AI prevention and control system for adverse reactions of centralized purchasing drugs based on real world data

Abstract

The application provides an AI prevention and control system for adverse reactions of centrally-purchased drugs based on real-world data, and relates to the technical field of artificial intelligence, comprising: a data collection module for collecting real-world data of centrally-purchased drugs; a signal identification module for identifying adverse reaction risk signals of drugs based on pre-trained AI models according to real-world data; a grading evaluation module for grading and evaluating adverse reaction risk signals of drugs; a strategy formulation module for formulating grading prevention and control strategies based on grading evaluation results; and a strategy execution module for executing grading prevention and control strategies. The application realizes automatic identification, grading evaluation and effective prevention and control of adverse reaction risks of centrally-purchased drugs, integrates multi-source data by using AI technology, can quickly and accurately find potential risks, and significantly reduces the occurrence of adverse reactions through grading management and strategy execution, thereby ensuring patient safety and improving the efficiency and accuracy of drug safety supervision.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an AI-based system for preventing and controlling adverse drug reactions in centrally procured drugs based on real-world data. Background Technology

[0002] Currently, with the continuous development and application of drugs, the widespread use of centrally procured drugs involving a large number of patients has led to increasingly prominent adverse reaction issues. Traditional monitoring methods rely on manual analysis and single data sources, making it difficult to quickly and accurately identify and address potential risks in multi-source data, resulting in insufficient efficiency and accuracy.

[0003] Therefore, there is an urgent need for a prevention and control system based on real-world data and combined with artificial intelligence technology. This system can automatically identify risks, assess hazards in a tiered manner, and formulate targeted strategies to improve drug safety supervision capabilities and protect patients' health. Summary of the Invention

[0004] One of the objectives of this invention is to provide an AI-based system for preventing and controlling adverse drug reactions in centrally procured drugs based on real-world data, in order to address the problems mentioned in the background section.

[0005] The AI-based adverse drug reaction prevention and control system for centralized procurement drugs, based on real-world data, provided in this embodiment of the invention includes:

[0006] The data collection module is used to collect real-world data on drugs procured through centralized procurement.

[0007] The signal recognition module is used to identify adverse drug reaction risk signals based on pre-trained AI models and real-world data.

[0008] The grading assessment module is used to grade and assess adverse drug reaction risk signals.

[0009] The strategy formulation module is used to formulate tiered prevention and control strategies based on the results of the tiered assessment.

[0010] The strategy execution module is used to execute tiered prevention and control strategies.

[0011] Optionally, the data collection module collects real-world data on drugs procured through centralized procurement, including:

[0012] Collect historical usage data of centrally procured drugs from multiple data sources;

[0013] Data cleaning is performed on historical usage data to obtain real-world data.

[0014] Optionally, the multiple data sources include at least: medical institutions, pharmaceutical companies, and regulatory platforms.

[0015] Optionally, the methods for cleaning historical usage data include at least: missing value imputation and outlier removal.

[0016] Optionally, the pre-training steps for the AI ​​model include:

[0017] A machine learning model was obtained by using a large amount of real-world sample data of centrally procured drugs labeled with adverse drug reaction risk signals.

[0018] Optionally, the grading assessment module performs a grading assessment of adverse drug reaction risk signals, including:

[0019] Based on pre-defined grading and assessment rules with multiple grading dimensions, adverse drug reaction risk signals are graded and assessed.

[0020] Optionally, the multiple grading dimensions include at least: severity, frequency of occurrence, and preventability.

[0021] Optionally, the grading assessment results include: low risk, medium risk, and high risk.

[0022] Optionally, the strategy formulation module formulates a tiered prevention and control strategy based on the tiered assessment results, including:

[0023] The corresponding tiered prevention and control strategy is determined from the pre-set tiered prevention and control strategy library based on the tiered assessment results.

[0024] Optionally, the strategy execution module executes a tiered prevention and control strategy, including:

[0025] A tiered prevention and control strategy is implemented through the centralized procurement drug adverse reaction platform.

[0026] The present invention has achieved the following beneficial effects:

[0027] By constructing an AI-based prevention and control system based on real-world data, the system enables automatic identification, hierarchical assessment, and effective prevention and control of adverse reaction risks associated with centrally procured drugs. By integrating multi-source data using AI technology, potential risks can be quickly and accurately identified, and the occurrence of adverse reactions can be significantly reduced through hierarchical management and strategy execution, thereby ensuring patient safety and improving the efficiency and accuracy of drug safety supervision.

