Prescription auditing method and system based on regional remote prescription auditing platform

Abstract

The invention provides a prescription auditing method and system based on a regional remote prescription auditing platform. The method comprises the following steps: receiving an original prescription data set uploaded by a regional medical institution; obtaining dynamic state information of the audit node, and customizing a task allocation strategy; task complexity is set, and a processing strategy is refined; rechecking the abnormal result, and outputting a rechecking result; and performing multi-level publishing on the reexamination result, and generating a regional drug use risk analysis report according to the reexamination result. According to the method, the core problems of uneven resource distribution, unreliable auditing process, low exception handling efficiency and the like in remote prescription auditing in the prior art are successfully solved, and a comprehensive and practical solution is provided for optimization of a remote prescription auditing system in a regional medical network.

Description

Technical Field

[0001] The present invention belongs to the field of remote prescription review platforms, and in particular, relates to a prescription review method and system based on a regional remote prescription review platform. Background Art

[0002] With the rapid development of informatization in the medical industry, the construction of regional medical networks has become an important direction for promoting the sharing of medical resources and improving the level of medical services. In this context, the remote prescription review platform, as an important part of the regional medical service system, has gradually been widely used. The main function of the remote prescription review platform is to connect multiple medical institutions in the region through the network, upload the prescriptions issued by doctors to the platform, and the review subjects in the platform (including manual review and intelligent review) will check the rationality of the prescriptions, thereby ensuring the safety of patients' medication. However, there are still many problems in the actual application of the existing remote prescription review technology, which restricts its further development and promotion.

[0003] Most of the current remote prescription review systems adopt a centralized processing architecture, that is, all prescription data is uniformly uploaded to the central platform for processing. This architecture was relatively effective in early telemedicine practices because of its simple logic and low initial deployment cost. However, with the explosive growth of medical data and the increase in the demand for complex prescription review, this centralized architecture has gradually exposed obvious defects. First, the central platform is prone to become a performance bottleneck of the system: when the prescription volume of medical institutions in a certain area surges, the central platform may have tasks piled up due to insufficient processing capacity, resulting in a significant decrease in review efficiency. Secondly, due to the lack of a task scheduling mechanism, medical resources in the region have not been used evenly, and the review resources of some institutions have been idle for a long time, while other institutions are unable to complete the review tasks in time due to overload. In addition, the centralized review system lacks flexibility in dealing with complex prescription reviews, and usually relies on a single subject (such as AI or manual review) to make decisions, and the judgment of a single subject may be inaccurate, especially for rare disease medications, personalized patient treatment plans, or special drug interaction risks.

[0004] On the other hand, existing technologies have also failed to fully consider the dynamic characteristics of regional medical environments. For example, in medical practices in different regions, due to differences in drug use habits, patient medical history, and local disease spectra, general review rules may be difficult to adapt to local needs. In addition, existing systems generally lack an efficient review and tracking mechanism, resulting in inefficient handling of abnormal prescriptions. In the prescription review process, due to the lack of a multi-subject collaboration mechanism, the advantages of manual review and AI review have not been effectively integrated, which has increased labor costs and reduced the overall credibility of the review. In summary, the existing remote prescription review technology has shown obvious shortcomings when faced with high loads, complex review tasks, and multi-subject collaboration requirements, and a new solution is urgently needed to meet the above challenges. Summary of the invention

[0005] The purpose of the present invention is to propose a prescription review method and system based on a regional remote prescription review platform, which successfully solves the core problems encountered in the prior art in remote prescription review, such as uneven resource allocation, unreliable review process and low efficiency of exception handling, and provides a comprehensive and practical solution for the optimization of remote prescription review systems in regional medical networks.

[0006] In order to achieve the above-mentioned object, a first aspect of the present invention provides a prescription review method based on a regional remote prescription review platform, the method comprising:

[0007] S1. Receive the original prescription data set uploaded by the regional medical institutions, then extract the complexity features of each prescription in the original prescription data set to obtain the feature vector F, calculate the complexity score of each prescription using the weighted linear model according to the feature vector F, and finally sort all prescription data in descending order according to the complexity score, give priority to tasks with high scores, and form a task queue;

[0008] S2. Obtain the dynamic status information of the review node, combine it with the task queue, design the allocation cost function to calculate the optimal matching comprehensive cost of the task and the review node, the lowest value means the match is optimal, and then customize the task allocation strategy according to the result of the allocation cost function;

[0009] S3. Set the task complexity, build a deep learning model to pre-screen the task allocation queue, and refine the corresponding processing strategy for each task based on the task complexity and the pre-screening result. Finally, merge each refined processing strategy into a final decision score to determine whether the current task has passed the review or needs to be modified.

[0010] S4, reviewing the abnormal results of the deep learning model, feeding back the review results to the deep learning model to optimize the deep learning model, and outputting the review results;

[0011] S5. Publish the review results at multiple levels, generate a regional medication risk analysis report based on the review results, and dynamically adjust the classification standards of task complexity based on the statistical characteristics of abnormal tasks based on the review results. Finally, design a closed-loop optimization mechanism for system performance based on the tasks in the audit process, the review results, and the optimized deep learning model.

[0012] Wherein, the S3 specifically includes:

[0013] Design task complexity The tasks are classified into:

[0014] Indicates low-complexity tasks, which are mainly reviewed by AI and do not require human intervention;

[0015] It indicates tasks of medium complexity, which require the participation of pharmacists;

[0016] Indicates highly complex tasks that require doctor review and final decision making;

[0017] Among them, θ1 is the low difficulty threshold, θ2 is the difficulty threshold;

[0018] Feeding standardized prescription text and patient information into deep learning models Generate review score and annotate suspicious points Among them, the set of annotation doubtful points including drug interaction points, dosage abnormalities and contraindications or allergy risks; for complex tasks Introducing special regularization terms Enhance the meticulousness of suspicious point annotations, as shown below:

[0019]

[0020] Among them, f AI represents a deep learning function, is a set of labeled doubt points The regularization term is calculated as follows:

[0021]

[0022] Among them, λ is the regularization weight, which controls the impact of suspicious point annotation on the score; Confidence(p) is the AI ​​confidence of suspicious point p, and the suspicious point with the lowest confidence will have the largest deduction on the score;

[0023] For task complexity The task is to have pharmacists participate in the review and form manual review scores For task complexity The task is reviewed by the doctor and the doctor generates the score

[0024] The review scores of multiple subjects are integrated into the final decision score For judging tasks Whether it has passed the review or needs to be modified is indicated as follows:

[0025]

[0026] Among them, α, β, and γ are dynamic weights based on task complexity. and node reliability Adjustment; for low-complexity tasks, α has the largest weight; for high-complexity tasks, β and γ have the largest weights;

[0027] Score the final decision The threshold value θ is not reached R Mark the task as abnormal and record a detailed set of doubts Abnormal tasks are forwarded to a higher-level review panel.

