conclusion. For example, it may replace drugs that may cause adverse reactions, adjust the dosage or method of medication, etc., and finally obtains the optimized target medication recommendation conclusion for medical staff to use as a reference for adjusting medication. It can capture actual problems in the medication process in real time, and avoid the risks of continuous medication by dynamically adjusting the medication recommendation conclusions. At the same time, adverse reaction information can be used to incrementally fine-tune the medication recommendation AI model, further improving the model's ability to cope with complex medication scenarios, realizing continuous optimization of medication recommendations, and ensuring the safety and efficacy of medication for medical patients.

[0167] Furthermore, after administering medication to the patients based on the recommended medication, information on adverse reactions is collected, including the type, severity, duration, whether medication was discontinued, and intervention measures. This information is then quantified according to a pre-defined scoring standard to obtain an adverse reaction score. The quantification formula is:

[0168] ;

[0169] in, The number of times the medication is effective. This refers to the number of adverse reactions after medication. Generally, an adverse reaction score > 95 indicates that the system performs well in terms of adverse reactions to the recommended medication; preferably, the adverse reaction score should be greater than 98. In summary, when reliability > 95, robustness > 0.9, and adverse reaction score > 98, the overall reliability of the system's target medication recommendations is considered reliable. Even when reliability > 95, robustness > 0.9, and adverse reaction score > 98 are simultaneously met, the system still needs to maintain regular monitoring of the above three parameters (reliability score, robustness score, and adverse reaction score) and continuous fine-tuning training. The monitoring method is as follows: Figure 6 As shown. Ideally, the reliability score is 100, the robustness score is 1, and the adverse reaction score is 100.

[0170] In this embodiment, the target medication recommendation conclusion, evaluation result, evaluation reason, correct conclusion, and medication effect are stored in the medication recommendation conclusion storage database.

[0171] The system-generated target medication recommendations, the evaluation results of each medication recommendation item by doctors or experts with certain professional titles, the reasons for the evaluation when an item is judged as unreliable by experts, the correct conclusions marked by experts for unreliable items, and the medication effects reported by medical patients based on the target medication recommendations are all uniformly stored in the medication recommendation conclusion storage database. The database manages this data by dimension, realizing the traceability of key data in the entire medication recommendation process. This provides complete data support for subsequent judgment of modules that need to be optimized in the system, incremental fine-tuning of the feature information extraction AI model and the medication recommendation AI model, and at the same time, it helps to continuously iterate and optimize system performance, ensure the accuracy and reliability of medication recommendation conclusions, and reduce medication risks.

[0172] Before using a large-scale AI model to extract medical features from the initial medical information of medical subjects, the model must first be constructed. During the construction phase, a pre-trained large-scale language model for medical information extraction (such as a finely tuned version of Qwen-Med IE, trained on tens of millions of Chinese electronic medical records, with a rich medical corpus and semantics) is selected. The IE fine-tuned model is used as a pre-trained model. Then, a certain number of domestic and foreign standard medical databases, specific format datasets, and medication recommendation conclusion databases are used to perform incremental fine-tuning and instruction fine-tuning to obtain a large-scale AI model for feature information extraction. Incremental fine-tuning mainly uses public datasets such as CMeEE (Chinese Medical Entity Recognition) and CMeIE (Medical Information Extraction), combined with hospital-owned medical records and physical examination information annotation data (especially various clinical expressions of rare diseases and rare drug allergies) to carry out NER / RE (Named Entity Recognition / Relation Extraction) fine-tuning. Instruction fine-tuning uses the Instruction-tuning format to construct multi-expression data. The input of training samples is context (structured input) + instruction (optional instructions), and the output is response (fixed JSON). The initial instruction fine-tuning sample size is around 10,000, and subsequent fine-tuning is determined based on the model's output format and content performance.

[0173] Similarly, before inputting relevant information into the medication recommendation AI model, the model also needs to be built: the construction phase is based on a pre-trained large language model for medical question answering (such as Med-GLM), and incremental fine-tuning and instruction fine-tuning are performed to obtain the medication recommendation AI model; incremental fine-tuning prioritizes the use of domestic ADR (adverse drug reaction) / allergic reaction related databases such as the labeled National Adverse Drug Reaction Monitoring System and Pharmacopoeia / Instruction Manual Database, and secondly uses labeled international databases such as SIDER and MedDRA. Instruction fine-tuning is similar to the feature information extraction AI model, using the Instruction-tuning format to construct data with multiple expression modes. The input and output formats of the training samples are the same, with the initial sample size being around 10,000, and subsequent fine-tuning is determined based on the performance of the output format.

