Intelligent anesthetic injection control system

Abstract

The invention relates to the field of intelligent control, and particularly discloses an anesthetic injection intelligent control system which predicts and generates personalized initial injection parameters by deeply integrating multi-source heterogeneous data such as patient individual information, surgical sites and anesthetic types. Through high-frequency collection of physiological and process signals such as tissue pressure and injection displacement, the system can accurately calculate the current actual pressure, the pressure change trend, the real-time flow rate and the total injected amount. The real-time monitoring data is continuously compared and evaluated with the generated personalized threshold value and the preset target so as to intelligently judge whether the current injection is smooth or not, whether abnormal resistance exists or not or whether the current injection is close to the target or not, and a mark reflecting the current injection state is generated according to the judgment result; and finally, the system dynamically adjusts the output of the injection pump according to the state flag in combination with a preset control logic and an optimization algorithm, and adaptively corrects the current target flow velocity.

Description

Technical Field

[0001] The present application relates to the field of intelligent control, and more specifically, to an intelligent control system for anesthetic injection. Background Art

[0002] Anesthesia is an indispensable and key component of modern medicine. The accurate and safe injection of anesthetics is crucial to ensuring the success of surgery, alleviating patient pain, and avoiding complications. Traditional anesthetic injection methods often rely on the experience of medical staff or simple constant-rate injection pumps. This method cannot fully take into account the individual differences of each patient, the specific conditions of the surgical site, and the characteristics of the anesthetic itself. Different patients have different physiological states, and the absorption and distribution of the same anesthetic in different tissues or parts also vary. Manual or constant-rate injection cannot be fine-tuned according to these real-time changes, which may result in insufficient anesthetic dosage affecting the anesthetic effect, or excessive dosage increasing the risk of toxic side effects and complications. In addition, changes in tissue pressure during the injection process are important indicators for assessing whether the injection is smooth and whether there is resistance or leakage. Traditional injection methods often lack real-time monitoring and intelligent response to tissue pressure. Once pressure abnormalities occur, it is difficult to automatically adjust the injection strategy, which may cause local tissue damage or injection failure.

[0003] While some automated anesthetic delivery devices are currently available, such as syringe pumps based on set fixed rates or simple preset programs, most of these devices remain at the level of executing preset instructions and lack true intelligence and adaptive capabilities. Some more advanced systems may incorporate basic pressure monitoring capabilities, which can pause or slow down when pressure exceeds a fixed threshold. However, these systems are still based on static, non-personalized rules and are unable to make dynamic, optimal adjustments based on the patient's specific condition and real-time tissue response.

[0004] Therefore, an optimized intelligent control system for anesthetic injection is desired. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems, the present application is proposed. The embodiment of the present application provides an intelligent control system for anesthetic push injection, which predicts and generates personalized initial push injection parameters by deeply integrating multi-source heterogeneous data such as patient individual information, surgical site and anesthetic type. During the actual push injection process, the system can accurately calculate the current actual pressure, pressure change trend, real-time flow rate and total amount injected by high-frequency acquisition of physiological and process signals such as tissue pressure and push injection displacement. These real-time monitoring data will be continuously compared and evaluated with the personalized thresholds and preset targets generated above, so as to intelligently judge whether the current push injection is smooth, whether there is abnormal resistance or whether it is close to the target, and generate a sign reflecting the current injection status accordingly. Finally, the system dynamically adjusts the output of the push pump based on this status sign, combined with the preset control logic and optimization algorithm, and adaptively corrects the current target flow rate. This closed-loop strategy from personalized initial setting to real-time intelligent decision-making and adaptive adjustment ensures the safety of the anesthetic push injection process.

[0006] According to one aspect of the present application, an intelligent control system for anesthetic injection is provided, comprising:

[0007] A data input module is used to obtain basic patient information, surgical site, and anesthetic type input by the operator through the user interface;

[0008] An initial injection parameter generation module is used to input basic patient information, surgical site, and anesthetic type into a machine learning model to obtain personalized initial injection parameters. The personalized initial injection parameters include a recommended initial flow rate, a comfortable pressure threshold, a safe pressure threshold, and a target injection volume.

[0009] a bolus injection data acquisition module, configured to, after issuing a start injection instruction, cause the controller to start the bolus injection pump motor to use the recommended initial flow rate as the current target flow rate to start bolus injection, and to collect pressure signals and displacement signals at high frequency through the data acquisition module during the bolus injection process;

[0010] An injection parameter calculation module is used to calculate the current pressure, pressure change rate, real-time flow rate and cumulative injection volume based on the collected pressure signal and displacement signal;

[0011] an injection status recognition module, configured to compare the current pressure with a comfortable pressure threshold and a safe pressure threshold, compare the pressure change rate with a preset gradient threshold, and compare the accumulated injected volume with a target total injection volume, so as to generate an injection status flag;

[0012] The target flow rate update control module is used to make an adaptive flow rate decision based on the injection status flag to obtain a new target flow rate.

[0013] Compared with the existing technology, the present application provides an intelligent control system for anesthetic push injection, which predicts and generates personalized initial push injection parameters by deeply integrating multi-source heterogeneous data such as patient individual information, surgical site and anesthetic type. During the actual push injection process, by high-frequency acquisition of physiological and process signals such as tissue pressure and push injection displacement, the system can accurately calculate the current actual pressure, pressure change trend, real-time flow rate and total amount injected. These real-time monitoring data will be continuously compared and evaluated with the personalized thresholds and preset targets generated above, so as to intelligently judge whether the current push injection is smooth, whether there is abnormal resistance or whether it is close to the target, and generate a sign reflecting the current injection status accordingly. Finally, based on this status sign, the system dynamically adjusts the output of the push pump and adaptively corrects the current target flow rate in combination with the preset control logic and optimization algorithm. This closed-loop strategy from personalized initial setting to real-time intelligent decision-making and adaptive adjustment ensures the safety of the anesthetic push injection process and solves the key pain points of the existing technology in terms of insufficient personalization and lack of real-time adaptive ability to make intelligent decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0015] Figure 1 4 is a system block diagram of an intelligent anesthetic injection control system according to an embodiment of the present application.

[0016] Figure 2 Schematic diagram of data flow of an intelligent control system for anesthetic injection according to an embodiment of the present application.

[0017] Figure 3 This is a block diagram of an initial injection parameter generation module in an intelligent anesthetic injection control system according to an embodiment of the present application.

[0018] Figure 4 This is a block diagram of a unit for strengthening local semantic associations between information in an intelligent control system for anesthetic injection according to an embodiment of the present application. DETAILED DESCRIPTION

[0019] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described herein.

[0020] As used in this application and the claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not intended to refer to the singular but may include the plural. Generally speaking, the terms "comprises" and "include" only indicate the inclusion of the steps and elements specifically identified, and these steps and elements do not constitute an exclusive list. A method or apparatus may also include other steps or elements.

[0021] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are only illustrative, and different aspects of the system and method can use different modules.

[0022] Flowcharts are used in this application to illustrate the operations performed by the systems according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the various steps may be processed in reverse order or simultaneously, as needed. Furthermore, other operations may be added to these processes, or one or more operations may be removed from these processes.

[0023] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described herein.

[0024] It should be understandable that, because traditional methods rely on the experience of medical staff or simple constant-speed push injections, it is difficult to accurately set initial parameters based on individual patient differences (such as physiological status), specific surgical conditions (such as site), and the characteristics of the anesthetic itself. During the push injection process, it is impossible to perform dynamic feedback and intelligent regulation based on changes in real-time physiological signals such as tissue pressure, and it is difficult to cope with changes in push resistance or abnormal conditions, and it is impossible to optimize the push injection strategy in real time. Even for some automated equipment, its control logic is mostly based on static, non-personalized rules, lacking true intelligent decision-making capabilities, and unable to effectively utilize multi-source complex information (such as patient information, surgical information, real-time feedback) for dynamic and optimal adjustments, making it difficult to ensure the safety, efficiency, and patient comfort of the push injection process.

[0025] To address the aforementioned technical issues, the technical solution of this application constructs a closed-loop system capable of intelligent, personalized, and adaptive control of the entire anesthetic bolus infusion process, namely, an intelligent anesthetic bolus infusion control system. First, to overcome the shortcomings of traditional methods, which rely on experience and lack personalized initial settings, this solution introduces a machine learning model. This model fully integrates key data such as the patient's individual information, surgical site, and the specific type of anesthetic before the bolus infusion begins. Beyond simple data input, the model utilizes a series of advanced data processing and feature engineering techniques to deeply explore the complex connections and interaction patterns hidden among this heterogeneous information. By fine-tuning the encoding and association modeling of this multi-source information, particularly capturing and strengthening the local semantic connections between the information, the system accurately predicts and outputs a set of highly personalized initial bolus parameters, including a recommended starting flow rate, a comfortable pressure threshold adapted to the patient's physiological state, and an upper pressure limit to ensure safety. These personalized parameters lay a solid foundation for subsequent precise bolus infusion, ensuring that the bolus infusion process is tailored to the patient's actual needs from the outset.

