AI-based one-stop unmanned medicament preparation dosing system and method

Abstract

The invention discloses an AI-based one-stop unmanned medicament preparation dosing system and method. The method comprises the following steps: S1, collecting prescription information; s2, constructing a corresponding bionic organelle reaction cavity; s3, setting an initial process parameter group to obtain a first round of medicament sample; s4, collecting multi-channel response data; s5, generating a response track sequence and a response track diagram based on the multi-channel response data; s6, pairing the response track sequence with the target drug effect response track to construct a target matching subgraph; s7, executing an asynchronous reverse graph path traversal algorithm, reversely searching a response path meeting the minimum reverse cost, and outputting a new round of process parameter group; and S8, performing medicament sample preparation based on the new round of process parameter group, and repeating the steps S3 to S7 until the matching error between the target matching subgraph and the response trajectory graph is lower than a preset tolerance threshold. According to the invention, bionic reaction modeling and an asynchronous inverse graph path traversal algorithm are fused, and one-stop unmanned medicament intelligent preparation, dosing and control are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent automated drug preparation, and in particular to an AI-based one-stop unmanned drug preparation and dosing system and method. Background Art

[0002] The advancement of precision medicine, targeted biological therapy, and personalized drug delivery is placing higher demands on automation, intelligence, and precise control of the pharmaceutical preparation process. Traditional drug dosing processes typically rely on manual operation or automated equipment with pre-set rules, encompassing multiple steps such as drug proportioning, mixing, injection, and reaction observation. While a relatively mature industrial system has been established for conventional batch drug production, numerous technical bottlenecks remain in scenarios requiring flexible control and rapid iteration, such as personalized therapy, experimental screening, and targeted drug efficacy verification.

[0003] Current mainstream automated dosing systems rely on fixed process parameter settings and passive data collection. They lack the ability to dynamically simulate drug responses in biological microenvironments, particularly in reconstructing the targeted action of drugs at the organelle level. This makes it difficult for existing systems to automatically generate optimal preparation plans based on specific drugs, prescriptions, and targeting mechanisms. They also lack the ability to monitor the efficacy of preparations in real-time within in vivo simulations, let alone intelligently optimize and feedback control the preparation pathway.

[0004] Some systems attempt to incorporate AI for drug parameter prediction or data-driven modeling, but most focus on upstream processes like molecular property assessment and efficacy prediction, becoming disconnected from the actual formulation process and struggling to achieve closed-loop control of the entire "prescription-preparation-reaction-feedback" process. Regarding drug reaction validation, traditional reaction chamber models are mostly general-purpose microfluidic structures that lack the sophisticated biomimetic capabilities of organelles. They are unable to faithfully replicate the diffusion and response characteristics of drugs within specific cellular substructures (such as mitochondria and the Golgi apparatus), making it difficult to optimize process parameters for specific targets.

[0005] Furthermore, response data collection typically remains at the simple sensor reading stage, failing to structure and integrate it into a graphical trajectory suitable for AI optimization. This inability to model the causal chain between pH, potential, and fluorescence multi-channel signals, and the inability to reverse-derive process parameters from efficacy trajectories, furthermore, lacks the ability to derive process parameters from efficacy trajectories. Existing path optimization methods, such as genetic algorithms, grid search, or reinforcement learning, suffer from issues such as high-dimensionality, uninterpretability, high iteration costs, and unstable convergence, making them difficult to deploy in real time in actual pharmaceutical production.

[0006] In terms of describing pharmacodynamic targets, most existing methods use static indicators or interval value targets, such as blood drug concentration range, dissolution limit, etc., which are difficult to express the continuously changing expected pharmacodynamic response trajectory, and thus cannot achieve dynamic target driving of the preparation process.

[0007] Therefore, how to provide an AI-based one-stop unmanned drug preparation and dosing system and method is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0008] One purpose of the present invention is to propose an AI-based one-stop unmanned drug preparation and dosing system and method. The present invention integrates bionic organelle reaction modeling, graph structure path optimization and AI-driven process control methods, and systematically realizes automatic closed-loop control of the entire process from prescription analysis to drug preparation. It constructs a response trajectory graph that can match dynamic efficacy targets, and generates a new round of process parameter groups through an asynchronous reverse graph path traversal algorithm. It has the significant advantages of strong personalization, high response accuracy and high degree of preparation automation, and effectively breaks through the problem that traditional dosing systems cannot provide real-time feedback and optimization for the dynamic reactions of drug targets.

[0009] The AI-based one-stop unmanned drug preparation and dosing method according to an embodiment of the present invention includes the following steps: S1. Collect prescription information, parse it to obtain drug structural parameters, dosage ratio, target organelle type and target pharmacodynamic response trajectory, and generate drug-organelle mapping data; S2. constructing a corresponding bionic organelle reaction cavity according to the drug-organelle mapping data; S3. Setting an initial process parameter set, preparing a medicine based on the initial process parameter set, and obtaining a first-round medicine sample; S4, injecting the first round of drug samples into the bionic organelle reaction chamber, and collecting multi-channel response data generated by the first round of drug samples in the bionic organelle reaction chamber using three types of response sensors: pH, potential, and fluorescence; S5. generating a response trajectory sequence based on the multi-channel response data, and constructing a response trajectory graph based on the response trajectory sequence; S6. Pairing the response trajectory sequence with the target drug efficacy response trajectory to build a graph, and constructing a target matching subgraph based on the node similarity measurement and time series consistency; S7, executing an asynchronous reverse graph path traversal algorithm, reversely searching for a response path that satisfies the minimum reverse cost based on the target matching subgraph, and outputting a new round of process parameter groups; S8. Prepare the pharmaceutical sample based on the new set of process parameters, and repeat steps S3 to S7 until the matching error between the target matching subgraph and the response trajectory graph is lower than a preset tolerance threshold.