[0174] Next, we will test two models that have undergone double fine-tuning. During the trial phase, experts will manually evaluate and quantify the drug recommendation conclusions from nearly N months / N thousand times. The evaluation and quantification values ​​will be used to determine whether the model performance meets the standards. If it meets the standards, it will be launched online. If it does not meet the standards, the model will be optimized according to the main types of the conclusion items that did not meet the standards (the AI ​​model for feature information extraction will be optimized for contraindicated drug issues, and the AI ​​model for drug recommendation will be optimized for non-contraindicated drug issues).

[0175] Therefore, firstly, at the model construction level, the feature information extraction AI model, through incremental and instructional fine-tuning (standardized output format) of the pre-trained large language model for medical information extraction, can accurately extract key medical feature information such as basic information, drug allergy history, and symptom information of medical subjects. The medication recommendation AI model, through targeted incremental and instructional fine-tuning of the pre-trained large language model for medical question answering, possesses professional medication risk reasoning and recommendation capabilities. The accurate adaptation of the two models lays a reliable foundation for medication recommendation. Secondly, at the medication recommendation process level, by combining drug sensitivity test results, multi-dimensional data (medical feature information, drug database information), and the drug library associated with the RAG middleware, standardized target medication recommendation conclusions covering recommended drugs, medication risks, medication details, and contraindications can be generated. This improves medication accessibility and reduces the workload of medical staff in making judgments. Furthermore, in terms of model optimization and data management, the initial stage of pre-launch involves quantifying model performance through expert evaluation, accurately identifying models to be optimized based on the types of items that did not meet the standards, and fine-tuning them using labeled data. In the pre-stabilization stage, the medication feedback module captures adverse reactions and dynamically adjusts recommendation conclusions. Simultaneously, a database storing medication recommendation conclusions ensures traceability of key data throughout the entire process (recommendation conclusions, evaluation results, medication effects, etc.), providing data support for continuous model optimization. Finally, in terms of overall application value, the combination of model accuracy, process standardization, dynamic optimization, and data traceability significantly improves the accuracy and credibility of medication recommendations, effectively mitigates medication risks, ensures medication safety and efficacy for patients, and simultaneously improves the efficiency of medical services.

[0176] Furthermore, embodiments of this application also provide a computer device. Figure 7 This is a structural diagram of a computer device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0177] Figure 7 This is a schematic diagram of a computer device provided in an embodiment of this application. Specifically, it may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the AI-based large model-based drug sensitivity testing recommendation method disclosed in any of the foregoing embodiments.

[0178] In this embodiment, the power supply 23 is used to provide operating voltage for the various hardware devices on the computer device; the communication interface 24 can create a data transmission channel between the computer device and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0179] The processor 21 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 21 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 21 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 21 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, the processor 21 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0180] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored on it include operating system 221, computer program 222 and data 223, etc., and the storage method can be temporary storage or permanent storage.

[0181] The operating system 221 manages and controls the various hardware devices and computer programs 222 on the computer device to enable the processor 21 to perform calculations and processing on the massive amounts of data 223 in the memory 22. It can be Windows, Unix, Linux, etc. The computer program 222, in addition to including a computer program capable of performing the AI-based large-scale model-based drug sensitivity testing recommendation method disclosed in any of the foregoing embodiments, may further include computer programs capable of performing other specific tasks. The data 223 may include data received by the computer device from external devices, as well as data collected by its own input / output interface 25.

[0182] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned disclosed method for recommending drug use for drug sensitivity testing based on an AI large model. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0183] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0184] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly in hardware, software modules executed by a processor, or a combination of both. The software module may be located in random access memory (RAM), memory, read-only memory (ROM), electrically programmable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), register, hard disk, removable disk, CD-ROM (Compact Disc Read-Only Memory), or any other form of storage medium known in the art.

[0185] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0186] The above provides a detailed description of the drug recommendation system, method, device, and medium for drug sensitivity testing based on an AI large model provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only intended to help understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A drug recommendation system for drug sensitivity testing based on an AI large-scale model, characterized in that, Applied to computer devices, including: The medical patient information collection module is used to collect medical feature information from the initial medical information of the medical patient using a feature information extraction AI model; wherein, the medical feature information includes at least one of the following: basic information of the medical patient and current and past drug allergy information and disease information; The drug recommendation module is used to determine preliminary drug recommendation results based on the drug sensitivity test results of the medical subject and a preset expert database; wherein, the preset expert database is constructed based on industry standards for drug use. The medication recommendation module is used to input the medical feature information, the preliminary drug recommendation results, and the drug information already available in the drug database into the medication recommendation AI model, and output the target medication recommendation conclusion; wherein, the target medication recommendation conclusion includes at least the recommended drug, the medication risk information of the recommended drug, and medication details.