[0026] During the actual bolus injection process, the system goes beyond simply executing initial settings and instead incorporates a real-time feedback and control mechanism. By acquiring real-time physiological and process signals, such as tissue pressure and bolus displacement, at high frequency, the system accurately calculates the current bolus status, including actual pressure, pressure trends, real-time flow rate, and total volume injected. This real-time monitoring data is continuously compared with pre-set, personalized thresholds, particularly those generated by an initial machine learning model. Based on this real-time status assessment, the system intelligently determines whether the bolus is proceeding smoothly, encountering abnormal resistance, or approaching the target. Based on this decision, it generates a marker reflecting the current injection status. Ultimately, based on this injection status marker, the system, combined with pre-set control logic and optimization algorithms, dynamically adjusts the bolus pump output and adaptively adjusts the current target flow rate. This closed-loop control strategy, from personalized initial settings to real-time feedback, intelligent decision-making, and adaptive adjustment, ensures that the anesthetic bolus injection process remains safe, effective, and comfortable, significantly improving patient anesthesia safety. This addresses the technical limitations of existing technologies, such as insufficient personalization, lack of real-time adaptive capabilities, and inability to effectively utilize complex multi-source information for intelligent decision-making.

[0027] In the technical solution of the present application, an intelligent control system for anesthetic injection is proposed. Figure 1 4 is a system block diagram of an intelligent anesthetic injection control system according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of the intelligent control system for anesthetic injection according to the embodiment of the present application. Figure 1 and Figure 2As shown, the anesthetic bolus injection intelligent control system 100 according to the embodiment of the present application includes: a data input module 110, which is used to obtain the patient's basic information, surgical site and anesthetic type input by the operator through the user interface; an initial injection parameter generation module 120, which is used to input the patient's basic information, surgical site and anesthetic type into the machine learning model to obtain personalized initial injection parameters, and the personalized initial injection parameters include a recommended initial flow rate, a comfortable pressure threshold, a safe pressure threshold and a target injection volume; a bolus injection data acquisition module 130, which is used to start the bolus injection pump motor after issuing the start injection instruction to use the recommended initial flow rate as the current target flow rate The push injection is started, and during the push injection process, the pressure signal and the displacement signal are collected at high frequency through the data acquisition module; the injection parameter calculation module 140 is used to calculate the current pressure, pressure change rate, real-time flow rate and cumulative injected volume based on the collected pressure signal and displacement signal; the injection state recognition module 150 is used to compare the current pressure with the comfort pressure threshold and the safety pressure threshold, compare the pressure change rate with the preset gradient threshold, and compare the cumulative injected volume with the target total injection volume to generate an injection state flag; the target flow rate update control module 160 is used to make an adaptive flow rate decision based on the injection state flag to obtain a new target flow rate.

[0028] In the aforementioned intelligent anesthetic bolus control system 100, the data input module 110 is used to obtain basic patient information, surgical site, and anesthetic type input by the operator through the user interface. It should be understood that these three types of information, namely, basic patient information, surgical site, and anesthetic type, collectively constitute a set of key variables that influence anesthetic demand, metabolism, distribution, and patient response. These are essential inputs for subsequent machine learning models to perform accurate predictions and generate personalized parameters.

[0029] Specifically, in a specific example of the present application, first, the system provides a structured user input interface, which is usually integrated into the anesthesia workstation or an independent control device display screen. During the injection preparation stage, the operator (such as an anesthesiologist or nurse) will enter information according to the interface guidance. For the patient's basic information, the interface will set corresponding fields, such as for entering age, weight, gender, height, and important past medical history, allergy history, main diagnosis and other text or options. This information can be entered manually through the keyboard, or selected through standardized controls such as drop-down menus and check boxes to ensure the standardization and consistency of the data. Some advanced systems may allow automatic retrieval of some existing patient basic information from the hospital information system (HIS) or electronic medical record system (EMR) by scanning the patient's wristband barcode or NFC recognition, and the operator will then check and supplement it. This can be regarded as an automated information acquisition auxiliary means based on network and database interaction.

[0030] Secondly, for surgical site information, the user interface typically provides a preset list of surgical sites that conform to clinical classification standards for the operator to select, such as through a multi-level linked menu that accurately identifies specific anatomical regions. For certain special or uncommon sites, an "Other" option may also be provided, accompanied by a text box for manual description. In more advanced systems, surgical site selection can even be linked with the surgical scheduling system or image navigation system to automatically obtain some information or assist in selection.

[0031] Again, regarding the type of anesthetic, the interface will provide a list of anesthetic types currently approved for use by medical institutions, from which the operator needs to select the specific anesthetic name to be used in this operation. In addition to the name of the anesthetic, you may also need to enter or select relevant parameters such as the concentration and specifications of the anesthetic. This information is crucial for calculating the total amount of injection and flow rate. Similarly, this link can also be integrated with the pharmacy management system or smart medicine cabinet system to automatically enter information such as anesthetic type and batch number by scanning the drug barcode to ensure accuracy and simplify operations.

[0032] In the above-mentioned intelligent anesthetic bolus control system 100, the initial injection parameter generation module 120 is used to input the patient's basic information, surgical site, and anesthetic type into a machine learning model to obtain personalized initial injection parameters. The personalized initial injection parameters include a recommended initial flow rate, a comfortable pressure threshold, a safe pressure threshold, and a target injection volume. It should be understood that since the three types of information, namely, the patient's basic information, the surgical site, and the anesthetic type, are the most critical factors in determining anesthetic demand and bolus characteristics, there are complex and nonlinear relationships between them that are difficult to express with simple rules. The machine learning model can learn these complex patterns from a large amount of historical data, thereby generating an initial bolus recommendation that best suits each patient's individual situation, laying a personalized foundation for subsequent real-time adaptive control. Therefore, the patient's basic information, surgical site, and anesthetic type are input into the machine learning model to obtain personalized initial injection parameters.

[0033] Figure 3 FIG. 1 is a block diagram of a carbon emission activity data feature extraction module in an anesthetic injection intelligent control system according to an embodiment of the present application. Figure 3As shown, in an embodiment of the present application, the initial injection parameter generation module 120 includes: a patient information encoding unit 121, which is used to perform semantic embedding encoding on the patient's basic information to obtain a patient's basic information semantic embedding encoding vector; a surgical information encoding unit 122, which is used to perform one-hot encoding on the surgical site and the anesthetic type to obtain a surgical site one-hot embedding encoding vector and an anesthetic type one-hot embedding encoding vector; an information association transfer unit 123, which is used to calculate the basic information-surgical site sparse association transfer matrix and the basic information-anesthetic type sparse transfer association matrix based on the patient's basic information semantic embedding encoding vector, the surgical site one-hot embedding encoding vector and the anesthetic type one-hot embedding encoding vector; an information local semantic association encoding unit 124, which is used to respectively calculate the basic information-surgical site sparse association transfer matrix and the basic information-anesthetic type sparse transfer association matrix from the basic information-surgical site sparse association transfer matrix and the basic information-anesthetic type sparse transfer association matrix. The information-anesthetic type sparse transfer association matrix extracts semantic features to obtain the basic information-surgical site local semantic association coding vector and the basic information-anesthetic type local semantic association coding vector; the information local semantic association enhancement unit 125 is used to perform local semantic association enhancement on the basic information-surgical site local semantic association coding vector and the basic information-anesthetic type local semantic association coding vector to obtain an enhanced basic information-surgical site local semantic association coding vector and an enhanced basic information-anesthetic type local semantic association coding vector; the feature decoding and regression unit 126 is used to perform feature decoding and regression on the fused features after performing feature fusion on the enhanced basic information-surgical site local semantic association coding vector and the enhanced basic information-anesthetic type local semantic association coding vector to obtain personalized initial injection parameters. In the above-mentioned anesthetic injection intelligent control system 100, the patient information encoding unit 121 and the surgical information encoding unit 122 are used to perform semantic embedding encoding on the patient's basic information to obtain the patient's basic information semantic embedding coding vector, and perform one-hot encoding on the surgical site and anesthetic type to obtain a one-hot embedding coding vector for the surgical site and a one-hot embedding coding vector for the anesthetic type. It should be understood that, considering that the basic information of patients often contains rich descriptive texts and numerical values ​​with inherent semantic associations, the surgical site and anesthetic type are finite, discrete categorical variables. Due to the different nature of these two types of data, different encoding strategies need to be adopted to maximize the retention and expression of their respective characteristics. Therefore, in the technical solution of the present application, the basic information of patients is semantically embedded to obtain a semantic embedding encoding vector of the basic information of patients, and the surgical site and anesthetic type are one-hot encoded to obtain a one-hot embedding encoding vector of the surgical site and a one-hot embedding encoding vector of the anesthetic type. By adopting different encoding methods, the original input information of different types and structures such as the basic information of patients and the surgical site and anesthetic type are uniformly converted into a high-dimensional numerical vector representation, making it a format that can be processed by the machine learning model. Specifically, the semantic embedding encoding of the basic information of patients is to capture and quantify the complex semantic relationships and potential patterns contained in the patient information.For example, through semantic embedding, similar disease descriptions or combinations of signs can be closer in the vector space, allowing the model to understand the subtle differences in patient status and learn patterns related to injection parameters from these subtle differences. This encoding method aims to compress unstructured or semi-structured patient information into a dense, semantically rich vector to provide more insightful input for personalized decision-making. The purpose of one-hot encoding of surgical sites and anesthetic types is to represent these discrete categorical information in a simple, clear and unbiased manner. One-hot encoding creates a unique binary vector for each category, ensuring that the model treats different surgical sites or anesthetic types as independent entities, avoiding the incorrect size or order relationships that may be introduced by simple numerical encoding, and effectively converting categorical information into discrete features suitable for model processing.