[0010] Optionally, the S1 specifically includes: S11. Obtain prescription information, including structured electronic prescription files and natural language format prescription text; S12. Parse the prescription information, extract the drug name, single dose, and frequency of use, and determine the corresponding dosage ratio; S13. Retrieving corresponding molecular structure information from a drug database according to the drug name to generate drug structure parameters, wherein the drug structure parameters include molecular composition, functional group characteristics, charge state, and hydrophilicity index; S14. Analyze the targeted therapeutic site information involved in the prescription and determine the corresponding targeted organelle type based on the known mechanism of action of the drug; S15. Retrieving drug-related in vivo kinetic reference data based on the drug's use and duration of action requirements to construct a target pharmacodynamic response trajectory, where the target pharmacodynamic response trajectory is composed of pharmacodynamic variables at multiple time nodes; S16. Correlating and integrating the drug structural parameters, dosage ratio, targeted organelle type, and target pharmacodynamic response trajectory to generate drug-organelle mapping data.

[0011] Optionally, the bionic organelle reaction chamber is constructed based on the drug-organelle mapping data, and the bionic organelle reaction chamber includes: A closed microcavity structure is filled with a simulation buffer corresponding to the target organelle type in the drug-organelle mapping data, wherein the simulation buffer is composed of pH value, osmotic pressure and ions; A permeable membrane structure, which mimics the selective permeability characteristics of the targeted organelle membrane and has the ability to control the diffusion rate of small molecules and establish membrane potential; The liquid injection interface is connected to the medicine preparation module and is used to receive the quantitative injection of each round of medicine samples; Three types of response sensor interfaces, connected to pH response sensors, potential response sensors, and fluorescence detectors, respectively, for collecting multi-channel response data generated by drug samples in the bionic organelle reaction chamber; The data output port is used to transmit the multi-channel response data to the response mapping module in real time.

[0012] Optionally, the initial process parameter group includes the feeding sequence, feeding speed, mixing rate, mixing time, reaction temperature and pH control target value. The initial process parameter group is input into the drug preparation module, and the drug preparation module is controlled to execute the raw material addition, mixing, reaction and stabilization treatment sub-processes in sequence to obtain the first round of drug samples.

[0013] Optionally, the S5 specifically includes: S51. Constructing an original response data sequence indexed by time nodes based on the multi-channel response data; S52. Generate a response state unit for each time node. The response state unit is a six-dimensional vector containing three types of signal values: pH, potential, and fluorescence, and the first-order change rate of each signal value, which are arranged in time sequence to form a response trajectory sequence; S53, performing fragmentation processing on the response trajectory sequence to extract signal mutation segments, wherein the signal mutation segments satisfy the requirement that the signal change rate exceeds a set driving threshold and have a start and end boundary; S54, taking each signal mutation segment as a graph node of the response trajectory graph, wherein the graph node attributes include a starting state vector, an ending state vector, a dominant signal channel identifier, and a duration of the change; S55. Construct directed edges of the response trajectory diagram based on the trigger relationship between different signal mutation segments. The trigger relationship is established according to the causal order of the successive signals, indicating that the termination state vector of the predecessor signal mutation segment activates the starting state vector of the successor signal mutation segment within the set time delay range, and outputs the response trajectory diagram.

[0014] Optionally, the S6 specifically includes: S61, extracting each time node in the response trajectory sequence in chronological order, and performing a structured comparison with each time node in the target drug efficacy response trajectory to form an initial node pair set; S62, calculating the node similarity of each initial node pair, wherein the node similarity is determined based on the numerical deviation of the three signal values ​​of pH, potential and fluorescence at the same time node; S63, screening node pairs whose node similarity is within a set similarity threshold range and whose time node sequence of the response trajectory sequence is consistent with the time node sequence of the target drug efficacy response trajectory, to form a matching path; S64. Based on the matching path, extract a minimum directed subgraph including all matching nodes and directly connected edges from the response trajectory graph to construct a target matching subgraph.

