A method and system for intelligent auditing of the whole process based on intravenous drug orders

Abstract

The application provides a full-process intelligent auditing method and system based on intravenous medication orders, which comprises the following steps: acquiring intravenous medication orders and multidimensional medical record data, generating order feature vectors and patient feature vectors; a rule engine performs static compliance determination to screen the first order; an unsupervised model is used to detect hidden risks to determine physiological parameters that need to be dynamically monitored; physiological parameters are collected in real time, and dynamic detection is performed through the rule engine; corresponding medication control operations are performed according to the dynamic detection results, the unsupervised model is incrementally corrected, and the prediction ability of the unsupervised model is iteratively optimized. Through the triple auditing mechanism of static auditing, dynamic auditing and hidden risk investigation, the application realizes full-process, high-precision and individualized intelligent auditing of intravenous medication orders, effectively solving the core drawbacks of the existing static auditing of intravenous medication orders, such as limited auditing range, low precision, inability to identify hidden risks, lack of dynamic adaptation and iterative optimization ability.

Description

Technical Field

[0001] This application relates to the field of intelligent review of the entire process, and more specifically, to a method and system for intelligent review of the entire process based on intravenous medication orders. Background Technology

[0002] Intravenous administration is one of the most commonly used methods of drug delivery in clinical treatment. Due to its direct entry into the bloodstream and rapid onset of action, it is widely used in the treatment of various diseases. However, the safety requirements for intravenous medication are extremely high. The rationality and compliance of medical orders are directly related to the patient's life safety. If a medical order is irregular or incompatible, it can easily lead to adverse drug reactions, allergic reactions, or even endanger the patient's life. Therefore, the review of intravenous medication orders is a crucial link in the safe management of clinical medication use.

[0003] Currently, the mainstream clinical method for reviewing intravenous medication orders is still mainly static review. This type of review mainly performs a one-time verification of the order after it is written and before the patient takes the medication, using a pre-set rule base. Its review logic is relatively simple, and it can only identify and remove orders with obvious violations. It cannot achieve dynamic control over the entire process of intravenous medication, and the review accuracy is difficult to meet the needs of individualized clinical medication. It has drawbacks such as limited review scope and low review accuracy, making it difficult to achieve precise control over the entire process of intravenous medication from order writing to medication execution, and thus failing to fully guarantee the safety of clinical intravenous medication.

[0004] Therefore, there is an urgent need for a method that can achieve intelligent review of intravenous medication orders with high accuracy throughout the entire process. Summary of the Invention

[0005] To address the problems existing in current technologies, this application provides a method and system for end-to-end intelligent review of intravenous medication orders. The specific solution is as follows: A fully intelligent review method based on intravenous medication orders includes: The system acquires and preprocesses intravenous medication orders to be reviewed and multi-dimensional medical record data of patients to generate order feature vectors and patient feature vectors including static physiological features. The preset rule engine performs static compliance determination on the intravenous medication orders based on the static physiological characteristics and medical order feature vectors, and selects the statically compliant medical orders as the first medical orders; The first medical order is subjected to latent risk detection by a pre-trained unsupervised model, and the first physiological parameter that needs to be dynamically monitored is determined based on the type of latent risk detected. The system collects the first physiological parameters of the patient during the intravenous medication process according to the intravenous medication order in real time, extracts temporal and statistical features from them, and performs dynamic detection through the rule engine to determine whether the patient's current state is suitable for continuing intravenous medication according to the intravenous medication order. Based on the dynamic detection results, corresponding medication control operations are executed. At the same time, the dynamic detection results, time-series features, and statistical features are used as feedback data to incrementally correct the unsupervised model, thereby achieving iterative optimization of the unsupervised model's predictive capabilities.

[0006] In some specific embodiments, the multidimensional medical record data includes baseline data of the patient's liver and kidney function, stage data of underlying diseases, data on drug allergy history, data on adverse reactions to previous intravenous medications, and physiological data on body weight and body surface area.

[0007] In some specific embodiments, the preset rule engine includes a layered and decoupled static explicit rule base and a dynamic adaptive rule base; the static compliance judgment combines baseline data of liver and kidney function, basic disease staging data and physiological data of weight and body surface area from multi-dimensional medical record data to perform individualized explicit compliance verification of drug dosage and infusion rate. At the same time, it verifies drug contraindications based on drug allergy history data and avoids high-risk drug combinations based on past adverse reactions to intravenous drugs, achieving the first layer of precise filtering of explicit risks and eliminating intravenous drug orders with clear violation logic and insufficient suitability.

[0008] In some specific embodiments, the latent risk detection specifically includes: The medical order feature vector, patient feature vector, and pre-set drug pharmacology feature library are fused together and input into the pre-trained unsupervised model. The model's built-in dual-branch detection logic is used to mine individualized pharmacokinetic abnormality risk, potential risk of delayed adverse reaction, and implicit conflict risk of cross-disease combination drug use, and outputs implicit risk detection results including implicit risk type, risk confidence level, and risk association features.

[0009] In some specific embodiments, the latent risk detection further includes: Based on the pre-defined criteria for classifying latent risks, and combining the risk confidence level, probability of occurrence, and clinical harm of latent risks, the detected latent risks are classified into three levels: low, medium, and high. For high-risk latent risks, real-time verification of patient individual characteristics and drug interaction characteristics is strengthened. For medium-risk latent risks, the focus is on verifying the compatibility between drug dosage and patient physiological baseline. For low-risk latent risks, routine characteristic monitoring is performed.

[0010] In some specific embodiments, the statistical features include the mean, coefficient of variation, extreme values, and frequency of anomalies of the first physiological parameter. The time-series features are extracted using a sliding window algorithm to capture the short-term fluctuation trend, peak-to-trough difference, and time-series change point features of the first physiological parameter. The dynamic detection specifically includes: The extracted temporal and statistical features are compared with the static physiological feature baseline; the difference data are verified for compliance by combining the dynamic thresholds corresponding to the drug pharmacological characteristics and dosing cycle in the dynamic adaptation rule base; and the trend prediction algorithm is used to predict the changing trend of the first physiological parameter based on historical time series data to identify potential dynamic abnormality risks in advance.

[0011] In some specific embodiments, the dynamic detection frequency is adaptively adjusted according to the latent risk level throughout the entire intravenous medication execution cycle. When abnormal physiological parameters are detected, the authenticity of the abnormal parameters is first verified by backtracking through time-series trends, and then the correlation between the abnormality and intravenous medication is verified by combining the drug's duration of action with the patient's underlying disease status to avoid misjudgment.

[0012] In some specific embodiments, the pre-trained unsupervised model includes a feature fusion branch and an anomaly detection branch. The feature fusion branch is used to deeply fuse the medical order feature vector, the patient feature vector, and the drug pharmacological features, and strengthens the correlation weight between individualized physiological features and drug action features through an attention mechanism to suppress interference from irrelevant features. The anomaly detection branch uses a combination of density clustering algorithm and isolation forest algorithm. First, density clustering is used to perform preliminary clustering of the fused features to divide normal medication feature clusters and potential anomaly clusters. Then, isolation forest is used to perform fine screening of potential anomaly clusters.

[0013] In some specific embodiments, it also includes: If the patient completes the intravenous medication as prescribed and the dynamic monitoring shows no abnormalities throughout the process, the entire review process ends. If there are abnormalities in the dynamic monitoring results during the medication process, the prescription will be adjusted, and the adjusted prescription will be used as the intravenous medication prescription to be reviewed again. The above steps will be repeated to form a closed-loop review process.

