Artificial intelligence-based antibacterial substance management information system

Abstract

This invention discloses an artificial intelligence-based antimicrobial substance management information system, relating to the field of medical information. It obtains multiple historical antimicrobial grade sequences sorted by time from historical antimicrobial treatment data and generates a gradient sequence based on the difference between adjacent time steps. The system calculates the length of continuous gradients using a sliding window, reflecting the trend of continuous escalation or deterioration of the antimicrobial grade. In the self-attention layer of the Transformer network, time steps with larger continuous gradient lengths are assigned higher attention weights, thereby highlighting key change segments in the antimicrobial grade sequence. The Transformer network is trained based on the historical antimicrobial grade sequences to obtain an antimicrobial grade prediction model. This invention can utilize the characteristics of antimicrobial grade changes to achieve intelligent prediction and optimized management of future antimicrobial drug grades.

Description

Technical Field

[0001] This invention relates to the field of medical information, and more specifically, to an artificial intelligence-based information system for managing antimicrobial substances. Background Technology

[0002] Antibiotics are a core component of clinical treatment for bacterial infections, and their safety, timing, and classification directly impact treatment efficacy. Compared to other drug types, antibiotics exhibit a distinct hierarchical structure, typically including unrestricted-use, restricted-use, and special-use antibiotics. Antibiotic therapy often involves dynamic adjustments between different antibiotic classes; when a lower-class antibiotic fails to achieve sufficient efficacy, a higher-class antibiotic is escalated, and after the condition stabilizes, the antibiotic is downgraded to a lower-class antibiotic to reduce the risk of resistance and side effects. This sequential pattern of hierarchical shifts is a key characteristic that distinguishes antibiotic therapy from other drug treatments.

[0003] In practical applications, the historical records of antimicrobial drug use are scattered across different patients and time periods, resulting in significant differences and uncertainties between antimicrobial grading sequences. This makes it difficult for human experience to accurately grasp future trends in antimicrobial grading and to predict future demand for antimicrobial substances. Furthermore, there is currently no unified subdivision standard within antimicrobial grading; different institutions use significantly different classification methods, making it difficult for internal classification schemes to remain stable and effective across different scenarios. This makes it challenging to build a unified antimicrobial substance management system.

[0004] With the accumulation of a large amount of historical antimicrobial treatment data, how to make full use of the grade jump characteristics existing in the antimicrobial grade sequence, extract hidden patterns from complex sequences, and use them to predict future antimicrobial grades or antimicrobial substance requirements is an important problem that needs to be solved. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide an antimicrobial substance management information system based on artificial intelligence, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] An artificial intelligence-based information system for managing antimicrobial substances, comprising:

[0008] The historical antimicrobial rank sequence construction module is used to obtain multiple historical antimicrobial rank sequences sorted by time from historical antimicrobial treatment data;

[0009] A gradient sequence generation module is used to generate a gradient sequence based on the difference in antibacterial level between adjacent time steps, wherein the gradient value of each time step in the gradient sequence is the difference between the antibacterial level of the current time step and the antibacterial level of the previous time step.

[0010] The continuous gradient length calculation module is used to count the number of consecutive gradient values ​​with the same sign that appear before each time step in the gradient sequence based on a sliding window, and to take the corresponding number as the continuous gradient length of the corresponding time step, with the sign including positive and negative.

[0011] The antibacterial level prediction model construction module is used to make time steps with larger continuous gradient lengths obtain higher weights in the attention weight calculation in the self-attention layer of the Transformer network, and to train the Transformer network based on the historical antibacterial level sequence to obtain the antibacterial level prediction model.

[0012] Preferably, the system further includes:

[0013] The antimicrobial substance management prediction module is used to predict the antimicrobial drugs for one or more future time steps based on the antimicrobial level prediction model when a current antimicrobial level sequence is received; or to predict and formulate a procurement plan for the total demand of each antimicrobial drug for one or more future time steps based on the antimicrobial level prediction model when multiple current antimicrobial level sequences are received.

[0014] Preferably, the antimicrobial grades in the antimicrobial grade sequence include unrestricted-use antimicrobial drugs, restricted-use antimicrobial drugs, and special-use antimicrobial drugs.

[0015] Preferably, the antibacterial grade prediction model construction module is further used for:

[0016] Multiple internal classification schemes were constructed to further classify antimicrobial drugs within the categories of unrestricted use, restricted use, and special use.