[0015] Optionally, the S7 specifically includes: S71, in the target matching subgraph In the graph, define each node The response state energy function is: ; in, Representing graph nodes The response state energy value, Representing graph nodes pH value, Representing graph nodes The potential value, Representing graph nodes The fluorescence signal value, 、 and represents the weight coefficient; S72, for directed edges , define the inverse cost function as: ; in, Represents a connection graph node To graph node Directed edges The reverse cost value, Representing graph nodes The associated process parameter vector, Representing graph nodes The associated process parameter vector, represents the L1 norm, represents the process disturbance penalty factor; S73, in the target matching subgraph Execute asynchronous reverse graph path traversal algorithm in the response trajectory to terminate the node To the starting node Perform reverse path traversal to search for the response path that satisfies the minimum reverse cost ; S74. During the reverse path traversal, if there is an edge The reverse cost value Greater than the preset jump threshold , then the reverse jump mechanism is triggered, and the jump candidate nodes that meet the reverse jump conditions are searched in the target matching subgraph. : Condition 1: Node Timestamp Less than the current node Timestamp ,Right now ; Condition 2: Node To the current node The reverse side of The reverse cost value Less than or equal to the jump threshold ,Right now ; Condition 3: Time difference Less than the maximum hop delay window ,Right now ; Among all the candidate jump nodes that meet the three conditions at the same time, select the node that minimizes the reverse cost. , and the edge Incorporate response pathways; S75. Response path with minimum reverse cost The process parameter vectors corresponding to all nodes above are interpolated by spline to generate a new round of process parameter groups; S76: Input the new round of process parameter group as a control instruction into the medicine preparation module to replace the initial process parameter group to execute medicine sample preparation.

[0016] Optionally, the matching error between the target matching subgraph and the response trajectory graph specifically includes spatially matching the nodes in the target matching subgraph with the corresponding nodes in the response trajectory graph, calculating the difference value between each pair of nodes using Euclidean distance, and performing weighted summation on all difference values.

[0017] The AI-based one-stop unmanned drug preparation and dosing system according to an embodiment of the present invention includes the following modules: Prescription parsing module, used to extract drug structural parameters, dosage ratio, target organelle type and target pharmacodynamic response trajectory, and generate drug-organelle mapping data; The biomimetic reaction chamber module is used to construct a biomimetic organelle reaction chamber based on drug-organelle mapping data and collect pH, potential and fluorescence response data; The pharmaceutical preparation module is used to complete the raw material addition, mixing, reaction and stabilization sub-processes according to the process parameter group and output the pharmaceutical sample; Response mapping module, used to construct response trajectory sequence and response trajectory graph, and generate target matching subgraph; The path optimization module is used to execute the asynchronous reverse graph path traversal algorithm in the target matching subgraph to obtain the response path with the minimum reverse cost and generate a new round of process parameter groups; The iterative control module is used to evaluate the matching error and determine whether to proceed to the next round of drug preparation.

[0018] The beneficial effects of the present invention are: (1) The present invention can extract drug structural parameters, dosage ratios, target organelle types, and target pharmacodynamic response trajectories from structured or natural language prescriptions, generate drug-organelle mapping data, realize intelligent initialization of the preparation process, and significantly improve the adaptability and responsiveness of drug preparation.

[0019] (2) By constructing a bionic organelle reaction cavity that matches the prescription information, the simulated response of the drug in the microphysiological environment is achieved. The three types of sensing channels, pH, potential, and fluorescence, are combined to collect multi-dimensional physiological signals, providing real and measurable data support for the efficacy trajectory.

[0020] (3) By modeling the drug reaction process through the response trajectory graph, a target matching subgraph construction method driven by node similarity and time consistency is introduced for the first time to achieve structured expression and precise matching of complex dynamic drug efficacy targets.

[0021] (4) An asynchronous reverse path search method based on the reverse cost function and jump tolerance control mechanism is proposed, which significantly improves the optimality and global stability of the response path and provides efficient algorithm support for the reverse derivation of process parameters.

[0022] (5) The system has a complete error evaluation and iterative update mechanism, which can continuously optimize the process parameters between the response trajectory map and the target trajectory until the matching error is lower than the preset threshold, thus realizing a truly one-stop, unmanned closed-loop control of intelligent drug preparation. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is the overall flow chart of the AI-based one-stop unmanned drug preparation and dosing method proposed in the present invention; Figure 2 This is a schematic diagram of the structure of the AI-based one-stop unmanned drug preparation and dosing system proposed in the present invention. DETAILED DESCRIPTION

[0024] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0025] refer to Figure 1 The AI-based one-stop unmanned drug preparation and dosing method includes the following steps: S1. Collect prescription information, parse it to obtain drug structural parameters, dosage ratio, target organelle type and target pharmacodynamic response trajectory, and generate drug-organelle mapping data; S2. constructing a corresponding bionic organelle reaction cavity according to the drug-organelle mapping data; S3. Setting an initial process parameter set, preparing a medicine based on the initial process parameter set, and obtaining a first-round medicine sample; S4, injecting the first round of drug samples into the bionic organelle reaction chamber, and collecting multi-channel response data generated by the first round of drug samples in the bionic organelle reaction chamber using three types of response sensors: pH, potential, and fluorescence; S5. generating a response trajectory sequence based on the multi-channel response data, and constructing a response trajectory graph based on the response trajectory sequence; S6. Pairing the response trajectory sequence with the target drug efficacy response trajectory to build a graph, and constructing a target matching subgraph based on the node similarity measurement and time series consistency; S7, executing an asynchronous reverse graph path traversal algorithm, reversely searching for a response path that satisfies the minimum reverse cost based on the target matching subgraph, and outputting a new round of process parameter groups; S8. Prepare the pharmaceutical sample based on the new set of process parameters, and repeat steps S3 to S7 until the matching error between the target matching subgraph and the response trajectory graph is lower than a preset tolerance threshold.

