A medicine sales management method based on artificial intelligence

Abstract

The application relates to the technical field of data management, and discloses a medicine sales management method based on artificial intelligence, which comprises the following steps: collecting medicine sales data of multiple heterogeneous sources to obtain a medicine sales data set, training a prediction model of medicine sales demand according to the medicine sales data set and medicine sales influencing factors, predicting the medicine sales demand of a future preset period by using the prediction model to obtain a prediction result, establishing a real-time transaction monitoring mechanism according to the prediction result to detect and give an early warning of the abnormality of medicine transaction behaviors, and constructing a blockchain tracing system to establish an unalterable medicine sales record chain and an intelligent medicine classification system, so that different categories of medicines are classified and stored to obtain an allocation control strategy. The application can significantly improve the medicine circulation efficiency, deeply excavate the correlation between the sales law and external factors, accurately capture the change characteristics of the medicine market demand, optimize the inventory cost, and enhance the market supervision compliance.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of data management, in particular to a medicine sales management method based on artificial intelligence. BACKGROUND

[0002] With the rapid development of modern medicine circulation system, especially in the aspects of digital transformation, supply chain optimization and intelligent management, it is increasingly urgent to accurately predict medicine sales demand, monitor transaction abnormalities in real time and trace the whole chain. Although the traditional medicine sales management method has made certain progress in basic data collection, inventory management and sales statistics, it still cannot systematically solve the core problems such as standardization integration of multi-source heterogeneous data, deviation analysis of prediction model and actual sales data, and cross-regional medicine category collaborative control, and it is difficult to meet the fine management needs of different regions, different channels and different categories of medicines.

[0003] Therefore, how to integrate multi-source heterogeneous data, deeply mine the correlation between sales rules and external factors, accurately capture the characteristics of changes in medicine market demand, and have a strong adaptive medicine sales intelligent management method has become a problem to be solved. SUMMARY

[0004] The application provides a medicine sales management method based on artificial intelligence, which solves the technical problems that it is difficult to integrate multi-source heterogeneous data, deeply mine the correlation between sales rules and external factors, accurately capture the characteristics of changes in medicine market demand, and have a strong adaptability.

[0005] The application provides a medicine sales management method based on artificial intelligence, which comprises the following steps:

[0006] Collecting medicine sales data of multiple heterogeneous sources, and cleaning, standardizing and storing the multiple medicine sales data to obtain a medicine sales data set;

[0007] According to the medicine sales data set and the medicine sales influencing factors, a prediction model of medicine sales demand is trained and constructed;

[0008] The prediction model predicts the medicine sales demand in a future preset period to obtain a prediction result;

[0009] According to the prediction result, a real-time transaction monitoring mechanism is established to detect and warn the abnormality of medicine transaction behavior;

[0010] Based on the prediction result and the real-time transaction monitoring mechanism, a blockchain tracing system is constructed to establish an unalterable medicine sales record chain;

[0011] According to the prediction result and the blockchain traceability system, an intelligent medicine classification system is constructed to obtain a fine inventory allocation control strategy;

[0012] The allocation control strategy divides the medicine categories into sub-regions to obtain a plurality of sub-regional medicine categories.

[0013] According to the sub-regional medicine category data combined with the prediction result, a medicine sales investigation data is obtained.

[0014] According to the medicine sales investigation data, deviation analysis, correlation analysis and causality analysis are performed on the actual sales data and the prediction result of the sub-regional medicine categories to obtain an analysis result.

[0015] According to the analysis result, the interference factors of the emerging influence factors affecting the sales data are screened out to obtain accurate emerging influence factors.

[0016] The accurate emerging influence factors are fed back to the prediction model for parameter optimization and model iteration.

[0017] As a preferred scheme of the present application, a real-time transaction monitoring mechanism is established, which comprises:

[0018] Based on the prediction result, a plurality of risk level evaluation threshold ranges are set.

[0019] The risk score is obtained by comparing the deviation degree of the actual sales transaction data and the prediction data.

[0020] The current risk level is obtained by comparing the score result with the plurality of risk level evaluation threshold ranges.

[0021] Further comprising a multi-level anomaly detection architecture constructed according to the plurality of risk level evaluation threshold ranges, to perform multi-level anomaly violation detection on the prediction result, and obtain a detection result; the detection result indicates that the sales data contains false or untruthful information content caused by sales behavior through improper means or violation; (medicine over-purchasing, abnormal price fluctuation, non-normal time transaction)

[0022] The first layer matches a plurality of preset abnormal sales transaction rules through a rule engine; the abnormal sales behavior is screened out through the plurality of abnormal sales transaction rules to obtain an abnormal sales behavior.

[0023] The second layer performs abnormal quantification risk scoring on the abnormal sales behavior; if there are at least two abnormal sales behaviors, the risk level evaluation is performed on the risk score according to the plurality of risk level evaluation threshold ranges to obtain an evaluation result; the corresponding level of early warning mechanism is triggered according to the evaluation result to obtain a hierarchical early warning.

[0024] The third layer dynamically adjusts the pre-warning threshold range based on the hierarchical pre-warning feedback, so as to accurately adjust the risk level boundary of multi-level abnormal violation detection.

[0025] As a preferred scheme of the present application, a blockchain traceability system is constructed, comprising:

[0026] Based on the prediction results and the real-time transaction monitoring mechanism, the drug sales data are in chain connection with the production batch information, the logistics track and the quality inspection report, which are not tamperable;

[0027] According to the chain connection, a distributed ledger architecture is established, and the whole life cycle information of each drug transaction is stored in the form of blocks;

[0028] Further comprising constructing an intelligent contract mechanism, when abnormal sales behavior is detected, triggering a traceability query and responsibility positioning procedure;

[0029] Through the traceability query and responsibility positioning procedure, the rapid traceability query of the drug whole life cycle sales record and the responsibility positioning of the abnormal sales behavior are performed.

[0030] As a preferred scheme of the present application, an intelligent drug classification system is constructed, comprising:

[0031] Based on the prediction results and multiple dimension influence factors, the drugs are intelligently classified and placed in multiple sub-regional drug categories with different demand characteristics; the multiple dimension influence factors include but are not limited to sales geographical position, sales channel, drug category, etc.;

[0032] The sub-regional drug categories are dynamically divided according to the drug sales trend data to obtain multiple trend characteristic drug categories; wherein, by analyzing the historical sales data time series of the sub-regional drug categories, and by trend analysis, the drug categories with rising trend, falling trend, stable trend and fluctuation trend in the current period are identified, and the drug categories with different trend characteristics are classified into different trend partitions respectively;

[0033] For the drug categories in the rising trend partition, an aggressive expansion inventory strategy is formulated to increase the safety stock and the replenishment frequency, and to preferentially allocate high-quality shelf positions and promotion resources;

[0034] For the drug categories in the falling trend partition, a conservative contraction inventory strategy is formulated to reduce the inventory level and the order quantity, and to strengthen the clearance promotion and the substitute product recommendation;

[0035] For the drug categories in the stable trend partition, a balanced maintenance inventory strategy is formulated to maintain the existing inventory level and the replenishment rhythm;

[0036] A flexible adaptive inventory strategy is formulated for the drug categories in the fluctuation trend partition, and a dynamic safety stock and an elastic replenishment mechanism are set; the fluctuation trend partition indicates that the sales data fluctuates frequently and obviously, and has an unstable trend;

[0037] Based on the drug categories of multiple trend characteristics, the sub-regional drug categories are analyzed according to the efficacy of the drugs, so as to establish an efficacy associated category matching mechanism;

[0038] The efficacy associated category matching mechanism identifies the drug category combinations containing synergistic therapeutic effect, complementary functional effect and joint drug use advantage by constructing a drug efficacy associated knowledge graph, and forms an efficacy associated category matrix.

[0039] The efficacy associated category matrix comprises:

[0040] In the rising trend partition, when the sales of the core efficacy drugs increase, the partition configuration and sales data of the efficacy associated categories are promoted synchronously, and the sales data of the associated categories are synergistically increased by the main categories;

[0041] In the falling trend partition, when the sales of the core efficacy drugs decline, the configuration structure of the efficacy associated categories is adjusted in time, and the sales demand is diverted through alternative category conversion and efficacy upgrade;

[0042] Based on the efficacy associated category matrix, a cross-partition efficacy associated category linkage mechanism is established, when the sales data of the core efficacy drugs in a certain partition abnormally fluctuate, the sales warning of the associated efficacy categories in other partitions is triggered, and the sales supply chain of the main and auxiliary drugs is coordinated.

[0043] As a preferred scheme of the present application, the sub-regional drug category data are investigated in combination with the prediction results, including:

[0044] The basic data of the drug categories in each sub-region are collected to obtain the sales key indicators; the sales key indicators include actual sales volume, inventory level, turnover rate, customer group characteristics, sales linkage coefficient, abnormal event propagation time delay, risk diffusion range and cross-regional demand conduction strength quantitative indicators;

[0045] The collected actual sales key indicators are compared and matched with the prediction results to identify the sub-regions with prediction deviation anomalies and the drug categories in the current sub-regions;

[0046] For the drug categories with prediction deviation anomalies, the sales environment, market competition status and consumer behavior change sales influencing factors of the drug categories in the current sub-region are investigated in depth;

[0047] The sales influencing factors in all the investigation information are integrated to form structured drug sales investigation data;

[0048] According to the drug sales research data mining, the sales correlation mode, abnormal sales behavior propagation path and risk diffusion law among sub-regions are mined to identify emerging factors and potential risks affecting drug sales;

[0049] The emerging influence factors affecting drug sales form an influence network related to each other;

[0050] The influence network identifies, quantifies and predicts the complex correlation between multiple factors to analyze and prospectively warn the changes in the drug sales environment.