[0014] A fully intelligent review system based on intravenous medication orders, comprising: The input unit is used to acquire intravenous medication orders to be reviewed and multi-dimensional medical record data of patients, and to preprocess them to generate a medication order feature vector and a patient feature vector including static physiological features. The static review unit is used to preset the rule engine to perform static compliance judgment on the intravenous medication order based on the static physiological characteristics and the medical order feature vector, and select the statically compliant medical orders as the first medical orders; The latent detection unit is used to perform latent risk detection on the first medical order using a pre-trained unsupervised model, and to determine the first physiological parameter that needs to be dynamically monitored based on the risk type of the detected latent risk. The dynamic detection unit is used to collect the first physiological parameters of the patient during the intravenous medication process according to the intravenous medication order in real time, extract temporal features and statistical features from them, and perform dynamic detection through the rule engine to determine whether the patient's current state is suitable to continue intravenous medication according to the intravenous medication order; The output unit is used to execute corresponding medication control operations based on the dynamic detection results, and simultaneously use the dynamic detection results, time-series features, and statistical features as feedback data to incrementally correct the unsupervised model, thereby achieving iterative optimization of the unsupervised model's predictive capabilities.

[0015] Beneficial Effects: This application proposes a full-process intelligent review method and system based on intravenous medication orders. Through a triple review mechanism of static review, dynamic review, and hidden risk screening, it achieves full-process, high-precision, and personalized intelligent review of intravenous medication orders, providing comprehensive protection for clinical medication safety. First, a rule engine is used to accurately filter explicit risks before medication. Then, a pre-trained unsupervised model is used to mine hidden risks not covered by the static rule base and configure targeted dynamic monitoring parameters, overcoming the limitations of static review. Simultaneously, dynamic detection is achieved by collecting physiological parameters in real time during medication, extracting temporal and statistical features, realizing real-time risk management during medication and avoiding dynamic safety hazards. The dynamic detection data is used as feedback to incrementally correct the unsupervised model, achieving iterative optimization to adapt to changes in clinical scenarios, ultimately forming a closed-loop review process. This improves the accuracy and comprehensiveness of order review, reduces the workload of medical staff, increases review efficiency, effectively reduces the incidence of adverse drug reactions, and balances safety and practicality.

[0016] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the entire intelligent review process for this application; Figure 2 This is a schematic diagram illustrating the principle of the intelligent review method for the entire process of this application; Figure 3 This is a schematic diagram of the implicit risk detection process for this application; Figure 4 This is a schematic diagram of the dynamic detection process of this application; Figure 5 This is a schematic diagram of the full-process intelligent review system modules for this application.

[0019] Figure reference numerals: 1-Input unit; 2-Static audit unit; 3-Hidden detection unit; 4-Dynamic detection unit; 5-Output unit. Detailed Implementation

[0020] 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 a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0021] This application proposes a full-process intelligent review method based on intravenous medication orders, effectively addressing the core shortcomings of existing static review methods for intravenous medication orders, such as limited review scope, low accuracy, inability to identify hidden risks, and lack of dynamic adaptation and iterative optimization capabilities. It achieves full-process, high-precision, and personalized intelligent review of intravenous medication orders, providing comprehensive protection for clinical medication safety. A flowchart illustrating the full-process intelligent review process based on intravenous medication orders is attached. Figure 1 As shown in the attached diagram, the principle is as follows: Figure 2 As shown, the specific solution is as follows: A fully intelligent review method based on intravenous medication orders includes: 101. Obtain the intravenous medication orders to be reviewed and the patient's multi-dimensional medical record data and preprocess them to generate the order feature vector and the patient feature vector including static physiological features; 102. The preset rule engine performs static compliance judgment on intravenous medication orders based on static physiological characteristics and medical order feature vectors, and selects statically compliant medical orders as the first medical order. 103. Use a pre-trained unsupervised model to detect latent risks in the first medical order, and determine the first physiological parameter that needs to be dynamically monitored based on the risk type of the detected latent risks. 104. Real-time collection of the first physiological parameters of the patient during the intravenous medication process according to the intravenous medication order, extraction of temporal and statistical features, and dynamic detection through a rule engine to determine whether the patient's current state is suitable for continuing intravenous medication according to the intravenous medication order; 105. Execute corresponding medication control operations based on dynamic detection results, and use dynamic detection results, time series features, and statistical features as feedback data to incrementally correct the unsupervised model, thereby achieving iterative optimization of the unsupervised model's predictive capabilities.

[0022] This application discloses a full-process intelligent review method based on intravenous medication orders, which breaks through the limitations of the existing single static review of intravenous medication orders and constructs a closed-loop review mechanism for the entire process from pre-medication to post-medication and from data acquisition to model optimization. Through multi-step collaborative linkage, it achieves accurate review of intravenous medication orders and comprehensive protection of medication safety.

[0023] The core function of step 101 is to provide accurate and standardized basic data support for subsequent full-process review. The intravenous medication orders to be reviewed are the core object of the review, containing all core instruction information related to intravenous medication, such as drug type, dosage, infusion rate, administration cycle, solvent type, and route of administration. This information directly determines the rationality of the medication regimen and is the core basis for the review. The patient's multi-dimensional medical record data is key to achieving individualized review. It contains various basic data that reflect the patient's individual physiological state and medication suitability, supporting consideration of individual patient differences during subsequent review and avoiding inaccuracies caused by uniform review standards. Preprocessing is a crucial step connecting the raw data with subsequent review stages. Its core principle is to solve problems such as inconsistent formats, disorganized information, and invalid interfering data in the original medical orders and medical record data. This is because the original medical orders are mostly unstructured text manually entered by medical staff, and the medical record data also contains various scattered examination reports, medical history records, and other non-standardized information, which cannot be directly recognized and processed by the computer review module.

[0024] The final generated prescription feature vector is used to accurately characterize the core features of intravenous medication orders, while the patient feature vector, containing static physiological characteristics, is used to characterize the individual patient's physiological state. Together, they provide a data foundation for subsequent static compliance assessment and hidden risk detection, avoiding review errors caused by non-standard or incomplete data. The prescription feature vector is a quantitative representation of the core information of intravenous medication orders, with each vector dimension corresponding to a key feature of the prescription, accurately reflecting the details of the medication regimen. The patient feature vector, containing the patient's static physiological characteristics, is a quantitative representation of the individual patient's physiological state, accurately reflecting the patient's basic physical condition. Together, they provide a high-quality data foundation for subsequent static compliance assessment and hidden risk detection.

[0025] Step 102 is the explicit risk filtering step before medication, which is also the first line of defense for the entire process review. Its core purpose is to quickly identify and eliminate intravenous medication orders with clear violations, reduce the medication safety risks caused by obviously non-compliant orders, reduce the workload of subsequent review steps, and ensure that the orders entering the subsequent steps have basic compliance.

[0026] The pre-defined rule engine is the core execution module of this step. Its core principle is based on clinically recognized compliance guidelines for intravenous medication. It pre-builds a standardized audit rule system that can automatically compare, verify, and judge the input feature vectors without manual intervention, thus improving audit efficiency and accuracy. The core logic of static compliance judgment is to combine the medical order feature vector and patient feature vector generated in step 101, and use the rule engine to call the pre-defined audit rules to comprehensively verify the compliance of intravenous medication orders. It focuses on explicit violations that clearly pose safety risks and have unified judgment standards, such as drug dosage exceeding the clinically safe range, drug solvent selection not meeting pharmacological requirements, and administration rate exceeding the patient's tolerance range. Through this standardized verification process, statically compliant and qualified medical orders are selected as the first medical orders. Here, the first medical orders specifically refer to those that have undergone explicit risk filtering before medication, do not have any clear violation logic, and have basic medication safety. The core function of this step is to complete basic safety control before medication, eliminate obviously non-compliant medical orders in advance, and prevent such medical orders from entering the medication process and causing safety accidents. At the same time, it provides qualified audit targets for the subsequent implicit risk detection stage, ensuring that subsequent audits can focus on more hidden and complex risk points, and improve the pertinence and efficiency of the entire process audit.