[0026] The present invention realizes automatic control of the entire process from drug analysis, reaction response to parameter iterative optimization by establishing a complete closed-loop preparation process with prescription information as the entry point. Unlike traditional static preparation schemes, this scheme can build a customized bionic reaction system based on the molecular structure, dosage requirements and target location information of the drug, and combine the structured matching of the response trajectory and the efficacy target trajectory to realize graphical modeling and path reverse planning of the preparation path, thereby significantly improving the intelligence, individual adaptability and dynamic response accuracy of drug preparation. The system has closed-loop iteration capabilities and can continuously optimize process parameters on the basis of ensuring dynamic matching of drug efficacy, realizing a truly unmanned one-stop drug preparation process.

[0027] In this embodiment, S1 specifically includes: S11. Obtain prescription information, including structured electronic prescription files and natural language format prescription text; S12. Parse the prescription information, extract the drug name, single dose, and frequency of use, and determine the corresponding dosage ratio; S13. Retrieving corresponding molecular structure information from a drug database according to the drug name to generate drug structure parameters, wherein the drug structure parameters include molecular composition, functional group characteristics, charge state, and hydrophilicity index; S14. Analyze the targeted therapeutic site information involved in the prescription and determine the corresponding targeted organelle type based on the known mechanism of action of the drug; S15. Retrieving drug-related in vivo kinetic reference data based on the drug's use and duration of action requirements to construct a target pharmacodynamic response trajectory, where the target pharmacodynamic response trajectory is composed of pharmacodynamic variables at multiple time nodes; S16. Correlating and integrating the drug structural parameters, dosage ratio, targeted organelle type, and target pharmacodynamic response trajectory to generate drug-organelle mapping data.

[0028] By parsing structured electronic prescriptions and natural language prescription text, the team extracted parameters including drug structural characteristics, dosage ratios, targeted organelle types, and target efficacy trajectories, breaking down the information barriers between original prescription information and manufacturing execution. The proposed structural analysis and database retrieval strategy ensures the comprehensiveness and accuracy of information extraction. The generated drug-organelle mapping data provides a data foundation for biomimetic reaction construction and response evaluation, avoiding the uncertainty associated with manually set manufacturing parameters and achieving a customized manufacturing starting point based on the drug's inherent properties.

[0029] In this embodiment, the construction of the bionic organelle reaction chamber is based on the drug-organelle mapping data, and the bionic organelle reaction chamber includes: A closed microcavity structure is filled with a simulation buffer corresponding to the target organelle type in the drug-organelle mapping data, wherein the simulation buffer is composed of pH value, osmotic pressure and ions; A permeable membrane structure, which mimics the selective permeability characteristics of the targeted organelle membrane and has the ability to control the diffusion rate of small molecules and establish membrane potential; The liquid injection interface is connected to the medicine preparation module and is used to receive the quantitative injection of each round of medicine samples; Three types of response sensor interfaces, connected to pH response sensors, potential response sensors, and fluorescence detectors, respectively, for collecting multi-channel response data generated by drug samples in the bionic organelle reaction chamber; The data output port is used to transmit the multi-channel response data to the response mapping module in real time.

[0030] The designed biomimetic organelle reaction chamber is constructed based on targeted mapping data of drugs and organelles. By simulating the physical environment of specific organelles within the chamber, it simulates the reaction process of the drug within specific subcellular structures. The reaction chamber integrates multi-channel sensors and has dynamic data acquisition capabilities, providing accurate dynamic reaction data support for response trajectory construction and reverse optimization. This allows the system to perform feedback optimization based on realistic simulation results, significantly improving the physiological relevance of response data and the accuracy of drug efficacy prediction.

[0031] In this embodiment, the initial process parameter group includes the feeding sequence, feeding speed, mixing rate, mixing time, reaction temperature and pH control target value. The initial process parameter group is input into the drug preparation module, and the drug preparation module is controlled to execute the raw material addition, mixing, reaction and stabilization treatment sub-processes in sequence to obtain the first round of drug samples.

[0032] In this embodiment, the S5 specifically includes: S51. Constructing an original response data sequence indexed by time nodes based on the multi-channel response data; S52. Generate a response state unit for each time node. The response state unit is a six-dimensional vector containing three types of signal values: pH, potential, and fluorescence, and the first-order change rate of each signal value, which are arranged in time sequence to form a response trajectory sequence; S53, performing fragmentation processing on the response trajectory sequence to extract signal mutation segments, wherein the signal mutation segments satisfy the requirement that the signal change rate exceeds a set driving threshold and have a start and end boundary; S54, taking each signal mutation segment as a graph node of the response trajectory graph, wherein the graph node attributes include a starting state vector, an ending state vector, a dominant signal channel identifier, and a duration of the change; S55. Construct directed edges of the response trajectory diagram based on the trigger relationship between different signal mutation segments. The trigger relationship is established according to the causal order of the successive signals, indicating that the termination state vector of the predecessor signal mutation segment activates the starting state vector of the successor signal mutation segment within the set time delay range, and outputs the response trajectory diagram.