[0028] Furthermore, the original prescription data set is represented as D Rx ={d1,d2,…,d i ,d n}, where each piece of data d i Includes: Prescription text content T i , Patient Information i Additional notes from the doctor C i ;

[0029] After receiving the original prescription data set uploaded by the regional medical institution, the original prescription data set is standardized, including: mapping of drug names, unification of dosage units and verification of patient information;

[0030] After standardization, each piece of data is converted into a triple form: in, Represents the standardized data and obtains the standardized prescription data

[0031] The extraction of complexity features from each prescription in the original prescription data set to obtain a feature vector F specifically includes:

[0032] For each prescription Extracted from the number of drug interactions f1, the number of dosage abnormalities f2, and the number of contraindications or allergic conflicts f3, each prescription Extract a feature vector:

[0033] F i =[f1,f2,f3]

[0034] Among them, f1, f2, and f3 represent the number of drug interactions, the number of dosage anomalies, and the number of contraindication conflicts, respectively;

[0035] The complexity score S of each prescription is calculated using the weighted linear model based on the feature vector F. i , specifically:

[0036] S i =w1f1+w2f2+w3f3

[0037] Among them, w1, w2, and w3 are preset weights, reflecting the impact of each feature on the audit complexity;

[0038] The above method arranges all prescription data in descending order according to the complexity score, prioritizes tasks with high scores, and forms a task queue. Rx , specifically:

[0039]

[0040] in, Indicates that tasks are arranged from high complexity to low complexity.

[0041] Furthermore, the dynamic status information of the review node includes the current task load of node j, the processing capacity of node j, the communication delay of node j and the task success rate of node j, and initializes the availability index A of each node j j , which is used to reflect the real-time processing capability of the node, is expressed as follows:

[0042]

[0043] Among them, A j The highest value indicates that the node has the lightest load and the strongest task processing capability.

[0044] Furthermore, the allocation cost function is expressed as follows:

[0045]

[0046] in, Representation Task The comprehensive cost assigned to node j has the lowest optimal value; λ1, λ2 represent weight parameters, which are used to adjust the importance of communication delay and complexity adaptation items; α represents the reliability adjustment factor, which is used to enlarge or reduce The impact of j =L j / C j Represents the load utilization rate of node j, which is used for balance control to avoid overload; τ j represents the communication delay of node j; L j represents the current task load of node j, C j represents the processing capacity of node j; represents the task success rate of node j;

[0047] The task allocation strategy specifically includes:

[0048] For each task Traverse all nodes j∈{1,2,…,m}, according to Calculate the allocation cost;

[0049] Select the node j with the smallest cost * , complete the task allocation, expressed as follows:

[0050]

[0051] Select the node j with the smallest cost * , complete the task assignment:

[0052]

[0053] Repeat until the task queue is Rx All tasks in the assignment are completed.

[0054] Furthermore, the S2 also includes designing a dynamic adjustment mechanism to deal with the node task backlog, specifically including:

[0055] When the load utilization rate of node j is U j >β, where β is the preset threshold that triggers task transfer;

[0056] Select the task with the lowest complexity score from the task set of node j and recalculate And transfer to the node j′ with the second best cost; at the same time, dynamically adjust the value of λ2 to improve the priority of complexity adaptability in task transfer.

[0057] Furthermore, the task complexity Pharmacists participate in the review of the following tasks:

[0058] Pharmacists receive a collection of doubts annotated by AI and rating Check and correct each item, including: removing misreporting labels, adding missing labels, and modifying drug dosage or combination recommendations;

[0059] Pharmacists form manual review scores Defined as:

[0060]

[0061] in, is the set of doubts after adjustment by the pharmacist, and Adjust is the correction function;

[0062] The complexity of the task The tasks are reviewed by doctors, including:

[0063] Doctor receives pharmacist's review result and patient information The doctor-generated score is based on the patient's personalized medical history for final review The patient's personalized medication risk is calculated by combining specific weights as follows:

[0064]

[0065] Among them, g Doc Represents the final judgment function;

[0066] The preliminary review output results include: tasks that have passed the review and their scores, tasks that have not passed the review and their modification suggestions, and abnormal tasks and doubt records.

[0067] Furthermore, the S4 specifically includes:

[0068] Accept the task set marked as abnormal in S3 Abnormal , where the abnormal task set Task Abnormal Include each task The following information: Final review score Indicates the comprehensive score of multi-subject review; doubt point collection Record the specific issues marked by the review; mark the reasons Describes the triggering conditions for a task to be marked as abnormal;

[0069] For abnormal task set Task Abnormal Categorize and sort by severity priority to get the review task queue Review ; Wherein, the classification includes high-risk abnormal tasks, medium-risk abnormal tasks and low-risk abnormal tasks;

[0070] According to the review task queue Review Dynamically assign tasks to expert review teams based on priority and task type;

[0071] During the expert review stage, Doubts Check each item and generate review decisions At the same time, the review process and results of each task are recorded in the exception log Abnormal ; Wherein, the review decision Including score correction after review Corrected doubt set and the task passes the status The abnormal log Log Abnormal Include the task unique identifier Original review score Correction rating Review expert's annotation opinions and final task status

[0072] Feedback the review results to the deep learning model for optimization, and update and generate a new regional medication rule library At the same time, according to the characteristic distribution of high-risk abnormal tasks, the classification and priority calculation method of abnormal tasks are adjusted.

[0073] Furthermore, the severity priority ranking is calculated as follows:

[0074]

[0075] in, represents the priority score; Indicates the number of suspicious points marked as high risk; Indicates the total number of suspicious points; Indicates the comprehensive score; w1, w2 are weight parameters used to balance the impact of high-risk doubts and low scores; score according to priority Sort from high to low to form a priority review task queue Review ;

[0076] According to the review task queue Queue Review Dynamically assign tasks to the expert review team based on priority and task type, including:

[0077] According to each expert's field of expertise m , the suspicious type of the matching task Priority will be given to experts in related fields;

[0078] In the corresponding field, according to the current task load L of the expert m and task processing efficiency C m , the real-time review capability of the computing expert is expressed as follows:

[0079]

[0080] Among them, A m Indicates the review capability is empty; priority is given to assigning tasks to experts with the highest review capability among experts in related fields.