[0017] Under each internal grading scheme, an antimicrobial grade prediction model was trained, and the corresponding prediction error was evaluated on a validation dataset.

[0018] The internal grading scheme with the lowest prediction error was selected as the internal hierarchical structure of the antibacterial grade sequence for training and inference of the antibacterial grade prediction model.

[0019] Preferably, the internal ranking scheme with the lowest prediction error is obtained based on reinforcement learning, including:

[0020] An internal hierarchical scheme is used as a state representation for the reinforcement learning environment;

[0021] In each reinforcement learning iteration, a new state representation is generated, and the corresponding antibacterial grade prediction model is trained based on the new state representation; the new state representation is obtained by changing the number or sorting method of the internal sub-grades in the internal grading scheme.

[0022] The negative value of the prediction error of the antibacterial level prediction model on the validation dataset is used as the reward value for reinforcement learning.

[0023] The reinforcement learning policy is updated based on the reward value to guide the policy to generate state features with lower prediction errors in subsequent iterations.

[0024] Preferably, the construction process of the internal hierarchical scheme includes:

[0025] Extract similar segments from historical antimicrobial grade sequences. The antimicrobial drugs in the similar segments are arranged continuously and all belong to the same major category of antimicrobial grade. The same major category of antimicrobial grade includes any one of the following: unrestricted use antimicrobial drugs, restricted use antimicrobial drugs, and special use antimicrobial drugs.

[0026] For each major category, all similar segments within the major category are clustered to obtain multiple clusters. From each cluster, one or more typical segments are referenced to extract an internal hierarchical scheme as a candidate scheme for the corresponding cluster. The number of segments contained in each cluster is counted, and the candidate schemes of the clusters with more segments are ranked higher. The top N candidate schemes are used as candidate internal hierarchical schemes for the corresponding major category.

[0027] Preferably, the antibacterial grade prediction model construction module adjusts the attention weights based on the continuous gradient length in the self-attention layer of the Transformer network in the following way: linearly mapping the embedding vector of each time step in the input sequence to obtain the corresponding query vector, key vector, and value vector, and calculating the value vector at any time step. t Time step s When assigning attention weights, the time step... s The continuous gradient length is mapped to a weight coefficient according to a preset monotonically increasing function. The weight coefficient is multiplied by the attention weight calculated from the query vector and the key vector so that the time step with a larger continuous gradient length gets a higher weight in the attention weight calculation. The monotonically increasing correspondence is based on any one of a linear increasing function, an exponential increasing function, or a piecewise increasing function.

[0028] Preferably, the training process of the antibacterial grade prediction model construction module includes:

[0029] The historical antimicrobial grade sequence, gradient sequence, and continuous gradient length are used as training inputs.

[0030] In each layer of the Transformer network's self-attention structure, the attention weights are amplified or reduced based on the continuous gradient length.

[0031] The network parameters are adjusted based on the error between the target output and the Transformer network output, using training labels that correspond to the actual antibacterial level at the next time step.

[0032] The Transformer network is trained iteratively through multiple rounds until its prediction performance on the validation set meets the preset convergence condition.

[0033] Preferably, one of the historical antimicrobial grade sequences corresponds to the antimicrobial medication records of a patient ID over a continuous period of time.

[0034] Preferably, the patient ID is anonymized in the antimicrobial substance management information system.

[0035] The advantage of this invention over existing technologies lies in its in-depth analysis of the antimicrobial drug hierarchy sequence. Leveraging the inherent hierarchical structure and escalation / demotion characteristics of antimicrobial therapy, it incorporates continuous gradient length as a crucial descriptor of sequence changes and introduces it into the self-attention layer of the Transformer network. A larger continuous gradient length indicates a sustained adjustment in the antimicrobial hierarchy in the same direction, and its changes provide stronger indications of future trends, thus warranting a higher weight in the attention layer. The Transformer model excels at capturing long dependencies and sequence patterns; by incorporating the continuous gradient length feature, it further highlights key trend segments in antimicrobial hierarchy changes, thereby improving prediction accuracy.