[0034] Specifically, basic patient information (such as age, weight, medical history, and physiological indicators) often contains a rich mix of numerical data and textual data with complex internal logic and context. Simple numerical or isolated feature representations fail to capture the deeper semantic connections and patterns underlying this information, such as the potential association between a history of hypertension and increased cardiovascular risk, or the comprehensive health status indicated by a specific combination of physiological indicators. Semantic embedding techniques, particularly neural network-based models (such as Word2Vec, GloVe, and FastText, or more advanced Transformer architectures such as BERT and its medical variants BioBERT and ClinicalBERT), can map this high-dimensional, sparse, and even unstructured patient information into a low-dimensional, dense vector space. In this space, semantically similar patient features or conditions are more closely represented in their vectors. A specific implementation involves the following steps: First, preprocess the patient's basic information. Textual information may require tokenization, stop word removal, and lemmatization, while numerical information may require normalization or standardization. Next, the words in the text are converted into word vectors. Pooling (e.g., average pooling, max pooling) or more complex sequence encoders (e.g., LSTM, GRU, or Transformer encoders) are used to generate a semantic embedding vector for the entire patient information. This vector automatically extracts and generates a semantic embedding vector for the patient's basic information that captures key semantic features. This vector is designed to summarize the patient's comprehensive physiological status and potential risk factors for model reference.

[0035] Furthermore, surgical site and anesthetic type are inherently discrete, finite categorical variables. For example, a surgical site might be a specific category such as "maxillary sinus" or "near the mandibular canal," while an anesthetic type might be a specific type such as "lidocaine" or "articaine." These categories typically do not have an inherent ordinal relationship (i.e., "maxillary sinus" is not numerically larger or smaller than "near the mandibular canal"). Using simple integer encoding (e.g., 0, 1, 2, etc.) can lead machine learning models to misinterpret an ordinal relationship between these values, introducing bias. One-hot encoding effectively avoids this problem. The specific implementation involves first counting all possible surgical site categories and all possible anesthetic type categories. Then, for each specific surgical site instance, it is represented as a binary vector with a length equal to the total number of surgical site categories, where the category position corresponding to the instance is 1 and all other positions are 0. Similarly, the anesthetic type is processed similarly to generate its one-hot embedding encoding vector. For example, if there are three surgical site categories, A, B, and C, the one-hot encoding for site A is [1, 0, 0], site B is [0, 1, 0], and site C is [0, 0, 1]. This encoding ensures that each category is treated as an independent, orthogonal feature, clearly indicating to the model that they are mutually exclusive classification options with no implicit order or size relationship. While one-hot encoding may increase feature dimensionality, its simplicity, intuitiveness, and unbiased nature make it a common and effective method for handling such categorical variables.

[0036] In the aforementioned intelligent anesthetic injection control system 100, the information-to-information association transfer unit 123 is used to calculate the basic information-surgical site sparse association transfer matrix and the basic information-anesthetic type sparse transfer association matrix based on the patient's basic information semantic embedding encoding vector, the surgical site one-hot embedding encoding vector, and the anesthetic type one-hot embedding encoding vector. It should be understood that although the patient's basic information has been encoded as the patient's basic information semantic embedding encoding vector, and the surgical site and anesthetic type have been encoded as the surgical site one-hot embedding encoding vector and the anesthetic type one-hot embedding encoding vector, this merely converts different types of data into numerical formats and does not fully capture the potential and complex interactions between these heterogeneous information. Simply concatenating their encoding features is insufficient to reveal the complex interaction patterns between them, nor is it sufficient to reveal the deep conditional dependencies such as how a patient's specific condition affects their response to or need for a particular anesthetic at a specific surgical site. For example, a patient with a specific history of cardiovascular disease will have a physiological state (represented by a semantic embedding vector) that will have a different impact on the metabolism and distribution of a particular anesthetic at a well-supplied surgical site (represented by a one-hot vector) than an average patient. This effect is not a simple addition of patient information or site / anesthetic information itself, but rather a "transfer" or "association" effect. Therefore, a sparse association transfer matrix of basic information-surgical site and a sparse association transfer matrix of basic information-anesthetic type are calculated to explicitly model and quantify the strength or pattern of association between this patient state and specific site / anesthetic category.

[0037] More specifically, in an embodiment of the present application, the information association transfer unit is used to: calculate the transfer matrix of the patient's basic information semantic embedding coding vector relative to the surgical site one-hot embedding coding vector to obtain a basic information-surgical site sparse association transfer matrix; calculate the transfer matrix of the patient's basic information semantic embedding coding vector relative to the anesthetic type one-hot embedding coding vector to obtain a basic information-anesthetic type sparse transfer association matrix.

[0038] In the aforementioned intelligent anesthetic injection control system 100, the information-local semantic association encoding unit 124 is used to extract semantic features from the basic information-surgical site sparse association transfer matrix and the basic information-anesthetic type sparse association transfer matrix, respectively, to obtain a basic information-surgical site local semantic association encoding vector and a basic information-anesthetic type local semantic association encoding vector. It should be understood that while the basic information-surgical site and basic information-anesthetic type sparse association transfer matrices provide a preliminary quantification of the strength of associations between individual patient characteristics and surgical site and anesthetic type, respectively, these matrices may be high-dimensional and still contain raw, fine-grained association information. More importantly, the true determinant for predicting initial injection parameters is often not the association between any two single features, but rather the "joint association pattern" or "local semantic association" within a specific "local" region, such as the "joint association pattern" or "local semantic association" between a patient's specific symptoms or signs and a specific surgical site or group of related surgical sites, or a specific anesthetic drug or group of related drugs. Simply using the entire matrix as input may introduce noise, and the model will have difficulty automatically capturing these key local patterns. Text convolutional neural networks (TextCNNs) are naturally good at extracting local invariance and combinational features from sequence or similar two-dimensional structure inputs (similar to capturing n-gram features in text), and are therefore very suitable for scanning and identifying important local patterns or "association templates" in these association matrices. Therefore, in the technical solution of the present application, the basic information-surgical site sparse association transfer matrix and the basic information-anesthetic type sparse transfer association matrix are further input into the text convolutional neural network model to obtain the basic information-surgical site local semantic association coding vector and the basic information-anesthetic type local semantic association coding vector. By utilizing the local feature extraction capability of the text convolutional neural network model, higher-level, discriminative local semantic association features are learned and abstracted from the structured association transfer matrix. By sliding convolution kernels of different sizes on the transfer matrix, the network can capture the association strength pattern between certain specific combination features in the patient's basic information and certain specific categories in the surgical site or anesthetic type, such as the interaction pattern of a certain pharmacological component during anesthesia of a certain anatomical region in patients with a specific disease history. Through convolution and pooling operations, these local patterns are integrated and compressed into more compact feature representations: the basic information-surgical site local semantic association encoding vector and the basic information-anesthetic type local semantic association encoding vector. These vectors are designed to concentrate the most predictive local interaction information in the transfer matrix and filter out redundant or irrelevant associations, thereby forming a more refined and expressive feature representation.In particular, in an embodiment of the present application, the information local semantic association encoding unit is used to: input the basic information-surgical site sparse association transfer matrix and the basic information-anesthetic type sparse transfer association matrix into the text convolutional neural network model respectively to obtain the basic information-surgical site local semantic association encoding vector and the basic information-anesthetic type local semantic association encoding vector.