[0033] When constructing the response trajectory sequence, the system not only extracts time indexes and state vectors from the original signal sequence but also introduces signal mutation segments as graph nodes, defining node attributes based on multi-channel change rates. By extracting the causal triggering relationships between segments and establishing a directed graph structure, this not only preserves timing information but also explicitly expresses the dynamic drive chain between multi-channel signals. This trajectory modeling approach lays the structural foundation for subsequent target matching, path planning, and graph traversal, serving as a critical bridge from data to graph-based intelligent reasoning.

[0034] In this embodiment, S6 specifically includes: S61, extracting each time node in the response trajectory sequence in chronological order, and performing a structured comparison with each time node in the target drug efficacy response trajectory to form an initial node pair set; S62, calculating the node similarity of each initial node pair, wherein the node similarity is determined based on the numerical deviation of the three signal values ​​of pH, potential and fluorescence at the same time node; S63, screening node pairs whose node similarity is within a set similarity threshold range and whose time node sequence of the response trajectory sequence is consistent with the time node sequence of the target drug efficacy response trajectory, to form a matching path; S64. Based on the matching path, extract a minimum directed subgraph including all matching nodes and directly connected edges from the response trajectory graph to construct a target matching subgraph.

[0035] By introducing a matching mechanism that incorporates dual constraints of node similarity and temporal consistency, false matches caused by single-channel deviations or temporal mismatches can be effectively avoided, ensuring that matching results are both physiologically meaningful and logically reasonable. The generation of a target matching subgraph not only establishes a structural comparison between the response trajectory and the efficacy target but also limits the scope of the reverse search graph, improving the computational efficiency of the graph traversal and the stability of path convergence. This is a core prerequisite for achieving accurate reverse path derivation.

[0036] In this embodiment, the S7 specifically includes: S71, in the target matching subgraph In the graph, define each node The response state energy function is: ; in, Representing graph nodes The response state energy value, Representing graph nodes pH value, Representing graph nodes The potential value, Representing graph nodes The fluorescence signal value, 、 and represents the weight coefficient; S72, for directed edges , define the inverse cost function as: ; in, Represents a connection graph node To graph node Directed edges The reverse cost value, Representing graph nodes The associated process parameter vector, Representing graph nodes The associated process parameter vector, represents the L1 norm, represents the process disturbance penalty factor; S73, in the target matching subgraph Execute asynchronous reverse graph path traversal algorithm in the response trajectory to terminate the node To the starting node Perform reverse path traversal to search for the response path that satisfies the minimum reverse cost ; S74. During the reverse path traversal, if there is an edge The reverse cost value Greater than the preset jump threshold , then the reverse jump mechanism is triggered, and the jump candidate nodes that meet the reverse jump conditions are searched in the target matching subgraph. : Condition 1: Node Timestamp Less than the current node Timestamp ,Right now ; Condition 2: Node To the current node The reverse side of The reverse cost value Less than or equal to the jump threshold ,Right now ; Condition 3: Time difference Less than the maximum hop delay window ,Right now ; Among all the candidate jump nodes that meet the three conditions at the same time, select the node that minimizes the reverse cost. , and the edge Incorporate response pathways; S75. Response path with minimum reverse cost The process parameter vectors corresponding to all nodes above are interpolated by spline to generate a new round of process parameter groups; S76: Input the new round of process parameter group as a control instruction into the medicine preparation module to replace the initial process parameter group to execute medicine sample preparation.

[0037] By proposing a response state energy function and an inverse cost function, the authors combined physiological signal intensity differences with process parameter perturbations to form inverse path costs, constructing an interpretable and computable path optimization evaluation criterion. An asynchronous inverse graph path traversal algorithm, combined with a jump mechanism, supports tolerance jump searches on costly paths, breaking through the local optimality limitations of traditional path search. Spline interpolation generates continuously executable process instructions, enabling dynamic reconstruction and rapid deployment of optimal parameter paths. This is a key algorithmic module for implementing intelligent feedback closed-loop control.

[0038] In this embodiment, the matching error between the target matching subgraph and the response trajectory graph specifically includes spatial position matching of nodes in the target matching subgraph and corresponding nodes in the response trajectory graph, calculating the difference value between each pair of nodes using Euclidean distance, and performing weighted summation of all difference values.

[0039] In path iteration control, weighted error calculation is performed based on the Euclidean distance between corresponding nodes in the target matching subgraph and the response trajectory graph, enabling a quantitative measurement of the difference between the preparation result and the target efficacy. This not only provides an iterative termination condition but also provides guidance for improving path accuracy, avoiding blind optimization iterations or misjudging convergence results, and ensuring that formulation iterations have clear convergence targets and evaluation benchmarks.

[0040] refer to Figure 2 , an AI-based one-stop unmanned drug preparation and dosing system, including the following modules: Prescription parsing module, used to extract drug structural parameters, dosage ratio, target organelle type and target pharmacodynamic response trajectory, and generate drug-organelle mapping data; The biomimetic reaction chamber module is used to construct a biomimetic organelle reaction chamber based on drug-organelle mapping data and collect pH, potential and fluorescence response data; The pharmaceutical preparation module is used to complete the raw material addition, mixing, reaction and stabilization sub-processes according to the process parameter group and output the pharmaceutical sample; Response mapping module, used to construct response trajectory sequence and response trajectory graph, and generate target matching subgraph; The path optimization module is used to execute the asynchronous reverse graph path traversal algorithm in the target matching subgraph to obtain the response path with the minimum reverse cost and generate a new round of process parameter groups; The iterative control module is used to evaluate the matching error and determine whether to proceed to the next round of drug preparation.