[0027] Step 103 is a core and crucial step in overcoming the limitations of existing static audits and improving audit accuracy. Its core function is to uncover hidden medication risks that static compliance judgments cannot identify, while providing precise targeting for dynamic monitoring during subsequent medication use. This addresses the core drawback of existing static audits, which can only identify explicit risks but cannot address hidden risks.

[0028] The pre-trained unsupervised model is the core execution module of this step. Its core principle is to enable the model to identify abnormal medication patterns not covered by static rules through prior model training. Pre-training specifically refers to training the model with a large amount of historical intravenous medication case data before it is put into actual use. This historical data includes both normal medication cases and various hidden risk abnormal cases. Through training, the model can learn the characteristic differences between normal medication and hidden risk medication, and has the ability to autonomously identify hidden risks. Moreover, it does not require manual annotation of cases, and can adapt to the scenario in clinical practice where there are diverse types of hidden risks that are difficult to fully annotate in advance.

[0029] The specific process of latent risk detection involves inputting the feature vector of the first medical order and the patient feature vector selected in step 102 into a pre-trained unsupervised model. The model then uses its learned anomaly detection patterns to perform in-depth analysis and judgment on the first medical order, identifying any latent medication risks. These latent risks differ from the explicit risks in step 102; they lack standardized judgment rules, are more concealed, and have a higher probability of occurrence that is more difficult to predict. Examples include the risk of pharmacokinetic abnormalities due to individual patient's abnormal drug metabolism, and the risk of delayed adverse reactions that may result from long-term drug infusion.

[0030] After detecting latent risks, the first physiological parameter to be dynamically monitored is determined based on the type of latent risk detected. The core principle is that different types of latent risks correspond to different changes in the patient's physiological indicators during medication. For example, the risk of delayed-onset allergic reaction corresponds to abnormal fluctuations in physiological indicators such as heart rate and blood pressure, while the risk of pharmacokinetic abnormalities corresponds to changes in indicators related to liver and kidney function. Targeted determination of the first physiological parameter enables precise targeting of subsequent dynamic monitoring, avoiding the waste of resources and inaccurate monitoring caused by indiscriminate monitoring. This allows subsequent dynamic monitoring to focus on physiological indicators directly related to latent risks, ensuring that risk signals can be captured in a timely manner and providing a clear monitoring direction for dynamic management during medication.

[0031] Step 104 is a dynamic risk management step in the medication process. Its core function is to make up for the core deficiency of the existing static review, which only focuses on the pre-medication period and does not focus on the medication process. It realizes dynamic safety management throughout the medication process, timely identifies dynamic risks caused by factors such as drug effects and individual patient reactions, and ensures patient medication safety.

[0032] The core logic of this process revolves around three key steps: real-time data acquisition, feature extraction, and dynamic detection. Real-time acquisition involves collecting the first physiological parameters of the patient during intravenous medication administration as prescribed. Specifically, real-time acquisition refers to continuously collecting these first physiological parameters at regular time intervals throughout the entire intravenous medication cycle. This ensures timely capture of changes in the patient's physiological state during medication administration, preventing missed risk assessments due to untimely data collection. The collected first physiological parameters are the targeted physiological indicators corresponding to the latent risk types determined in step 103, ensuring the relevance and practicality of the data collection.

[0033] Extracting temporal and statistical features is key to achieving accurate dynamic risk assessment. The core principle is that real-time values ​​of a single physiological parameter cannot fully reflect the changing trends and overall condition of a patient's physiological state. For example, normal blood pressure at a certain moment does not mean that the patient's blood pressure has not fluctuated abnormally during medication, nor does it mean that the overall blood pressure level meets the medication requirements. However, temporal and statistical features can complement each other and comprehensively and accurately characterize the patient's physiological state during medication.

[0034] Temporal features are used to capture the fluctuation patterns of physiological parameters over time, reflecting the changing trends of physiological parameters, such as the fluctuation range of blood pressure between 10 and 20 minutes after medication and the changing trend of heart rate over time. Statistical features are used to quantify the overall distribution and abnormalities of physiological parameters, reflecting the overall level and dispersion of physiological parameters, such as the mean, coefficient of variation, extreme values, and frequency of abnormalities of blood pressure over a certain period of time.

[0035] After extracting these features, dynamic detection is performed through a rule engine. The core principle is that the rule engine calls preset dynamic review rules, combines the extracted temporal and statistical features, compares them with the patient's static physiological baseline before medication, and judges the suitability of the patient's current physiological state with the intravenous medication order. This determines whether the patient is suitable to continue taking the medication according to the current intravenous medication order. The core function of this step is to identify dynamic risks in the medication process in real time, realize full safety control of the medication process, avoid safety hazards caused by changes in the patient's physiological state during medication, and ensure the safety of the patient throughout the entire medication process.

[0036] Step 105 is the core step in completing the closed-loop management of the entire process and realizing iterative optimization of the review accuracy. Its core function is to deal with the risks detected dynamically in a timely manner, and at the same time, to continuously improve the accuracy of the unsupervised model in detecting implicit risks through the feedback iteration mechanism, so that the entire review method can continuously adapt to the changes in the actual clinical medication scenarios.

[0037] The principle behind implementing corresponding medication control operations based on dynamic monitoring results is to take appropriate safety management measures based on the level of risk detected dynamically, thereby ensuring patient medication safety. For example, if the dynamic monitoring results show no abnormalities, the current intravenous medication order is continued; if a minor abnormality is detected, appropriate intervention measures are taken and medication is continued; if a serious abnormality is detected, medication is immediately terminated to prevent further risk escalation. Through this targeted medication control operation, risks are managed in a timely manner, ensuring patient medication safety.

[0038] Using dynamic detection results, temporal features, and statistical features as feedback data, the unsupervised model is incrementally corrected. The principle is that the feedback data includes real-world risk cases and patient physiological response data from clinical medication use. This data reflects the risk characteristics in actual clinical medication scenarios and is more timely and targeted than the historical data used during model pre-training. Iteratively updating the unsupervised model using this feedback data allows it to continuously learn new latent risk features, adjust its recognition parameters and patterns, and continuously improve the accuracy of latent risk detection. This incremental correction method does not require retraining the model; it only uses newly added feedback data for iterative optimization. This saves time and resources on model updates and allows the model to continuously adapt to changes in clinical medication scenarios, achieving iterative optimization of the unsupervised model's predictive capabilities.

[0039] In some specific embodiments, multidimensional medical record data includes baseline data on the patient's liver and kidney function, stage data of underlying diseases, data on drug allergy history, data on adverse reactions to previous intravenous medications, and physiological data on body weight and body surface area.

[0040] The patient's baseline liver and kidney function data refers to the basic indicators of liver and kidney function obtained through clinical examinations before this intravenous medication. The liver and kidneys are the main organs for drug metabolism and excretion, and their functional status directly determines the patient's ability to metabolize and tolerate drugs. Patients with different levels of liver and kidney function have significant differences in their tolerance to the same drug dose and the probability of adverse reactions. The role of this data is to provide core evidence for determining whether the dosage and type of drug are suitable for the individual patient during the subsequent review process.