[0028] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0029] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0030] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0031] Figure 1 This is a schematic diagram of the AI-based adverse drug reaction prevention and control system for centralized procurement drugs based on real-world data, as described in an embodiment of the present invention.

[0032] Figure 2 This is another schematic diagram of the AI-based adverse drug reaction prevention and control system for centralized procurement drugs based on real-world data in an embodiment of the present invention. Detailed Implementation

[0033] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0034] The research and development approach of this application is to construct an AI-based risk control system based on real-world data (RWD) to automatically identify, classify, assess, and effectively control the adverse reaction risks of centrally procured drugs. The system aims to utilize AI technology combined with multi-source data to quickly and accurately identify potential risks, and through tiered management and strategy execution, reduce the occurrence of adverse drug reactions, ensure patient safety, and improve the efficiency and accuracy of drug safety supervision.

[0035] Figure 1 This application provides a schematic diagram of an AI-based adverse drug reaction prevention and control system for centralized procurement drugs based on real-world data, as an example. Figure 1 As shown, the system includes:

[0036] Data collection module 100 is used to collect real-world data on drugs procured through centralized procurement. The data collection module 100 collects real-world data on drugs procured through centralized procurement, including:

[0037] Historical usage data of centrally procured drugs will be collected from multiple data sources. These multiple data sources include at least: medical institutions (such as electronic medical record systems and drug usage records of hospitals and clinics), pharmaceutical companies (such as drug sales and adverse reaction report data of pharmaceutical companies), and regulatory platforms (such as databases or adverse reaction monitoring systems of drug regulatory authorities).

[0038] Historical usage data is cleaned to obtain real-world data. The methods for cleaning historical usage data include at least: missing value imputation (using interpolation, mean imputation, or similarity-based imputation methods to fill missing parts of the data) and outlier removal (detecting and removing outliers using statistical methods such as Z-score or IQR to prevent them from interfering with subsequent analysis).

[0039] Here's an implementation example: Drug A, procured through centralized procurement, is used in multiple hospitals, and the system needs to collect relevant data. The data collection module 100 extracts drug A's usage records from a hospital's electronic medical record system, obtains sales batch information from the pharmaceutical company, and downloads adverse reaction reports from the regulatory platform. It discovers missing dosage data for some patients, and the system fills in the missing data based on the average of the patient's age and weight (e.g., the average adult dose is 10mg daily). A record showing a dosage of 1000mg / day is detected, far exceeding the normal range (10-20mg / day), and is identified as an outlier and removed using the IQR method. Finally, cleaned real-world data is obtained, including fields such as patient age, dosage, and adverse reaction descriptions.

[0040] The signal recognition module 200 is used to identify adverse drug reaction risk signals based on real-world data and a pre-trained AI model. The pre-training steps of the AI ​​model include:

[0041] A machine learning model was obtained by using a large amount of real-world sample data of centrally procured drugs labeled with adverse drug reaction risk signals.

[0042] This study utilizes large-scale real-world drug sample data labeled with adverse drug reaction risk signals to construct and apply an AI detection model. First, the data is collected and standardized (including cleaning, standardized encoding such as MedDRA / ATC, feature engineering, and partitioning into training / validation / test sets). The Transformer architecture (suitable for sequence / text data) or CNN architecture (suitable for structured data) is selected, and the model is built using PyTorch or TensorFlow. The model is trained on the training set, and hyperparameters are fine-tuned using the validation set. Regularization (L1 / L2), Dropout, and early stopping strategies are applied to prevent overfitting. The model is fine-tuned for specific adverse drug reaction identification tasks (e.g., specific drug monitoring). The model performance is rigorously evaluated on independent test sets, with key metrics including precision, recall, F1 score, and AUC-ROC, ensuring its reliable identification of statistically significant risk signals. Deploy the pre-trained model to analyze new data and output specific, interpretable signals of potential ADR risks, such as reporting a statistically significant association between drug X and symptom Y (providing OR / RR and CI) or an abnormally high incidence of symptom Z in a specific population compared to the expected baseline.