[0039] In the aforementioned intelligent anesthetic injection control system 100, the information local semantic association enhancement unit 125 is configured to enhance the local semantic associations of the basic information-surgical site local semantic association encoding vector and the basic information-anesthetic type local semantic association encoding vector to obtain enhanced basic information-surgical site local semantic association encoding vector and enhanced basic information-anesthetic type local semantic association encoding vector. It should be understood that although the previous step extracts preliminary local semantic association encoding vectors from the basic information-surgical site sparse association transfer matrix and the basic information-anesthetic type sparse transfer association matrix using a text convolutional neural network, and these vectors capture certain local interaction patterns between individual patient information and surgical site or anesthetic type, these preliminary extracted features may still be insufficiently refined, may contain redundant information, or fail to fully reveal the deep, complex internal structure and dependencies between these local associations. In particular, given the complexity of real-world clinical data and the nonlinear nature of associations, local features derived from simple stacking or preliminary convolution may not capture the most predictive, higher-level, context-aware association information. There is a need to deeply structure and refine these preliminary local association features in order to overcome the possible limitations of the original local features in terms of expressive power, so that they can more accurately reflect the key information that affects the personalized parameters of anesthetic injection. Based on this, in the technical solution of the present application, the basic information-surgical site local semantic association coding vector and the basic information-anesthetic type local semantic association coding vector are further subjected to local semantic association enhancement to obtain enhanced basic information-surgical site local semantic association coding vector and enhanced basic information-anesthetic type local semantic association coding vector. Specifically, the processing method of local semantic association enhancement aims to parse the original basic information-surgical site local semantic association coding vector and the basic information-anesthetic type local semantic association coding vector into more basic local components through feature decomposition, and to explicitly model and refine the complex relationship between the local components of these basic information-surgical site local semantics and basic information-anesthetic type local semantics through feature distillation, filter out noise, and enhance the expression of key information. Ultimately, feature reconstruction intelligently aggregates this refined local information into an enhanced basic information-surgical site local semantic association encoding vector and an enhanced basic information-anesthesia type local semantic association encoding vector that comprehensively and powerfully represent the deep associations between the individual patient and the surgical site and anesthetic type. The core goal of this process is to improve the discriminative power and robustness of the features, making them more suitable for downstream personalized parameter regression prediction.

[0040] Figure 4 FIG. 1 is a block diagram of a characteristic expression unit for carbon emission activity data to be calculated in an anesthetic injection intelligent control system according to an embodiment of the present application. Figure 4As shown, in an embodiment of the present application, the information local semantic association enhancement unit 125 includes: a basic information-surgical site semantic deep understanding sub-unit 1251, which is used to perform a structured decomposition of the basic information-surgical site local semantic association coding vector based on one-dimensional convolutional coding to decompose the basic information-surgical site local semantic association coding vector into a set of basic information-surgical site local semantic initial local implicit feature vectors; a basic information-surgical site semantic association analysis sub-unit 1252, which is used to generate a basic information-surgical site local semantic geometric topology fine-grained association mask matrix based on the geometric affinity between any two basic information-surgical site local semantic initial local implicit feature vectors in the set of basic information-surgical site local semantic initial local implicit feature vectors; a basic information-surgical site association semantic feedback reconstruction sub-unit 1253, which is used to interactively refine the set of basic information-surgical site local semantic initial local implicit feature vectors based on the basic information-surgical site local semantic geometric topology fine-grained association mask matrix to obtain an enhanced basic information-surgical site local semantic association coding vector.

[0041] Specifically, the basic information-surgical site semantic deep understanding subunit 1251 is used to perform a structured decomposition based on one-dimensional convolutional coding on the basic information-surgical site local semantic association coding vector to decompose the basic information-surgical site local semantic association coding vector into a set of basic information-surgical site local semantic initial local implicit feature vectors, which can be expressed as follows:

[0042] Conv l×1 (X)={h1,h2,...,h i ,...,h n}

[0043] Among them, X is the basic information-surgical site local semantic association encoding vector, Conv l×1 (·) is a one-dimensional convolutional coding process with a step size of l, h1,h2,h i ,h j ,h n They are respectively the 1st, 2nd, i-th, j-th and n-th basic information-surgical site local semantics initial local latent feature vectors in the set of basic information-surgical site local semantics initial local latent feature vectors.

[0044] It should be understood that although the basic information-surgical site local semantic association encoding vector mentioned above has captured the local interaction pattern between the patient's individual information and the surgical site, this single encoding vector itself may still be a relatively macroscopic complex containing multiple fine-grained semantic components. In order to conduct deeper structural processing and refinement to reveal the more basic structural patterns within these local associations and the complex dependencies between them, it is necessary to first "disassemble" this overall encoding vector and make explicit the multiple semantic sub-patterns hidden within it that may be intertwined with each other. Specifically, in the technical solution of the present application, the "basic information-surgical site local semantic association encoding vector" is scanned and parsed through the sliding window and multi-convolution kernel mechanism of a one-dimensional convolutional neural network (1DCNN). 1DCNN is good at capturing local patterns in sequence data or one-dimensional vectors, and each convolution kernel can focus on extracting a specific type of local structural information. In this way, the purpose is to refine and deconstruct the original, relatively concentrated basic information-surgical site local semantic association encoding vector into a set of more elementary and more specific basic information-surgical site local semantic initial local implicit feature vectors. Each such initial local latent feature vector of basic information-surgical site local semantics can be viewed as a representation of a smaller-scale, more specific semantic segment within the basic information-surgical site local semantics association encoding vector. This decomposition process aims to transform the original features from a holistic representation into a distributed, structured representation consisting of multiple local feature bases, thereby decoupling the multiple potential patterns hidden within the overall association information (for example, the risk signal associated with a specific patient sign and a specific anatomical structure at the surgical site) and making them explicit as a set of local feature units that can be further analyzed independently and interactively modeled.

[0045] Specifically, the basic information-surgical site semantic association analysis subunit 1252 is used to generate a basic information-surgical site local semantic geometric topology fine-grained association mask matrix based on the geometric affinity between any two basic information-surgical site local semantic initial local implicit feature vectors in the set of basic information-surgical site local semantic initial local implicit feature vectors, including: calculating the geometric affinity between any two basic information-surgical site local semantic initial local implicit feature vectors in the set of basic information-surgical site local semantic initial local implicit feature vectors to obtain a basic information-surgical site local semantic geometric association topology matrix composed of multiple basic information-surgical site local semantic geometric affinities; inputting the basic information-surgical site local semantic geometric association topology matrix into a gated mask function to obtain a basic information-surgical site local semantic geometric topology fine-grained association mask matrix.

[0046] Specifically, the geometric affinity between any two basic information-surgical site local semantic initial local implicit feature vectors in the set of basic information-surgical site local semantic initial local implicit feature vectors is calculated to obtain a basic information-surgical site local semantic geometric association topology matrix composed of multiple basic information-surgical site local semantic geometric affinities, which can be expressed as follows:

[0047]

[0048] Among them, ‖·‖ represents the norm of the vector, γ is the trainable association weight, s ij h i and h j The basic information-surgical site local semantic geometric affinity between them is the eigenvalue of the (i, j) position in the basic information-surgical site local semantic geometric association topology matrix.

[0049] It should be understood that while each vector in the set of initial local latent feature vectors of the basic information-surgical site local semantics represents a local semantic segment within the original basic information-surgical site associated semantics, they do not exist in isolation but rather collectively form a complex semantic network. These local semantic segments, for example, the local feature representing "advanced patient age" and the local feature representing "rich surgical site blood supply," may have inherent, non-explicit correlations or similarities. Simply possessing these discrete local features is insufficient; the system needs to understand how they interrelate and influence each other to more comprehensively grasp the overall anesthetic risk and needs. Therefore, by calculating the geometric affinity between any two initial local latent feature vectors of the basic information-surgical site local semantics, the system can capture the proximity or interdependence strength of these basic information-surgical site local semantic segments on the underlying feature manifold structure. Organizing all these pairwise geometric affinity values ​​forms the "basic information-surgical site local semantics geometric association topology matrix." This matrix is ​​like a map that clearly depicts the interconnection relationships and strengths between all decomposed local semantic features. It explicitly models the intrinsic structure and geometric topology between these local semantic information.

[0050] Specifically, the basic information-surgical site local semantic geometry association topology matrix is ​​input into the gated mask function to obtain the basic information-surgical site local semantic geometry topology fine-grained association mask matrix, which is expressed as follows:

[0051] M=Sigmoid(W g S+b g )

[0052] Among them, S is the basic information-surgical site local semantic geometry association topology matrix composed of multiple basic information-surgical site local semantic geometry affinities, M is the basic information-surgical site local semantic geometry topology fine-grained association mask matrix, W g is the gated mask weight matrix, b g is the gated mask bias matrix, and Sigmoid is the Sigmoid function.