[0041] Underlying disease staging data refers to information on the specific developmental stage of a patient's underlying disease, such as the disease stage in a diabetic patient or the classification in a hypertension patient. Different stages of the underlying disease affect the patient's overall physiological state and response to intravenous medication. For example, patients in the acute phase of an underlying disease have lower drug tolerance and higher medication risks compared to patients in remission. This data is used to support subsequent review processes, allowing for the assessment of the rationality of medication regimens in conjunction with the patient's underlying disease status, thus achieving individualized medication review.

[0042] Drug allergy history data refers to whether a patient has a history of allergies to various drugs, including the types of drugs allergic to and the manifestations of allergic reactions. Intravenous medications enter the bloodstream directly, and allergic reactions occur more quickly and are more harmful. The core function of this data is to provide key evidence for subsequent static compliance assessments, to quickly identify medical orders with drug allergy risks, to preemptively eliminate medical orders related to allergic drugs, and to avoid the occurrence of severe allergic reactions. It is an important basic data for medication safety audits.

[0043] Past adverse drug reaction data refers to records of various adverse drug reactions that occurred when patients previously received intravenous medications, including the corresponding drugs, dosages, reaction manifestations, and treatment methods. The principle behind this data is that the probability of adverse reactions occurring again when a patient has previously experienced adverse reactions to a particular drug or medication regimen is significantly higher. Its purpose is to provide a reference for subsequent review, assist in identifying potentially high-risk medication regimens, and provide individual characteristic support for the detection of latent risks, thereby improving the accuracy of latent risk identification.

[0044] Physiological data on body weight and body surface area refer to the quantitative data of a patient's actual weight and body surface area. Core medication parameters such as the dosage and infusion rate of intravenous drugs usually need to be adjusted individually based on the patient's weight and body surface area. Patients with different weights and body surface areas have significantly different tolerance doses for drugs. The purpose of this data is to provide a quantitative basis for verifying the dosage and infusion rate of drugs in subsequent static compliance assessments, ensuring that medication parameters are adapted to individual patients, avoiding ineffective medication or adverse reactions due to excessive or insufficient dosage, and further improving the accuracy of the review.

[0045] In some specific embodiments, the preset rule engine includes a layered and decoupled static explicit rule base and a dynamic adaptive rule base. The static compliance judgment combines baseline data of liver and kidney function, basic disease staging data and physiological data of weight and body surface area from multi-dimensional medical record data to perform individualized explicit compliance verification of drug dosage and infusion rate. At the same time, it verifies drug contraindications based on drug allergy history data and avoids high-risk drug combinations based on past adverse reactions to intravenous drugs, achieving the first layer of precise filtering of explicit risks and eliminating intravenous drug orders with clear violation logic and insufficient suitability.

[0046] The pre-defined rule engine employs a layered and decoupled architecture, specifically divided into a static explicit rule base and a dynamic adaptation rule base. Layered decoupling means that the two rule bases are independent of each other in terms of functional division, data storage, and execution logic, and do not interfere with each other. The static explicit rule base stores clearly defined and fixed drug compliance standards in clinical practice, covering unchanging review rules such as pharmacopoeia requirements, clinical drug use guidelines, and basic drug contraindications, specifically used for static compliance determination before medication. The dynamic adaptation rule base stores dynamic judgment standards related to the real-time status of the medication process, used for dynamic detection in subsequent medication processes. This layered and decoupled architecture facilitates independent rule updates and maintenance while ensuring the accuracy of rule calls at different review stages, avoiding review errors caused by rule confusion.

[0047] Static compliance determination is the core application of the static explicit rule base. Its execution process combines multi-dimensional patient medical record data to achieve individualized verification, rather than using a uniform and universal audit standard. Specifically, the verification work is completed through two-way comparison of multi-dimensional data and medical order information. It combines baseline data of liver and kidney function, basic disease stage data, and physiological data of weight and body surface area to verify the dosage and infusion rate of medication in the medical order.

[0048] Liver and kidney function determines the ability to metabolize and excrete drugs, the stage of underlying diseases reflects the patient's overall physiological tolerance, and body weight and body surface area are the core basis for quantifying drug dosage. The above physiological data vary among different individuals, and the corresponding safe drug dosage and infusion rate are also different. Individualized verification is to match the drug dosage and infusion rate in the doctor's order with the patient's own physiological data to determine whether it meets the individualized safety range of the patient, rather than just meeting general clinical standards. This avoids the problem that the doctor's order meets general standards but is not suitable for the individual patient.

[0049] Meanwhile, static compliance assessments will verify medication contraindications based on drug allergy history data. Drug allergy is a key explicit risk to safety in intravenous medication. By comparing the types of drugs in the prescription with the patient's drug allergy history, prescriptions containing allergenic drugs can be directly identified and blocked, preventing the occurrence of severe allergic reactions from the source. It will also avoid high-risk drug combinations based on past adverse reaction data of intravenous medication. If a patient has experienced adverse reactions to a drug or combination of drugs in the past, the probability of risk will be greatly increased when used again. By comparing such historical data, high-risk drug combinations for the patient can be accurately identified and eliminated.

[0050] Through the aforementioned multi-dimensional and individualized explicit compliance verification, this embodiment can achieve the first layer of precise filtering of explicit risks. It removes all intravenous medication orders that clearly violate clinical medication guidelines or that meet general standards but are incompatible with the individual patient's physiological condition. Only orders that are basically compliant and suitable for the individual patient are retained for subsequent review. This process not only strengthens the security of static review but also improves the level of individualization of review, reduces the review pressure in subsequent stages, and makes up for the review loopholes caused by traditional static review relying solely on general rules and ignoring individual patient differences. It builds a solid first line of defense for intelligent review of intravenous medication orders before medication, and improves the accuracy and reliability of the overall review.

[0051] The rule engine is not a single program module, but a complete system composed of three core modules: a rule base, an inference engine, and a data adaptation layer, as well as a visual rule management interface. The rule base is the core carrier for storing various intravenous medication review rules, responsible for storing static explicit rules and dynamic adaptation rules in a hierarchical manner according to function. The inference engine is the core of rule execution, responsible for calling rules in the rule base, matching input data, and completing automated judgment. The data adaptation layer is the bridge connecting the rule engine with external data (medical order feature vectors, patient feature vectors, real-time physiological parameter features, etc.), responsible for data format conversion and legality verification. The rule management interface provides a visual operation entry point for medical staff or maintenance personnel, supporting management operations such as adding, modifying, and deleting rules.

[0052] In some specific embodiments, latent risk detection specifically includes: fusing the feature vectors of medical orders and patients with a pre-defined drug pharmacology feature library, inputting the fusion into a pre-trained unsupervised model, and using the model's built-in dual-branch detection logic to separately mine individualized pharmacokinetic abnormality risks, potential risks of delayed adverse reactions, and latent conflict risks of cross-disease combination therapy, outputting latent risk detection results that include latent risk type, risk confidence level, and risk association features. The latent risk detection process is attached. Figure 3 As shown.

[0053] The feature fusion process integrates prescription feature vectors, patient feature vectors, and a pre-built drug pharmacology feature library. Prescription feature vectors represent the specific intravenous medication regimen, patient feature vectors represent the patient's individual physiological and medical history, and the drug pharmacology feature library is a pre-constructed collection of professional data including drug metabolism pathways, mechanisms of action, pharmacological properties, and compatibility characteristics. The principle of feature fusion is to integrate the core information from the medication regimen, the individual patient, and the drug itself into a complete set of comprehensive feature vectors. A single-dimensional feature cannot fully reflect the fit between the medication regimen and the individual patient; only by combining these three types of features can a comprehensive and complete analytical foundation be provided for identifying latent risks, avoiding missed detections due to missing feature information.