[0043] Continuing the above implementation example: A deep neural network model is trained using data from centralized procurement drugs tagged with adverse reactions over the past 5 years (e.g., 100,000 records, of which 5,000 are marked as abnormal liver function). The model learns the association between abnormal liver function and specific medication patterns (e.g., high dose, elderly). When cleaned data on drug A is input into the model, it finds that the reported frequency of abnormal liver function for this drug is 8% in people over 65 years of age, significantly higher than the 2% in the normal population, identifying it as a potential risk signal. The output signal is: Risk signal: Drug A may cause abnormal liver function in the elderly.

[0044] The grading assessment module 300 is used to grade and assess adverse drug reaction risk signals. The grading assessment module 300 performs grading and assessment of adverse drug reaction risk signals, including:

[0045] Based on pre-defined grading assessment rules across multiple dimensions, adverse drug reaction risk signals are graded and assessed. These grading dimensions include at least: severity (the degree of harm caused by the adverse reaction, such as mild discomfort, hospitalization, or life-threatening conditions), frequency (the incidence rate of the adverse reaction, such as <1% being rare, 1-5% being common, and >5% being highly prevalent), and preventability (whether the risk can be reduced by adjusting the medication regimen or other measures). The risk level corresponding to the assessment results of adverse drug reaction risk signals under different grading dimensions can be pre-set as needed, and this level serves as the final grading assessment result.

[0046] Continuing with the above implementation example: Assess the risk signals of drug A causing abnormal liver function in the elderly. Abnormal liver function requiring medical intervention but not fatal is rated as Grade 3 (out of 5). An 8% incidence rate is rated as high (>5%). Risk can be reduced by lowering the dosage, rated as high. A combined risk level of Grade 3 severity, high frequency of occurrence, and high preventability corresponds to a moderate risk level.

[0047] The strategy formulation module 400 is used to formulate tiered prevention and control strategies based on the tiered assessment results. The tiered assessment results include: low risk, medium risk, and high risk. The strategy formulation module 400 formulates tiered prevention and control strategies based on the tiered assessment results, including:

[0048] The corresponding tiered prevention and control strategy is determined from the pre-set tiered prevention and control strategy library based on the tiered assessment results.

[0049] A pre-configured tiered prevention and control strategy library containing response measures corresponding to different risk levels. For example:

[0050] Low risk: Strengthen monitoring.

[0051] Intermediate risk: Adjust medication regimen and strengthen monitoring.

[0052] High risk: Suspend use and launch an investigation.

[0053] Continuing with the above implementation example: A strategy is developed for a moderate risk signal for drug A. The assessment result is moderate risk; the system selects a suggestion from its strategy library to adjust the drug dosage and enhance monitoring. The output control strategy is: for patients over 65 years of age, adjust the drug A dosage from 10 mg / day to 5 mg / day and monitor liver function weekly.

[0054] The strategy execution module 500 is used to execute the hierarchical prevention and control strategy. The strategy execution module 500 executes the hierarchical prevention and control strategy, including:

[0055] A tiered prevention and control strategy is implemented through the centralized procurement drug adverse reaction platform.

[0056] Utilize the adverse reaction platform for centrally procured drugs (such as an online monitoring system) to publish and implement strategies. Based on the strategy content, trigger corresponding measures, such as sending notifications, updating medication guidelines, or initiating investigations.

[0057] Continuing with the above implementation example: Implement the intermediate-risk strategy for drug A. The strategy implementation module 500 sends a notification to the healthcare institution via the platform, requesting that the dosage of drug A be adjusted to 5 mg / day for patients aged 65 and older, and that liver function be monitored weekly. The healthcare institution receives the notification and implements the strategy.

[0058] The aforementioned system, by constructing an AI-based prevention and control system based on real-world data, enables the automatic identification, hierarchical assessment, and effective prevention and control of adverse reaction risks associated with centrally procured drugs. By integrating multi-source data using AI technology, it can quickly and accurately identify potential risks and significantly reduce the occurrence of adverse reactions through hierarchical management and strategy execution, thereby ensuring patient safety and improving the efficiency and accuracy of drug safety supervision.

[0059] In some embodiments, the AI-based adverse drug reaction prevention and control system for centralized procurement drugs, based on real-world data, further includes:

[0060] Human-machine collaborative dynamic intervention module 600, such as Figure 2 As shown, it includes:

[0061] The context-aware unit 601 is used to capture a multi-dimensional dynamic parameter set in the execution context of the tiered prevention and control strategy in real time. The multi-dimensional dynamic parameter set includes: real-time status data of the medical environment, patient group characteristic evolution data, and drug usage feedback stream.