[0081] Furthermore, the S5 specifically includes:

[0082] Task based on review results Final , generate audit reports for each task and publish them in layers, including:

[0083] for Generates a brief report of the task, including the task identifier score and passing records;

[0084] for Tasks, combined with a set of doubts Provide specific modification suggestions and file them in the medical institution’s feedback system;

[0085] for Tasks, combined with a set of doubts Provide specific modification suggestions and file them in the medical institution’s feedback system;

[0086] According to the abnormal log Log Abnormal And the score distribution of all tasks, generate a regional medication risk analysis report Risk ; The regional drug risk analysis report Risk Includes statistics of high-risk drugs, performance evaluation of audit models, and performance analysis of nodes and experts;

[0087] Utilize the newly added annotation data during the review process, which includes the task scores and doubt correction Retraining Deep Learning Models Improve the accuracy of annotations and update the review rule base

[0088] According to the review results, the characteristic distribution of abnormal tasks is statistically analyzed, and the classification standards of task complexity are dynamically adjusted, including: middle The distribution is clustered and the complexity score thresholds θ1 and θ2 are redefined;

[0089] Record the rule base after each optimization and Model The version number of the audit system is generated every quarter. Perfomance , dynamically update the node task allocation strategy and perform closed-loop optimization.

[0090] In another aspect of the present invention, a prescription review system based on a regional remote prescription review platform is provided, the system comprising:

[0091] The original prescription collection module is used to receive the original prescription data set uploaded by the regional medical institutions, and then extract the complexity features of each prescription in the original prescription data set to obtain the feature vector F, calculate the complexity score of each prescription using the weighted linear model according to the feature vector F, and finally sort all prescription data in descending order according to the complexity score, give priority to tasks with high scores, and form a task queue;

[0092] The prescription allocation module is used to obtain the dynamic status information of the prescription review node, form a task queue based on it, design an allocation cost function to calculate the optimal matching comprehensive cost of the task and the prescription review node, and the lowest value means the match is optimal, and then customize the task allocation strategy based on the result of the allocation cost function;

[0093] The prescription review module is used to set the task complexity and build a deep learning model to pre-review the task allocation queue. It combines the task complexity and the pre-review results to perform corresponding detailed processing strategies for each task. Finally, each detailed processing strategy is integrated into the final decision score to determine whether the current task has passed the review or needs to be modified.

[0094] A prescription review module, used to review the abnormal results of the deep learning model, feed back the review results to the deep learning model to optimize the deep learning model, and output the review results;

[0095] The system optimization module is used to publish the review results at multiple levels, generate regional drug risk analysis reports based on the review results, and dynamically adjust the classification standards of task complexity based on the statistical characteristics of abnormal tasks based on the review results. Finally, based on the tasks in the audit process, the review results and the optimized deep learning model, a closed-loop optimization mechanism for system performance is designed;

[0096] The steps of executing the prescription review module specifically include:

[0097] Design task complexity The tasks are classified into:

[0098] like Indicates low-complexity tasks, which are mainly reviewed by AI and do not require human intervention;

[0099] like It indicates tasks of medium complexity, which require the participation of pharmacists;

[0100] like Indicates highly complex tasks that require doctor review and final decision making;

[0101] Among them, θ1 is the low difficulty threshold, θ2 is the difficulty threshold;

[0102] Feeding standardized prescription text and patient information into deep learning models Generate review score and annotate suspicious points Among them, the set of annotation doubtful points including drug interaction points, dosage abnormalities and contraindications or allergy risks; for complex tasks Introducing special regularization terms Enhance the meticulousness of suspicious point annotations, as shown below:

[0103]

[0104] Among them, f AI represents a deep learning function, is a set of labeled doubt points The regularization term is calculated as follows:

[0105]

[0106] Among them, λ is the regularization weight, which controls the impact of suspicious point annotation on the score; Confidence(p) is the AI ​​confidence of suspicious point p, and the suspicious point with the lowest confidence will have the largest deduction on the score;

[0107] For task complexity The task is to have pharmacists participate in the review and form manual review scores For task complexity The task is reviewed by the doctor and the doctor generates the score

[0108] The review scores of multiple subjects are integrated into the final decision score For judging tasks Whether it has passed the review or needs to be modified is indicated as follows:

[0109]

[0110] Among them, α, β, and γ are dynamic weights based on task complexity. and node reliability Adjustment; for low-complexity tasks, α has the largest weight; for high-complexity tasks, β and γ have the largest weights;

[0111] Score the final decision The threshold value θ is not reached R Mark the task as abnormal and record a detailed set of doubts Abnormal tasks are forwarded to a higher-level review panel.

[0112] The beneficial technical effects of the present invention are at least as follows:

[0113] (1) The present invention introduces a dynamic load balancing mechanism in the regional medical network, monitors the task load and resource availability of each node in real time, and dynamically allocates review tasks based on the complexity of the prescription and the processing capacity of the node. Through this mechanism, the performance bottleneck and uneven resource utilization caused by the centralized processing architecture in the prior art are avoided, thereby significantly improving the efficiency of prescription review and resource utilization.

[0114] (2) The present invention designs a multi-subject collaborative decision-making model that organically combines AI review, pharmacist review, and doctor review, and comprehensively utilizes the professional capabilities of the three parties. During the review process, multiple opinions are integrated through a consistency algorithm to generate more accurate and reliable review results. This mechanism makes up for the defect that a single subject decision may produce deviations in the existing technology, and has significant advantages in dealing with complex or personalized prescriptions.

[0115] (3) The present invention further designs an abnormal review and tracking mechanism. For prescriptions with disagreements or abnormal situations in collaborative decision-making, the system will automatically submit them to the superior review team. At the same time, the system records all review processes to form a traceable review log. This mechanism not only improves the processing efficiency of abnormal prescriptions, but also enhances the transparency and security of the system, overcoming the problems of slow processing of abnormal prescriptions and lack of transparency in the review process in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0116] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0117] Figure 1 The present invention is a flowchart of a prescription review method based on a regional remote prescription review platform according to an embodiment of the present invention.

[0118] Figure 2 The present invention is a framework diagram of a prescription review system based on a regional remote prescription review platform according to an embodiment of the present invention. DETAILED DESCRIPTION

[0119] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0120] like Figure 1 As shown, an embodiment of the present invention provides a prescription review method based on a regional remote prescription review platform, and the method includes the following steps S1-S5:

[0121] S1. Receive the original prescription data set uploaded from the regional medical institutions, then extract the complexity features of each prescription in the original prescription data set to obtain the feature vector F, calculate the complexity score of each prescription using the weighted linear model based on the feature vector F, and finally sort all prescription data in descending order according to the complexity score, give priority to tasks with high scores, and form a task queue.