[0054] The fused integrated feature vector is input into the pre-trained unsupervised model, which is an intelligent recognition module specifically trained for identifying latent risks of intravenous medication. It relies on autonomously learning the feature distribution rules to achieve risk determination, without relying on a large number of manually labeled samples. It can adapt to the actual scenario in clinical practice where the types of latent risks are complex and the labeled samples are insufficient.

[0055] The built-in dual-branch detection logic of the model is key to improving the effectiveness of latent risk identification. This logic works collaboratively through two complementary detection branches to comprehensively analyze the fused features from different analytical dimensions, overcoming the limitations of single detection logic in terms of limited coverage and single identification dimensions. This ensures that all types of latent risks can be effectively captured. Through this dual-branch detection logic, three core types of latent risks can be accurately identified. These include: Individualized pharmacokinetic abnormality risk (the risk that a patient's absorption, metabolism, and excretion of drugs deviates from standard norms due to their own physiological characteristics; this type of risk has no fixed explicit rules and can only be identified through characteristic matching analysis between the individual and the drug); potential risk of delayed adverse reactions (the risk that has no obvious manifestation in the early stages of medication but will gradually appear with continuous drug infusion; static review cannot predict such delayed risks before medication); and latent conflict risk of cross-disease combination therapy (the risk that when a patient has multiple underlying diseases, the combination therapy regimen may meet the standards for a single disease, but the potential adverse effects may arise from the synergistic effect across diseases; this type of conflict also cannot be identified through conventional static rules).

[0056] After completing risk detection, the model outputs comprehensive results including latent risk types, risk confidence levels, and risk association features. Latent risk types clarify the specific types of risks, providing guidance for subsequent monitoring. Risk confidence levels quantify the likelihood of a risk occurring, directly reflecting its urgency. Risk association features pinpoint the specific medications or individual patient characteristics that trigger the risk, providing direct reference for determining subsequent dynamic monitoring parameters.

[0057] High-risk latent risks typically correspond to situations with high confidence levels, high probability of occurrence, and serious clinical harm. For example, patients may have significantly abnormal baseline liver and kidney function data, be prescribed high-alert intravenous drugs that are mainly metabolized by the liver and kidneys, or have individualized pharmacokinetic abnormality risks detected by the model. Once these risks occur, they can directly lead to drug accumulation and poisoning in the body, causing organ damage or even endangering the patient's life.

[0058] Medium-risk latent risks correspond to situations with moderate risk confidence, a certain probability of occurrence, and relatively obvious clinical harm but no immediate fatal risk. For example, the patient's weight and body surface area are slightly lower than the normal standard, the prescribed medication dosage is close to the normal safe upper limit, and the model detects a risk of deviation between the drug dosage and the patient's physiological baseline. This risk may cause moderate adverse reactions such as nausea and vomiting, and slight abnormal fluctuations in blood pressure. Although it will not directly endanger life, it requires timely attention and adjustment.

[0059] Low-risk latent risks correspond to situations with low confidence levels, extremely low probability of occurrence, and minor clinical harm, typically presenting only as transient, mild discomfort. For example, if a patient's underlying disease is in a stable remission phase, all physiological indicators are normal, and the medication regimen is standardized and reasonable, the model may only detect very slight short-term fluctuations in physiological parameters, possibly resulting in transient, mild dizziness or slight irritation at the puncture site—symptoms without substantial clinical harm. No special intervention is required; routine monitoring is sufficient. This type is classified as low-risk latent risk. Through the above classification and corresponding examples, subsequent verification and monitoring strategies can be more aligned with clinical practice, enabling precise and differentiated management of different risk levels.

[0060] In some specific embodiments, latent risk detection further includes: classifying detected latent risks into three levels—low, medium, and high—based on a preset latent risk classification standard, combined with the risk confidence level, probability of occurrence, and clinical harm of the latent risk; for high-risk latent risks, strengthening real-time verification of patient individual characteristics and drug interaction characteristics; for medium-risk latent risks, focusing on verifying the compatibility between drug dosage and patient physiological baseline; and for low-risk latent risks, performing routine characteristic monitoring. The process is attached. Figure 3 As shown, a tiered and differentiated control strategy is implemented for identified hidden risks. By scientifically classifying risk levels and matching corresponding verification and monitoring methods, precise focus and efficient handling of hidden risks of varying severity can be achieved.

[0061] Based on pre-defined unified clinical standards, and comprehensively considering three core dimensions—risk confidence level, actual probability of occurrence, and potential clinical harm—all detected latent risks are uniformly classified into three levels: low, medium, and high risk. Risk confidence level characterizes the reliability of the model's assessment of the risk's existence; probability of occurrence is the predicted likelihood of the risk's occurrence based on individual patient characteristics and medication regimen; and clinical harm corresponds to the extent of damage to the patient's body should the risk occur. The combination of these three factors allows for an objective and comprehensive measurement of the actual danger of latent risks, ensuring that the classification results align with actual clinical safety management needs.

[0062] After classifying the risks, different verification and monitoring strategies are adopted for each level of latent risk, forming the core logic of hierarchical management. For high-risk latent risks, which have a high probability of occurrence and significant clinical harm, directly endangering patient safety, an enhanced verification approach is adopted. This focuses on real-time and detailed verification of individual patient characteristics and drug interaction characteristics. By closely monitoring the compatibility between the patient's physiological state and drug effects, subtle abnormal changes are promptly detected, preventing high-risk risks from escalating into actual safety incidents. For medium-risk latent risks, which have a certain probability of occurrence and clinical harm, although less urgent than high-risk risks, they still require close attention. Therefore, the core verification focuses on the compatibility between drug dosage and the patient's physiological baseline. By accurately verifying whether the drug dosage matches the patient's current physiological condition, risks are avoided due to dosage mismatch, thus achieving effective management of medium-risk risks. For low-risk latent risks, which have a low probability of occurrence and minor clinical harm, no excessive management resources are required; only routine characteristic monitoring is necessary.

[0063] In some specific embodiments, statistical features include the mean, coefficient of variation, extreme values, and frequency of anomalies of the first physiological parameter. Temporal features are extracted using a sliding window algorithm to capture the short-term fluctuation trend, peak-to-trough difference, and temporal change point features of the first physiological parameter. Dynamic detection specifically includes: comparing the extracted temporal and statistical features with the static physiological feature baseline; performing compliance verification on the discrepancies by combining dynamic thresholds corresponding to drug pharmacological characteristics and dosing cycles from a dynamic adaptation rule base; and predicting the changing trend of the first physiological parameter based on historical time-series data using a trend prediction algorithm to identify potential dynamic anomaly risks in advance. The dynamic detection process is attached. Figure 4 As shown. The static physiological baseline refers to the set of basic physiological indicators measured before the start of this intravenous medication, under conditions of rest, stability, no drug intervention, and relatively stable condition. This baseline serves as the core reference standard for dynamic monitoring during subsequent medication administration to determine whether physiological parameters are abnormal, rather than a universally accepted clinical normal reference range. This baseline is derived from the patient's multi-dimensional medical record data before medication and vital sign data collected immediately before medication, including key physiological values ​​related to medication safety such as heart rate, blood pressure, respiration, body temperature, basic liver and kidney function indicators, and blood oxygen saturation. It is entirely determined based on the patient's individual physiological state, possessing highly individualized attributes and accurately reflecting the patient's normal physiological level before receiving this intravenous medication.