[0036] This invention further considers the lack of a unified subdivision scheme within antimicrobial grading. By constructing multiple candidate internal grading schemes and training models separately, the optimal scheme is selected using the prediction error of the models on the validation dataset. This allows the internal grading structure to automatically adapt to the changing patterns of real antimicrobial sequences, improving the model's applicability and generalization ability. Reinforcement learning is used to optimize the internal grading scheme, enabling continuous improvement through model feedback. Clustering is used to extract common similar segments from historical antimicrobial grading sequences, automatically identifying potential grading structures from the data, making the candidate schemes more closely resemble real clinical usage characteristics.

[0037] This invention can be used not only to predict the future trend of a single antimicrobial grade sequence, but also to predict the total demand of multiple current antimicrobial grade sequences in the future time period. It can be further used to formulate plans for the allocation, replenishment and procurement of antimicrobial substances, thereby realizing the intelligentization and automation of the antimicrobial substance management information system. Attached Figure Description

[0038] Figure 1 This is an overall schematic diagram of the system of the present invention;

[0039] Figure 2 This is a schematic diagram of the gradient of the present invention;

[0040] Figure 3 This is a schematic diagram illustrating the construction of the internal hierarchical scheme of this invention;

[0041] Figure 4 This is a schematic diagram of the reinforcement learning method of the present invention. Detailed Implementation

[0042] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0043] like Figure 1 As shown, the system of the present invention specifically includes:

[0044] The historical antimicrobial rank sequence construction module is used to obtain multiple historical antimicrobial rank sequences sorted by time from historical antimicrobial treatment data;

[0045] A gradient sequence generation module is used to generate a gradient sequence based on the difference in antibacterial level between adjacent time steps, wherein the gradient value of each time step in the gradient sequence is the difference between the antibacterial level of the current time step and the antibacterial level of the previous time step.

[0046] The continuous gradient length calculation module is used to count the number of consecutive gradient values ​​with the same sign that appear before each time step in the gradient sequence based on a sliding window, and to take the corresponding number as the continuous gradient length of the corresponding time step, with the sign including positive and negative.

[0047] The antibacterial level prediction model construction module is used to make time steps with larger continuous gradient lengths obtain higher weights in the attention weight calculation in the self-attention layer of the Transformer network, and to train the Transformer network based on the historical antibacterial level sequence to obtain the antibacterial level prediction model.

[0048] In a more specific embodiment, the data of this invention originates from the electronic medical record database of the hospital information system. A historical antimicrobial grade sequence corresponds to the antimicrobial medication records of a patient ID during a continuous period of hospitalization or treatment. Considering the sensitivity and privacy protection requirements of medical data, the system first performs strict anonymization processing before the data enters the model training process. In some embodiments, a secure hash algorithm can be used to encrypt the patient ID. Specifically, the SHA-256 algorithm is used to convert the original patient identification identifier into a fixed-length hash string, ensuring that the patient's true identity cannot be deduced in reverse during subsequent analysis, while preserving the temporal correlation of the same patient's data.

[0049] This invention's system, based on the "Administrative Measures for the Clinical Application of Antimicrobial Drugs," classifies antimicrobial drugs into three categories: unrestricted-use antimicrobial drugs, restricted-use antimicrobial drugs, and special-use antimicrobial drugs. In the initial data mapping, these three categories are encoded as the values ​​1, 2, and 3, respectively. In this way, unstructured medical order text is transformed into time-series data ordered by time. ,in Representing the Antibacterial grade at time step.

[0050] In this invention, whether it's the original broad category sorting or the optimized, finer-grained hierarchical coding system discussed later, the corresponding gradient sequence can be obtained. For time steps... Its gradient value The calculation formula is as follows: ,in This represents the antibacterial level value at the current time step. This is the value from the previous time step. This gradient value reflects the direction of change in drug intensity.

[0051] Furthermore, after obtaining the aforementioned gradient sequence, this invention identifies continuously escalating or de-escalating segments in the antibacterial treatment process using a continuous gradient length calculation module, such as... Figure 2 The image shown is a simple example of a gradient. The continuous gradient length calculation module is based on a sliding window mechanism, which calculates the gradient length step by step from the gradient sequence at each time step. t The number of consecutive gradient values ​​with the same sign that have appeared previously. For example, for the current time step. t If at time step t -1、 t If all gradient values ​​in the sequence -2… are positive, it indicates that the antibacterial level shows a continuous upward trend over multiple consecutive time steps. In this case, the length of the continuous gradient is the number of these consecutive positive gradients. Similarly, if all consecutive gradients are negative, it indicates a continuous downward trend. The sliding window updates the statistical results at each time step, thus generating a corresponding length of continuous gradient for each time step. To illustrate this more clearly, assume there is a gradient sequence of data: The corresponding generated sequence of continuous gradient lengths is as follows: the first non-zero positive gradient has a length of 1; the second positive gradient, preceded by a positive gradient, has a length of 2; the third positive gradient, preceded by two positive gradients, has a length of 3; the first negative gradient, interrupting the positive sequence and re-counting, has a length of 1; the second negative gradient, preceded by a negative gradient, has a length of 2. The result is... .