[0122] Specifically, the original prescription data set uploaded by the regional medical institutions is received, denoted as D Rx ={d1,d2,…,d n}, where each piece of data d i include:

[0123] T i : Prescription text content, including drug name, dosage, frequency of use and other information.

[0124] P i : Patient information, including age, medical history, allergy records, and other characteristics.

[0125] C i : Additional notes from the physician, such as special medication instructions.

[0126] Furthermore, in order to improve the integrity and consistency of the data, the following standardization processing is performed:

[0127] Drug name mapping: based on regional drug knowledge base Standardize drug names into international coding forms to eliminate spelling errors or common name interference.

[0128] Dosage unit unification: Standardize dosage units (such as "milligrams" and "tablets") into a unified format to facilitate subsequent calculations.

[0129] Patient Information Verification: Checks the completeness of allergy records and medical history information to ensure all key fields are valid.

[0130] After standardization, each piece of data is converted into a triple form:

[0131]

[0132] Get standardized prescription data

[0133] Furthermore, for each prescription Complexity features are extracted from the following three dimensions:

[0134] Number of drug interactions f1: from the drug knowledge base Query the drug combinations {t1, t2, …} in the prescription and calculate the logarithm of potential interactions, for example, f1 = |Interactions({t1, t2, …})|.

[0135] Abnormal dosage number f2: combined with regional medication rule library Check whether there are any drug items with too high or too low dosage in the prescription, and count the abnormal number.

[0136] Contraindications or allergy conflicts number f3: According to patient information P i, match drug contraindication symptoms or allergy conflict items, and count the total number of violations.

[0137] Through the above calculation, a feature vector is extracted for each prescription:

[0138] F i =[f1,f2,f3]

[0139] Among them, f1, f2, and f3 represent the number of drug interactions, the number of dosage anomalies, and the number of contraindication conflicts, respectively.

[0140] Furthermore, according to the extracted feature vector F i , calculate the complexity score S of each prescription i , using the following weighted linear model:

[0141] S i =w1f1+w2f2+w3f3

[0142] Among them, w1, w2, and w3 are preset weights, reflecting the impact of each feature on the audit complexity.

[0143] For example, w1=0.5 (drug interaction has the highest weight), w2=0.3 (dose abnormality is second), and w3=0.2 (contraindications are relatively low).

[0144] Further, according to the complexity score S i , sort all prescription data in descending order, give priority to tasks with high scores, and form a task queue:

[0145]

[0146] in Indicates that tasks are arranged from high complexity to low complexity.

[0147] Furthermore, the output priority sorted task queue Queue Rx , as the input for the next step of dynamic task allocation and load balancing.

[0148] S2. Obtain the dynamic status information of the review node, combine it with the task queue, design the allocation cost function to calculate the optimal matching comprehensive cost of the task and the review node, and the lowest value means the match is optimal. Then customize the task allocation strategy according to the result of the allocation cost function.

[0149] Specifically, the priority task queue output from step 1 As input to this step. Each task include:

[0150] represents the complexity score, which is used to measure the audit difficulty of the task; Indicates standardized prescription text; P ik Indicates patient information; Indicates doctor's note.

[0151] Furthermore, the dynamic status information of the review node is obtained at the same time:

[0152] L j represents the current task load of node j (number of assigned tasks); C j represents the processing capacity of node j (the number of tasks that can be processed per unit time); τ j represents the communication delay of node j (reflecting the response speed of the node network); Represents the task success rate of node j (based on the accuracy evaluation of historical tasks).

[0153] Furthermore, the availability index A of each node j is initialized j , which is used to reflect the real-time processing capability of the node:

[0154]

[0155] Among them, A j The higher the value, the lighter the load on the node and the stronger its task processing capability. This calculation ensures that the load and processing capability are dynamically updated to support subsequent allocation.

[0156] Furthermore, in order to optimize the task allocation logic for the scenario of this patent, an allocation cost function is defined To achieve the task The optimal match with node j. The function contains the following key factors:

[0157] Node load adjustment term: based on the real-time availability of node j A j As the core, ensure that tasks are assigned to high-availability nodes first;

[0158] Complexity adaptation items: High-complexity tasks need to be matched with high-reliability nodes first To reduce audit risk;

[0159] Communication delay correction term: Consider the network delay τ of the node j ,avoid allocating tasks to high-latency nodes;

[0160] Balance control item: Add a regularization item to prevent some nodes from being overloaded.

[0161] Furthermore, the allocation cost function of the comprehensive design is:

[0162]

[0163] in, Representation Task The comprehensive cost assigned to node j, the lower the value, the better; λ1, λ2 are weight parameters used to adjust the importance of communication delay and complexity adaptation items; α is the reliability adjustment factor used to amplify or reduce The impact of j =L j / C j Indicates the load utilization rate of node j, which is used for balance control to avoid overload.

[0164] Furthermore, the task allocation strategy is as follows:

[0165] Calculate the cost: for each task Traverse all nodes j∈{1,2,…,m}, according to Calculate the allocation cost.

[0166] Select the optimal node: select the node j with the smallest cost * , complete the task assignment:

[0167]

[0168] Update Node Status: Tasks After allocation, the node's L is updated in real time j and A j , ensuring that the next task is assigned based on the latest status.

[0169] This process is repeated until the task queue Queue Rx All tasks in the assignment are completed.

[0170] Furthermore, in order to cope with the possible task backlog of nodes, a dynamic adjustment mechanism is designed:

[0171] When the load utilization rate U of a node j j >β (β is a preset threshold, for example, β=0.85), triggering task transfer;

[0172] Select the task with the lowest complexity score from the task set of node j and recalculate and transfer to the node j′ with the second best cost;

[0173] At the same time, the value of λ2 is dynamically adjusted to enhance the priority of complexity adaptability in task transfer.

[0174] Furthermore, the assigned task list Task is output Node , record the task set of each node:

[0175] Task Node ={(j1,{d i1 ,d i2,…}),(j2,{d ik ,…}),…}

[0176] The task set of each node j is directly used as the input of multi-agent collaborative review in the next step.

[0177] S3. Set the task complexity and build a deep learning model to pre-screen the task allocation queue. Combine the task complexity and pre-screening results to perform corresponding detailed processing strategies for each task. Finally, merge each detailed processing strategy into a final decision score to determine whether the current task has passed the review or needs to be modified.

[0178] Specifically, to enhance audit efficiency, according to the complexity of the task To grade the tasks:

[0179] Low-complexity tasks are mainly reviewed by AI, without human involvement;

[0180] Medium-complexity tasks require the participation of pharmacists;

[0181] Highly complex tasks require doctor review and final decision.