[0064] Statistical characteristics are a quantitative representation of the overall state and abnormalities of the primary physiological parameter within a continuous monitoring period. Specifically, they include the parameter mean, coefficient of variation, extreme values, and frequency of abnormalities. The parameter mean reflects the overall average level of the primary physiological parameter during the monitoring period, used to determine whether the patient's physiological state remains within a stable baseline range. The coefficient of variation measures the dispersion of the physiological parameter, directly reflecting the stability during parameter monitoring. Extreme values ​​are the highest and lowest values ​​of the physiological parameter within the monitoring period, used to identify extreme abnormal values. The frequency of abnormalities counts the number of times the physiological parameter deviates from the normal range, used to determine the frequency of abnormalities. These statistical characteristics comprehensively reflect the stable state of the primary physiological parameter from an overall quantitative perspective, compensating for the deficiency that a single real-time value cannot reflect the overall change pattern of the parameter, and providing a stable and reliable quantitative basis for dynamic detection.

[0065] Temporal features are extracted using a sliding window algorithm. This algorithm segments and analyzes continuously collected physiological parameter data according to fixed time intervals to obtain the short-term fluctuation trend, peak-to-trough difference, and temporal change point features of the first physiological parameter. The short-term fluctuation trend reflects the upward, downward, or stable changes of the physiological parameter within a short period. The peak-to-trough difference quantifies the amplitude of fluctuations in the physiological parameter within a short period. The temporal change point features accurately pinpoint the time points and numerical changes of sudden and drastic changes in the physiological parameter. These temporal features capture the dynamic change patterns of physiological parameters from the perspective of time variation, and can identify gradual changes or sudden anomalies that cannot be detected by static values ​​alone, providing a key basis for the accurate identification of dynamic risks.

[0066] Dynamic detection achieves comprehensive and accurate risk identification through a three-tiered, progressive logic. First, the extracted temporal and statistical features are compared with the patient's static physiological baseline. This baseline represents the patient's normal physiological indicators before medication; this comparison quickly identifies deviations in physiological parameters after medication compared to the baseline state, clarifying whether abnormal changes have occurred. Second, dynamic thresholds corresponding to drug pharmacological characteristics and dosing cycles in a dynamic adaptation rule base are used for compliance verification of the discrepancies. Drug pharmacological characteristics determine the normal fluctuation range of physiological parameters after medication, and the dosing cycle corresponds to the parameter change patterns at different medication stages. The dynamic thresholds are stage-specific judgment standards set to fit these characteristics, rather than fixed universal thresholds. This verification accurately distinguishes whether deviations in physiological parameters are normal fluctuations under drug influence or abnormal risk fluctuations. A trend prediction algorithm predicts the subsequent trend of the first physiological parameter based on historical time-series data. This algorithm analyzes the collected temporal feature patterns to infer future changes in physiological parameters, enabling early identification of potential deterioration trends before parameters exceed abnormal thresholds. This provides proactive warnings of dynamic abnormal risks, preventing the risk from escalating and causing safety issues.

[0067] By combining quantitative statistical features with dynamic temporal features, along with a three-layer progressive dynamic detection logic, dynamic detection breaks through the limitations of traditional single real-time numerical judgment. It comprehensively analyzes the changes in the patient's physiological state during medication from three dimensions: overall quantification, temporal fluctuation, and trend prediction. It can accurately identify existing overt abnormalities and predict potential latent dynamic risks in advance. At the same time, combined with a dynamic adaptation rule base, it realizes individualized and phased threshold judgment, further improving the accuracy and clinical adaptability of dynamic detection. It forms an effective connection with the aforementioned latent risk classification and control link, providing scientific and timely judgment support for subsequent medication control operations.

[0068] In some specific embodiments, the dynamic detection frequency is adaptively adjusted according to the latent risk level throughout the entire intravenous medication execution cycle. When abnormal physiological parameters are detected, the authenticity of the abnormal parameters is first verified by backtracking through time-series trends, and then the correlation between the abnormality and intravenous medication is verified by combining the drug's duration of action with the patient's underlying disease status to avoid misjudgment.

[0069] Throughout the entire intravenous medication administration cycle, the dynamic monitoring frequency adaptively adjusts based on the previously categorized low-risk, medium-risk, and high-risk latent risks. The core principle is that different levels of latent risk correspond to different safety control priorities. For high-risk latent risks, the monitoring frequency is automatically increased, shortening the time interval between data collection and analysis to achieve high-density real-time monitoring of physiological parameters, ensuring that risk signals are captured immediately. For medium-risk latent risks, a moderate monitoring frequency is used to balance monitoring efficiency while ensuring risk control. For low-risk latent risks, a conventionally low monitoring frequency is maintained, without requiring excessive monitoring resources. This adaptive adjustment method enables the rational allocation of monitoring resources, prioritizing monitoring effectiveness in high-risk scenarios while improving the overall operational efficiency of the dynamic monitoring process.

[0070] When abnormal physiological parameters are detected during dynamic monitoring, they are not immediately identified as a medication risk. Instead, a two-step verification process is used to rule out potential misjudgments. The first step verifies the authenticity of the abnormal parameter through time-series trend backtracking. Time-series trend backtracking involves retrieving previously collected time-series data and fluctuation characteristics of the physiological parameter to determine whether the abnormal value is a genuine anomaly resulting from continuous changes or a false anomaly caused by accidental factors such as equipment interference or instantaneous fluctuations. This verifies the validity of the anomaly from a data perspective, avoiding incorrect judgments caused by external interference. The second step verifies the correlation between the abnormality and intravenous medication by combining the drug's duration of action with the patient's underlying disease status. The drug's duration of action corresponds to different time stages of drug onset, peak blood drug concentration, and metabolic excretion, which can determine whether the time of the abnormality matches the drug's action cycle. The patient's underlying disease status is used to distinguish whether the abnormality is caused by intravenous medication or by non-medication-related abnormalities resulting from the progression of the patient's underlying disease. Through these two steps of verification, abnormal physiological parameters caused by non-medication factors can be effectively eliminated, the false judgment rate of dynamic detection can be greatly reduced, and the subsequent medication control operations can be based on real and effective medication risks. This can not only deal with real medication abnormalities in a timely manner, but also prevent misjudgments from interfering with the normal diagnosis and treatment process.

[0071] In some specific embodiments, the pre-trained unsupervised model includes a feature fusion branch and an anomaly detection branch. The feature fusion branch is used to deeply fuse the medical order feature vector, patient feature vector, and drug pharmacological features, and strengthens the correlation weight between individualized physiological features and drug action features through an attention mechanism to suppress interference from irrelevant features. The anomaly detection branch uses a combination of density clustering algorithm and isolated forest algorithm. First, density clustering is used to perform preliminary clustering of the fused features to divide normal medication feature clusters and potential anomaly clusters. Then, isolated forest is used to perform fine screening of potential anomaly clusters.

[0072] The pre-trained unsupervised model employs a dual-branch architecture of feature fusion and anomaly detection. The core principle is to separate feature processing and anomaly identification into two specialized modules, each with its own function yet working collaboratively. This avoids the efficiency and accuracy degradation caused by a single module handling multiple tasks, ensuring the model's overall implicit risk identification capability. The feature fusion branch is specifically responsible for the deep integration and optimization of multi-dimensional features. Its core function is to deeply fuse medical order feature vectors, patient feature vectors, and drug pharmacological features. Unlike simple feature concatenation, this branch uses an attention mechanism to intelligently allocate feature weights. The principle is that different features contribute significantly differently to implicit risk identification. The attention mechanism automatically learns the correlation between various features and implicit risks, strengthening the correlation weights between individualized physiological features and drug action features, while suppressing irrelevant redundant features. This generates a fused feature vector focused on the core risk identification dimensions, preventing missed or false positives due to feature redundancy and interference from the outset, ensuring that the subsequent anomaly detection branch analyzes only effective features.