[0052] The physical significance of the continuous gradient length is that if a patient upgrades antibiotics multiple times in succession, the continuous gradient length is large, indicating that the condition may be continuously deteriorating or that drug-resistant bacteria may be emerging. This trend should be given high attention when predicting future medication.

[0053] In this embodiment of the invention, the antibacterial grade prediction model is constructed based on a Transformer network. The self-attention mechanism in the Transformer network models the correlation within the input sequence through the query vector Q, key vector K, and value vector V. For the first... t At each time step, the embedding vector is linearly mapped to generate the corresponding query vector. Key vector Sum value vector The query vector expresses the information that we want to focus on at the current time step, the key vector expresses the feature descriptions of other time steps in the sequence, and the value vector ultimately represents the content that needs to be weighted and aggregated.

[0054] Any two time steps in the self-attention layer t and s Attention weights are calculated according to the following formula:

[0055] ;

[0056] in, The dimension of the key vector is used to scale the dot product result.

[0057] To enable the model to perceive the continuous gradient length feature in the antibacterial grade sequence, this invention introduces weighting coefficients based on the traditional attention formula. For any time step t, the continuous gradient length... The system uses a preset monotonically increasing function to map it to the weight coefficient of that time step:

[0058] ;

[0059] in, For time step t The corresponding weight coefficients indicate that the larger the continuous gradient length, the more significant its trend information, and the higher the attention it should receive in the attention mechanism. The function can adopt a monotonically increasing relationship, such as a linearly increasing function, an exponentially increasing function, or a piecewise increasing function, to ensure that a larger continuous gradient length corresponds to a larger weight coefficient.

[0060] After introducing this weighting coefficient, the calculation process for the attention weights is adjusted as follows: for each time step... t Pay attention to time steps s The attention weight, multiplied by the value from the time step s Weighting coefficients The corrected attention weights are obtained as follows:

[0061] ;

[0062] Where t represents the current time step of interest, and s represents the time step of interest being observed. For the query vector at time step t, The key vector at time step s transpose, The dimension of the key vector. These are the weight coefficients obtained by mapping from continuous gradient lengths.

[0063] In this calculation method, the weighting coefficients only apply to the time step. s This is because in the self-attention structure of the Transformer, each attention weight α(t,s) essentially represents the current time step. t Time step s The degree of correlation. In other words, the weight term adjusts the importance of the information being followed, not the importance of the party initiating the following. Therefore, the length of the continuous gradient related to the strength of the trend should be applied to the key vector side corresponding to the time series segment being followed, i.e., using... To zoom in or out right The response enabled the timing of the upgrades or downgrades to be more pronounced during periods of significant escalation or downgrade trends. s Above, self-attention can give a higher level of focus.

[0064] Final output vector The weighted sum is then calculated by combining the corrected weights with the corresponding value vectors:

[0065] ;

[0066] It corresponds to s The value vector of the step.

[0067] Through the above mechanism, the model can automatically amplify the influence of continuously upgraded or downgraded segments, making these segments occupy a more important position in the final prediction, which helps to improve the prediction accuracy of antimicrobial grade change trends.

[0068] In some embodiments, the Transformer network structure of this invention follows a stacked encoder architecture. Each encoder layer consists of a self-attention sublayer and a feedforward network sublayer, with residual connections between sublayers and layer normalization modules ensuring training stability. The input antimicrobial grade sequence first passes through an embedding layer, mapping the discrete grade encoding to a continuous vector representation; subsequently, positional encoding is introduced, explicitly expressing the order of different time steps in the sequence within the model. The self-attention sublayer executes according to the aforementioned improved attention calculation method, introducing weight coefficients based on continuous gradient lengths to assign higher weights to segments with prominent trends in the sequence. The feedforward network sublayer consists of two linear transformations and activation functions, used to extract more abstract temporal features. The Transformer encoder in this invention can be stacked 2 to 4 layers to fully capture the dynamic patterns of antimicrobial drug changes while balancing model capacity and training efficiency. Finally, the high-dimensional features output by the encoder are mapped to the antimicrobial grade prediction result for the next time step through a linear classification head, achieving modeling and inference of the sequence trend.