[0041] The resulting one-stop, unmanned drug preparation and dosing system integrates a prescription parsing module, a biomimetic reaction chamber module, a drug preparation module, a response mapping module, a path optimization module, and an iterative control module, forming a complete closed-loop process from data-driven response acquisition to graph modeling and process optimization. The system, requiring no human intervention, automatically executes drug preparation, response evaluation, and path iteration based on prescriptions, significantly reducing experimental and formulation costs and improving efficiency. It is suitable for multiple application scenarios, including intelligent pharmaceutical manufacturing, personalized medication, and high-throughput drug efficacy screening, and has strong potential for technology promotion and industrialization.

[0042] Example 1: To verify the feasibility of this invention, it was applied to a study on the intelligent preparation of a mitochondrial-targeted drug combination formulation. This study, conducted jointly by the pharmacy department and experimental center of a large medical institution, aimed to address issues such as low formulation precision, unstable efficacy response, and lack of personalized control in traditional manual formulation processes.

[0043] The samples selected for the experiment were from several clinical cases with real prescription backgrounds, totaling 18 sets of prescription data, covering both structured electronic medical records and unstructured text formats. The drugs involved primarily focused on combination anti-cancer therapies, primarily including liposome doxorubicin and cisplatin injection, commonly used drugs for mitochondrial-targeted therapies. Traditional procedures rely heavily on pharmacists' experience, including manual weighing, mixing, injection, and recording of response changes, making them incapable of meeting the requirements for refined control and intelligent response feedback.

[0044] After receiving prescription information, the system's analysis module automatically extracts the drug's molecular structure, dosage frequency, target organ information, and target response trajectory. The system then uses an internal drug database to perform structural modeling based on the drug's molecular functional groups, charge state, and hydrophilicity / hydrophobicity parameters. The analysis success rate is 100%, and the structural deviation error is controlled within 1.6%.

[0045] The system then constructed a biomimetic mitochondrial reaction chamber, simulating the typical pH, potential gradient, and ion concentration of the mitochondrial environment. It also configured a membrane structure to simulate real-world membrane permeability. The drug preparation module executed the sample preparation process based on a default initial set of process parameters, including the order of dosing (cisplatin followed by doxorubicin), a stirring rate of 600 rpm, a reaction temperature set to body temperature, and a pH control range of 7.2 ± 0.05.

[0046] After preparation, the drug sample is injected into the biomimetic reaction chamber. The system collects dynamic response signals using three types of response sensors (pH, potential, and fluorescence), constructing them into a response trajectory sequence and a response trajectory graph. In the graph matching module, the system compares the response trajectory with the target efficacy trajectory node by node, extracting the target matching subgraph using a node similarity metric and temporal consistency constraints.

[0047] The path optimization module executes an asynchronous reverse graph path traversal algorithm, outputting a set of optimized process parameters in each iteration. The system supports up to eight rounds of automatic process iterations. In actual testing, the optimal path converges between the fifth and sixth rounds, with the average matching error for all response trajectories kept below 0.25.

[0048] During the test, the system compared the average single-round preparation time, response trajectory matching, process parameter consistency error and response fragment recognition accuracy of the manual method and the AI ​​system on 18 groups of prescription samples. The results showed that this system showed obvious advantages in most dimensions.

[0049] Table 1 Performance comparison between AI intelligent drug preparation system and traditional manual methods

[0050] Based on analysis of Table 1 above, six representative drug preparation samples were selected for this comparative experiment. The proposed AI-powered drug preparation system and traditional manual methods were used to test drug configuration and response behavior. A cross-evaluation across multiple key metrics validated the system's performance advantages in real-world scenarios.

[0051] First, looking at the drug structure complexity scores, all samples ranged from 6.8 to 8.1, demonstrating the challenges of sample selection and the system's adaptability in handling complex drug structures. For example, A03, despite a complexity score of 8.1, was still able to be prepared with high precision by the AI ​​system in 44 minutes, demonstrating the system's ability to analyze high-dimensional drug features.

[0052] In terms of preparation efficiency, the AI ​​system's average single-round preparation time is controlled between 39 and 44 minutes, compared to the 70 to 78 minutes of traditional manual methods, a time savings of over 40%. For example, sample A01 took only 39 minutes to prepare using the AI ​​system, while manual preparation took 73 minutes, a significant efficiency improvement.

[0053] Response trajectory matching is a key indicator of whether a drug's biological reaction process is close to the preset target trajectory. The AI ​​system's matching degree consistently exceeded 88%, with a maximum of 93.5%. Manual methods generally achieved matching degrees below 62%, with the lowest dropping to 56.4%. This indicates that traditional methods exhibit significant deviations in multi-channel response control. The AI ​​system, however, can automatically match efficacy trajectories and optimize preparation parameters, significantly improving the consistency of drug responses.