[0062] The context-aware unit 601 is responsible for real-time capture and analysis of multi-dimensional dynamic parameter sets during the implementation of the tiered prevention and control strategy. Through comprehensive monitoring of the medical environment, patient groups, and drug usage feedback, it constructs a dynamic profile of the current implementation context. The multi-dimensional dynamic parameter set can be obtained through the API interfaces of the information management systems of medical institutions, pharmaceutical companies, and regulatory platforms. The multi-dimensional dynamic parameter set includes the following three key dimensions:

[0063] Real-time status data of the medical environment: reflects the current resources and operational status of medical institutions, such as hospital bed occupancy rate and workload index of medical staff.

[0064] Evolutionary data on patient population characteristics: capturing dynamic trends in patient populations, such as age distribution and the proportion of underlying diseases.

[0065] Drug usage feedback stream: Real-time updates on the effects and adverse reactions of drugs after use, such as the number and type of adverse reaction reports.

[0066] The context-aware unit 601 captures multi-dimensional dynamic parameters such as the state of the medical environment, the evolution of patient group characteristics, and drug usage feedback in real time to construct a dynamic profile of the execution context, providing the system with an accurate real-time data foundation and ensuring the accuracy of subsequent decisions.

[0067] Here is an implementation example: Taking a certain centrally procured drug B as an example, assume its tiered prevention and control strategy is to reduce the dosage and strengthen monitoring for elderly patients. The context-aware unit 601 collects the following data in real time during the strategy execution process:

[0068] Real-time status data of the medical environment: The current bed occupancy rate of a certain tertiary hospital is 92%, and the workload index of medical staff is 8.5 (out of 10), indicating that resources are highly strained.

[0069] Evolutionary data on patient population characteristics: In the past week, the proportion of patients over 70 years old admitted to the hospital has increased from 25% to 40%, and 60% of elderly patients have hypertension or diabetes.

[0070] Drug usage feedback stream: In the past 48 hours, the regulatory platform received 5 adverse reaction reports related to drug B (3 of which were dizziness and 2 were hypotension), an increase from the previous week's daily average of 2.

[0071] The expert intervention decision-making unit 602 is used to generate an expert intervention urgency index based on a deep contextual analysis model and a multi-dimensional dynamic parameter set. Specifically, the deep contextual analysis model constructs a dynamic risk prediction matrix by integrating historical characteristics of critical events related to the implementation of prevention and control strategies with multi-dimensional dynamic parameters, and calculates the expert intervention urgency index based on this dynamic risk prediction matrix.

[0072] The expert intervention decision-making unit 602, based on a deep contextual analysis model, generates an expert intervention urgency index according to a multi-dimensional dynamic parameter set. Its core lies in quantifying the urgency of the current situation through a dynamic risk prediction matrix, thereby determining whether expert intervention is necessary. The specific steps are as follows:

[0073] Data input: Receives a multi-dimensional dynamic parameter set provided by the context-aware unit 601.

[0074] Model Construction: The deep contextual analysis model integrates critical event characteristics (such as hospitalizations due to adverse reactions) from the implementation of historical prevention and control strategies with multidimensional dynamic parameters to construct a dynamic risk prediction matrix. Matrix elements include risk factors, occurrence probabilities, and impact levels. When determining these matrix elements, the similarity between the multidimensional dynamic parameters and critical event characteristics is calculated. When the similarity exceeds a preset similarity threshold (e.g., 80%), the risk represented by the critical event characteristics is used as the risk factor, the similarity as the occurrence probability, and the magnitude of the historical impact of the critical event characteristics as the impact level.

[0075] Index Calculation: Based on the dynamic risk prediction matrix, a weighted calculation is performed to obtain the expert intervention urgency index. The expert intervention urgency index ranges from 0 to 1, with a higher value indicating a greater need for expert intervention.