[0069] The training process of the Transformer model in this invention employs a supervised learning model. Input data includes: historical antibacterial rank sequences. gradient sequence and continuous gradient length sequences Each layer of the Transformer network's self-attention structure embeds the aforementioned weight amplification logic. Training labels. This represents the antibacterial level at the actual next time step. A weighted cross-entropy loss function is used to address the imbalance issue between different levels of samples. The error between the target output and the Transformer network output is adjusted using a backpropagation algorithm to adjust the network parameters. The system has a convergence condition: training stops when the loss value on the validation set no longer decreases for 5 consecutive epochs, or when the accuracy reaches 90% or higher.

[0070] In a further embodiment, to address the problem that the traditional three-level classification (unrestricted, restricted, special) is too coarse-grained and fails to capture subtle changes in the condition, this invention also systematically introduces an internal grading scheme optimization mechanism. This mechanism aims to further subdivide the levels within the aforementioned three categories, for example, subdividing the restricted use level into restricted-low intensity, restricted-medium intensity, restricted-high intensity, etc., thereby improving the sensitivity of the prediction model.

[0071] In some embodiments, the internal hierarchical scheme can be constructed manually, while in others it can be constructed through reinforcement learning or clustering.

[0072] like Figure 3The diagram illustrates candidate scheme generation based on clustering. The system extracts similar segments from the entire historical dataset. A similar segment is defined as a continuous time series whose values ​​all belong to the same major category. For example, in the sequence ...1,2,2,2,3..., 2,2,2 is a similar segment belonging to the restricted-use category. In some embodiments, a threshold can be set for the length of similar segments, such as considering only segments consisting of three or more similar drugs as similar segments. For each major category, the system collects all similar segments and performs cluster analysis using the K-Means algorithm. Feature vectors are constructed based on the drug's chemical structure, pharmacokinetic parameters, and average clinical dosage. Each cluster formed after clustering represents a potential subtype of drug use pattern. The system extracts one or more typical segments from each cluster as references to form multiple candidate internal grading schemes. The selection of typical segments or combinations of multiple typical segments can be made manually by medical experts. Given that clustering has already been completed, this selection balances the convenience of the clustering algorithm with the rich experience of medical experts.

[0073] In a further embodiment, to select the most representative regimens, the system counts the number of fragments contained in each cluster. A higher number of fragments indicates a more common clinical use pattern, and the corresponding grading regimen is ranked higher. The system ultimately retains the highest-ranked regimens. In some embodiments, several schemes are used as a candidate pool. The value range is usually set to 5 to 10.

[0074] like Figure 4 As shown, in a further embodiment, the system of the present invention can also employ a reinforcement learning framework to select the optimal solution. The selection of the internal hierarchical solution is viewed as a Markov decision process. In this process, the state... Defined as the current internal hierarchical scheme configuration; Action Defined as a fine-tuning of the current scheme, including increasing or decreasing the number of sub-levels within a major category, or adjusting the sorting of sub-levels; rewards The initial state is set to the negative of the prediction error of the prediction model trained under the current scheme on the validation dataset. In some embodiments, the initial state can be set to the candidate scheme obtained based on clustering as described above.

[0075] The specific error evaluation metric used is root mean square error. If the error on the validation set is... Then reward In each iteration, the agent generates a new state representation—a new hierarchical structure—based on the current policy, and remaps historical data using this structure to train a temporary, lightweight prediction model to evaluate the error. Through multiple iterations, the policy network is continuously updated, guiding the system to converge toward the state with the lowest prediction error. The final selected internal hierarchical scheme with the lowest prediction error will serve as the system's fixed internal hierarchical structure for subsequent formal model training and inference. In other words, based on the fundamental ranking of unrestricted-use, restricted-use, and special-use antimicrobial drugs, the drugs within each category are further ranked to obtain the overall hierarchical ranking.