[0182] Hierarchical task queues are assigned to corresponding audit paths to ensure accurate matching of task processing and audit resources.

[0183] Furthermore, all tasks are first passed through the deep learning model Conduct a preliminary review, including:

[0184] enter:

[0185] Output: AI-generated review score and annotate suspicious points include:

[0186] Drug interaction points (e.g., high-risk combinations);

[0187] Abnormal dosage (e.g. too high / too low);

[0188] Contraindications or allergy risks (combined calculate).

[0189] Furthermore, for complex tasks Introducing special regularization terms Enhance the meticulousness of suspicious point annotations:

[0190]

[0191] in, is a set of labeled doubt points The regularization term is calculated as follows:

[0192]

[0193] Among them, λ is the regularization weight, which controls the impact of suspicious point annotation on the score; Confidence(p) is the AI ​​confidence of suspicious point p. The lower the confidence, the greater the deduction of the suspicious point on the score. The goal of this design is to optimize the AI ​​pre-examination of complex tasks and reduce the misleading caused by low-confidence annotation.

[0194] Furthermore, for task complexity Tasks that pharmacists participate in reviewing:

[0195] Pharmacists receive a collection of doubts annotated by AI and rating Check and correct each item:

[0196] Remove false positive annotations;

[0197] Add missing annotations;

[0198] Modify medication dosage or combination recommendations.

[0199] Furthermore, pharmacists form manual review scores Defined as:

[0200]

[0201] in, It is a collection of doubts adjusted by the pharmacist.

[0202] Furthermore, for task complexity The tasks are reviewed by doctors:

[0203] Doctor receives pharmacist's review result and patient details Conduct final review based on the patient's personalized medical history;

[0204] Doctor-generated ratings Calculate the patient's personalized medication risk by combining specific weights:

[0205]

[0206] Furthermore, the review scores of multiple subjects are integrated into the final decision score For judging tasks Whether it has passed the review or needs to be modified. The fusion model is:

[0207]

[0208] Among them, α, β, and γ are dynamic weights based on task complexity. and node reliability Adjustment; for low-complexity tasks, α has a heavy weight; for high-complexity tasks, β and γ have a heavy weight.

[0209] Furthermore, the final score The threshold value θ is not reached R Mark the task as abnormal and record a detailed set of doubts Abnormal tasks are forwarded to a higher-level review panel.

[0210] Furthermore, the audit results are output as decisions for each task, which may include the following:

[0211] Review and approve tasks and their ratings;

[0212] Review failed tasks and make modification suggestions;

[0213] Abnormal tasks and doubtful points are recorded for use by the review module.

[0214] S4. Review the abnormal results of the deep learning model, feed back the review results to the deep learning model to optimize the deep learning model, and output the review results.

[0215] Specifically, for the abnormal task set Task Abnormal Categorize and prioritize by severity:

[0216] Classification basis: Combined with the complexity score of the task and doubt collection The proportion of high-risk annotations in the data can be divided into the following three categories:

[0217] High-risk abnormal tasks: including critical drug interactions, dosage exceeding the limit and other serious problems;

[0218] Medium-risk abnormal tasks: tasks involving a large number of suspicious points but not endangering patient safety;

[0219] Low-risk anomaly tasks: contain only a small number of suspicious annotations, which may be false positives.

[0220] Priority Calculation: Defining Severity Scores For sorting:

[0221]

[0222] in, Indicates the number of suspicious points marked as high risk; Indicates the total number of suspicious points; represents the comprehensive score; w1 and w2 represent weight parameters, which are used to balance the impact of high-risk doubts and low scores. Sort from high to low to form a priority review task queue Review .

[0223] Further, according to the review task queue Review Dynamically assign tasks to expert review teams based on priority and task type:

[0224] Selection of review experts: The system records each expert’s professional fieldε m (such as kidney disease drugs, hepatotoxic drugs, etc.), matching task suspicious type Priority will be given to experts in related fields;

[0225] Expert load balancing: Based on the current task load L of the expert m and task processing efficiency C m , real-time review capabilities of computing experts:

[0226]

[0227] Prioritize the assignment of tasks to experts with higher review capabilities. Each review team receives its assigned task set and starts the detailed review process.

[0228] Furthermore, during the expert review stage, the tasks Doubts Check each item and generate review decisions Include the following information:

[0229] Correction of scores after review Reflect the experts' adjustments to the original review results;

[0230] Corrected doubt set Delete the false positives and add the missing problems;

[0231] Task pass status Indicates whether the task has passed the review.

[0232] Furthermore, the review process and results of each task are recorded in the exception log. Abnormal ,include:

[0233] Unique task identifier

[0234] Original review score Correction rating

[0235] Review expert’s annotation opinions;

[0236] Final task status

[0237] Log data is used for subsequent optimization of audit rules and system tracking analysis.

[0238] Furthermore, the results of the exception review are fed back to the audit system in the following ways:

[0239] Optimize AI model: collect new doubts added during review Adjustments to expert annotations As training data to improve AI models Ability to mark suspicious points;

[0240] Update the review rule base: Update the rules related to drug interactions, dosage specifications, etc., and generate a new regional medication rule base

[0241] Improve task classification standards: Adjust the classification and priority calculation methods of abnormal tasks based on the characteristic distribution of high-risk abnormal tasks.

[0242] Furthermore, the reviewed task results are summarized and output, including:

[0243] Review tasks and revise scores;

[0244] Tasks that failed the review and the experts’ suggestions for revision;

[0245] Abnormal Log Abnormal , as the basic data for tracking analysis.

[0246] This step compensates for the potential misjudgment problem in multi-subject audits through the dynamic allocation of review experts and closed-loop feedback mechanism, and provides a clear direction for the optimization of system rules. The logging and tracking of results support the transparency and traceability of the system, while enhancing the robustness of the audit system in long-term operation.

[0247] S5. Publish the review results at multiple levels, generate a regional medication risk analysis report based on the review results, and dynamically adjust the classification standards of task complexity based on the statistical characteristics of abnormal tasks based on the review results. Finally, design a closed-loop optimization mechanism for system performance based on the tasks in the audit process, the review results, and the optimized deep learning model.

[0248] Specifically, receive the final review result set Task for each task from the review output of step 4 Final and exception log Abnormal , specifically including: It indicates the comprehensive score after review; represents the corrected doubt set; Indicates the task passing status, indicating whether the task has passed the review; A unique identifier for each task, used to track its audit process and logging.

[0249] At the same time, receive the intermediate data accumulated from step 1 to step 4, including: review score Marking doubtful points The processing time and accuracy of each entity in the audit process will be used for subsequent system optimization.