[0051] The influence network specifically includes:

[0052] The prescription electronic trend driven by the popularization of digital medicine and the development of telemedicine services form a synergistic effect, which jointly promotes the digital transformation of drug sales channels and impacts the traditional drugstore sales mode, and this digital transformation process directly catalyzes the in-depth application of artificial intelligence assisted diagnosis technology;

[0053] Based on the improvement of digital medical infrastructure, the application of artificial intelligence assisted diagnosis technology and the deep integration of the prescription electronic trend, the precise medical demand and personalized drug sales mode are promoted, and the popularization of the precise medical mode in turn accelerates the improvement of health management consciousness in the whole society;

[0054] The preventive drug demand growth catalyzed by the improvement of health management consciousness is exponentially amplified through the social media health information dissemination channel, and the two promote each other and reshape the drug category structure, forming a new drug consumption ecology oriented to prevention and health care;

[0055] The social media health information dissemination has a guiding influence on the consumer drug selection behavior, and forms a positive feedback cycle with the improvement of health management consciousness, and this cycle effect further strengthens the market acceptance and user stickiness of telemedicine services;

[0056] The development of telemedicine services improves the timeliness and convenience of drug distribution, which is interdependent with multiple factors such as the popularization of digital medicine, the prescription electronic trend, the application of artificial intelligence assisted diagnosis technology, and finally builds an online and offline integrated drug sales ecological system characterized by technology driven, demand oriented and service integrated. The ecological system realizes intelligent collaborative optimization of the whole chain of drug sales through a multi-factor linkage mechanism.

[0057] As a preferred scheme of the present application, deviation analysis, correlation analysis and causal relationship analysis are performed on the actual sales data and prediction results of the sub-regional drug categories to obtain analysis results, including:

[0058] The absolute deviation, relative deviation and deviation distribution characteristics of the actual sales data and the prediction results of each sub-region are obtained to identify the deviation abnormal drug categories and time nodes, and a deviation analysis result is obtained;

[0059] Based on the deviation analysis result, the correlation coefficient of the sales data between different sub-regions and different drug categories is obtained to obtain a correlation analysis result of the sales data between drug categories;

[0060] Based on the correlation analysis result, the root cause of the prediction deviation is identified through a causal inference algorithm, including the causal relationship of external environmental factors, model parameter settings, and data quality problems, to obtain a causal relationship analysis result;

[0061] Integrate the three analysis results to form a comprehensive analysis result including deviation feature description, correlation relationship graph, and causal influence chain;

[0062] According to the comprehensive analysis result combined with the drug sales research data, the emerging influencing factors affecting drug sales are analyzed in depth and the interference factors of the emerging influencing factors are screened out to obtain accurate emerging influencing factors.

[0063] As a preferred scheme of the present application, the emerging influencing factors affecting drug sales are analyzed in depth and the interference factors of the emerging influencing factors are screened out, including:

[0064] A multi-dimensional factor analysis framework is established to quantitatively evaluate the emerging influencing factors through three dimensions of influence intensity, action time delay and duration cycle to form an influence factor matrix;

[0065] An interference factor identification system is constructed based on the influence factor matrix to identify and screen out the interference factors in the emerging influencing factors;

[0066] Among them, the emerging influencing factors are subjected to outlier detection to obtain an outlier detection result;

[0067] The outlier detection result identifies the abnormal fluctuation points in the policy environment factor data;

[0068] When the medical insurance directory adjustment frequency data exceeds the preset abnormal value of the historical mean standard deviation, it is marked as a potential interference factor and is subjected to audit confirmation, and the confirmed abnormal data is processed by using the median replacement method, and the processing result is fed back to the influence factor matrix for weight adjustment;

[0069] Based on the outlier detection result, redundant variables in the social and economic factors are analyzed and screened out;

[0070] When the correlation coefficient between the income level of residents in the social and economic factors and the health consciousness degree exceeds the preset threshold value, it is identified as a high correlation redundant variable, combined with the influence strength in the influence factor matrix, the redundant variable main factor with large influence strength is retained, and the redundant variable secondary factor is removed; To avoid multiple collinearity interference, and update the screening result to the influence factor matrix at the same time.

[0071] As a preferred scheme of the present application, it further comprises:

[0072] By identifying the low variance interference term in the emerging influence factor, when the variance of the new drug research and development progress data in the preset time range is less than the preset value of the total variance, a low information amount interference factor is obtained.

[0073] The low information amount interference factor is removed from the model input variable to obtain a removal result.

[0074] The removal result is transmitted to the dynamic monitoring mechanism for recording;

[0075] Based on the removal result, through time series stationarity test, the non-stationary interference sequence in the competition pattern factor is screened out, and the ADF test is performed on the market concentration change data. When the P value is greater than the preset value, it is determined as a non-stationary sequence interference factor.

[0076] The non-stationary sequence interference factor is eliminated by difference transformation or detrending processing to obtain a processing result.

[0077] The processing result is fed back to the interference factor processing verification program to obtain an accurate interference factor processing result.

[0078] As a preferred scheme of the present application, it further comprises:

[0079] A dynamic monitoring mechanism for interference factors is established, based on the identification results of various interference factors, and the interference factor detection threshold is set as a double standard that the influence weight is lower than the preset value and the prediction contribution is less than the preset value.

[0080] When a certain interference factor meets the interference factor standard for a continuous preset time, it is automatically removed from the influence factor matrix, and the weight redistribution of the influence factor matrix is triggered at the same time.

[0081] A interference factor processing verification program is constructed to backtest the influence factor set after the interference factors are screened out.

[0082] Among them, the emerging influence factors before and after the interference factors are screened out are respectively input into the prediction model to obtain two output results.

[0083] According to the two output results, the root mean square error improvement degree is obtained, when the error improvement degree exceeds the preset improvement value, it is confirmed that the interference factor is screened out effectively, and the new influencing factor set after screening out the interference factor is input into the multi-dimensional factor analysis framework for the next round of evaluation, otherwise it is fed back to the dynamic monitoring mechanism to re-evaluate the screening standard, forming a closed-loop optimization mechanism to ensure the continuous accuracy of the influence factor identification and interference factor screening.

[0084] In a second aspect, a computer-readable storage medium is used to store computer-readable instructions capable of running the above-mentioned artificial intelligence-based drug sales management method when the computer-readable instructions are read by a computer.

[0085] The beneficial effects of the present application are as follows: through automatic collection, cleaning and standardization processing of multi-source heterogeneous data, the data integration time is greatly shortened, and the efficiency is significantly improved; based on the intelligent demand prediction of the LSTM-XGBoost fusion prediction model, the prediction period is compressed from weekly to daily, the prediction accuracy is significantly improved, the amount of manual intervention is greatly reduced, and the automation level of drug sales management is effectively improved; through the intelligent drug classification system and cross-trend partition inventory coordination linkage mechanism, the drug circulation turnover speed is greatly improved, the supply chain response time is significantly shortened, and the systematic optimization of the whole chain circulation efficiency is realized.

[0086] By constructing a multi-dimensional feature vector, the time characteristics, policy characteristics, correlation characteristics and other multi-dimensional sales influencing factors are deeply mined to accurately identify the characteristics of changes in drug market demand; the correlation strength and direction of the sales data between different sub-regions are analyzed by using Pearson correlation coefficient and Spearman correlation coefficient, and the correlation identification accuracy is significantly improved; the Granger causality test and cointegration test methods are used to analyze the causal relationship and long-term equilibrium relationship of the sales data between regions, the action path and strength of external factors such as policy change lag effect, seasonal factor periodicity influence and sudden event impact effect are successfully identified, and a dynamic influencing factor knowledge graph is established to realize comprehensive insight and deep analysis of the drug sales law.

[0087] Based on the vector autoregressive model, the dynamic influence path and transmission mechanism of external shocks on the sales of each sub-region are identified, the demand change early warning accuracy is significantly improved; the root cause of prediction error is analyzed by using the causal inference algorithm, the mutual influence relationship and regional demand transmission mechanism of the sales data between sub-regions are mined, the cross-regional synergistic effect is identified, and the demand fluctuation prediction accuracy is greatly improved; a factor influence strength model is established to assign weight coefficients and influence time delay parameters to each key influencing factor, and real-time monitoring and accurate prediction of market demand change characteristics are realized.