[0076] In a further embodiment of the present invention, the reinforcement learning network may adopt a deep Q-learning-based framework, the core of which consists of a state representation module, a policy network, a target network, and an experience replay module.

[0077] In this invention, the state representation module encodes the current internal hierarchical scheme into a fixed-dimensional state vector. This vector contains the number of sub-levels within each major category (unrestricted-use antimicrobial drugs, restricted-use antimicrobial drugs, and special-use antimicrobial drugs), their ordering, and the typical segment features represented by each sub-level. In this way, hierarchical systems of different complexities can be uniformly mapped to a vector space that can be processed by reinforcement learning networks.

[0078] The policy network employs a multi-layer fully connected neural network structure, including an input layer, two hidden layers, and an output layer. The hidden layers incorporate the ReLU activation function to enhance non-linear expressiveness. The output layer corresponds to a set of optional actions A, including: increasing the number of sub-levels within a given class, decreasing the number of sub-levels, and adjusting the sub-level ranking. The output of the policy network is the Q-value for each action, representing the expected long-term reward obtained by performing that action under the current ranking scheme.

[0079] During training, all state transition samples are stored in the experience replay module. The experience pool stores the results of each iteration in the form of state-action-reward-next state. During policy network training, small batches of samples are randomly sampled from the experience pool to make the sample distribution more uniform, thereby reducing training bias caused by temporal correlation and improving the convergence stability of reinforcement learning. In each iteration, the policy network generates the optimal action based on the current state vector. After the action is executed, historical data is remapped and a lightweight prediction model is trained. The reward value R is obtained from the prediction error of the model on the validation set. The policy network updates its parameters using the reward signal, gradually approaching the level classification scheme that produces the lowest prediction error through multiple iterations. The finally converged reinforcement learning policy is used to automatically determine the optimal internal hierarchy structure.

[0080] The trained antimicrobial level prediction model will be deployed in the hospital's antimicrobial substance management prediction module, supporting two core business scenarios.

[0081] Scenario 1 is single-patient medication prediction. The system receives the current antimicrobial grade sequence for a specific patient, inputs it into the model, and outputs the most likely antimicrobial grade to be used in one or more future time steps, such as the next 24 hours or 48 hours. This can help doctors assess the course of the disease in advance. If the model predicts that the grade will jump from the restricted level to the special level, the system will automatically issue an early warning, prompting the doctor to conduct bacterial culture or expert consultation in advance.

[0082] Scenario two involves forecasting overall regional demand and formulating procurement plans. In this model, the system receives the current antimicrobial level sequence of all inpatients in the entire hospital or a specific department. The model then projects the medication level for each patient in a specific future period, such as the next week.

[0083] Assuming a department has 100 patients, the model predicts that within the next week, 20 will remain in unrestricted condition, 50 will be upgraded to restricted condition, and 30 will require special condition. The system combines average daily dosage data for different levels of medication to calculate the overall demand for each type of drug and automatically generates a procurement list. It recommends that the pharmacy department stockpile the corresponding quantity of special-grade antibacterial substances in advance to avoid shortages of emergency medications and reduce the backlog of unnecessary drugs, thus achieving refined inventory management.

[0084] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An artificial intelligence-based information system for managing antimicrobial substances, characterized in that, include: The historical antimicrobial rank sequence construction module is used to obtain multiple historical antimicrobial rank sequences sorted by time from historical antimicrobial treatment data; A gradient sequence generation module is used to generate a gradient sequence based on the difference in antibacterial level between adjacent time steps, wherein the gradient value of each time step in the gradient sequence is the difference between the antibacterial level of the current time step and the antibacterial level of the previous time step. The continuous gradient length calculation module is used to count the number of consecutive gradient values ​​with the same sign that appear before each time step in the gradient sequence based on a sliding window, and to take the corresponding number as the continuous gradient length of the corresponding time step, with the sign including positive and negative. The antibacterial level prediction model construction module is used to make time steps with larger continuous gradient lengths obtain higher weights in the attention weight calculation in the self-attention layer of the Transformer network, and to train the Transformer network based on the historical antibacterial level sequence to obtain the antibacterial level prediction model. The antibacterial grade prediction model construction module, within the self-attention layer of the Transformer network, adjusts the attention weights based on the continuous gradient length in the following manner: It linearly maps the embedding vector at each time step of the input sequence to obtain the corresponding query vector, key vector, and value vector, and calculates the value vector at any time step. t Time step s When assigning attention weights, the time step... s The continuous gradient length is mapped to a weight coefficient according to a preset monotonically increasing function. The weight coefficient is multiplied by the attention weight calculated from the query vector and the key vector so that the time step with a larger continuous gradient length gets a higher weight in the attention weight calculation. The monotonically increasing correspondence is based on any one of a linear increasing function, an exponential increasing function, or a piecewise increasing function.