[0076] Continuing with the above implementation example: Historical critical event characteristics show that when bed occupancy exceeded 90%, the proportion of elderly patients exceeded 35%, and there were 3 new adverse reaction reports per day, a mass hypotension event occurred, resulting in the hospitalization of 5 patients. Among the multidimensional dynamic parameters, the current bed occupancy rate is 92%, the proportion of elderly patients is 40%, and there are 3 new adverse reaction reports per day. Using Euclidean distance, the similarity between the historical critical event characteristics and the multidimensional dynamic parameters is approximately 96.2%, and the impact magnitude of the historical critical event characteristics is 0.8. Therefore, when the system generates a dynamic risk prediction matrix, the risk factors are bed resource stress of 92% (counted as occupancy rate), occurrence probability of 96.2%, and impact magnitude of 0.8. The weight of the risk factor is 0.4, the weight of the occurrence probability is 0.3, and the weight of the impact magnitude is 0.3. The weighted calculation yields an expert intervention urgency index of 0.8966 (0.92×0.4+0.962×0.3+0.8×0.3).

[0077] The expert intervention decision-making unit 602 is based on a deep contextual analysis model. It integrates the characteristics of historical emergency events with real-time parameters to construct a dynamic risk prediction matrix and calculate the urgency index of expert intervention. This enables a dynamic quantitative assessment of the current risk and provides a scientific basis for expert intervention.

[0078] The adaptive expert matching unit 603 is used to match the optimal domain expert and connect it to the execution context when the expert intervention urgency index exceeds the preset index threshold, based on a multi-dimensional dynamic parameter set and the dynamic capability evaluation results of experts in each domain in the expert knowledge graph.

[0079] When the urgency index for expert intervention exceeds a preset threshold (e.g., 0.6), the expert matching process is initiated. The expert knowledge graph maintains a dynamic database containing dynamic capability assessment results for experts in each domain, such as professional background (domain expertise), historical performance (success rate, processing timeliness), and current availability (online status). The expert capability requirements of the execution scenario, reflected by a multi-dimensional dynamic parameter set, are determined. Cosine similarity is used to calculate the matching degree between the dynamic capability assessment results of each domain expert and the expert capability requirements; the expert with the highest matching degree is selected as the optimal domain expert. Experts are then connected to the execution scenario via communication interfaces (e.g., video conferencing or real-time chat).

[0080] When the urgency index of expert intervention exceeds the threshold, the adaptive expert matching unit 603 quickly matches the optimal domain expert based on multidimensional dynamic parameters and expert knowledge graph, ensuring that appropriate expert resources are accessed at critical moments, thereby improving the timeliness and effectiveness of intervention.

[0081] The cognitive state monitoring unit 604 is used to analyze the decision-making behavior trajectory of the optimal domain expert after accessing the execution context in real time, and extract the cognitive state evolution feature vector from the decision-making behavior trajectory. The cognitive state evolution feature vector includes: decision response delay, operation path deviation, and historical behavior pattern matching degree.

[0082] The cognitive state monitoring unit 604 extracts cognitive state evolution feature vectors by analyzing the expert's decision-making behavior trajectory in the execution context in real time, thereby assessing their decision-making ability and workload level. The cognitive state evolution feature vectors include:

[0083] Decision response delay: The time it takes for an expert to make a decision after receiving information, reflecting the speed of response.

[0084] Operational path deviation: The degree of difference between expert operation and standard procedure (a standard expert intervention process set in advance based on risk factors), which measures the standardization of execution.

[0085] Historical behavior pattern matching: The similarity between current behavior and the expert's past successful cases, assessing consistency.

[0086] The cognitive state evolution feature vector can be obtained by collecting behavioral logs (such as click records and input timestamps) from the smart terminals used by experts and combining them with vector embedding technology.

[0087] The cognitive state monitoring unit 604 analyzes the expert's decision-making behavior trajectory in real time, extracts cognitive state evolution feature vectors including decision response delay, operation path deviation, and historical behavior pattern matching degree, assesses the expert's decision-making ability and workload level, and provides key basis for personalized assistance.

[0088] Continuing with the above implementation example: After expert A accesses the drug B scenario, the system monitors their behavior:

[0089] Decision response delay: After receiving system prompts, the average decision time is 35 seconds, which is higher than its historical average of 25 seconds.

[0090] Operation path deviation: Skipping the step of recording the reason for adjustment when adjusting the dose results in a deviation of 20%.

[0091] Historical behavioral pattern matching: The current decision matches 75% of past successful cases (such as dose reduction and monitoring).