[0088] By intelligent drug classification system and multi-dimensional partition control strategy, the inventory proportion of unsalable drugs is significantly reduced, and the capital turnover rate is greatly improved; based on the ABC-XYZ classification method combined with the trend partition characteristics, the multi-dimensional inventory allocation control strategy is implemented, the rising trend partition is implemented, the conservative contraction type strategy is executed in the falling trend partition, the balanced maintenance type strategy is adopted in the stable trend partition, and the flexible adaptive type strategy is constructed in the fluctuating trend partition, the inventory holding cost is greatly reduced, and the shortage rate is significantly controlled; the cross-trend partition inventory coordination linkage mechanism and the inventory replacement scheme based on the efficacy correlation are established, the overall inventory turnover efficiency is significantly improved, and the inventory cost is significantly optimized and the economic benefit is maximized.

[0089] The "rule + algorithm" double-layer abnormal transaction monitoring mechanism is constructed, the accuracy of abnormal transaction identification is significantly improved, and the investigation cost is greatly reduced; the unalterable drug sales record chain is established by using the block chain technology, the sales data is chained in association with the production batch information, the logistics track and the quality inspection report, the data tracing requirement in the "Drug Circulation Supervision and Management Method" is completely met, and the quality problem response time is greatly shortened; the intelligent contract mechanism is established, when the abnormal transaction or the prediction deviation exceeds the threshold value, the tracing query and the responsibility positioning are automatically triggered, the whole process supervision efficiency is significantly improved, and the transparent supervision and compliance guarantee of the whole life cycle of the drug are realized.

[0090] By inputting the research results of key influencing factors in the form of structured data into the prediction model, the online adaptive updating of the model is realized by using the incremental learning technology, the prediction accuracy is continuously improved, and the effect is significant; the model performance monitoring mechanism is established, the prediction accuracy improvement effect is evaluated regularly, and the model architecture is dynamically adjusted and the algorithm is intelligently optimized according to the evaluation results, so that the system can adapt to the change of market environment and maintain long-term stable high-precision prediction ability; the inventory allocation decision support system is constructed, the trend prediction, cost optimization and risk assessment three core algorithm modules are integrated, personalized inventory parameter suggestions are generated for each drug category, and precise and intelligent inventory allocation control is realized. BRIEF DESCRIPTION OF DRAWINGS

[0091] Figure 1 It is a kind of drug sales management method flow chart based on artificial intelligence provided in the embodiment of the application;

[0092] Figure 2 It is the allocation control strategy method flow chart provided in the embodiment of the application. DETAILED DESCRIPTION

[0093] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that discussions of these implementations are merely provided to enable better understanding of the subject matter described herein, and can include various process steps, components, or structures that can be changed, modified, or eliminated. Various examples can omit, substitute, or add various process steps or components. Also, it should be understood that instead of being performed at the same time, the processes or steps can be performed sequentially or in different orders. Furthermore, some examples can be described as being performed by a single device or component, while other examples can be described as being performed by a plurality of devices or components.

[0094] As shown in Figure 1 An artificial intelligence-based pharmaceutical sales management method includes the following steps:

[0095] Step 1: Collecting a plurality of heterogeneous sources of pharmaceutical sales data, and performing cleaning, standardization processing and storage on the plurality of pharmaceutical sales data to obtain a pharmaceutical sales data set;

[0096] Step 2: Training a prediction model for pharmaceutical sales demand according to the pharmaceutical sales data set and pharmaceutical sales influencing factors; the prediction model predicts the pharmaceutical sales demand in a future preset period to obtain a prediction result;

[0097] Step 3: Establishing a real-time transaction monitoring mechanism according to the prediction result, and performing abnormal detection and abnormal early warning on pharmaceutical transaction behavior;

[0098] Step 4: Based on the prediction result and the real-time transaction monitoring mechanism, a blockchain traceability system is constructed to establish an unalterable pharmaceutical sales record chain;

[0099] Step 5: According to the prediction result and the blockchain traceability system, an intelligent pharmaceutical classification system is constructed to obtain a fine-grained inventory allocation control strategy.

[0100] As shown in Figure 2 Further, the allocation control strategy includes:

[0101] Step 51: The allocation control strategy divides the various categories of pharmaceuticals into sub-regions to obtain a plurality of sub-regional pharmaceutical categories;

[0102] Step 52: Investigating according to the sub-regional pharmaceutical category data combined with the prediction result to obtain pharmaceutical sales investigation data;

[0103] Step 53: According to the pharmaceutical sales investigation data, deviation analysis, correlation analysis and causal relationship analysis are performed on the actual sales data and the prediction result of the sub-regional pharmaceutical categories to obtain an analysis result;

[0104] Step 54: According to the analysis results, further analyze and exclude the interference factors of the emerging influencing factors of sales data to obtain accurate emerging influencing factors; and feed the accurate emerging influencing factors back to the prediction model for parameter optimization and model iteration.

[0105] Further, the prediction model is a fusion model which is obtained by weighted average method The sub-models of integrated long short-term memory network (LSTM), extreme gradient boosting (XGBoost) and autoregressive integrated moving average model (ARIMA) output; wherein the weight coefficients of each sub-model are dynamically determined based on the root mean square error (RMSE) on the validation set, and the specific calculation formula is: Wherein i, j respectively represent LSTM, XGBoost, ARIMA model; the weight is recalculated at each model retraining, and satisfies .

[0106] The blockchain traceability system is constructed, including: adopting Hyperledger Fabric alliance chain framework, network nodes including drug regulatory departments, production enterprises and circulation enterprises; the consensus mechanism adopts Byzantine fault tolerance (PBFT) algorithm; the smart contract is written in Go language, wherein the preset abnormal trigger condition is: when the risk score output by the real-time transaction monitoring mechanism is greater than 55 points or the prediction deviation continuously exceeds 25% for a specified time length, the traceability query contract is automatically executed; the traceability query contract traces the whole life cycle record of the drug batch number by querying the Merkle tree structure of the on-chain block, and calls the responsibility positioning algorithm to automatically generate the responsibility report according to the transaction timestamp, digital signature and link metadata.

[0107] The abnormal detection rule library in the real-time transaction monitoring mechanism, the rules of which are generated and updated in the following way: the initial rule set is defined by business experts according to historical violation cases; subsequently, frequent abnormal patterns are mined using association rule learning (such as Apriori algorithm) by analyzing historical transaction data, and the rules are added to the rule library after being confirmed by experts; each rule is configured with confidence and support threshold, and is given a dynamic weight, which is periodically adjusted according to the historical triggering accuracy of the rule.

[0108] The trend analysis in the construction of the intelligent medicine classification system includes: using statistical significance test (such as Mann-Kendall trend test) to judge the trend of the sales data time series; for each medicine category, calculate the sales Sen's slope of the consecutive n months (n≥3), when the slope is greater than 0 and the p value is less than 0.05, it is determined to be an upward trend; when the slope is less than 0 and the p value is less than 0.05, it is determined to be a downward trend; when the p value is greater than or equal to 0.05, it is determined to be a stable trend; when the sequence unit root test (ADF test) is determined to be non-stationary and cannot be determined by trend test, it is determined to be a fluctuating trend.

[0109] It also includes interference factor screening of emerging factors affecting medicine sales: using variance inflation factor (VIF) detection and screening redundant variables with multiple collinearity higher than the preset threshold (VIF>10); using low variance filtering method, removing low information features with variance less than 1% of the total sum of all feature variances; unit root test (ADF test) is performed on the time series type influence factor, and non-stationary sequence is screened out, and the stationary sequence is standardized.

[0110] The smart contract (SmartContract) in the blockchain traceability system contains the core TraceabilityQuery function. The execution logic of this function is as follows:

[0111] Listen to the message queue output by the real-time transaction monitoring module, when the risk score is greater than 55 or the prediction deviation rate of a certain category is more than 25-28% for t hours in a row, automatically call the contract.

[0112] According to the medicine batch number in the trigger event, traverse the blockchain ledger, compare the medicine information (such as production batch, logistics order number) stored in the block body, and use Merkle tree structure to quickly locate and retrieve all related transaction blocks.

[0113] The responsibility locating algorithm (ResponsibilityLocator) embedded in the contract analyzes the retrieved on-chain data, automatically determines the most likely link and responsible subject according to the transaction type, link (production, logistics, sales), digital signature, and timestamp anomaly, and generates a structured responsibility report, which is stored on the chain and notified to the relevant regulatory nodes.

[0114] The trend judgment in the intelligent medicine classification system uses statistical methods to enhance objectivity:

[0115] Mann-Kendall non-parametric trend test method is used to analyze the historical sales time series of medicine categories. Calculate Z statistic and p value, when p value is less than significance level α (usually set to 0.05), it is considered that there is a significant trend.

[0116] The magnitude of the trend is calculated using Sen's slope estimation method. Sen's slope is a robust slope estimation method that is not sensitive to outliers in the time series.

[0117] The trend is comprehensively determined in combination with the significance test result and the Sen's slope value: when p<0.05 and Sen's slope>0, it is determined that there is a significant upward trend; when p<0.05 and Sen's slope<0, it is determined that there is a significant downward trend; when p≥0.05, it is determined that there is no significant trend (stable trend); when the sequence is determined to be non-stationary by the ADF test, but the Mann-Kendall test does not show a significant monotonic trend, it is determined to be a fluctuating trend.

[0118] The data preprocessing process in the emerging influencing factor affecting drug sales aims to screen out interference factors, and includes the following steps:

[0119] The variance inflation factor (VIF) of all continuous variables is obtained. Variables with a VIF value greater than 10-15 are identified as having serious multicollinearity, and more representative variables are retained according to business knowledge, and the remaining redundant variables are removed.