2. The artificial intelligence-based antimicrobial substance management information system according to claim 1, characterized in that, The system also includes: The antimicrobial substance management prediction module is used to predict the antimicrobial drugs for one or more future time steps based on the antimicrobial level prediction model when a current antimicrobial level sequence is received; or to predict and formulate a procurement plan for the total demand of each antimicrobial drug for one or more future time steps based on the antimicrobial level prediction model when multiple current antimicrobial level sequences are received.

3. The artificial intelligence-based antimicrobial substance management information system according to claim 1, characterized in that, The antimicrobial grade sequence includes unrestricted-use antimicrobial drugs, restricted-use antimicrobial drugs, and special-use antimicrobial drugs.

4. The artificial intelligence-based antimicrobial substance management information system according to claim 3, characterized in that, The antibacterial grade prediction model construction module is also used for: Multiple internal classification schemes were constructed to further classify antimicrobial drugs within the categories of unrestricted use, restricted use, and special use. Under each internal grading scheme, an antimicrobial grade prediction model was trained, and the corresponding prediction error was evaluated on a validation dataset. The internal grading scheme with the lowest prediction error was selected as the internal hierarchical structure of the antibacterial grade sequence for training and inference of the antibacterial grade prediction model.

5. The artificial intelligence-based antimicrobial substance management information system according to claim 4, characterized in that, The internal ranking scheme with the lowest prediction error is obtained based on reinforcement learning and includes: An internal hierarchical scheme is used as a state representation for the reinforcement learning environment; In each reinforcement learning iteration, a new state representation is generated, and the corresponding antibacterial grade prediction model is trained based on the new state representation; the new state representation is obtained by changing the number or sorting method of the internal sub-grades in the internal grading scheme. The negative value of the prediction error of the antibacterial level prediction model on the validation dataset is used as the reward value for reinforcement learning. The reinforcement learning policy is updated based on the reward value to guide the policy to generate state features with lower prediction errors in subsequent iterations.

6. The artificial intelligence-based antimicrobial substance management information system according to claim 4, characterized in that, The process of constructing the internal hierarchical scheme includes: Extract similar segments from historical antimicrobial grade sequences. The antimicrobial drugs in the similar segments are arranged continuously and all belong to the same major category of antimicrobial grade. The same major category of antimicrobial grade includes any one of the following: unrestricted use antimicrobial drugs, restricted use antimicrobial drugs, and special use antimicrobial drugs. For each major category, all similar segments within the major category are clustered to obtain multiple clusters. From each cluster, one or more typical segments are referenced to extract an internal hierarchical scheme as a candidate scheme for the corresponding cluster. The number of segments contained in each cluster is counted, and the candidate schemes of the clusters with more segments are ranked higher. The top N candidate schemes are used as candidate internal hierarchical schemes for the corresponding major category.

7. The artificial intelligence-based antimicrobial substance management information system according to claim 1, characterized in that, The training process of the antibacterial grade prediction model construction module includes: The historical antimicrobial grade sequence, gradient sequence, and continuous gradient length are used as training inputs. In each layer of the Transformer network's self-attention structure, the attention weights are amplified or reduced based on the continuous gradient length. The network parameters are adjusted based on the error between the target output and the Transformer network output, using training labels that correspond to the actual antibacterial level at the next time step. The Transformer network is trained iteratively through multiple rounds until its prediction performance on the validation set meets the preset convergence condition.

8. The artificial intelligence-based antimicrobial substance management information system according to claim 1, characterized in that, One of the aforementioned historical antimicrobial grade sequences corresponds to a patient ID's antimicrobial medication records over a continuous period of time.

9. The artificial intelligence-based antimicrobial substance management information system according to claim 8, characterized in that, The patient ID is anonymized in the antimicrobial substance management information system.

Images