[0250] Further, according to the review results Task Final , generate audit reports for each task and publish them by category:

[0251] Directly publish through tasks: For Generates a brief report of the task, including the task identifier score and passing records.

[0252] Generate modification suggestions for failed tasks: Tasks, combined with a set of doubts Provide specific modification suggestions and file them in the medical institution’s feedback system.

[0253] Record the review comments of abnormal tasks: Generate a review log for abnormal tasks and record them in detail Review score Final state and expert opinions for system tracking and analysis.

[0254] Furthermore, according to the exception log Abnormal And the score distribution of all tasks, generate a regional medication risk analysis report Risk ,include:

[0255] Statistics of high-risk drugs: by analyzing Count frequently occurring high-risk drug combinations and their corresponding patient groups;

[0256] Performance evaluation of audit mode: based on and Evaluate the accuracy and efficiency of multi-subject collaborative audits;

[0257] Performance analysis of nodes and experts: Combine the historical task processing records of each node and expert to generate a node performance report for optimizing task allocation strategies.

[0258] Furthermore, the review results are fed back to the audit rule base and AI models Optimize rules and model performance:

[0259] Rule base optimization: through analysis The new doubts are updated to update the drug interaction rules and contraindication check rules to form a new rule base

[0260] Model retraining: Using the newly added annotated data during the review process (task scoring and doubt correction ), retrain the AI ​​model Improve the accuracy of annotations:

[0261]

[0262] in, Represents the optimized model.

[0263] Furthermore, the characteristic distribution of abnormal tasks is statistically analyzed based on the review results, and the classification criteria of task complexity are dynamically adjusted:

[0264] For abnormal tasks middle The distribution is clustered and the complexity score thresholds θ1 and θ2 are redefined;

[0265] Update the logic of task allocation and multi-agent collaborative review to make resource allocation more accurate.

[0266] Furthermore, based on the task data, review results and optimized model in the audit process, a closed-loop optimization mechanism for system performance is designed:

[0267] Record the rule base after each optimization and Model version number to ensure that the optimization process is traceable;

[0268] Generate audit system performance analysis report every quarter Performance , perform trend analysis on task processing time, audit accuracy, and number of abnormal tasks;

[0269] Dynamically update node task allocation strategies to improve task processing efficiency.

[0270] Furthermore, the following results are finally output:

[0271] Audit reports issued to medical institutions, including suggestions for changes to tasks that passed and failed;

[0272] Regional Medication Risk Analysis Report Risk ;

[0273] Audit rule base and AI models An updated version of;

[0274] System Performance Analysis Report Performance , as reference data for long-term optimization.

[0275] This step builds a closed-loop feedback system through result publishing, rule updating and model optimization to ensure continuous improvement of the audit, while the report generation of regional risk analysis supports the optimal allocation of medical resources and provides support for the scalability and long-term operation of the system.

[0276] like Figure 2 As shown, in another embodiment of the present invention, a prescription review system based on a regional remote prescription review platform is provided, and the system includes:

[0277] The original prescription collection module 101 is used to receive the original prescription data set uploaded by the regional medical institutions, then extract the complexity feature of each prescription in the original prescription data set to obtain the feature vector F, calculate the complexity score of each prescription using the weighted linear model according to the feature vector F, and finally sort all the prescription data in descending order according to the complexity score, give priority to processing the tasks with high scores, and form a task queue;

[0278] The prescription allocation module 102 is used to obtain the dynamic state information of the prescription review node, form a task queue based on the information, design an allocation cost function to calculate the optimal matching comprehensive cost of the task and the prescription review node, and the lowest value means the matching is optimal, and then customize the task allocation strategy based on the result of the allocation cost function;

[0279] The prescription review module 103 is used to set the task complexity, build a deep learning model to pre-review the task allocation queue, and refine the corresponding processing strategy for each task based on the task complexity and the pre-review result. Finally, each refined processing strategy is integrated into a final decision score to determine whether the current task has passed the review or needs to be modified;

[0280] A prescription review module 104 is used to review the abnormal results of the deep learning model, feed back the review results to the deep learning model to optimize the deep learning model, and output the review results;

[0281] System optimization module 105 is used to publish the review results at multiple levels, generate a regional medication risk analysis report based on the review results, and dynamically adjust the classification standards of task complexity based on the statistical characteristics of abnormal tasks according to the review results. Finally, based on the tasks in the audit process, the review results and the optimized deep learning model, a closed-loop optimization mechanism for system performance is designed;

[0282] The steps of executing the prescription review module specifically include:

[0283] Design task complexity The tasks are classified into:

[0284] like Indicates low-complexity tasks, which are mainly reviewed by AI and do not require human intervention;

[0285] like It indicates tasks of medium complexity, which require the participation of pharmacists;

[0286] like Indicates highly complex tasks that require doctor review and final decision making;

[0287] Among them, θ1 is the low difficulty threshold, θ2 is the difficulty threshold;

[0288] Feeding standardized prescription text and patient information into deep learning models Generate review score and annotate suspicious points Among them, the set of annotation doubtful points including drug interaction points, dosage abnormalities and contraindications or allergy risks; for complex tasks Introducing special regularization terms Enhance the meticulousness of suspicious point annotations, as shown below:

[0289]

[0290] Among them, f AI represents a deep learning function, is a set of labeled doubt points The regularization term is calculated as follows:

[0291]

[0292] Among them, λ is the regularization weight, which controls the impact of suspicious point annotation on the score; Confidence(p) is the AI ​​confidence of suspicious point p, and the suspicious point with the lowest confidence will have the largest deduction on the score;

[0293] For task complexity The task is to have pharmacists participate in the review and form manual review scores For task complexity The task is reviewed by the doctor and the doctor generates the score

[0294] The review scores of multiple subjects are integrated into the final decision score For judging tasks Whether it has passed the review or needs to be modified is indicated as follows:

[0295]

[0296] Among them, α, β, and γ are dynamic weights based on task complexity. and node reliability Adjustment; for low-complexity tasks, α has the largest weight; for high-complexity tasks, β and γ have the largest weights;

[0297] Score the final decision The threshold value θ is not reached R Mark the task as abnormal and record a detailed set of doubts Abnormal tasks are forwarded to a higher-level review panel.

[0298] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of the present invention. In practical applications, technicians in this field can select part or all of them according to actual needs to achieve the purpose of the present embodiment, and no limitation is made here.

[0299] In addition, for technical details that are not described in detail in this embodiment, reference can be made to the parameter operation method provided in any embodiment of the present invention, and will not be repeated here.