[0073] The anomaly detection branch is specifically responsible for anomaly identification of the fused feature vectors. Its core principle is to employ a combined approach of density clustering and isolated forest algorithms to achieve a two-layer anomaly detection logic of "coarse screening + fine screening," rather than relying on a single algorithm. The density clustering algorithm first performs preliminary clustering of the fused features. Its principle is to divide clusters based on the density distribution of feature vectors. Feature vectors from normal medication use cases will form high-density clusters of normal medication use features, while feature vectors from anomaly cases with latent risks, deviating from the normal distribution, will form low-density clusters of potential anomalies. This step quickly narrows down the anomaly detection range, reducing the workload of subsequent fine screening. The isolated forest algorithm then performs fine screening on the potential anomaly clusters formed by density clustering. Its principle is to accurately identify true anomalies from a small number of anomaly samples, eliminating false anomalies misjudged during clustering, ensuring that only feature vectors with genuine latent risks are located. This combined detection method balances detection efficiency and accuracy, ensuring both comprehensive anomaly detection and improved accuracy.

[0074] For pre-trained unsupervised models used for latent risk detection, Isolation Forest or Denoising Autoencoder (DAE) are preferred as they are well-suited to anomaly detection scenarios in the medical field. Isolation Forest is suitable for quickly identifying anomalous latent risk samples within high-dimensional features, while Denoising Autoencoder excels at learning normal medication patterns from complex doctor's orders and patient fusion features, thereby accurately locating latent risks deviating from these patterns. Both models do not require manual annotation of training data, making them suitable for the diverse types of latent risks and the scarcity of labeled samples in clinical settings. The model training process revolves around the core requirement of latent risk detection in intravenous medication.

[0075] First, a large-scale historical clinical intravenous medication case data was collected, covering normal medication cases with no risk and abnormal cases containing various hidden risks. The corresponding medical order features, multi-dimensional patient medical record features, and drug pharmacological features were extracted from the cases to form the original training dataset. The original dataset was preprocessed, including missing value imputation, feature standardization, and outlier removal. At the same time, feature fusion was used to integrate the three types of features—medical orders, patients, and drugs—into a unified high-dimensional feature vector to ensure the standardization and validity of the input data. Based on the selected model architecture, a basic model is built. If an isolated forest is chosen, a reasonable number of trees (usually 100-200) and a sample size (adapted to medical data scale, set to 256-512) are set. Multiple isolated trees are constructed by randomly selecting features and split points. Abnormal samples (latent risk cases) will be isolated more quickly because their features deviate from the normal distribution. If a denoising autoencoder is chosen, a network structure of input layer, hidden layer, and output layer is built. The dimension of the input layer is consistent with the dimension of the fused feature vector. The hidden layer extracts core features by progressively reducing dimensionality, and the output layer reconstructs the input features. During training, slight Gaussian noise is added to the input features to force the model to learn the core distribution pattern of normal medication features. The preprocessed training dataset is input into the built model for iterative training. The isolated forest continuously builds isolated trees and calculates the isolated path length of each sample. The shorter the path, the higher the probability of being identified as an abnormal sample. The denoising autoencoder uses the reconstruction error as the loss function (using mean squared error loss) and optimizes the network parameters through backpropagation so that the reconstruction error of the model for normal samples approaches minimum, while the reconstruction error for latent risk abnormal samples is significantly higher than that for normal samples. Clinical validation is incorporated into the training process. A small number of labeled cases with known latent risks are selected as the validation set. By adjusting model hyperparameters (such as tree depth in the isolated forest, number of hidden layer nodes in the denoising autoencoder, and learning rate), the precision and recall of the model are balanced to avoid overfitting or underfitting. After training, the model is incrementally pre-trained and optimized by incorporating the latest clinical medication case data to continuously update the model's ability to identify latent risk features. This ensures that the model adapts to changes in scenarios such as updates to clinical medication guidelines and the addition of new drug types. Ultimately, a pre-trained unsupervised model that can accurately identify various latent risks in intravenous medication orders is obtained. In practical applications, this model can determine the existence of latent risks by calculating the isolated path length (isolated forest) or reconstruction error (denoising autoencoder) of the input samples and output the corresponding risk confidence, providing a basis for subsequent risk level classification and dynamic monitoring parameter determination.

[0076] In some specific embodiments, the process also includes: if the patient completes the intravenous medication according to the intravenous medication order and there are no abnormalities in the dynamic monitoring throughout the process, then the entire process review ends; if there are abnormalities in the dynamic monitoring results during the medication process, then after the order is adjusted, the adjusted order is used as the intravenous medication order to be reviewed again, and the above steps are repeated to form a closed-loop review process.

[0077] If the patient completes the entire intravenous medication procedure according to the prescribed intravenous medication order, and all dynamic monitoring results throughout the medication process are normal, it indicates that the intravenous medication order not only passed the pre-administration static compliance assessment and implicit risk detection, but also adapted to the patient's real-time physiological state during actual medication administration, without triggering any dynamic risks, and fully meets the individual patient's medication safety requirements. At this point, the entire process review ends. The core principle of this approach is based on actual clinical diagnosis and treatment logic. The review process is terminated only after confirming that there is no risk throughout the medication administration process, avoiding meaningless redundant operations, effectively improving review efficiency, and conforming to the routine clinical principles for handling compliant and risk-free medication.

[0078] If abnormalities are detected in the dynamic monitoring results during medication administration, it indicates that the current intravenous medication order is not compatible with the patient's real-time physiological state, or that there are unidentified dynamic medication risks. In this case, temporary intervention for the abnormality alone cannot fundamentally guarantee subsequent medication safety. It is necessary to first make targeted adjustments to the original intravenous medication order, and then re-issue the adjusted order as an intravenous medication order awaiting review. The entire review process described above must be repeated, including acquiring and preprocessing the new order and the patient's updated multi-dimensional medical record data; completing static compliance judgment through a pre-set rule engine; using a pre-trained unsupervised model to detect implicit risks and determine dynamic monitoring parameters; collecting physiological parameters during medication administration in real time to complete dynamic monitoring; and executing medication control operations and optimizing the unsupervised model based on the monitoring results. The core principle of this approach is to ensure, through a standardized re-review process, that the revised medical orders are not merely modified based on human experience, but undergo a full-process process of explicit risk filtering, implicit risk discovery, and dynamic adaptability verification. This mechanism eliminates the possibility that the revised medical orders may still contain unidentified risks, and brings the medical order adjustment process under the control of the entire review process.

[0079] A fully intelligent review system based on intravenous medication orders, the system's modules are illustrated in the attached diagram. Figure 5 As shown, the system includes: Input unit 1 is used to acquire intravenous medication orders to be reviewed and multi-dimensional medical record data of patients and preprocess them to generate a medical order feature vector and a patient feature vector including static physiological features. Static review unit 2 is used to preset the rule engine to perform static compliance judgment on the intravenous medication order based on the static physiological characteristics and the medical order feature vector, and select the statically compliant medical orders as the first medical order; The latent detection unit 3 is used to perform latent risk detection on the first medical order through a pre-trained unsupervised model, and determine the first physiological parameter that needs to be dynamically monitored based on the risk type of the detected latent risk. The dynamic detection unit 4 is used to collect the first physiological parameters of the patient during the intravenous medication process according to the intravenous medication order in real time, extract temporal features and statistical features from them, and perform dynamic detection through the rule engine to determine whether the patient's current state is suitable to continue intravenous medication according to the intravenous medication order; Output unit 5 is used to execute corresponding medication control operations based on the dynamic detection results, and at the same time use the dynamic detection results, time series features and statistical features as feedback data to incrementally correct the unsupervised model, thereby realizing iterative optimization of the prediction capability of the unsupervised model.