[0120] The variance of all features is obtained, and features with a variance value less than 1-2% of the total sum of all feature variances are removed. These features have too small a change in value and are considered to be low-information noise.

[0121] For time series factors such as "new drug research and development progress index" and "market concentration", ADF unit root test is performed. If the test result shows that the sequence is a non-stationary sequence (p value is greater than 0.05), the sequence is first-order differentiated or detrended until it passes the stationarity test to eliminate the interference of non-stationary sequences as input to the prediction model.

[0122] Specifically, the drug sales management system based on artificial intelligence constructed in the embodiment of the application adopts a distributed architecture design, and the specific configuration is as follows:

[0123] The server uses 2-4 Intel Xeon Gold 6248 (16-24 core) processors, 64-256 GB of memory as the main computing node, and is equipped with high-performance SSD hard disks to ensure data read-write efficiency; the database cluster uses a 3-5 node HBase cluster architecture, each node has a storage capacity of 1-5 TB, and uses RAID5 configuration to ensure data security and reliability; equipped with NVIDIA Tesla V100 or A100 GPU acceleration cards for AI model training and inference calculation, with GPU memory of 16-80 GB, supporting parallel computing to improve model training efficiency.

[0124] The operating system adopts CentOS 7.6 or above, providing a stable Linux running environment; the database system selects HBase 2.4.9 or above, supporting distributed storage and real-time query of large-scale data; the AI computing framework adopts TensorFlow 2.6-2.10 and XGBoost 1.5-1.7, respectively used for building deep learning models and gradient boosting models; the stream processing engine adopts Apache Flink 1.12-1.16, supporting real-time data stream processing and low-latency computing; the blockchain platform adopts Hyperledger Fabric 2.2-2.5, building a consortium chain environment to ensure data tamper resistance.

[0125] A four-layer software system is built based on the C / S architecture, including a basic support layer (database, middleware, network communication), a platform service layer (data service, algorithm service, blockchain service), a core business layer (prediction unit, monitoring unit, traceability unit, classification unit), and an application layer (user interface, API interface, report system). The layers interact with each other through standardized interfaces to ensure the modularity and scalability of the system.

[0126] First, deploy each software module component on the server side, configure API interface permissions and data encryption certificates, and establish a secure data transmission channel; then import historical sales data for the past 2-5 years (about 2000-8000 million records), with a data volume of about 300-800 GB, and use batch import to complete data migration; complete the pre-training process of the prediction model, with a training time of about 24-72 hours, and generate initial model parameters; configure the abnormal transaction rule library, with pre-set 10-30 types of judgment rules and threshold parameters for violation patterns, including:

[0127] Large transaction abnormal pattern: transactions with a single transaction amount exceeding 2-5 times the historical average value, with a threshold set at the historical average value + 2σ to 3σ;

[0128] Frequent transaction abnormal pattern: the same customer transacts more than 20-50 times within 24 hours, or the cumulative transaction amount within 7 days exceeds 1.5-3 times the monthly average;

[0129] Price abnormal pattern: the sales price of a drug deviates from the market average by more than ±20%-30%, or the price difference of the same drug in different channels exceeds 10%-20%;

[0130] Inventory abnormal pattern: inventory data does not match sales data, with a difference rate exceeding 3%-8% of abnormal cases;

[0131] Time abnormal pattern: large transactions outside business hours (9 pm to 7 am), with a single transaction amount exceeding 3000-8000 yuan;

[0132] Geographical anomaly pattern: cross-regional transactions where the customer's purchase location is more than 300-800 kilometers away from the registered address;

[0133] Drug combination anomaly pattern: purchasing a large number of drugs with the same efficacy in a short period of time, with the quantity exceeding 5-15 times the normal drug usage;

[0134] Payment anomaly pattern: splitting a single large transaction using multiple payment methods, or the proportion of cash transactions exceeding 60%-90% of total transactions;

[0135] Customer behavior anomaly pattern: new registered customers with first transaction amount exceeding 5000-15000 yuan, or dormant customers suddenly activating for large transactions;

[0136] Prescription drug violation pattern: selling prescription drugs without a prescription, or violating the rules of matching prescriptions with sold drugs;

[0137] Data integrity anomaly pattern: missing rate of key fields exceeding 5%-15%, or abnormal records with data format not conforming to standard specifications;

[0138] Supplier anomaly pattern: expired supplier qualifications, abnormal fluctuations in supply prices exceeding 15%-25%, or unauthorized supplier transactions;

[0139] Return anomaly pattern: return rate exceeding 10%-20% or return amount exceeding 8%-15% of sales;

[0140] System operation anomaly pattern: a large number of data modification operations in a short period of time, with single user modifying more than 500-2000 records in 1 hour;

[0141] Compliance anomaly pattern: sales behavior violating drug management regulations, such as selling expired drugs, counterfeit drugs, or unapproved imported drugs, etc. The threshold values of various anomalies can be dynamically adjusted according to actual business conditions.

[0142] Example 1

[0143] In the embodiments of the present application, a complete data collection and processing system is constructed based on the actual needs of drug sales management. Specifically:

[0144] First, the collection targets include drug sales transaction records, inventory change data, supplier information, customer purchase behavior, price fluctuation information, regulatory compliance records, and other multi-dimensional data; the collection range includes heterogeneous data sources such as medical institutions at all levels, drug retail enterprises, drug wholesale enterprises, e-commerce platforms; data interface technology and crawler technology are used to establish connection channels with each data source.

[0145] Real-time acquisition of drug sales transaction data through API interface; acquisition of inventory management system data through database connection; collection of warehouse environment data using sensor networks; acquisition of market price information through third-party data services; automated collection of network public data by equipment at regular intervals; manual entry of sales data from special channels to ensure comprehensive and accurate data collection.

[0146] Each data source is verified and checked. This collection method ensures the diversity and integrity of the data, reflecting both real-time drug sales and adapting to different business scenarios. Factors such as data timeliness, accuracy, and compliance are considered during collection; multi-source data collection allows the prediction model to more accurately analyze drug sales trends, especially when responding to market changes, allowing rapid access to key information. At the same time, this collection method reduces the problem of data silos, facilitating subsequent data integration work.

[0147] Based on the results of multi-source heterogeneous data collection, each data source will be assigned to the corresponding data cleaning task. The specific task will be allocated according to the data quality and characteristics of the data source to ensure the comprehensiveness of data processing. The collected drug sales data is cleaned and standardized through various technical means:

[0148] Sales transaction data: including transaction time, drug code, quantity, amount, purchaser information, sales channel, and other core fields;

[0149] Inventory data: including drug inventory, safety inventory threshold, restocking records, expired drug disposal records, etc.

[0150] Supplier data: including supplier qualifications, supply capacity, price system, cooperation history, and other information;

[0151] Customer data: including customer type, purchase preferences, credit status, geographic distribution, and other characteristics;

[0152] Regulatory data: including drug approval number, quality inspection report, adverse reaction report, recall record, etc.

[0153] According to the five types of data collected, the information of each business scenario is integrated into an electronic drug sales data archive. Each archive will include data from different sources, formats, and types, ensuring that data for each drug category can be stored, updated, and managed uniformly; this electronic archive records the detailed sales trajectory of each drug category. The included drug sales data set includes basic information about the drug, sales history, market performance, and other information, as well as special management information for special drugs (such as controlled drugs, high-risk drugs, etc.).

[0154] Through data cleaning processing, duplicate records, incorrect information, and missing data are removed to ensure data quality. The data deduplication process filters out duplicate transaction records by comparing unique identifiers such as drug codes, transaction times, and customer IDs, and merges them. Data correction is achieved by establishing a standard library of drug basic information to automatically identify and correct spelling errors in drug names, specifications, and manufacturers. Outlier detection identifies extreme data points such as price anomalies and sales anomalies through statistical analysis methods, and determines the processing method according to business rules. Missing value imputation uses time series interpolation, regression prediction, and other methods to supplement missing sales data.

[0155] Furthermore, the data collection program is started daily at a fixed time (e.g., 7:00-9:00 am) to automatically obtain the sales data of the previous day from various pharmacies, hospitals, and distributors through a pre-set API interface. The data collection volume is approximately 50-200 million records per day, and the response time for a single API call is controlled within 100-500 ms. At the same time, a web crawler is started to crawl the sales data from cooperating e-commerce platforms. The crawler uses a distributed architecture and supports concurrent processing.

[0156] Data standardization processing uses multi-modal AI technology based on the BERT pre-training model to intelligently identify and standardize information such as drug names. The system calls the BERT model to perform synonym mapping on drug names (e.g., acetaminophen tablets and paracetamol tablets); uses regular expressions to extract and format key fields such as specifications, quantities, and prices, improving field extraction accuracy; and supports automatic identification and processing of structured data (e.g., Excel tables) and unstructured data (e.g., PDF sales sheets).

[0157] The cleaned and converted standardized data is stored in an HBase distributed database, with a single node supporting 500-2000 data writes per second, a data delay controlled within 30-100 ms, and the entire data processing process completed within 10-30 minutes. A data quality monitoring mechanism is established to monitor data integrity, accuracy, and consistency indicators in real time, improving data cleaning accuracy.