[0300] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0301] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0302] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory / random access memory, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0303] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A prescription review method based on a regional remote prescription review platform, characterized in that: The method comprises: S1. Receive the original prescription data set uploaded by the regional medical institutions, then extract the complexity features of each prescription in the original prescription data set to obtain the feature vector F, calculate the complexity score of each prescription using the weighted linear model according to the feature vector F, and finally sort all prescription data in descending order according to the complexity score, give priority to tasks with high scores, and form a task queue; S2. Obtain the dynamic status information of the review node, combine it with the task queue, design the allocation cost function to calculate the optimal matching comprehensive cost of the task and the review node, the lowest value means the match is optimal, and then customize the task allocation strategy according to the result of the allocation cost function; S3. Set the task complexity, build a deep learning model to pre-screen the task allocation queue, and refine the corresponding processing strategy for each task based on the task complexity and the pre-screening result. Finally, merge each refined processing strategy into a final decision score to determine whether the current task has passed the review or needs to be modified. S4, reviewing the abnormal results of the deep learning model, feeding back the review results to the deep learning model to optimize the deep learning model, and outputting the review results; S5. Publish the review results at multiple levels, generate a regional medication risk analysis report based on the review results, and dynamically adjust the classification standards of task complexity based on the statistical characteristics of abnormal tasks based on the review results. Finally, design a closed-loop optimization mechanism for system performance based on the tasks in the audit process, the review results, and the optimized deep learning model. Wherein, the S3 specifically includes: Design task complexity The tasks are classified into: Indicates low-complexity tasks, which are mainly reviewed by AI and do not require human intervention; It indicates tasks of medium complexity, requiring the participation of pharmacists; Indicates highly complex tasks that require doctor review and final decision making; Among them, θ1 is the low difficulty threshold, θ2 is the difficulty threshold; Feeding standardized prescription text and patient information into deep learning models Generate review score and annotate suspicious points Among them, the set of annotation doubtful points including drug interaction points, dosage abnormalities and contraindications or allergy risks; for complex tasks Introducing special regularization terms Enhance the meticulousness of suspicious point annotations, as shown below: Among them, f AI represents a deep learning function, is a set of labeled doubt points The regularization term is calculated as follows: Among them, λ is the regularization weight, which controls the impact of suspicious point annotation on the score; Confidence(p) is the AI ​​confidence of suspicious point p, and the suspicious point with the lowest confidence will have the largest deduction on the score; For task complexity The task is to have pharmacists participate in the review and form manual review scores For task complexity The task is reviewed by the doctor and the doctor generates the score The review scores of multiple subjects are integrated into the final decision score For judging tasks Whether it has passed the review or needs to be modified is indicated as follows: Among them, α, β, and γ are dynamic weights based on task complexity. and node reliability Adjustment; for low-complexity tasks, α has the largest weight; for high-complexity tasks, β and γ have the largest weights; Score the final decision The threshold value θ is not reached R Mark the task as abnormal and record a detailed set of doubts Abnormal tasks are forwarded to a higher-level review panel.

2. A prescription review method based on a regional remote prescription review platform according to claim 1, characterized in that: The original prescription data set is represented as D Rx ={d1,d2,…,d i ,d n }, where each piece of data d i Includes: Prescription text content T i , Patient Information i Additional notes from the doctor C i ; After receiving the original prescription data set uploaded by the regional medical institution, the original prescription data set is standardized, including: mapping of drug names, unification of dosage units and verification of patient information; After standardization, each piece of data is converted into a triple form: in, Represents the standardized data and obtains the standardized prescription data The complexity feature extraction of each prescription in the original prescription data set to obtain a feature vector F specifically includes: For each prescription Extracted from the number of drug interactions f1, the number of dosage abnormalities f2, and the number of contraindications or allergic conflicts f3, each prescription Extract a feature vector: <h2 style=";text-align:left;direction:ltr">F<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> (f1,f2,f3) Among them, f1, f2, and f3 represent the number of drug interactions, the number of dosage anomalies, and the number of contraindication conflicts, respectively; The complexity score S of each prescription is calculated using the weighted linear model based on the feature vector F. i , specifically: S i =w1f1+w2f2+w3f3 Among them, w1, w2, and w3 are preset weights, reflecting the impact of each feature on the audit complexity; The above method arranges all prescription data in descending order according to the complexity score, prioritizes tasks with high scores, and forms a task queue. Rx , specifically: in, Indicates that tasks are arranged from high complexity to low complexity.

3. The prescription review method based on the regional remote prescription review platform according to claim 1 is characterized in that: The dynamic status information of the review node includes the current task load of node j, the processing capacity of node j, the communication delay of node j and the task success rate of node j, and initializes the availability index A of each node j j , which is used to reflect the real-time processing capability of the node, is expressed as follows: Among them, A j The highest value indicates that the node has the lightest load and the strongest task processing capability.

4. A prescription review method based on a regional remote prescription review platform according to claim 3, characterized in that: The allocation cost function is expressed as follows: in, Indicates the task The comprehensive cost assigned to node j has the lowest optimal value; λ1, λ2 represent weight parameters, which are used to adjust the importance of communication delay and complexity adaptation items; α represents the reliability adjustment factor, which is used to enlarge or reduce The impact of j =L j / C j Represents the load utilization rate of node j, which is used for balance control to avoid overload; τ j represents the communication delay of node j; L j represents the current task load of node j, C j represents the processing capacity of node j; represents the task success rate of node j; The task allocation strategy specifically includes: For each task Traverse all nodes j∈{1,2,…,m}, according to Calculate the allocation cost; Select the node j with the smallest cost * , complete the task allocation, expressed as follows: Select the node j with the smallest cost * , complete the task assignment: Repeat until the task queue is full. Rx All tasks in the assignment are completed.

5. A prescription review method based on a regional remote prescription review platform according to claim 4, characterized in that: S2 also includes designing a dynamic adjustment mechanism to deal with the node task backlog, specifically including: When the load utilization rate of node j is U j >β, where β is the preset threshold that triggers task transfer; Select the task with the lowest complexity score from the task set of node j and recalculate And transfer to the node j′ with the second best cost; at the same time, dynamically adjust the value of λ2 to improve the priority of complexity adaptability in task transfer.

6. The prescription review method based on the regional remote prescription review platform according to claim 1 is characterized in that: The complexity of the task Pharmacists participate in the review of the following tasks: Pharmacists receive a collection of doubts annotated by AI and rating Check and correct each item, including: removing misreporting labels, adding missing labels, and modifying drug dosage or combination recommendations; Pharmacists form manual review scores Defined as: in, is the set of doubts after adjustment by the pharmacist, and Adjust is the correction function; The complexity of the task The tasks are reviewed by doctors, including: Doctor receives pharmacist's review result and patient information The doctor-generated score is based on the patient's personalized medical history for final review The patient's personalized medication risk is calculated by combining specific weights as follows: Among them, g Doc Represents the final judgment function; The preliminary review output results include: tasks that have passed the review and their scores, tasks that have not passed the review and their modification suggestions, and abnormal tasks and doubt records.