[0092] The cognitive state evolution feature vector is [35, 0.2, 0.75]. The analysis shows that expert A has a high cognitive load (increased time delay) and a slightly decreased decision quality (increased deviation).

[0093] The hierarchical auxiliary strategy generation unit 605 is used to match auxiliary strategies based on the cognitive state evolution feature vector and decouple the auxiliary strategies into a sequence of autonomous sub-strategy units with temporal dependency logical relationships. Each sub-strategy unit in the autonomous sub-strategy unit sequence satisfies the autonomous understanding time constraint: the autonomous understanding time predicted based on the capability profile of the optimal domain expert for its sub-strategy unit does not exceed a preset tolerance time threshold.

[0094] The hierarchical auxiliary policy generation unit 605 matches personalized auxiliary policies based on the expert's cognitive state evolution feature vector and decouples them into a sequence of autonomous sub-policy units with temporal-dependent logical relationships. The process is as follows:

[0095] Strategy matching: Based on the cognitive state evolution feature vector, a suitable auxiliary strategy is selected from a strategy library (such as gradual guidance or strong intervention). The strategy library contains pre-set auxiliary strategies corresponding to different cognitive state evolution feature vectors.

[0096] Strategy decoupling: Decompose the auxiliary strategy into multiple sub-strategy units with temporal dependency logic. Temporal dependency logic means that these sub-strategy units need to be executed sequentially according to their order in the auxiliary strategy. Each sub-strategy unit has a clear expected strategy auxiliary goal (such as providing data or suggesting solutions).

[0097] Time Constraint: Based on expert capability profiles (such as historical response speed), predict the autonomous understanding time for each sub-strategy unit to ensure it does not exceed a preset tolerance time threshold (such as 30 seconds). The time taken for experts with different expert capability profiles to autonomously understand different sub-strategy units can be determined in advance through experiments.

[0098] The hierarchical auxiliary strategy generation unit 605 generates and decouples a sequence of auxiliary strategy units with temporal dependencies based on the expert's cognitive state, ensuring that the autonomous understanding time of each sub-strategy is controllable, and balancing the auxiliary effect with the expert's workload.

[0099] Continuing with the above implementation example: For expert A's feature vector [35, 0.2, 0.75], the system matches a progressive decision-making guidance strategy, which is decomposed as follows:

[0100] Sub-strategy unit 1: Provide a summary of adverse reaction trends (time prediction 10 seconds).

[0101] Sub-strategy unit 2: Suggested dose adjustment plan (time prediction 15 seconds).

[0102] Sub-strategy unit 3: Prompt for handling low blood pressure risk (estimated time: 20 seconds).

[0103] All times were less than the threshold of 30 seconds, thus satisfying the constraint.

[0104] The intelligent intervention control unit 606 is used to sequentially activate each sub-strategy unit in the autonomous sub-strategy unit sequence according to the extended time window sequence. Each extended time window in the extended time window sequence corresponds to one sub-strategy unit, and the duration of the extended time window is the sum of the autonomous understanding time of the corresponding sub-strategy and a preset buffer time.

[0105] Also used for:

[0106] For each activated sub-strategy unit, based on the sub-strategy unit, corresponding strategy assistance is provided to the optimal domain expert, and based on the updated extracted cognitive state evolution feature vector, the deviation of the optimal domain expert from the expected strategy assistance target of the sub-strategy unit is analyzed. When the deviation is lower than the preset deviation threshold, the sub-strategy unit is immediately terminated and the process jumps to the next extended time window.

[0107] When the number of consecutively terminated sub-strategy units reaches a threshold N, all subsequent strategy assistance to the optimal domain expert is completely terminated. Here, N is a positive integer, and its value is determined by the closest approximation between the strategy assistance effect represented by the consecutively terminated sub-strategy units and the overall assistance effect of the assistance strategy, which is the closest to a preset proximity threshold.