[0053] It should be understood that although the previously constructed "basic information-surgical site local semantic geometric association topology matrix" explicitly represents the strength of associations between individual local semantic features, this raw basic information-surgical site local semantic geometric association topology matrix may contain noise (i.e., some accidental, non-realistic weak associations are calculated), or not all calculated associations are equally important for the final prediction of anesthesia parameters. For example, some associations between basic information-surgical site local semantic features may be redundant, or the importance of some associations may change in specific contexts (such as specific patient types or specific surgical stages). Directly using the raw basic information-surgical site local semantic geometric association topology matrix for subsequent feature distillation may introduce these noisy or unimportant associations, thereby affecting the quality of the final enhanced features and the decision accuracy of the model. It is worth noting that the gated mask function (such as simple gating based on sigmoid activation or more complex gating mechanisms such as gated recurrent units (GRUs) or long short-term memory networks (LSTMs)) acts as an "attention regulator." Its goal is to dynamically and nonlinearly learn and generate a "mask" based on the information in the association topology matrix itself or other relevant contextual information. This mask does not simply retain or discard associations, but rather fine-tunes and modulates the strength of each association in the original basic information-surgical site local semantic geometric association topology matrix. It can enhance those basic information-surgical site semantic associations that are considered more important and more discriminative, while suppressing those associations that may be noise, redundant, or less critical in the current context. In this way, the connections that are truly valuable for understanding the complex interactions between patients and surgical sites are highlighted, and potential interference information is filtered out. This refined fine-grained association mask matrix will serve as a more precise guide for subsequent feature distillation, allowing the transfer and fusion of information between local features to be more focused on key paths, thereby hopefully generating enhanced feature representations that are more expressive and more robust to noise, providing support for more accurate decision-making for the intelligent control system of anesthetic injection.

[0054] Specifically, the basic information-surgical site association semantics feedback reconstruction subunit 1253 is used to interactively refine the set of basic information-surgical site local semantics initial local implicit feature vectors based on the basic information-surgical site local semantics geometric topology fine-grained association mask matrix to obtain an enhanced basic information-surgical site local semantics association coding vector, including: based on the basic information-surgical site local semantics geometric topology fine-grained association mask matrix, performing neighborhood-aware distillation on each basic information-surgical site local semantics initial local implicit feature vector in the set of basic information-surgical site local semantics initial local implicit feature vectors to obtain a set of basic information-surgical site local semantics distillation initial local implicit feature vectors; performing aggregation reconstruction based on the self-attention mechanism on the set of basic information-surgical site local semantics distillation initial local implicit feature vectors to obtain an enhanced basic information-surgical site local semantics association coding vector.

[0055] Specifically, based on the basic information-surgical site local semantic geometric topology fine-grained association mask matrix, each basic information-surgical site local semantic initial local implicit feature vector in the set of basic information-surgical site local semantic initial local implicit feature vectors is subjected to neighborhood-aware distillation to obtain a set of basic information-surgical site local semantic distillation initial local implicit feature vectors, which can be expressed as follows:

[0056]

[0057] Among them, W d is the distillation weight matrix, L is h i The scale of , ReLU is the ReLU activation function, is matrix multiplication, ⊙ is point multiplication by position, are the 1st, 2nd, i-th and n-th basic information-surgical site local semantic distillation initial local implicit feature vectors in the set of basic information-surgical site local semantic distillation initial local implicit feature vectors, {·;·} is the vector cascade, Distill the implicit aggregated feature vector for the basic information - local semantics of the surgical site.

[0058] Understandably, these initial local latent feature vectors of basic information-surgical site semantics are still relatively independent, or in other words, they have not yet fully utilized the inter-correlation information to optimize their own representation. To improve the expressive quality of each basic information-surgical site semantic feature, better reflect its contextual meaning in a complex semantic network, eliminate potential ambiguity, or supplement missing information, an explicit process is required to promote information exchange and mutual refinement between them. Specifically, through neighborhood-aware distillation, for each basic information-surgical site semantics initial local latent feature vector in the set, the system utilizes the precise connectivity and correlation strength (weights) provided by the basic information-surgical site semantics geometric topology fine-grained association mask matrix to identify its important "neighboring" local features. Then, through weighted aggregation or more complex message passing mechanisms (such as the aggregation operation in graph neural networks), the information of these neighboring basic information-surgical site semantics (modulated by the mask matrix weights) is incorporated into the currently updated basic information-surgical site semantics initial local latent feature vector. This process is "neighborhood-aware" because it explicitly considers the local structure and mutual influence between features; it is also "distillation" because it aims to extract the essence from relevant neighborhood information, remove redundancy, and strengthen the core semantics, making the representation of each local feature purer and more powerful. For example, a local feature representing "patient's poor cardiac function" may absorb information from the strongly related local feature "surgical site is sensitive to circulatory perturbations" through neighborhood-aware distillation, making its own representation more comprehensive reflecting this joint risk.

[0059] Specifically, the set of initial local latent feature vectors of basic information-surgical site local semantic distillation is aggregated and reconstructed based on the self-attention mechanism to obtain the enhanced basic information-surgical site local semantic association encoding vector, which is expressed as follows:

[0060]

[0061] Among them, W q 、W k and W v are the basic information-surgical site local semantic query weight matrix, the basic information-surgical site local semantic key weight matrix, and the basic information-surgical site local semantic value weight matrix, respectively. Q, K, and V are the basic information-surgical site local semantic query vector, the basic information-surgical site local semantic key vector, and the basic information-surgical site local semantic value vector, respectively. Z is the scale of K, softmax is the softmax activation function, and v enhanced To enhance basic information - local semantic association encoding vector of surgical site.

[0062] It should be understood that although the initial local latent feature vectors from the distilled basic information-surgical site local semantics already contain rich contextual information and are more coordinated with each other, they are still multiple scattered, fine-grained semantic units. To drive subsequent anesthesia parameter prediction or decision-making, the system often requires a single, highly summarized, and comprehensive overall feature representation that can fully represent the complex interactive relationship between "basic patient information" and "surgical site." Simply averaging or concatenating these distilled local feature vectors may lose the dynamic, long-range dependencies between them, and it is also impossible to adaptively determine the contribution of each local feature to the final overall representation based on the context. Therefore, in the technical solution of this application, the set of initial local latent feature vectors from the distilled basic information-surgical site local semantics is further aggregated and reconstructed based on the self-attention mechanism. This means that based on the overall context formed by all current local features, the system can automatically determine which distilled local semantic segments are more critical to forming the most informative final overall representation and assign them greater influence. It aims to capture the complex global dependencies (including long-range dependencies) between these refined basic information-surgical site local details, rather than just interactions within the neighborhood. The ultimate goal is to intelligently and dynamically "weave" this deeply refined and context-aware local information into a single, unified, and more expressive "enhanced basic information-surgical site local semantic association encoding vector" with higher information density. This enhanced feature vector, meticulously reconstructed through the self-attention mechanism, will serve as the final and most critical input feature for the intelligent anesthetic bolus control system to make subsequent precise predictions (such as predicting personalized initial bolus parameters), significantly improving the overall system's intelligence and decision-making accuracy.

[0063] Preferably, here, the spatial correlation distribution of the basic information-local semantic geometric association topology matrix of the surgical site on the implicit low-dimensional manifold will show nonlinear saturation phenomenon, so that the overall topological correlation structure of the basic information-local semantic geometric association topology matrix of the surgical site will collapse due to the nonlinear coupling effect, and this collapse will be more significant due to the polarization enhancement effect of the gated mask function, affecting the intrinsic geometric structure representation ability of the basic information-local semantic geometric topology fine-grained association mask matrix of the surgical site.

[0064] Based on this, for each eigenvalue m of the basic information-local semantic geometry topology fine-grained association mask matrix M of the surgical site i,j ∈M, first generate the gradient field components:

[0065]

[0066] in, represents the calculation gradient, To calculate the partial derivative, m r,s is the eigenvalue of the (r, s) position of the basic information - the fine-grained association mask matrix M of the local semantic geometric topology of the surgical site. This is used to correct the heterogeneity of the local nonlinear spatial structure and thus achieve fine-grained control of the geometric association field.

[0067] Then, the gradient field term is used as the external field driving term to perform mean field control of each eigenvalue:

[0068]

[0069] Where μ is the characteristic mean of all eigenvalues ​​of the basic information-local semantic geometry topology fine-grained association mask matrix M of the surgical site, and ω is the normalization coefficient of the constrained strong driving factor, that is, to adjust the scaling weight of the excessively strong external driving field, m′ i,j It is the enhanced basic information - the eigenvalues ​​of each position in the fine-grained association mask matrix of the local semantic geometric topology of the surgical site.