7. The prescription review method based on the regional remote prescription review platform according to claim 1 is characterized in that: The S4 specifically includes: Accept the task set marked as abnormal in S3 Abnormal , where the abnormal task set Task Abnormal Include each task The following information: Final review score Indicates the comprehensive score of multi-subject review; doubt point collection Record the specific issues marked by the review; mark the reasons Describes the triggering conditions for a task to be marked as abnormal; For abnormal task set Task Abnormal Categorize and sort by severity priority to get the review task queue Review ; Wherein, the classification includes high-risk abnormal tasks, medium-risk abnormal tasks and low-risk abnormal tasks; According to the review task queue Review Dynamically assign tasks to expert review teams based on priority and task type; During the expert review stage, Doubts Check each item and generate review decisions At the same time, the review process and results of each task are recorded in the exception log Abnormal ; Wherein, the review decision Including score correction after review Corrected doubt set and the task passes the status The abnormal log Log Abnormal Include the task unique identifier Original review score Correction rating Review expert's annotation opinions and final task status Feedback the review results to the deep learning model for optimization, and update and generate a new regional medication rule library At the same time, according to the characteristic distribution of high-risk abnormal tasks, the classification and priority calculation method of abnormal tasks are adjusted.

8. A prescription review method based on a regional remote prescription review platform according to claim 7, characterized in that: The severity priority order is calculated as follows: in, represents the priority score; Indicates the number of suspicious points marked as high risk; Indicates the total number of suspicious points; Indicates the comprehensive score; w1, w2 are weight parameters used to balance the impact of high-risk doubts and low scores; score according to priority Sort from high to low to form a priority review task queue Review ; According to the review task queue Queue Review Dynamically assign tasks to the expert review team based on priority and task type, including: According to each expert's field of expertise m , the suspicious type of the matching task Priority will be given to experts in related fields; In the corresponding field, according to the current task load L of the expert m and task processing efficiency C m , the real-time review capability of the computing expert is expressed as follows: Among them, A m Indicates review ability; tasks are assigned preferentially to experts with the highest review ability among experts in related fields.

9. The prescription review method based on the regional remote prescription review platform according to claim 7 is characterized in that: The S5 specifically includes: Task based on review results Final , generate audit reports for each task and publish them in layers, including: for Generates a brief report of the task, including the task identifier score and passing records; for Tasks, combined with a set of doubts Provide specific modification suggestions and file them in the medical institution’s feedback system; for Tasks, combined with a set of doubts Provide specific modification suggestions and file them in the medical institution’s feedback system; According to the abnormal log Log Abnormal And the score distribution of all tasks, generate a regional medication risk analysis report Risk ; The regional drug risk analysis report Risk Includes statistics of high-risk drugs, performance evaluation of audit models, and performance analysis of nodes and experts; Utilize the newly added annotation data during the review process, which includes the task scores and doubt correction Retraining Deep Learning Models Improve the accuracy of annotations and update the review rule base According to the review results, the characteristic distribution of abnormal tasks is statistically analyzed, and the classification standards of task complexity are dynamically adjusted, including: middle The distribution is clustered and the complexity score thresholds θ1 and θ2 are redefined; Record the rule base after each optimization and Model The version number of the audit system is generated every quarter. Performance , dynamically update the node task allocation strategy and perform closed-loop optimization.

10. A prescription review system based on a regional remote prescription review platform, characterized in that: The system comprises: The original prescription collection module is used to receive the original prescription data set uploaded by the regional medical institutions, and then extract the complexity features of each prescription in the original prescription data set to obtain the feature vector F, calculate the complexity score of each prescription using the weighted linear model according to the feature vector F, and finally sort all prescription data in descending order according to the complexity score, give priority to tasks with high scores, and form a task queue; The prescription allocation module is used to obtain the dynamic status information of the prescription review node, form a task queue based on it, design an allocation cost function to calculate the optimal matching comprehensive cost of the task and the prescription review node, and the lowest value means the match is optimal, and then customize the task allocation strategy based on the result of the allocation cost function; The prescription review module is used to set the task complexity and build a deep learning model to pre-review the task allocation queue. It combines the task complexity and the pre-review results to perform corresponding detailed processing strategies for each task. Finally, each detailed processing strategy is integrated into the final decision score to determine whether the current task has passed the review or needs to be modified. A prescription review module, used to review the abnormal results of the deep learning model, feed back the review results to the deep learning model to optimize the deep learning model, and output the review results; The system optimization module is used to publish the review results at multiple levels, generate regional drug risk analysis reports based on the review results, and dynamically adjust the classification standards of task complexity based on the statistical characteristics of abnormal tasks based on the review results. Finally, based on the tasks in the audit process, the review results and the optimized deep learning model, a closed-loop optimization mechanism for system performance is designed; The steps of executing the prescription review module specifically include: Design task complexity The tasks are classified into: like Indicates low-complexity tasks, which are mainly reviewed by AI and do not require human intervention; like It indicates tasks of medium complexity, which require the participation of pharmacists; like Indicates highly complex tasks that require doctor review and final decision making; Among them, θ1 is the low difficulty threshold, θ2 is the difficulty threshold; Feeding standardized prescription text and patient information into deep learning models Generate review score and annotate suspicious points Among them, the set of annotation doubtful points including drug interaction points, dosage abnormalities and contraindications or allergy risks; for complex tasks Introducing special regularization terms Enhance the meticulousness of suspicious point annotations, as shown below: Among them, f AI represents a deep learning function, is a set of labeled doubt points The regularization term is calculated as follows: Among them, λ is the regularization weight, which controls the impact of suspicious point annotation on the score; Confidence(p) is the AI ​​confidence of suspicious point p, and the suspicious point with the lowest confidence will have the largest deduction on the score; For task complexity The task is to have pharmacists participate in the review and form manual review scores For task complexity The task is reviewed by the doctor and the doctor generates the score The review scores of multiple subjects are integrated into the final decision score For judging tasks Whether it has passed the review or needs to be modified is indicated as follows: Among them, α, β, and γ are dynamic weights based on task complexity. and node reliability Adjustment; for low-complexity tasks, α has the largest weight; for high-complexity tasks, β and γ have the largest weights; Score the final decision The threshold value θ is not reached R Mark the task as abnormal and record a detailed set of doubts Abnormal tasks are forwarded to a higher-level review panel.

Images