[0158] Example 2

[0159] Based on the cleaned drug sales data set, an intelligent demand prediction model is constructed. Specifically:

[0160] In the feature engineering stage, key feature variables are extracted from the original data, including time features (seasonality, periodicity, trend), drug features (category, price, specification), market features (competition intensity, policy impact), and customer features (purchase frequency, preference analysis). Through correlation analysis, principal component analysis, and other methods, the most influential feature combination for sales prediction is selected.

[0161] Model construction adopts an ensemble learning approach, combining time series analysis, machine learning, and deep learning techniques. Specifically, ARIMA models capture trends and seasonal characteristics of time series; random forest models handle non-linear relationships and feature interactions; LSTM neural networks learn long-term dependencies; and Prophet models handle holidays and special events. The prediction results of multiple models are integrated through weighted fusion to improve prediction accuracy and stability.

[0162] The model training process uses a rolling time window approach, with historical sales data as the training set, gradually predicting and updating model parameters forward. Cross-validation and grid search are used to optimize model hyperparameters, ensuring the model's generalization ability. A model evaluation system is established, using root mean square error, mean absolute percentage error, and prediction accuracy to evaluate model performance.

[0163] Data set division and model selection: divide by time sequence (avoid future data leakage), such as using 2021-2022 data as the training set (60-70%), 2023 January-June as the validation set (15-20%), and 2023 July-December as the test set (15-25%).

[0164] Select appropriate models based on data characteristics:

[0165] Time series models: suitable for single-variable or strong time-dependent data, such as ARIMA (handles linear trends), Prophet (automatically identifies trends and seasonal effects, suitable for business personnel to quickly use);

[0166] Machine learning models: suitable for scenarios with multiple factors, such as random forests (handle non-linear relationships and overfitting), XGBoost (high precision, suitable for high-dimensional data);

[0167] Deep learning models: suitable for large amounts of data (such as tens of millions of records), such as LSTM (captures long and short-term time dependencies, suitable for online pharmacies with high-frequency data).

[0168] Determine the objective function, with the goal of minimizing the error between predicted and actual sales. Common loss functions include mean squared error (MSE) and mean absolute error (MAE);

[0169] Adjust parameters (such as the number of trees for random forests and the hidden layer dimension for LSTM) using grid search or Bayesian optimization, and evaluate the optimal parameters using the validation set:

[0170] Quantitative indicators: RMSE (root mean square error, reflecting error magnitude), MAE (mean absolute error, reflecting average deviation), R² (determination coefficient, closer to 1 indicates better fitting);

[0171] Qualitative analysis: Observe the consistency of the predicted curve with the actual curve trend (e.g. whether it accurately captures the sales peak before the Spring Festival).

[0172] Future period drug sales demand prediction: Based on the trained model, input future influencing factor data, output prediction results, and business adaptation.

[0173] Prescribed period: Set according to business needs (e.g. future 7 days, 30 days, 90 days), short-term for replenishment, long-term for production planning.

[0174] Input the influencing factor data of the period for which prediction is required:

[0175] Known factors: such as future holiday arrangements (e.g. Spring Festival in 2024), planned promotional activities (e.g. "Double 11" discount plan), fixed policies (e.g. no adjustment to the medical insurance directory in 2024);

[0176] Estimated factors: such as estimating future influenza incidence rate based on historical incidence rate trend (extrapolated using time series model), assuming competitor prices remain at current level (if no available plan).

[0177] Prediction logic: The model outputs sales based on input features (e.g. future 7-day week features, promotional plans, estimated incidence rate) and historical patterns:

[0178] Example: Prophet model decomposes trends (long-term growth / downturn), seasonal effects (monthly / quarterly cycles), and holiday effects, and outputs predicted values after superposition;

[0179] Example: LSTM model outputs sequence prediction by remembering past sales time patterns (e.g. higher sales on Fridays than on Thursdays) and combining future features.

[0180] Output form: Output by time period granularity (e.g. daily / weekly sales), with confidence intervals (e.g. 80-95% confidence interval, reflecting prediction uncertainty).

[0181] If the predicted sales are negative (unreasonable), adjust to 0; if a hospital is known to have a new procurement plan, appropriately adjust the predicted value for the corresponding area;

[0182] Aggregate prediction results by region / drug category (e.g. "total demand for cold medicine in all pharmacies in Beijing in the next 30 days").

[0183] Regularly verify prediction error with actual sales data, if error exceeds threshold (e.g. RMSE exceeds 20%), retrain model (add new data, supplement unconsidered influencing factors, such as sudden epidemic);

[0184] Dynamic update features (such as adding "online consultation volume" as a predictive feature of related drugs).

[0185] Specifically, it includes the feature engineering stage to construct 64-256 dimensional feature vectors, including time features (quarter, holiday, day of the week, etc.), policy features (medical insurance directory adjustment, drug approval policy changes, etc.), correlation features (sales correlation between cold medicine and thermometer, etc.), and economic indicator features (residents' income level, health awareness level, etc.). Feature selection is performed through Pearson correlation coefficient and mutual information, and the final combination of features that have the most impact on prediction is determined.

[0186] The LSTM neural network design adopts a 2-4 layer structure, with 64-128 hidden layer neurons, using the Adam optimizer, a learning rate of 0.1-0.3, a batch size of 16-64, and 50-80 training rounds. The network can capture the long-term dependencies and trend changes of drug sales data, and the accuracy of identifying long-term trends such as annual sales cycles has been improved.

[0187] XGBoost model parameter configuration, this model mainly handles nonlinear relationships and sudden factor influences, such as the impact of epidemics, policy adjustments, and competitor launches on sales. The impact of sudden events is effectively predicted.

[0188] ARIMA time series model uses an automatic parameter selection algorithm to determine the optimal (p, d, q) parameter combination through the AIC criterion. It is mainly used to capture the seasonal and periodic characteristics of sales data. Model integration uses a weighted fusion strategy, with LSTM weights of 0.3-0.5, XGBoost weights of 0.3-0.5, and ARIMA weights of 0.1-0.3. The final prediction accuracy is above 88%-95%.

[0189] Model training uses incremental learning techniques, automatically fine-tuning the fusion model with the latest sales data every day. Fine-tuning time is controlled within 20-60 minutes. A model performance monitoring mechanism is established, and when the prediction accuracy drops by more than 3%-8%, the model retraining program is automatically triggered.

[0190] Example 3

[0191] Based on the output results of the prediction model, a three-layer architecture real-time transaction monitoring mechanism is constructed. The specific implementation process is as follows:

[0192] The first layer rule engine performs real-time matching through a preset rule library of illegal transactions. The rule library contains 15-25 types of illegal patterns: drug over-purchasing rules (single purchase amount exceeding 2-5 times the historical average), abnormal price fluctuation rules (price deviation from market average by 25-35% or more), irregular time transaction rules (large transactions at night or during holidays), frequent return rules (return frequency exceeding 3-5 times within 7 days), cross-regional abnormal transaction rules (transactions beyond the normal distribution range), etc. Each rule has a clear trigger condition and weight coefficient, and the system monitors transaction records in real time. Once a violation pattern is matched, it is immediately marked.

[0193] The second layer of the abnormality scoring system quantitatively evaluates the marked abnormal behavior. The scoring algorithm considers factors such as rule weight, prediction deviation, and historical risk records to calculate a comprehensive risk score. The risk score ranges from 0 to 100, with 0-25 being low risk, 26-55 being medium risk, and 56-100 being high risk. Different levels of early warning are triggered based on risk levels: low-risk warning displays a yellow prompt on the monitoring panel; medium-risk warning sends an email notification to relevant personnel and generates a work order in the system; high-risk warning immediately notifies the manager by phone and automatically freezes related transactions.

[0194] The third layer of the dynamic optimization module continuously optimizes system parameters based on early warning feedback and detection accuracy. By analyzing the accuracy of historical warnings, it identifies false positives and false negatives, and dynamically adjusts rule weights and risk thresholds. A feedback learning mechanism is established, with manually confirmed warning results serving as training samples to continuously optimize the performance of the anomaly detection algorithm.

[0195] Specifically, the system uses Apache Flink real-time computing engine as a streaming processing platform, supporting real-time processing capacity of 5000-15000 transactions per second, with processing delay controlled within 50-200ms. The rule engine uses Drools rule engine, supporting dynamic rule configuration and hot updates.

[0196] The anomaly detection algorithm uses the Isolation Forest algorithm for unsupervised anomaly detection, with the following parameters: number of trees 50-200, sub-sample size 128-512, and abnormality ratio 3-8%. The algorithm can identify single-dimensional and multi-dimensional combined anomalies. The risk scoring algorithm considers factors such as rule weight (weight range 0.05-1.0), prediction deviation (deviation threshold 10-35%), and historical risk records (risk decay coefficient 0.8-0.95) to calculate a comprehensive risk score.

[0197] The early warning grading mechanism is designed as follows: 0-25 points for low risk (green), with prompt information displayed on the monitoring panel; 26-55 points for medium risk (yellow), sending email notifications to relevant personnel and generating work orders in the system; and 56-100 points for high risk (red), immediately notifying management personnel by phone and automatically freezing related transactions to prevent loss expansion.