[0080] Those skilled in the art will understand that the components of this application described above can be implemented using general-purpose computing systems. They can be centralized on a single computing system or distributed across a network of multiple computing systems. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage system for execution by the computing system. Alternatively, they can be fabricated as separate integrated circuit components, or multiple components or steps can be fabricated as a single integrated circuit component. Thus, this application is not limited to any particular combination of hardware and software.

[0081] Note that the above description is merely a preferred embodiment and the technical principles employed in this application. Those skilled in the art will understand that this application is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of this application. Therefore, although this application has been described in detail through the above embodiments, this application is not limited to the above embodiments. Many other equivalent embodiments may be included without departing from the concept of this application, and the scope of this application is determined by the scope of the appended claims.

[0082] The above disclosures are only a few specific implementation scenarios of this application. However, this application is not limited to these. Any variations that can be conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A fully intelligent review method based on intravenous medication orders, characterized in that, include: The system acquires and preprocesses intravenous medication orders to be reviewed and multi-dimensional medical record data of patients to generate order feature vectors and patient feature vectors including static physiological features. The preset rule engine performs static compliance determination on the intravenous medication orders based on the static physiological characteristics and medical order feature vectors, and selects the statically compliant medical orders as the first medical orders; The first medical order is subjected to latent risk detection by a pre-trained unsupervised model, and the first physiological parameter that needs to be dynamically monitored is determined based on the type of latent risk detected. The system collects the first physiological parameters of the patient during the intravenous medication process according to the intravenous medication order in real time, extracts temporal and statistical features from them, and performs dynamic detection through the rule engine to determine whether the patient's current state is suitable for continuing intravenous medication according to the intravenous medication order. Based on the dynamic detection results, corresponding medication control operations are executed. At the same time, the dynamic detection results, time-series features, and statistical features are used as feedback data to incrementally correct the unsupervised model, thereby achieving iterative optimization of the unsupervised model's predictive capabilities.

2. The end-to-end intelligent review method according to claim 1, characterized in that, The multidimensional medical record data includes baseline data on liver and kidney function, stage data of underlying diseases, data on drug allergy history, data on adverse reactions to previous intravenous medications, and physiological data on weight and body surface area.

3. The end-to-end intelligent review method according to claim 1, characterized in that, The preset rule engine includes a layered and decoupled static explicit rule base and a dynamic adaptive rule base. The static compliance judgment combines baseline data of liver and kidney function, basic disease staging data, and physiological data of body weight and body surface area from multi-dimensional medical record data to perform individualized explicit compliance verification of drug dosage and infusion rate. At the same time, it verifies drug contraindications based on drug allergy history data and avoids high-risk drug combinations based on past adverse reactions to intravenous drugs, achieving the first layer of precise filtering of explicit risks and eliminating intravenous drug orders with clear violation logic and insufficient suitability.

4. The end-to-end intelligent review method according to claim 1, characterized in that, The detection of latent risks specifically includes: The medical order feature vector, patient feature vector, and pre-set drug pharmacology feature library are fused together and input into the pre-trained unsupervised model. The model's built-in dual-branch detection logic is used to mine individualized pharmacokinetic abnormality risk, potential risk of delayed adverse reaction, and implicit conflict risk of cross-disease combination drug use, and outputs implicit risk detection results including implicit risk type, risk confidence level, and risk association features.

5. The end-to-end intelligent review method according to claim 1, characterized in that, The hidden risk detection also includes: Based on the pre-defined criteria for classifying latent risks, and combining the risk confidence level, probability of occurrence, and clinical harm of latent risks, the detected latent risks are classified into three levels: low, medium, and high. For high-risk latent risks, real-time verification of patient individual characteristics and drug interaction characteristics is strengthened. For medium-risk latent risks, the focus is on verifying the compatibility between drug dosage and patient physiological baseline. For low-risk latent risks, routine characteristic monitoring is performed.

6. The end-to-end intelligent review method according to claim 1, characterized in that, The statistical features include the mean, coefficient of variation, extreme values, and frequency of anomalies of the first physiological parameter. The time-series features are extracted using a sliding window algorithm to capture the short-term fluctuation trend, peak-to-trough difference, and time-series change point features of the first physiological parameter. The dynamic detection specifically includes: The extracted temporal and statistical features are compared with the static physiological feature baseline; the difference data are verified for compliance by combining the dynamic thresholds corresponding to the drug pharmacological characteristics and dosing cycle in the dynamic adaptation rule base; and the trend prediction algorithm is used to predict the changing trend of the first physiological parameter based on historical time series data to identify potential dynamic abnormality risks in advance.

7. The end-to-end intelligent review method according to claim 5, characterized in that, Throughout the entire course of intravenous medication, the frequency of dynamic monitoring is adaptively adjusted according to the level of latent risk. When abnormal physiological parameters are detected, the authenticity of the abnormal parameters is first verified by retrospective analysis of time-series trends. Then, the correlation between the abnormality and intravenous medication is verified by combining the drug's duration of action with the patient's underlying disease status to avoid misjudgment.

8. The end-to-end intelligent review method according to claim 1, characterized in that, The pre-trained unsupervised model includes a feature fusion branch and an anomaly detection branch; the feature fusion branch is used to deeply fuse the medical order feature vector, the patient feature vector, and the drug pharmacological features, and strengthens the correlation weight between individualized physiological features and drug action features through an attention mechanism to suppress interference from irrelevant features; The anomaly detection branch employs a combination of density clustering and isolated forest algorithms. First, density clustering is used to perform preliminary clustering of the fused features, dividing them into normal medication feature clusters and potential anomaly clusters. Then, isolated forest is used to perform fine screening of the potential anomaly clusters.

9. The end-to-end intelligent review method according to claim 1, characterized in that, Also includes: If the patient completes the intravenous medication as prescribed and the dynamic monitoring shows no abnormalities throughout the process, the entire review process ends. If there are abnormalities in the dynamic monitoring results during the medication process, the prescription will be adjusted, and the adjusted prescription will be used as the intravenous medication prescription to be reviewed again. The above steps will be repeated to form a closed-loop review process.

10. A fully intelligent review system based on intravenous medication orders, characterized in that, include: The input unit is used to acquire intravenous medication orders to be reviewed and multi-dimensional medical record data of patients, and to preprocess them to generate a medication order feature vector and a patient feature vector including static physiological features. The static review unit is used to preset the rule engine to perform static compliance judgment on the intravenous medication order based on the static physiological characteristics and the medical order feature vector, and select the statically compliant medical orders as the first medical orders; The latent detection unit is used to perform latent risk detection on the first medical order using a pre-trained unsupervised model, and to determine the first physiological parameter that needs to be dynamically monitored based on the risk type of the detected latent risk. The dynamic detection unit is used to collect the first physiological parameters of the patient during the intravenous medication process according to the intravenous medication order in real time, extract temporal features and statistical features from them, and perform dynamic detection through the rule engine to determine whether the patient's current state is suitable to continue intravenous medication according to the intravenous medication order; The output unit is used to execute corresponding medication control operations based on the dynamic detection results, and simultaneously use the dynamic detection results, time-series features, and statistical features as feedback data to incrementally correct the unsupervised model, thereby achieving iterative optimization of the unsupervised model's predictive capabilities.

Images