[0054] The AI ​​system maintained a consistent error in process parameters between 2.4% and 3.1%, significantly better than the 9.7% to 12.5% ​​achieved by manual methods. This demonstrates the system's significant advantages in parameter control accuracy and sample repeatability, mitigating the impact of human error on drug efficacy.

[0055] The accuracy of response segment recognition reflects the system's ability to identify key mutation points in physiological signals. The AI ​​system generally reaches above 95% in this regard, with the highest being 98.1%. However, manual processing capabilities are obviously insufficient, with some reaching only 74.9%. This shows that the system has higher sensitivity and resolution when processing nonlinear multidimensional signals.

[0056] This embodiment comprehensively demonstrates the feasibility and advantages of the present invention in practical applications by comparing and verifying the AI ​​intelligent drug preparation system with traditional manual methods under real prescription conditions. The system automatically parses prescriptions, constructs a bionic reaction environment, collects multi-channel response signals, and performs response trajectory matching and process optimization based on graph algorithms, ultimately achieving a closed-loop drug preparation process without human intervention. Experimental results show that the system is significantly superior to traditional methods in terms of average single-round preparation time, response trajectory matching, process parameter consistency error, and response fragment recognition accuracy. It can be seen that the present invention not only effectively solves the problems of low efficiency, uncontrollable response, and difficulty in personalization in traditional drug preparation, but also provides a new solution path for precision medicine and intelligent drug management, and has broad clinical application prospects and promotion value.

[0057] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. An AI-based one-stop unmanned drug preparation and dosing method, characterized in that: The steps include: S1. Collect prescription information, parse it to obtain drug structural parameters, dosage ratio, target organelle type and target pharmacodynamic response trajectory, and generate drug-organelle mapping data; S2. constructing a corresponding bionic organelle reaction cavity according to the drug-organelle mapping data; S3. Setting an initial process parameter set, preparing a medicine based on the initial process parameter set, and obtaining a first-round medicine sample; S4, injecting the first round of drug samples into the bionic organelle reaction chamber, and collecting multi-channel response data generated by the first round of drug samples in the bionic organelle reaction chamber using three types of response sensors: pH, potential, and fluorescence; S5. generating a response trajectory sequence based on the multi-channel response data, and constructing a response trajectory graph based on the response trajectory sequence; S6. Pairing the response trajectory sequence with the target drug efficacy response trajectory to build a graph, and constructing a target matching subgraph based on the node similarity measurement and time series consistency; S7, executing an asynchronous reverse graph path traversal algorithm, reversely searching for a response path that satisfies the minimum reverse cost based on the target matching subgraph, and outputting a new round of process parameter groups; S8. Prepare the pharmaceutical sample based on the new set of process parameters, and repeat steps S3 to S7 until the matching error between the target matching subgraph and the response trajectory graph is lower than a preset tolerance threshold.

2. The AI-based one-stop unmanned drug preparation and dosing method according to claim 1, characterized in that: Said S1 specifically includes: S11. Obtain prescription information, including structured electronic prescription files and natural language format prescription text; S12. Parse the prescription information, extract the drug name, single dose, and frequency of use, and determine the corresponding dosage ratio; S13. Retrieving corresponding molecular structure information from a drug database according to the drug name to generate drug structure parameters, wherein the drug structure parameters include molecular composition, functional group characteristics, charge state, and hydrophilicity index; S14. Analyze the targeted therapeutic site information involved in the prescription and determine the corresponding targeted organelle type based on the known mechanism of action of the drug; S15. Retrieving drug-related in vivo kinetic reference data based on the drug's use and duration of action requirements to construct a target pharmacodynamic response trajectory, where the target pharmacodynamic response trajectory is composed of pharmacodynamic variables at multiple time nodes; S16. Correlating and integrating the drug structural parameters, dosage ratio, targeted organelle type, and target pharmacodynamic response trajectory to generate drug-organelle mapping data.

3. The AI-based one-stop unmanned drug preparation and dosing method according to claim 1, characterized in that: The bionic organelle reaction chamber is constructed based on the drug-organelle mapping data, and the bionic organelle reaction chamber includes: A closed microcavity structure is filled with a simulation buffer corresponding to the target organelle type in the drug-organelle mapping data, wherein the simulation buffer is composed of pH value, osmotic pressure and ions; A permeable membrane structure, which mimics the selective permeability characteristics of the targeted organelle membrane and has the ability to control the diffusion rate of small molecules and establish membrane potential; The liquid injection interface is connected to the medicine preparation module and is used to receive the quantitative injection of each round of medicine samples; Three types of response sensor interfaces, connected to pH response sensors, potential response sensors, and fluorescence detectors, respectively, for collecting multi-channel response data generated by drug samples in the bionic organelle reaction chamber; The data output port is used to transmit the multi-channel response data to the response mapping module in real time.

4. The AI-based one-stop unmanned drug preparation and dosing method according to claim 1, characterized in that: The initial process parameter group includes the feeding sequence, feeding speed, mixing rate, mixing time, reaction temperature and pH control target value. The initial process parameter group is input into the drug preparation module, and the drug preparation module is controlled to execute the raw material addition, mixing, reaction and stabilization treatment sub-processes in sequence to obtain the first round of drug samples.