[0198] The reinforcement learning optimization uses the Q-learning algorithm to adjust the early warning threshold parameters based on artificial audit feedback. The learning rate is set to 0.05-0.2, the discount factor is 0.85-0.95, and the exploration rate gradually decays to 0.05-0.15. Through continuous learning, the system false positive rate has decreased, and the early warning accuracy has significantly improved.

[0199] Example 4

[0200] To ensure the non-tamperability and traceability of drug sales data, a traceability system based on blockchain technology is constructed. The specific implementation includes:

[0201] The distributed ledger architecture adopts the form of consortium chain, with key nodes such as drug regulatory departments, production enterprises, circulation enterprises, and medical institutions composing the blockchain network. Each node maintains a complete copy of the ledger and ensures data consistency through a consensus mechanism. The block structure includes a block header (timestamp, previous block hash, Merkle root) and a block body (transaction data, digital signature, smart contract execution result).

[0202] The smart contract mechanism sets automatic traceability queries and abnormal handling rules. When the system detects abnormal transactions, the smart contract automatically triggers the traceability program to query the complete chain information of the drug from production to sales. The contract has a pre-set responsibility locating algorithm that automatically identifies the responsible party based on the type and link of the anomaly and generates a responsibility report.

[0203] The data on-chain process adopts a batch submission method, with verified transaction data being packaged and chained every hour to reduce network burden while ensuring data timeliness. Sensitive information is encrypted before being chained to protect commercial privacy. A multi-verification mechanism is established, requiring at least three nodes to confirm each transaction data before writing it to the blockchain.

[0204] The visual query interface provides multi-dimensional traceability query functions. Users can quickly locate the full life cycle records of target drugs through drug batch numbers, production dates, and sales channels. The interface displays key information such as drug production information, quality inspection reports, circulation tracks, sales records, prediction matching degrees, and abnormal detection results, forming a complete traceability chain diagram.

[0205] Specifically, Hyperledger Fabric 2.2-2.5 is used as the underlying blockchain platform to build a consortium chain network composed of key nodes such as drug regulatory departments, production enterprises, distribution enterprises, and medical institutions. Each node maintains a complete copy of the ledger, and the PBFT (Byzantine Fault Tolerance) consensus mechanism is used to ensure data consistency, with a consensus time of 2-5 seconds.

[0206] The block structure design includes block header and block body: the block header contains metadata such as timestamp, previous block hash (SHA-256 algorithm), Merkle root hash, and block height; the block body contains drug transaction data, digital signature, smart contract execution result, prediction matching degree, anomaly detection result, and regulatory compliance status. Each block size is controlled within 0.5-2MB to ensure network transmission efficiency.

[0207] The smart contract mechanism is developed in Go language and sets up automated traceability query and anomaly handling rules. The contract has pre-set responsibility locating algorithms, including: anomaly type identification algorithm (based on rule matching and machine learning), responsibility link locating algorithm (tracking through timestamp and data flow), and responsibility subject confirmation algorithm (combined with digital signature and permission management). When the system detects an abnormal transaction or a prediction deviation exceeding the 25%-35% threshold range, the smart contract automatically triggers the traceability program to query the complete chain information of the drug from production to sales, and generates a responsibility report within 8-15 seconds.

[0208] The data on-chain optimization adopts a batch submission strategy, with verified transaction data being packed and chained every 30 minutes-2 hours, processing 500-8000 transaction records per batch, effectively reducing network burden. Commercial sensitive information is encrypted using AES-256 encryption algorithm before being chained, protecting commercial privacy. A multi-verification mechanism is established, requiring at least 2-5 node confirmations for each transaction data to be written to the blockchain, ensuring data reliability.

[0209] Query performance optimization is achieved through index mechanism and cache strategy, supporting fast query by prediction accuracy, anomaly type, time range, etc., with a query response time of 3-5 seconds. The query interface supports visual display, including time axis traceability graph, flowchart, and association graph.

[0210] Example 5

[0211] Based on the prediction results and multi-dimensional analysis, an intelligent drug classification management system is constructed. The specific implementation scheme is as follows:

[0212] Multi-dimensional classification first divides drugs into preliminary zones according to dimensions such as geographical location, sales channel, drug category, and prediction accuracy. The geographical dimension is divided according to provincial, municipal, and county administrative divisions; the channel dimension distinguishes between hospital, pharmacy, e-commerce, and clinic sales channels; the category dimension is classified according to therapeutic areas and pharmacological effects; and the accuracy dimension is classified according to historical prediction bias rates.

[0213] Trend analysis identifies the sales trend characteristics of each drug category through time series decomposition techniques. Rising trend identification: consecutive 3-6 month sales growth rates are positive and accelerating; falling trend identification: consecutive 2-4 month sales declines and gradually expanding declines; stable trend identification: sales fluctuations within ±6% and no obvious trend; fluctuating trend identification: sales changes exceeding 15%-20% and no regularity.

[0214] Different inventory strategies are developed for different trend characteristics. Rising trend drugs use an aggressive expansion strategy: increase safety stock by 20%-40%, change from monthly replenishment to weekly replenishment, and prioritize gold shelf positions; falling trend drugs use a conservative contraction strategy: extend the replenishment cycle and strengthen promotional clearance; stable trend drugs use a balanced maintenance strategy: maintain existing inventory parameters and regularly assess and adjust; fluctuating trend drugs use a flexible adaptation strategy: set dynamic safety stock and establish a rapid replenishment mechanism.

[0215] Efficacy correlation analysis establishes inter-category correlations through a drug knowledge graph. The knowledge graph contains information such as drug indications, contraindications, drug interactions, and combination therapy regimens. Graph neural network algorithms are used to mine implicit correlation patterns and identify drug combinations with synergistic therapeutic effects. A correlation coefficient matrix is established to quantify the correlation strength between different drugs.

[0216] Cross-zone linkage mechanism establishes an automated early warning and coordination system. When the sales of a core drug in a certain zone fluctuate abnormally, the system automatically analyzes the sales of its efficacy-related categories in other zones to predict potential chain reactions. Triggered warning mechanisms include inventory warnings, replenishment recommendations, and promotion strategy adjustments. Coordination mechanisms support cross-zone inventory allocation and resource reallocation.

[0217] Fine-tuned inventory allocation and control strategies also include: on the basis of the ABC-XYZ classification method, further integrating trend zone characteristics to achieve multi-dimensional precision inventory control. The classification algorithm combines fuzzy theory with traditional classification methods, using entropy weight method to determine classification weights, and subdividing drugs into 6-12 control categories, each with differentiated inventory parameters.

[0218] For the upward trend partition of drug categories, an aggressive expansion type inventory allocation control strategy is adopted: the safety stock coefficient is set to 1.5-2.0 times the standard demand, the replenishment frequency is increased to 2-3 times a week, a priority shelf location configuration mechanism is established, and a 20%-30% demand growth buffer stock is set. When the forecast shows a sustained upward trend, an automatic expansion replenishment mechanism is started, the replenishment trigger point is set to 70%-90% of the safety stock, and the replenishment amount is 1.8-2.0 times the economic order quantity. At the same time, a synergistic inventory increase scheme for efficacy-related categories is established, and when the inventory of core drugs increases, the inventory of efficacy-related categories is increased by 0.4-0.8 times simultaneously.

[0219] For the downward trend partition of drug categories, a conservative contraction type inventory allocation control strategy is implemented: the safety stock coefficient is adjusted to 0.5-0.8 times the standard demand, the replenishment cycle is extended to 2-3 weeks, and an intelligent clearance promotion trigger mechanism is established. When the inventory turnover days exceed 60 days, the price optimization algorithm is automatically started, the promotion discount range is 5%-15%, and the substitute product recommendation algorithm is started at the same time, recommending similar efficacy upward trend drugs. A stock gradient digestion scheme is constructed, and sales priority is set according to the proximity of shelf life, with priority given to the digestion of near-expiry drugs to maximize the preservation of inventory value.

[0220] For the stable trend partition of drug categories, a balanced maintenance type inventory allocation control strategy is implemented: maintain 1.0-1.2 times the standard demand safety stock, and maintain a fixed replenishment cycle every 10-14 days. A precise demand prediction-inventory matching mechanism is established, using moving average method to smooth short-term fluctuations, ensuring high matching of inventory level and actual demand. At the same time, an inventory efficiency monitoring system is established, maintaining an inventory turnover rate at the industry optimal level, with a target turnover rate of 8-12 times / year.

[0221] For the fluctuating trend partition of drug categories, a flexible adaptive inventory allocation control strategy is constructed: set the dynamic safety stock range to 0.8-1.8 times the standard demand, establish a two-level elastic replenishment mechanism, including basic replenishment (according to fixed cycle) and emergency replenishment (according to demand fluctuation trigger), when the demand fluctuation exceeds the preset threshold ±25%, the emergency replenishment program is automatically triggered, and the emergency replenishment amount is dynamically calculated according to the fluctuation amplitude. A multi-scenario inventory buffer pool is constructed, and corresponding inventory adjustment schemes are preset for different fluctuation modes such as seasonal fluctuations, sudden event impacts, and policy influences.