2. The drug recommendation system for drug sensitivity testing based on an AI large model according to claim 1, characterized in that, The recommended conclusions for the target medication also include contraindications.

3. The drug recommendation system for drug sensitivity testing based on an AI large model according to claim 2, characterized in that, Also includes: The model function application optimization module is used to obtain the comprehensive evaluation results of the target drug recommendation conclusions in the preset initial stage of online deployment, and to determine whether the feature information extraction AI model and the drug recommendation AI model meet the preset compliance conditions based on the comprehensive evaluation results. If the preset compliance conditions are not met, a non-compliance conclusion is generated. If the main item in the non-compliance conclusion is a contraindicated drug, the feature information extraction AI model is determined as a model to be optimized. If the main item in the non-compliance conclusion is not a contraindicated drug, the drug recommendation AI model is determined as a model to be optimized, and the model to be optimized is fine-tuned.

4. The drug recommendation system for drug sensitivity testing based on an AI large model according to claim 3, characterized in that, The model function application optimization module includes: The item evaluation unit is used to evaluate each drug recommendation item in the target drug recommendation conclusions within the preset initial stage of online launch, so as to obtain the evaluation result; wherein, the evaluation result includes credible, neutral and unreliable; The scoring unit for a single medication recommendation conclusion is used to quantify and average the evaluation results of each medication recommendation item in a single target medication recommendation conclusion to obtain the score of the single medication recommendation conclusion. The comprehensive evaluation result generation unit is used to calculate the average score of each of the single medication recommendation conclusions, and to determine the average score as the credibility of the medication recommendation conclusion. The robustness is determined based on the median absolute deviation of the scores of each of the single medication recommendation conclusions.

5. The drug recommendation system for drug sensitivity testing based on an AI large model according to claim 4, characterized in that, The model function application optimization module includes: The model incremental fine-tuning unit is used to obtain the judgment reasons and correct conclusions for the drug recommendation items whose judgment results are unreliable, and to perform incremental fine-tuning on the model to be optimized using the judgment reasons and correct conclusions.

6. The drug recommendation system for drug sensitivity testing based on an AI large model according to claim 5, characterized in that, Also includes: The medication feedback module is used to obtain the medication effects reported by each medical subject within a preset stable phase based on the target medication recommendation conclusion. If there is adverse reaction information in the medication effects, the target medication recommendation conclusion is adjusted using the adverse reaction information to obtain an optimized target medication recommendation conclusion.

7. The drug recommendation system for drug sensitivity testing based on an AI large model according to claim 6, characterized in that, The medication recommendation module is also used to store the target medication recommendation conclusion, evaluation result, evaluation reason, correct conclusion, and medication effect in the medication recommendation conclusion storage database.

8. The drug recommendation system for drug sensitivity testing based on an AI large model according to claim 1, characterized in that, The medical patient information collection module includes: The multimodal medical record digitization processing unit is used to parse and extract text information from the multimodal raw medical information of medical subjects to obtain initial medical information; wherein, the multimodal raw medical information is any one or more of the following: images, handwritten text, tables, and voice information; A feature information storage unit is used to store the medical feature information in a feature information database; wherein, the feature information database is a relational database.

9. The drug recommendation system for drug sensitivity testing based on an AI large model according to any one of claims 1 to 8, characterized in that, The feature information extraction AI model is obtained by incrementally fine-tuning and instruction fine-tuning a pre-trained large language model for medical information extraction.

10. The drug recommendation system for drug sensitivity testing based on an AI large model according to any one of claims 1 to 8, characterized in that, The medication recommendation AI model is obtained by incremental and instruction-based fine-tuning of a pre-trained large language model for medical question answering.

11. A method for recommending medications for drug sensitivity testing based on an AI large model, characterized in that, include: The AI ​​model for feature extraction collects medical feature information from the initial medical information of medical subjects; wherein, the medical feature information includes basic information of the medical subject and at least one of current and past drug allergy information and disease information; Preliminary drug recommendations are determined based on the drug sensitivity test results of the medical subjects and a pre-set expert database; wherein, the pre-set expert database is constructed based on industry standards for drug use. The medical feature information, the preliminary drug recommendation results, and the drug information already available in the drug database are input into the drug recommendation AI model, and the target drug recommendation conclusion is output. The target drug recommendation conclusion includes at least the recommended drug, the drug use risk information, and the drug use details.

12. A computer device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the drug recommendation method for drug sensitivity testing based on an AI large model as described in claim 11.

13. A computer-readable storage medium, characterized in that, Used to store a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the drug recommendation method for drug sensitivity testing based on an AI large model as described in claim 11.

Images