[0108] The intelligent intervention control unit 606 activates sub-strategy units according to the extended time window sequence and dynamically adjusts the intervention process. For each extended time window, the duration is the time required for the sub-strategy to autonomously comprehend itself plus a buffer time (e.g., 5 seconds). After activating a sub-strategy unit, the cognitive state evolution feature vector is updated, and the deviation is calculated (the cognitive state evolution feature vector reflects the latest cognitive state of the optimal domain expert; each sub-strategy unit has a set expected strategy assistance target, and the difference between the two can be calculated as the deviation). When the deviation is lower than a preset deviation threshold (e.g., 10%), it indicates that the optimal domain expert can autonomously complete the expected strategy assistance target, and the sub-strategy unit is immediately terminated, jumping to the next extended time window. If there are N consecutive terminations (e.g., 3 times), the intervention is terminated. The strategy assistance effect is pre-set through experiments when the expert autonomously comprehends the continuously terminated sub-strategy units. The overall assistance effect is pre-set for the assistance strategy. When the proximity between the two is closest to a preset proximity threshold (e.g., 0.8), it indicates that the optimal domain expert can autonomously achieve the overall assistance effect, and the intervention is terminated.

[0109] The intelligent intervention control unit 606 activates sub-strategy units according to the extended time window sequence and dynamically adjusts the intervention process based on the real-time performance of experts. When the expert performs well, unnecessary assistance is stopped, and when the expert performs poorly for a continuous period, the intervention is terminated, ensuring that the intervention is appropriate and efficient.

[0110] Continuing with the above implementation example: Set up an extended time window sequence for expert A:

[0111] Extended time window 1: Duration is 15 seconds (10+5). Within this window, sub-strategy unit 1 provides a trend summary. If the latency decreases to 30 seconds and the deviation is 5% (<10%), the process is terminated.

[0112] Extended time window 2: Duration is 20 seconds (15+5). Within this window, sub-strategy unit 2 provides suggested solutions. Deviation 15% (>10%), continue.

[0113] Extended time window 3: Duration is 25 seconds (20+5). Within this window, sub-strategy unit 3 will display a risk warning. If the deviation is 12% (>10%), continue.

[0114] The intervention is terminated when the intervention is interrupted 3 times consecutively (N=3).

[0115] In the AI-based adverse drug reaction prevention and control system for centralized drug procurement based on real-world data, the human-machine collaborative dynamic intervention module 600 is a key component. It aims to ensure the efficient implementation of tiered prevention and control strategies in complex and ever-changing medical environments through real-time monitoring and intelligent intervention. This module integrates functional units such as context awareness, expert intervention decision-making, adaptive expert matching, cognitive state monitoring, tiered auxiliary strategy generation, and intelligent intervention control. Through human-machine collaboration, it dynamically assesses risks, matches experts, and provides personalized assistance, thereby ensuring patient safety.

[0116] In the scenario of adverse reaction prevention and control of centralized procurement drugs, the application of the above-mentioned system is crucial. Because the use of centralized procurement drugs involves a large number of patients and multi-source data, the medical environment is complex and ever-changing. Traditional methods are difficult to deal with dynamic risks in real time. However, the human-machine collaborative dynamic intervention module 600 can capture multi-dimensional parameters in real time, dynamically assess risks, and match experts to provide personalized assistance, ensuring the efficient execution of prevention and control strategies, thereby effectively reducing the incidence of adverse reactions and improving regulatory efficiency.