[0222] When an abnormal inventory pressure occurs in a certain trend partition, the intelligent algorithm identifies the inventory redundancy space in other partitions to calculate the optimal inventory allocation scheme. The inventory allocation cost model considers transportation cost, time cost, opportunity cost and other factors to ensure the economy of the allocation decision. At the same time, an inventory substitution scheme based on efficacy correlation is established. When a drug is out of stock, several substitute drugs with similar efficacy are automatically recommended as preferred options.

[0223] An inventory allocation decision support system is constructed, integrating three core algorithm modules of trend prediction, cost optimization and risk assessment. The trend prediction module outputs trend prediction results based on the LSTM-XGBoost fusion model; the cost optimization module calculates the optimal inventory configuration using an integer programming algorithm; and the risk assessment module assesses inventory risk based on Monte Carlo simulation. The system generates personalized inventory parameter recommendations for each drug category, including key indicators such as optimal order point, economic order quantity, and maximum inventory limit.

[0224] Embodiment 6

[0225] The prediction effect is evaluated and optimized through multi-dimensional analysis methods, and a closed-loop feedback mechanism is established to continuously improve the prediction accuracy and business value of the system. The specific implementation process includes:

[0226] The bias analysis and deep diagnosis mechanism uses a multi-level bias analysis framework to obtain the prediction bias indicators of each sub-regional drug category.

[0227] The absolute deviation (MAE) reflects the direct gap between the predicted value and the actual value, which is wherein, is the th actual observation value, is the th predicted value, and is the total number of samples;

[0228] The relative deviation (MAPE) eliminates the dimensional influence and is convenient for comparison between different categories, which is wherein, is the th actual observation value, is the th predicted value, and is the total number of samples;

[0229] The root mean square error (RMSE) is more sensitive to large deviations, and the formula is wherein, is the th actual observation value, is the th predicted value, and is the total number of samples.

[0230] The bias distribution analysis uses kernel density estimation to identify the distribution pattern and outliers of the bias, and identifies outliers through the 3σ rule. A multi-level bias early warning mechanism is established: green safe zone (bias rate 1-10%), yellow attention zone (bias rate 10%-15%), orange warning zone (bias rate 15%-30%), and red danger zone (bias rate exceeding 30%). Corresponding treatment strategies are developed for different warning levels: yellow warning triggers parameter fine-tuning; orange warning starts model retraining; red warning conducts comprehensive model diagnosis and reconstruction.

[0231] Deep bias diagnosis includes time dimension bias analysis (identifying the time pattern of prediction accuracy), spatial dimension bias analysis (finding the impact of geographical factors on prediction accuracy), and category dimension bias analysis (analyzing the prediction difficulty differences of different drug types). A bias attribution algorithm is established to automatically identify the main factors causing prediction bias, including data quality problems, model adaptability problems, and external environmental changes.

[0232] The correlation analysis and association mining system uses multiple statistical methods to deeply mine the association between variables. Pearson correlation coefficient analyzes linear relationship, suitable for normally distributed data; Spearman correlation coefficient analyzes monotonic relationship, suitable for non-normally distributed data; Kendall tau correlation coefficient analyzes ordinal correlation, more robust to outliers. The correlation heat map visually displays the correlation strength between different sub-regions and drug categories. Correlation coefficients with absolute values greater than 0.7 are defined as strong correlation, 0.4-0.7 as moderate correlation, and less than 0.4 as weak correlation.

[0233] A dynamic correlation monitoring system is constructed to continuously monitor correlation changes using sliding window technology (window size 15-45 days) to timely discover the evolution trend of association. High correlation category combinations are identified to provide data support for joint promotion and inventory coordination. A correlation prediction model is established to predict the future trend of association, providing forward-looking guidance for strategy adjustment.

[0234] Multi-dimensional association analysis is implemented, including: geographical correlation analysis to identify sales linkage effects between different regions; time correlation analysis to find lag associations between time series; category correlation analysis to mine substitution and complementary relationships between drugs; channel correlation analysis to explore the synergistic effects of different sales channels. Compound association indicators are used to quantify multi-dimensional correlation strength, providing a basis for precision marketing and inventory optimization.

[0235] Causal relationship analysis and prediction mechanism uses various causal inference methods to identify the direction and strength of causal relationships between variables. Granger causality test identifies the causal relationship between time series, and rejects the null hypothesis when the test statistic F is greater than the critical value. Cointegration test (Johansen method) analyzes the long-term equilibrium relationship between variables, identifies common trends and adjustment speeds. Vector autoregression model (VAR) describes the dynamic relationship of multivariate system, and captures the mutual influence between variables.

[0236] Establish structured causal graph, use directed acyclic graph (DAG) to represent the causal relationship network between variables. Through causal discovery algorithm (such as PC algorithm, FCI algorithm), automatically learn the causal structure from observation data. Implement causal inference analysis, distinguish real causal relationship and false correlation, and provide reliable causal basis for decision-making.

[0237] Impulse response function analysis quantifies the dynamic impact of external shocks on the system: the duration and decay rate of policy change shocks (such as medical insurance policy adjustment); the lagged impact of seasonal factors (such as flu season) on related drug sales; the short-term impact and long-term impact of sudden events (such as epidemic outbreak). Establish shock transmission path diagram, identify key nodes and amplification mechanisms of impact propagation.

[0238] Model iteration optimization and adaptive learning mechanism based on analysis results to establish a continuous model optimization framework. Parameter optimization uses various advanced algorithms: grid search (GridSearch) for exhaustive parameter tuning; random search (RandomSearch) to improve search efficiency; Bayesian optimization (BayesianOptimization) based on prior knowledge to guide search direction; genetic algorithm (GeneticAlgorithm) simulates evolutionary process to find optimal solution; particle swarm optimization (PSO) searches global optimum through swarm intelligence.

[0239] Feature optimization uses multi-level feature selection strategy: filter method (such as chi-square test, mutual information) quickly selects relevant features; wrapper method (such as recursive feature elimination) evaluates the prediction performance of feature subset; embedded method (such as LASSO regression, Random Forest feature importance) selects features during model training. Establish dynamic evaluation mechanism of feature importance, update feature weight regularly, eliminate redundant features, and introduce new effective features.

[0240] Structural optimization adjusts model architecture according to prediction effect and business demand: neural network depth optimization (determine optimal number of layers by validation set error); integrated learning strategy optimization (adjust base learner number and weight); hybrid model architecture design (combine advantages of different types of models). Establish a balance mechanism between model complexity and performance, optimize model efficiency while ensuring prediction accuracy.

[0241] Adaptive learning mechanisms enable continuous evolution of models: online learning algorithms support real-time updates of models handling concept drift in data streams; incremental learning methods incorporate new data without retraining the entire model; transfer learning techniques migrate knowledge from trained models to new scenarios; multi-task learning optimizes multiple related prediction tasks simultaneously.

[0242] Establish A / B testing mechanisms and model evaluation systems, design strict controlled experiments to compare the effects of different optimization schemes. The A / B testing framework includes: experimental design (determine test indicators, sample allocation, test period); statistical analysis (hypothesis testing, confidence interval, effect size calculation); result interpretation (statistical significance, practical business value, potential risks). Establish a multi-dimensional evaluation index system: prediction accuracy indicators (RMSE, MAPE, SMAPE); business value indicators (inventory turnover rate improvement, sales growth rate, cost savings); system performance indicators (response time, concurrent processing capacity, stability).

[0243] Build a model performance monitoring and early warning system to monitor key indicators such as prediction accuracy, feature distribution, and prediction bias in real time. When model performance declines (accuracy drops by more than 3%-8% or bias exceeds threshold for 5-10 consecutive days), automatically trigger model retraining processes. Establish a model version management mechanism to support model rollback and A / B comparison, ensuring the safety and controllability of model updates.

[0244] Intelligent optimization decision support system integrates all the above analysis and optimization functions, providing visual optimization suggestions and decision support for management personnel. System outputs include: model performance diagnosis report, optimization strategy suggestion, risk assessment result, implementation path planning. Establish an optimization effect evaluation mechanism to track the actual effect of optimization measures and form a closed-loop continuous improvement system.

[0245] Through the above specific embodiments, the present application realizes intelligent management of drug sales data, and establishes a complete system from data collection, prediction analysis, real-time monitoring to traceability management. The entire system has the ability of self-learning, self-optimization and self-adaptation, which can continuously improve the prediction accuracy and business value, providing a scientific, efficient and intelligent technical solution for drug sales management. The system prediction accuracy, inventory turnover rate and anomaly detection accuracy are improved, providing important technical support for the digital transformation and intelligent upgrading of the pharmaceutical industry.

[0246] A computer readable storage medium for storing computer readable instructions that, when read by a computer, can execute an artificial intelligence-based drug sales management method:

[0247] Computer-readable media includes permanent and non-permanent, movable and non-movable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible by a computing device. According to the definition herein, computer-readable media does not include transitory media.

[0248] The above describes embodiments of the present application, but the embodiments are not limited to the specific implementation described above, which is merely illustrative and not restrictive. Those skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which are all within the protection of the embodiments.