[0070] In this way, under the action of the external field driving term as a high-order gradient, the nonlinear saturation of the geometric correlation distribution under the mean field is reversely driven, thereby compensating for the structural discretization effect caused by the correlation polarization enhancement by improving the statistical mean under the mean field, thereby improving the geometric structure representation fidelity of the basic information - the local semantic geometric topology fine-grained correlation mask matrix of the surgical site.

[0071] In the above-mentioned anesthetic injection intelligent control system 100, the feature decoding and regression unit 126 is used to perform feature decoding and regression on the fused features after the enhanced basic information-surgical site local semantic association coding vector and the enhanced basic information-anesthetic type local semantic association coding vector are feature fused to obtain personalized initial injection parameters. It should be understood that although the enhanced basic information-surgical site local semantic association coding vector and the enhanced basic information-anesthetic type local semantic association coding vector each deeply mine and characterize the complex local semantic associations between patient information and surgical site, and between patient information and anesthetic type, they are still relatively independent perspectives. However, the actual anesthetic injection decision, that is, the setting of personalized initial injection parameters (such as starting flow rate, comfortable pressure threshold, and safety pressure upper limit), must be the overall result of the interaction between the patient's characteristics, the characteristics of the surgical site, and the selected anesthetic properties. Simply relying on any one enhancement vector cannot fully reflect the complexity of this three-way interaction. Therefore, in order to obtain a feature representation that can fully represent the comprehensive situation of a patient receiving a specific anesthetic injection at a specific surgical site, it is necessary to first effectively integrate these two deep semantic information that focus on the "patient-site" and "patient-drug" interactions respectively.

[0072] Specifically, in a specific example of the present application, the feature decoding regression unit is used to: perform feature fusion on the enhanced basic information-surgical site local semantic association coding vector and the enhanced basic information-anesthetic type local semantic association coding vector to obtain the enhanced basic information-anesthetic surgical information local semantic fusion feature vector as a fusion feature; pass the enhanced basic information-anesthetic surgical information local semantic fusion feature vector through a decoder-based injection parameter controller to obtain personalized initial injection parameters. In particular, the fusion process (for example, by splicing, weighted summation, or more complex attention mechanism fusion) is intended to combine the deeply refined information contained in the two enhanced vectors about how patient specificity regulates its response pattern to the surgical site and anesthetic, respectively, to generate a new "enhanced basic information-anesthetic surgical information local semantic fusion feature vector". This fusion feature vector condenses the essence of all previously acquired and enhanced information. It is no longer a fragmented perspective, but a comprehensive, high-order semantic representation of all key influencing factors and their complex interactions in the complete scenario of "specific patient-specific surgical site-specific anesthetic". The effect is to produce a unified input feature with extremely high information density that can drive the prediction of the final parameters. It is worth mentioning that the "injection parameter controller" is essentially a decoder model that can perform regression tasks (such as a multi-layer perceptron, the decoding end of a recurrent neural network, or other specifically designed regression networks). It learns the complex nonlinear mapping relationship between the fused feature vector and the target injection parameters (multiple continuous values ​​such as the starting flow rate, comfortable pressure threshold, and safety pressure upper limit). This decoding regression process is an intelligent inference based on a large amount of historical data and clinical experience (obtained through model training and learning). It aims to accurately predict a set of initial injection parameters that are most suitable for the current patient, current surgery, and current anesthetic combination based on the unique situation reflected by the current fused features. These parameters overcome the drawbacks of traditional methods that rely on subjective experience or fixed settings, and fully consider the individual differences of patients, the specific needs of the surgery, and the pharmacological properties of anesthetics, thereby laying a solid and intelligent foundation for the subsequent safe, effective, and comfortable anesthetic injection process, and directly solve the core problems of insufficient personalization in the technical background and the lack of ability to comprehensively utilize complex information. In the above-mentioned anesthetic push injection intelligent control system 100, the push injection data acquisition module 130 is used to start the push injection pump motor after issuing the start injection instruction to recommend the initial flow rate as the current target flow rate to start the push injection, and collect the pressure signal and displacement signal at high frequency through the data acquisition module during the push injection process.It should be understandable that starting with the recommended initial flow rate is intended to put the results of intelligent analysis into practice, ensuring that the push injection is based on a comprehensive consideration of the patient, surgery, and drug characteristics from the very beginning; and high-frequency acquisition of pressure and displacement signals is the basis for building closed-loop feedback control, aiming to gain real-time insight into dynamic changes in the push injection process, providing immediate and accurate basis for subsequent intelligent adjustments, thereby making up for the lack of real-time monitoring and response capabilities in traditional methods.

[0073] Specifically, in a specific example of the present application, first, the controller receives an instruction to start the injection and a "recommended initial flow rate" parameter calculated by an upstream module. The controller, which typically has an embedded microprocessor or microcontroller unit, converts this numerical value representing the target flow rate into specific drive instructions for the injection pump motor (such as a precision stepper motor or DC servo motor). For example, if a syringe pump is driven by a stepper motor, the controller will calculate the stepping frequency or the number of steps per unit time that the motor should reach based on the recommended flow rate, and output a corresponding pulse sequence to the stepper motor through the motor drive circuit, causing it to rotate at a precisely controlled angular velocity, and then push the syringe piston through a screw-nut mechanism or similar transmission device to discharge the anesthetic at the set initial flow rate. If it is a DC servo motor, the controller will precisely control its speed by adjusting the voltage supplied to the motor or using a PWM (pulse width modulation) signal, similarly achieving a bolus injection at the recommended initial flow rate. This startup process ensures that the starting point of the bolus injection is based on personalized considerations, rather than blindly or empirically.

[0074] At the same time, once the bolus injection begins, the data acquisition module immediately goes into operation, continuously monitoring and recording two core physiological and physical parameters at a high frequency: pressure and displacement. Pressure signal acquisition is typically achieved by integrating a highly sensitive, high-response pressure sensor (e.g., a piezoresistive or capacitive pressure sensor based on MEMS technology) at the end of the injection line or near the syringe needle. This sensor converts the sensed real-time tissue pressure or line pressure into an electrical signal. This signal undergoes necessary signal conditioning, including amplification and filtering, before being converted to a digital signal by a high-speed analog-to-digital converter (ADC) at a preset high sampling frequency (e.g., tens of hertz to several kilohertz) for the controller to read and analyze. High-frequency acquisition captures details of rapid pressure changes, which is crucial for promptly identifying potential abnormalities such as excessive injection resistance, nerve contact, or vascular puncture. Displacement signal acquisition is directly correlated to the volume of anesthetic injected, and can be implemented in a variety of ways. For stepper motor-driven pumps, total displacement can be calculated by accurately recording the number of steps the motor has taken and combining the piston displacement corresponding to each step (calibrated using the pump's mechanical parameters). For servo motors, a rotary encoder mounted on the motor shaft or a linear encoder mounted directly on the piston push rod is usually used to measure displacement with high precision. The pulse signals or absolute position signals generated by these encoders are also read by the controller at high speed. High-frequency acquisition of displacement information not only ensures the precise control of anesthetic dosage, but also indirectly reflects the compliance of the tissue and the distribution of the drug when combined with pressure data. All of these high-frequency acquired data (pressure value sequence and displacement value sequence, both with time stamps) are sent to the controller's processing unit or specific data buffer in real time, providing continuous and intensive input for subsequent intelligent analysis and adaptive flow rate control algorithms, thus forming the perception basis for realizing dynamic optimization of injection strategies.

[0075] In the aforementioned intelligent anesthetic bolus control system 100, the injection parameter calculation module 140 is used to calculate the current pressure, pressure change rate, real-time flow rate, and cumulative injected volume based on the collected pressure and displacement signals. It should be understood that although the raw pressure and displacement signals are accurate, they are physical quantities and are insufficient to directly drive high-level decision-making. It is necessary to extract derived indicators with clear clinical significance and engineering control value from these continuous, potentially noisy raw data. These indicators are not only multi-dimensional real-time status feedback, but also the core basis for evaluating the current bolus status, predicting potential risks, and guiding subsequent adaptive adjustments. They enable the system to upgrade from simple sensor readings to an "understanding" of the bolus process, thereby achieving more refined and safer control.