5. The AI-based one-stop unmanned drug preparation and dosing method according to claim 1, characterized in that: The S5 specifically includes: S51. Constructing an original response data sequence indexed by time nodes based on the multi-channel response data; S52. Generate a response state unit for each time node. The response state unit is a six-dimensional vector containing three types of signal values: pH, potential, and fluorescence, and the first-order change rate of each signal value, which are arranged in time sequence to form a response trajectory sequence; S53, performing fragmentation processing on the response trajectory sequence to extract signal mutation segments, wherein the signal mutation segments satisfy the requirement that the signal change rate exceeds a set driving threshold and have a start and end boundary; S54, taking each signal mutation segment as a graph node of the response trajectory graph, wherein the graph node attributes include a starting state vector, an ending state vector, a dominant signal channel identifier, and a duration of the change; S55. Construct directed edges of the response trajectory diagram based on the trigger relationship between different signal mutation segments. The trigger relationship is established according to the causal order of the successive signals, indicating that the termination state vector of the predecessor signal mutation segment activates the starting state vector of the successor signal mutation segment within the set time delay range, and outputs the response trajectory diagram.

6. The AI-based one-stop unmanned drug preparation and dosing method according to claim 1, characterized in that: The S6 specifically includes: S61, extracting each time node in the response trajectory sequence in chronological order, and performing a structured comparison with each time node in the target drug efficacy response trajectory to form an initial node pair set; S62, calculating the node similarity of each initial node pair, wherein the node similarity is determined based on the numerical deviation of the three signal values ​​of pH, potential and fluorescence at the same time node; S63, screening node pairs whose node similarity is within a set similarity threshold range and whose time node sequence of the response trajectory sequence is consistent with the time node sequence of the target drug efficacy response trajectory, to form a matching path; S64. Based on the matching path, extract a minimum directed subgraph including all matching nodes and directly connected edges from the response trajectory graph to construct a target matching subgraph.

7. The AI-based one-stop unmanned drug preparation and dosing method according to claim 1, characterized in that: The S7 specifically includes: S71, in the target matching subgraph In the graph, define each node The response state energy function is: ； in, Representing graph nodes The response state energy value, Representing graph nodes pH value, Representing graph nodes The potential value, Representing graph nodes The fluorescence signal value, 、 and represents the weight coefficient; S72, for directed edges , define the inverse cost function as: ； in, Represents a connection graph node To graph node Directed edges The reverse cost value, Representing graph nodes The associated process parameter vector, Representing graph nodes The associated process parameter vector, represents the L1 norm, represents the process disturbance penalty factor; S73, in the target matching subgraph Execute asynchronous reverse graph path traversal algorithm in the response trajectory to terminate the node To the starting node Perform reverse path traversal to search for the response path that satisfies the minimum reverse cost ; S74. During the reverse path traversal, if there is an edge The reverse cost value Greater than the preset jump threshold , then the reverse jump mechanism is triggered, and the jump candidate nodes that meet the reverse jump conditions are searched in the target matching subgraph. : Condition 1: Node Timestamp Less than the current node Timestamp ,Right now ; Condition 2: Node To the current node The reverse side of The reverse cost value Less than or equal to the jump threshold ,Right now ; Condition 3: Time difference Less than the maximum hop delay window ,Right now ; Among all the candidate jump nodes that meet the three conditions at the same time, select the node that minimizes the reverse cost. , and the edge Incorporate response pathways; S75. Response path with minimum reverse cost The process parameter vectors corresponding to all nodes above are interpolated by spline to generate a new round of process parameter groups; S76: Input the new round of process parameter group as a control instruction into the medicine preparation module to replace the initial process parameter group to execute medicine sample preparation.

8. The AI-based one-stop unmanned drug preparation and dosing method according to claim 1, characterized in that: The matching error between the target matching subgraph and the response trajectory graph specifically includes spatially matching the nodes in the target matching subgraph with the corresponding nodes in the response trajectory graph, calculating the difference value between each pair of nodes using Euclidean distance, and performing weighted summation on all the difference values.

9. An AI-based one-stop unmanned pharmaceutical preparation and dosing system, which implements the AI-based one-stop unmanned pharmaceutical preparation and dosing method according to any one of claims 1 to 8, characterized in that: Includes the following modules: Prescription parsing module, used to extract drug structural parameters, dosage ratio, target organelle type and target pharmacodynamic response trajectory, and generate drug-organelle mapping data; The biomimetic reaction chamber module is used to construct a biomimetic organelle reaction chamber based on drug-organelle mapping data and collect pH, potential and fluorescence response data; The pharmaceutical preparation module is used to complete the raw material addition, mixing, reaction and stabilization sub-processes according to the process parameter group and output the pharmaceutical sample; Response mapping module, used to construct response trajectory sequence and response trajectory graph, and generate target matching subgraph; The path optimization module is used to execute the asynchronous reverse graph path traversal algorithm in the target matching subgraph to obtain the response path with the minimum reverse cost and generate a new round of process parameter groups; The iterative control module is used to evaluate the matching error and determine whether to proceed to the next round of drug preparation.

Images