Claims

1. A method for managing sales of pharmaceutical products based on artificial intelligence, characterized by, The method comprises the following steps: Collecting a plurality of heterogeneous source drug sales data, and performing cleaning, standardization processing and storage on the plurality of drug sales data to obtain a drug sales data set; According to the drug sales data set and the drug sales influencing factors, a prediction model of drug sales demand is trained and constructed; The prediction model predicts the drug sales demand in a future preset period to obtain a prediction result; According to the prediction result, a real-time transaction monitoring mechanism is established to detect and warn of abnormal drug transaction behavior; Based on the prediction result and the real-time transaction monitoring mechanism, a blockchain traceability system is constructed to establish an unalterable drug sales record chain; According to the prediction result and the blockchain traceability system, an intelligent drug classification system is constructed to obtain a fine-grained inventory allocation and control strategy; Wherein, the construction of the intelligent drug classification system comprises: Based on the prediction result and a plurality of dimension influencing factors, the drug categories are intelligently partitioned and placed to form a plurality of sub-regional drug categories with different demand characteristics; The sub-regional drug categories are dynamically divided according to drug sales trend data to obtain a plurality of drug categories with trend characteristics; wherein, by analyzing the historical sales data time series of the sub-regional drug categories, and by trend analysis, the drug categories with rising trend, falling trend, stable trend and fluctuating trend in the current period are identified, and the drug categories with different trend characteristics are classified into different trend partitions; Based on the plurality of drug categories with trend characteristics, the sub-regional drug categories are analyzed according to the drug efficacy to establish an efficacy associated category cooperation mechanism; Wherein, the efficacy associated category cooperation mechanism identifies drug category combinations containing synergistic therapeutic effect, complementary functional effect and joint drug use advantage by constructing a drug efficacy association knowledge graph to form an efficacy associated category matrix; Based on the efficacy associated category matrix, a cross-partition efficacy associated category linkage mechanism is established to trigger the sales warning of other partition associated efficacy categories when the sales data of the core efficacy drug in a partition abnormally fluctuates, so as to obtain the coordinated operation of the sales supply chain of the main drug and the auxiliary drug.

2. The method of claim 1, wherein the method is based on artificial intelligence. Establishing a real-time transaction monitoring mechanism comprises: Setting a plurality of risk level evaluation threshold ranges based on the prediction result; Risk scoring is performed by comparing the deviation degree of actual sales transaction data and prediction data to obtain a scoring result; According to the scoring result and the plurality of risk level evaluation threshold ranges, the current risk level is obtained; Further comprising constructing a multi-level abnormality detection architecture according to the plurality of risk level evaluation threshold ranges to perform multi-level abnormality violation detection on the prediction result to obtain a detection result; Wherein, the first layer matches a plurality of preset abnormal sales transaction rules through a rule engine; abnormality violation is immediately screened out through the plurality of abnormal sales transaction rules to obtain abnormal sales behavior; The second layer quantitatively scores the abnormality of the abnormal sales behavior, and if there are at least two abnormal sales behaviors, the risk score is sequentially evaluated according to the plurality of risk level evaluation threshold ranges to obtain an evaluation result; the corresponding level of the warning mechanism is triggered according to the evaluation result to obtain a hierarchical warning; The third layer is based on hierarchical early warning feedback, dynamically adjusts the early warning threshold range, and accurately adjusts the risk level boundary of multi-level abnormal violation detection. 3.The method of claim 1, wherein the method further comprises: A blockchain traceability system is constructed, including: ​ Based on the prediction results and real-time transaction monitoring mechanism, the drug sales data is linked with the production batch information, logistics track and quality inspection report in an unalterable chain; A distributed ledger architecture is established based on the chain connection, and the whole life cycle information of each drug transaction is stored in the form of blocks; An intelligent contract mechanism is also constructed, which triggers traceability query and responsibility positioning when detecting abnormal sales behavior; Through traceability query and responsibility positioning, the sales records of the whole life cycle of the drug are quickly traced and the responsibility of abnormal sales behavior is located.

4. The method of claim 1, wherein the method is based on artificial intelligence. According to the sub-regional drug category data combined with the prediction results, including: Collecting the basic data of drug categories in each sub-region to obtain key sales indicators; comparing and matching the collected actual sales key indicators with the prediction results to identify sub-regions with prediction deviation anomalies and drug categories in the current sub-region; For the drug categories with prediction deviation anomalies, the sales environment, market competition, and consumer behavior changes in the current sub-region are investigated in depth to identify the sales influencing factors; Integrating the sales influencing factors in all research information to form structured drug sales research data; According to the drug sales research data, the sales correlation mode, abnormal sales behavior transmission path and risk diffusion law among sub-regions are mined to identify emerging factors and potential risks that affect drug sales; The emerging influencing factors that affect drug sales form an interrelated influence network; The influence network identifies, quantifies and predicts the complex correlation between multiple factors to analyze and forewarn changes in the drug sales environment.

5. The method of claim 1, wherein the method is based on artificial intelligence. The actual sales data of the sub-regional drug categories and the prediction results are subjected to deviation analysis, correlation analysis and causal relationship analysis to obtain analysis results, including: Obtain the absolute deviation, relative deviation and deviation distribution characteristics of the actual sales data and prediction results of each sub-region to identify deviation abnormal drug categories and time nodes, and obtain deviation analysis results; Based on the deviation analysis results, obtain the correlation coefficients of sales data between different sub-regions and different drug categories to obtain the correlation analysis results of sales data between drug categories; Based on the correlation analysis results, the root causes of the prediction deviation are identified through causal inference algorithms, including external environmental factors, model parameter settings, and causal relationship influencing elements of data quality problems, to obtain causal relationship analysis results; Integrate the three analysis results to form comprehensive analysis results including deviation characteristic description, correlation relationship graph, and causal influence chain; According to the comprehensive analysis results combined with the drug sales research data, the emerging influencing factors affecting drug sales are analyzed in depth and the interference factors of the emerging influencing factors are screened out to obtain accurate emerging influencing factors.

6. The method of claim 5, wherein the AI-based medicine sale management method is characterized by, The emerging influencing factors affecting drug sales are analyzed in depth and the interference factors of the emerging influencing factors are screened out, including: A multi-dimensional factor analysis framework is established to quantitatively evaluate the emerging influencing factors through three dimensions of influence intensity, action time delay and duration period, and form an influence factor matrix; An interference factor identification system is constructed based on the influence factor matrix to identify and filter out interference factors from the emerging influencing factors; Among them, the abnormal value detection is performed on the emerging influencing factors to obtain an abnormal value detection result; The abnormal value detection result is obtained by identifying the abnormal fluctuation points in the policy environment factor data; When the medical insurance directory adjustment frequency data exceeds the preset abnormal value of the historical mean standard deviation, it is marked as a potential interference factor, and is audited and confirmed. The abnormal data confirmed is processed by the median replacement method, and the processing result is fed back to the influence factor matrix for weight adjustment; Based on the abnormal value detection result, redundant variables in the social and economic factors are analyzed and filtered out; When the correlation coefficient between the change of the income level of residents and the degree of health consciousness in the social and economic factors exceeds the preset threshold value, it is identified as a high-correlation redundant variable. Combined with the influence intensity in the influence factor matrix, the main interference factor of the redundant variable with strong influence intensity is retained, and the secondary interference factor of the redundant variable is removed. The screening result is updated to the influence factor matrix.

7. The method of claim 6, wherein the AI-based medicine sale management method is characterized by, Further comprising: By identifying low-variance interference items in the emerging influencing factors, when the variance of new drug research and development progress data in a preset time range is less than a preset value of the overall variance, a low-information-content interference factor is obtained; The low-information-content interference factor is removed from the model input variables to obtain a removal result; The removal result is transmitted to a dynamic monitoring mechanism for recording; Based on the removal result, non-stationary interference sequences in the competition pattern factors are filtered out through time series stationarity test. The ADF test is performed on the market concentration change data. When the P value is greater than a preset value, it is determined as a non-stationary sequence interference factor; The non-stationary sequence interference factor is eliminated by difference transformation or detrending processing to obtain a processing result; The processing result is fed back to the interference factor processing verification to obtain an accurate interference factor processing result.

8. The method of claim 7, wherein the AI-based medicine sale management method is characterized by, Further comprising: A dynamic monitoring mechanism for interference factors is established. Based on the identification results of various interference factors, the interference factor detection threshold is set as a double standard that the influence weight is lower than a preset value and the prediction contribution is less than a preset value; Among them, when a certain interference factor meets the interference factor standard for a continuous preset time, it is automatically removed from the influence factor matrix. The removal decision triggers the weight redistribution of the influence factor matrix; A interference factor processing verification is constructed to backtest the influence factor set after the interference factors are filtered out; Among them, the emerging influencing factors before and after the interference factors are filtered out are respectively input into the prediction model to obtain two output results; According to the root mean square error improvement degree of the two output results, when the error improvement degree exceeds a preset improvement value, it is confirmed that the interference factor filtering is effective. The emerging influencing factor set after the interference factors are filtered out is input into the multi-dimensional factor analysis framework for the next round of evaluation. Otherwise, it is fed back to the dynamic monitoring mechanism for reevaluation of the filtering standard.

9. A computer-readable storage medium, characterized in that, Computer readable instructions for storing, when read by a computer, are capable of running an artificial intelligence-based pharmaceutical sales management method as claimed in any one of claims 1-8.

Images