[0076] Specifically, in a specific example of this application, first, the current pressure is calculated. This is the most direct derived indicator. The raw electrical signal collected by the pressure sensor undergoes analog-to-digital conversion, resulting in a series of discrete digital values. These digital values ​​typically need to be converted using a pre-calibrated sensor characteristic curve (e.g., the correspondence between the sensor output voltage and the actual pressure) to obtain the actual pressure value expressed in units such as Pascal (Pa), millimeters of mercury (mmHg), or bar. In practical applications, to remove high-frequency noise that may be introduced during the acquisition process, the raw pressure signal is often filtered in real time, such as using a sliding average filter, exponential smoothing filter, or Kalman filter, to obtain a smoother and more realistic current pressure reading. This smoothed value is the current pressure. Next, the rate of pressure change (dP / dt) is calculated. The rate of pressure change (dP / dt) is a key indicator for assessing the dynamic trend of tissue resistance during a bolus injection and is important for warning of abnormal conditions such as nerve damage, vascular puncture, and even syringe blockage. Its calculation is typically based on continuously collected current pressure values. The simplest method is numerical differentiation: take two consecutive (or fixed-time interval) current pressure values, P(t) and P(t-Δt), and estimate the rate of change of pressure over time Δt using (P(t)-P(t-Δt)) / Δt. To improve robustness and reduce the effects of noise, more advanced methods use linear regression to fit the most recent N pressure points, then take the slope of the fitted line as the rate of change of pressure. Alternatively, digital filters (such as Savitzky-Golay filters) smooth the signal while calculating its derivative. High-frequency pressure changes, especially sudden, sharp increases or decreases, are crucial for triggering alarms or adjusting system strategies.

[0077] Next, calculate the real-time flow rate. The flow rate is the distance the syringe piston advances or the volume of liquid pushed out per unit time. Based on the displacement signal, the real-time flow rate can be calculated in two main ways: one is by numerically differentiating the high-frequency acquired displacement signal. Similar to the pressure change rate, continuous displacement points ΔD / Δt are calculated, where ΔD is the distance the piston moves in Δt time. In order to convert from displacement ΔD to volume flow rate, it is necessary to convert it in combination with the inner diameter or piston area of ​​the syringe (flow rate = (piston area * ΔD) / Δt). The second is to reversely calculate through control instructions. For precision-controlled syringe pumps, the known number of motor steps or speed can be directly mapped to the preset piston displacement rate, which is the theoretical real-time flow rate. However, in order to more accurately reflect the actual situation, it is usually combined with the data of the displacement encoder for correction or directly inferred from the displacement data. Similarly, filtering is necessary to eliminate noise and obtain a smooth real-time flow rate.

[0078] Finally, the cumulative injected volume is calculated. The cumulative injected volume is an important basis for anesthesia decision-making, reflecting the total amount of anesthetic the patient has received. This can be obtained by time-integrating the real-time flow rate, or more directly by accumulating the small displacement increments recorded by the displacement sensor during each sampling cycle and multiplying them by the piston cross-sectional area. For example, each time a new position value D(t) of the displacement sensor is read, the cumulative injected volume = cumulative injected volume (t-Δt) + piston area * (D(t)-D(t-Δt)). The system maintains an internal counter or accumulator to continuously update this value, and clears or saves it at the end of each bolus or when the system is stopped.

[0079] In the above-mentioned intelligent control system 100 for anesthetic injection, the injection state identification module 150 is used to compare the current pressure with the comfortable pressure threshold and the safe pressure threshold, compare the pressure change rate with the preset gradient threshold, and compare the cumulative injected amount with the target total injection amount to generate an injection state flag. It should be understood that the system needs to convert the continuously changing, quantified feedback data into discrete, semantically clear state judgments, so as to provide clear, concise and easy-to-process inputs for subsequent intelligent decision-making and control actions (such as flow rate adjustment, pause or stop injection). Therefore, the injection state flag is further generated. These generated injection state flags are like traffic lights, which can instantly indicate whether the current injection process is in a normal, warning or dangerous state, and whether the predetermined goal has been achieved. They are an indispensable abstract layer for realizing closed-loop adaptive control.

[0080] Specifically, in one specific example of this application, the current pressure is compared against a threshold to monitor in real time whether the injection pressure is within pre-set physiological and safety limits. The system maintains two key pressure thresholds: a comfort pressure threshold, which represents the upper limit of pressure that patients can typically tolerate while maintaining effective anesthesia. Exceeding this value may cause discomfort; and a safety pressure threshold, which is an absolute upper limit. Sustained excess of this value is highly likely to cause tissue damage, nerve compression, or other complications. During implementation, the logic unit within the controller continuously compares the calculated current pressure value against these two thresholds. For example, if the current pressure is less than or equal to the comfort pressure threshold, a "normal pressure" flag is generated; if the comfort pressure threshold is less than or equal to the current pressure and less than or equal to the safety pressure threshold, a "high pressure" or "near warning" flag is generated; and if the current pressure is greater than the safety pressure threshold, a "pressure exceeding limit" or "dangerous pressure" flag is generated. These flags directly reflect the safety and comfort of the pressure dimension.

[0081] Secondly, the comparison of the pressure change rate against the preset gradient threshold is intended to capture dynamic trends in injection pressure, particularly abnormally rapid changes, which often indicate that the needle tip may have contacted an undesirable structure (such as a nerve or blood vessel) or other unexpected events (such as a syringe blockage or tubing leak). The system presets one or more gradient thresholds, which define the normal range of the pressure change rate. During implementation, the calculated real-time pressure change rate is compared with this threshold. If |real-time pressure change rate| <= gradient threshold, a "stable pressure change" flag is generated; if the real-time pressure change rate exceeds the gradient threshold (positively exceeding the limit), a "sudden pressure rise" flag is generated, indicating the possibility of encountering high resistance; if the real-time pressure change rate is <- gradient threshold (negatively exceeding the limit, if rapid drops are also of concern), a "sudden pressure drop" flag may be generated, indicating the possibility of entering a low-resistance cavity or leaking. These flags provide the system with critical information about the dynamic stability of the bolus injection process.

[0082] Thirdly, the purpose of comparing the cumulative injected volume with the target total injection volume is to monitor the anesthetic administration process, ensure that the injection is completed as planned, and avoid overdose or underdose. The system will store the target total injection volume preset or given by the initial parameter calculation module. During implementation, the cumulative injected volume calculated in real time will be compared with the target total injection volume. If the cumulative injected volume is less than the target total injection volume, an "injection in progress" flag will be generated; if the cumulative injected volume is greater than or equal to the target total injection volume, a "target dose reached" or "injection completed" flag will be generated. This flag is directly related to the effectiveness of anesthesia and the accuracy of medication.

[0083] Ultimately, these single-dimensional status flags generated through comparison (e.g., normal pressure, sudden pressure increase, target dose reached, etc.) can be further combined or directly used as a multidimensional "injection status flag vector." In a more conventional embodiment, these comparison operations and flag generation are typically implemented through a series of conditional statements (if-then-else logic) or lookup tables in the controller software. For example, the controller periodically performs these comparisons and updates a set of internal Boolean or enumeration variables to represent the current status. If a network model is desired, a small classifier network can be designed (such as a simple multi-layer perceptron or decision tree equivalent network structure). Its inputs are the current pressure, pressure change rate, cumulative injection volume, and corresponding thresholds (or the thresholds can be internalized as part of the network learning). The network output is a directly encoded injection status flag (e.g., a one-hot encoded vector indicating which predefined state combination it belongs to). In this approach, the network needs to be trained with a labeled dataset (i.e., the correct state flags corresponding to different input combinations) to learn these comparison and judgment logic. In the above-mentioned intelligent control system 100 for anesthetic injection, the target flow rate update control module 160 is used to make adaptive flow rate decisions based on the injection status flag to obtain a new target flow rate. It should be understood that due to the static initial parameter setting or simple open-loop control, it is impossible to cope with the dynamic changes in individual patient reactions during the injection process and various unexpected situations that may occur. The injection status flag provides an instant snapshot of the current injection situation (such as whether the pressure is too high, whether it is close to the target dose, and whether there are abnormal pressure fluctuations). The system must dynamically adjust the injection strategy based on these "signal lights" and actively intervene, rather than passively waiting for adverse events to occur or manual intervention. This adaptive decision-making capability is the key to distinguishing intelligent systems from traditional injection equipment. It aims to ensure that the injection process is always carried out within a safe, effective and comfortable window and can respond to changes in a timely manner.