[0117] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A centralized drug procurement adverse reaction prevention and control system based on real-world data, characterized in that, include: The data collection module is used to collect real-world data on drugs procured through centralized procurement. The signal recognition module is used to identify adverse drug reaction risk signals based on pre-trained AI models and real-world data. The grading assessment module is used to grade and assess adverse drug reaction risk signals. The strategy formulation module is used to formulate tiered prevention and control strategies based on the results of the tiered assessment. The strategy execution module is used to execute tiered prevention and control strategies; The system also includes: a human-machine collaborative dynamic intervention module, which includes: The context-aware unit is used to capture a multi-dimensional dynamic parameter set in the execution context of the hierarchical prevention and control strategy in real time; wherein, the multi-dimensional dynamic parameter set includes: real-time status data of the medical environment, patient group characteristic evolution data, and drug use feedback stream; The expert intervention decision-making unit is used to generate an expert intervention urgency index based on a deep contextual analysis model and a multi-dimensional dynamic parameter set. The deep contextual analysis model constructs a dynamic risk prediction matrix by integrating the characteristics of historical prevention and control strategy execution crisis events with multi-dimensional dynamic parameters, and calculates the expert intervention urgency index based on the dynamic risk prediction matrix. The adaptive expert matching unit is used to match the optimal domain expert and connect it to the execution context when the urgency index of expert intervention exceeds the preset index threshold, based on a multi-dimensional dynamic parameter set and the dynamic capability evaluation results of experts in each domain in the expert knowledge graph. The cognitive state monitoring unit is used to analyze the decision-making behavior trajectory of the optimal domain expert after accessing the execution context in real time, and extract the cognitive state evolution feature vector from the decision-making behavior trajectory. The cognitive state evolution feature vector includes: decision response delay, operation path deviation, and historical behavior pattern matching degree. A hierarchical auxiliary strategy generation unit is used to match auxiliary strategies based on the cognitive state evolution feature vector and decouple the auxiliary strategies into a sequence of autonomous sub-strategy units with temporal dependency logical relationships. Each sub-strategy unit in the autonomous sub-strategy unit sequence satisfies the autonomous understanding time constraint condition: the autonomous understanding time of the sub-strategy unit predicted based on the ability profile of the optimal domain expert does not exceed a preset tolerance time threshold. The intelligent intervention control unit is used to sequentially activate each sub-strategy unit in the autonomous sub-strategy unit sequence according to the extended time window sequence; wherein, each extended time window in the extended time window sequence corresponds to a sub-strategy unit, and the duration of the extended time window is the sum of the autonomous understanding time of the corresponding sub-strategy and the preset buffer time; the intelligent intervention control unit is also used to: for each activated sub-strategy unit, provide corresponding strategy assistance to the optimal domain expert based on the sub-strategy unit, and analyze the deviation of the optimal domain expert from the expected strategy assistance target of the sub-strategy unit based on the updated extracted cognitive state evolution feature vector; when the deviation is lower than the preset deviation threshold, immediately stop the sub-strategy unit and jump to the next extended time window; when the number of consecutively stopped sub-strategy units reaches the threshold N, then completely terminate all subsequent strategy assistance to the optimal domain expert; wherein, N is a positive integer, and its value is based on: the strategy assistance effect represented by the continuously stopped autonomous understanding sub-strategy units, its closeness to the overall assistance effect of the assistance strategy, which is closest to the preset closeness threshold; The data collection module collects real-world data on drugs procured through centralized procurement, including: Collect historical usage data of centrally procured drugs from multiple data sources; Data cleaning is performed on historical usage data to obtain real-world data.

2. The AI-based adverse drug reaction prevention and control system for centralized procurement drugs based on real-world data as described in claim 1, characterized in that, The multiple data sources include at least: medical institutions, pharmaceutical companies, and regulatory platforms.

3. The AI-based adverse drug reaction prevention and control system for centralized procurement drugs based on real-world data as described in claim 1, characterized in that, The methods for cleaning historical usage data include at least: filling in missing values ​​and removing outliers.

4. The AI-based adverse drug reaction prevention and control system for centralized procurement drugs based on real-world data as described in claim 1, characterized in that, The pre-training steps for AI models include: A machine learning model was obtained by using a large amount of real-world sample data of centrally procured drugs labeled with adverse drug reaction risk signals.

5. The AI-based adverse drug reaction prevention and control system for centralized procurement drugs based on real-world data as described in claim 1, characterized in that, The grading assessment module performs grading assessments on adverse drug reaction risk signals, including: Based on pre-defined grading and assessment rules with multiple grading dimensions, adverse drug reaction risk signals are graded and assessed.

6. The AI-based adverse drug reaction prevention and control system for centralized procurement drugs based on real-world data as described in claim 5, characterized in that, The multiple grading dimensions include at least: severity, frequency of occurrence, and preventability.

7. The AI-based adverse drug reaction prevention and control system for centralized procurement drugs based on real-world data as described in claim 1, characterized in that, The tiered assessment results include: low risk, medium risk, and high risk.

8. The AI-based adverse drug reaction prevention and control system for centralized procurement drugs based on real-world data as described in claim 1, characterized in that, The strategy formulation module formulates tiered prevention and control strategies based on the tiered assessment results, including: The corresponding tiered prevention and control strategy is determined from the pre-set tiered prevention and control strategy library based on the tiered assessment results.

9. The AI-based adverse drug reaction prevention and control system for centralized procurement drugs based on real-world data as described in claim 1, characterized in that, The strategy execution module executes a tiered prevention and control strategy, including: A tiered prevention and control strategy is implemented through the centralized procurement drug adverse reaction platform.

Images