[0084] Specifically, in a specific example of the present application, the response to pressure exceeding the limit or sudden pressure increase: If the "injection status mark" indicates that the current pressure has exceeded the safety pressure threshold, or the pressure change rate shows a sudden increase in pressure (which may indicate that the needle tip touches the nerve or blood vessel wall, or encounters high resistance tissue), the decision logic will immediately generate a significantly reduced new target flow rate, and may even set the target flow rate to zero to pause the injection, and trigger an alarm at the same time. The specific reduction can be a preset fixed value, or a large proportion of the current flow rate. Adjustment when the pressure is high or close to the comfort threshold: If the mark indicates that the current pressure is close to or slightly exceeds the comfort pressure threshold, but is still within the safe range, and the pressure changes smoothly, the decision system may generate a moderately reduced new target flow rate to reduce patient discomfort and prevent further pressure increases. The reduction may be small, aiming for a smooth transition. Maintain or gradually increase the pressure steadily and within the normal range: If all pressure-related signs show normal results and the cumulative injected volume has not yet reached the target total volume, the decision-making system may maintain the current target flow rate, or under certain preset strategies (such as transitioning to the maintenance period after initial rapid induction), gradually increase the flow rate according to a predetermined plan or an allowed small increment (if the initial flow rate is low and the depth of anesthesia needs to be reached quickly). Processing when the target total injection volume is reached: Once the sign shows that the cumulative injected volume has reached or exceeded the target total volume, the decision-making system will usually set the new target flow rate to zero (stop bolus) or a very low maintenance flow rate, and may trigger a prompt that the injection is complete. Compensation for low pressure or slow flow rate: In certain specific cases, if the pressure is continuously too low and far from reaching the target dose, and there are no other adverse signs, the system may be designed to moderately increase the flow rate to ensure effective drug delivery.

[0085] In summary, the intelligent control system for anesthetic push injection according to the embodiment of the present application is explained, which predicts and generates personalized initial push injection parameters by deeply integrating multi-source heterogeneous data such as patient individual information, surgical site and anesthetic type. During the actual push injection process, the system can accurately calculate the current actual pressure, pressure change trend, real-time flow rate and total amount injected by high-frequency acquisition of physiological and process signals such as tissue pressure and push injection displacement. These real-time monitoring data will be continuously compared and evaluated with the personalized thresholds and preset targets generated above, so as to intelligently judge whether the current push injection is smooth, whether there is abnormal resistance or whether it is close to the target, and generate a sign reflecting the current injection status accordingly. Finally, the system dynamically adjusts the output of the push pump based on this status sign, combined with the preset control logic and optimization algorithm, and adaptively corrects the current target flow rate. This closed-loop strategy from personalized initial setting to real-time intelligent decision-making and adaptive adjustment ensures the safety of the anesthetic push injection process and solves the key pain points of the existing technology in terms of insufficient personalization and lack of real-time adaptive ability to make intelligent decisions.

[0086] While various embodiments of the present disclosure have been described above, the above descriptions are illustrative, non-exhaustive, and not intended to be limiting of the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, their practical applications, or improvements to existing technologies, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. An intelligent control system for anesthetic injection, characterized in that: include: A data input module is used to obtain basic patient information, surgical site, and anesthetic type input by the operator through the user interface; an initial injection parameter generation module, configured to input the patient's basic information, the surgical site, and the anesthetic type into a machine learning model to obtain personalized initial injection parameters, wherein the personalized initial injection parameters include a recommended initial flow rate, a comfortable pressure threshold, a safe pressure threshold, and a target injection volume; a bolus injection data acquisition module, configured to, after issuing a start injection instruction, cause the controller to start the bolus injection pump motor to start bolus injection at the recommended initial flow rate as the current target flow rate, and to collect pressure signals and displacement signals at high frequency through the data acquisition module during the bolus injection process; An injection parameter calculation module is used to calculate the current pressure, pressure change rate, real-time flow rate and cumulative injection volume based on the collected pressure signal and displacement signal; an injection status identification module, configured to compare the current pressure with the comfortable pressure threshold and the safe pressure threshold, compare the pressure change rate with a preset gradient threshold, and compare the accumulated injected volume with the target total injection volume, so as to generate an injection status flag; The target flow rate update control module is used to make an adaptive flow rate decision based on the injection status flag to obtain a new target flow rate.

2. The intelligent control system for anesthetic injection according to claim 1, characterized in that: The initial injection parameter generation module includes: A patient information encoding unit, configured to perform semantic embedding encoding on the patient basic information to obtain a patient basic information semantic embedding encoding vector; a surgical information encoding unit, configured to perform one-hot encoding on the surgical site and the anesthetic type to obtain a one-hot embedding encoding vector for the surgical site and a one-hot embedding encoding vector for the anesthetic type; An information association transfer unit is configured to calculate a basic information-surgical site sparse association transfer matrix and a basic information-anesthetic type sparse transfer association matrix based on the patient basic information semantic embedding coding vector, the surgical site one-hot embedding coding vector, and the anesthetic type one-hot embedding coding vector; an information local semantic association encoding unit, configured to extract semantic features from the basic information-surgical site sparse association transfer matrix and the basic information-anesthetic type sparse transfer association matrix to obtain a basic information-surgical site local semantic association encoding vector and a basic information-anesthetic type local semantic association encoding vector; an information local semantic association strengthening unit, configured to perform local semantic association strengthening on the basic information-surgical site local semantic association coding vector and the basic information-anesthetic type local semantic association coding vector to obtain an enhanced basic information-surgical site local semantic association coding vector and an enhanced basic information-anesthetic type local semantic association coding vector; A feature decoding and regression unit is used to perform feature decoding and regression on the fused features after performing feature fusion on the enhanced basic information-surgical site local semantic association coding vector and the enhanced basic information-anesthetic type local semantic association coding vector to obtain the personalized initial injection parameters.

3. The intelligent control system for anesthetic injection according to claim 2, characterized in that: The information association transfer unit is used to: Calculating a transfer matrix of the semantic embedding coding vector of the patient basic information relative to the one-hot embedding coding vector of the surgical site to obtain a basic information-surgical site sparse association transfer matrix; The transfer matrix of the semantic embedding coding vector of the patient basic information relative to the one-hot embedding coding vector of the anesthetic type is calculated to obtain the basic information-anesthetic type sparse transfer association matrix.

4. The intelligent control system for anesthetic injection according to claim 3, characterized in that: The local semantic association encoding unit between information is used to: input the basic information-surgical site sparse association transfer matrix and the basic information-anesthetic type sparse transfer association matrix into the text convolutional neural network model respectively to obtain the basic information-surgical site local semantic association encoding vector and the basic information-anesthetic type local semantic association encoding vector.

5. The intelligent control system for anesthetic injection according to claim 4, characterized in that: The information local semantic association strengthening unit includes: A basic information-surgical site semantic deep understanding subunit is used to perform a structured decomposition based on one-dimensional convolutional coding on the basic information-surgical site local semantic association coding vector to decompose the basic information-surgical site local semantic association coding vector into a set of basic information-surgical site local semantic initial local implicit feature vectors; A basic information-surgical site semantic association analysis subunit is configured to generate a basic information-surgical site local semantic geometric topology fine-grained association mask matrix based on the geometric affinity between any two basic information-surgical site local semantic initial local implicit feature vectors in the set of basic information-surgical site local semantic initial local implicit feature vectors; The basic information-surgical site association semantics feedback reconstruction subunit is used to interactively refine the set of initial local implicit feature vectors of the basic information-surgical site local semantics based on the basic information-surgical site local semantics geometric topology fine-grained association mask matrix to obtain the enhanced basic information-surgical site local semantics association encoding vector.

6. The intelligent control system for anesthetic injection according to claim 5, characterized in that: The basic information-surgical site semantic association analysis subunit is used to: Calculating the geometric affinity between any two basic information-surgical site local semantic initial local implicit feature vectors in the set of basic information-surgical site local semantic initial local implicit feature vectors to obtain a basic information-surgical site local semantic geometric association topology matrix composed of multiple basic information-surgical site local semantic geometric affinities; The basic information-surgical site local semantic geometry association topology matrix is ​​input into a gated mask function to obtain the basic information-surgical site local semantic geometry topology fine-grained association mask matrix.

7. The intelligent control system for anesthetic injection according to claim 6, characterized in that: The basic information-surgical site association semantic feedback reconstruction subunit is used to: Based on the basic information-surgical site local semantic geometric topology fine-grained association mask matrix, performing neighborhood-aware distillation on each basic information-surgical site local semantic initial local implicit feature vector in the set of basic information-surgical site local semantic initial local implicit feature vectors to obtain a set of basic information-surgical site local semantic distillation initial local implicit feature vectors; The set of initial local implicit feature vectors of the basic information-surgical site local semantic distillation is aggregated and reconstructed based on the self-attention mechanism to obtain the enhanced basic information-surgical site local semantic association encoding vector.

8. The intelligent control system for anesthetic injection according to claim 7, characterized in that: The feature decoding and regression unit is used to: Performing feature fusion on the enhanced basic information-surgical site local semantic association coding vector and the enhanced basic information-anesthetic type local semantic association coding vector to obtain an enhanced basic information-anesthetic surgical information local semantic fusion feature vector as a fusion feature; The enhanced basic information-anesthetic surgery information local semantic fusion feature vector is passed through a decoder-based injection parameter controller to obtain the personalized initial injection parameters.

Images