Intelligent infusion control system based on edge calculation

Abstract

The invention relates to the technical field of intelligent infusion, and discloses an intelligent infusion control system based on edge calculation, which comprises a data acquisition module, an intelligent infusion pump and an edge AI chip. According to the intelligent infusion control system based on edge calculation, the dosage error of manual calculation in premature infant intravenous nutrition support can be effectively reduced, multi-source data are processed in real time through the edge AI chip, the infusion formula and the infusion amount are dynamically adjusted in combination with reinforcement learning, the individual metabolism difference of the premature infant is rapidly adapted, and the premature infant intravenous nutrition support accuracy is improved. Hysteresis of traditional manual adjustment is avoided; meanwhile, by means of closed-loop adjustment and abnormity monitoring functions, risks are recognized in time, early warning is conducted, infusion safety is improved, medical care operation burdens are integrally relieved, and precision and reliability of premature infant vein nutrition supply are guaranteed.

Description

Technical Field

[0001] This invention relates to the field of intelligent infusion technology, specifically an intelligent infusion control system based on edge computing. Background Technology

[0002] Premature infants, due to their immature physiological functions and weak gastrointestinal digestive and absorptive capacity, often rely on intravenous nutrition to meet their energy and nutrient needs for growth and development. The precise delivery of intravenous nutrition is directly related to the life safety and long-term health prognosis of premature infants. Currently, in clinical practice, the formulation and adjustment of intravenous infusion plans for premature infants largely depend on manual calculations by medical staff. This requires combining multi-dimensional data such as the premature infant's weight, gestational age, past medical history, and real-time physiological indicators (e.g., blood glucose, electrolyte concentration), using empirical formulas or standard guidelines to calculate the proportions of various components in the nutrient solution (e.g., glucose, amino acids, fat emulsions) and the infusion volume per unit time. However, manual calculation is cumbersome and easily affected by subjective factors. On the one hand, data entry errors and incorrect formula application can lead to dosage errors, resulting in complications such as blood glucose fluctuations and metabolic disorders in premature infants, which can even be life-threatening in severe cases. On the other hand, premature infants exhibit significant individual differences, and their metabolic state changes dynamically with age. Manual adjustments to the plan often lag behind, making it difficult to match their immediate nutritional needs in real time.

[0003] To overcome the limitations of manual operation, some existing infusion devices possess basic data acquisition capabilities. These devices can acquire electronic health record data via interfaces or connect to biosensors to collect real-time physiological parameters. However, significant shortcomings remain in data processing and protocol adjustment. Some devices can only perform simple data aggregation and display, unable to perform in-depth fusion analysis of multi-source data, still requiring manual judgment and adjustment by medical staff. Other devices use cloud computing to process data and generate adjustment instructions, which, while possessing a certain degree of intelligence, suffer from data transmission delays. Premature infants have extremely high timeliness requirements for intravenous nutrition adjustments; delays may lead to missed optimal adjustment opportunities, increasing medical risks. Furthermore, existing intelligent infusion technologies lack dynamic optimization algorithms designed specifically for the physiological characteristics of premature infants. Most can only adjust formulas according to fixed templates, unable to achieve personalized, adaptive formula optimization based on real-time metabolic feedback from premature infants (such as blood glucose rate of change and nutrient consumption rate) through intelligent algorithms like reinforcement learning. This makes it difficult to meet the refined and dynamic intravenous nutritional support needs of premature infants. Therefore, developing an intelligent control system capable of real-time acquisition of multi-source data, efficient local processing, and dynamic optimization of infusion formulas has become a key direction for solving the current problems of accuracy and safety in intravenous nutrition supply for premature infants. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent infusion control system based on edge computing to solve the technical problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention discloses the following technical solution: an intelligent infusion control system based on edge computing, comprising: The intelligent infusion pump is communicatively connected to the data acquisition module; An edge AI chip is built into the intelligent infusion pump, and a reinforcement learning model runs in the edge AI chip; wherein, the reinforcement learning model performs feature extraction and fusion processing on the electronic health record data and real-time biosensor data, and dynamically generates infusion formula adjustment instructions, the infusion formula adjustment instructions including at least the nutrient composition ratio parameters of multiple nutrients and the infusion volume per unit time; The intelligent infusion pump automatically adjusts the nutrient composition ratio parameters of each component in the output intravenous nutrition solution and the total infusion volume per unit time according to the infusion formula adjustment command.

[0006] Preferably, the data acquisition module includes: The recording interface unit is configured to extract basic health data of the premature infant from the electronic health record system, the basic health data including weight, gestational age and daily nutritional requirements baseline. The biosensor interface unit is configured to receive real-time physiological parameters acquired by the biosensor. The data integration unit is configured to aggregate the basic health data and real-time physiological parameters into an input dataset in a unified format, wherein the input dataset includes at least metabolic indicators related to tolerance per unit time.

[0007] Preferably, the data acquisition module further includes a data preprocessing unit, which is configured to: Imput missing values ​​in the input dataset using interpolation; Abnormal data that exceeds the physiologically reasonable range is marked and corrected; Real-time physiological parameters of different dimensions are standardized to a preset range, with the nutrient metabolism rate data per unit time in the real-time physiological parameters being normalized separately.

[0008] Preferably, the data preprocessing unit is further configured as follows: Time series smoothing processing is performed on continuously collected blood glucose data; Calculate the correlation coefficient between the fluctuation range of blood glucose and the amount of nutrition infusion and metabolic indicators per unit time. The correlation coefficient is used as one of the input features for the reinforcement learning model to evaluate the rationality of the infusion volume.

[0009] Preferably, the edge AI chip includes: The data receiving unit is configured to receive the input dataset after it has been processed by the data acquisition module. The model computation unit is configured to load the reinforcement learning model and perform computation, and simultaneously calculate the co-optimization parameters of nutrient composition ratio parameters and unit time infusion volume. The instruction generation unit is configured to convert the model calculation results into executable infusion formula adjustment instructions, which include the nutrient component ratio parameters of each component and the corresponding infusion volume per unit time parameters.

[0010] Preferably, the step of loading the reinforcement learning model and performing operations by the model computation unit includes: The input dataset is segmented into features, with weight and gestational age from electronic health record data used as static features, and metabolic indicators from real-time biosensor data used as dynamic features. The weighting of data related to metabolic load per unit time in dynamic features is enhanced through attention mechanisms. The fused feature vector is iteratively processed by a reinforcement learning model to output the optimal nutrient composition ratio parameters and the corresponding safe infusion volume per unit time in the current state.

[0011] Preferably, the step of the model computing unit calculating the co-optimization parameters of nutrient component ratio parameters and unit time infusion volume includes: The metabolic stability range and the maximum tolerated infusion volume per unit time in historical infusion data are used as the dual-dimensional reward benchmark. The system correlates metabolic index deviations with infusion volume deviations in real time. When any deviation exceeds a preset threshold, the optimization parameters are corrected using a penalty function. A coordinated adjustment scheme is generated based on the corrected parameters. The scheme includes the infusion rate gradient of various nutrients and the matching relationship of total dose per unit time.

[0012] Preferably, the intelligent infusion pump includes: The multi-channel execution unit is configured to independently control the infusion flow rate of multiple nutrient solution components and synchronously match the nutrient component ratio parameters with the infusion volume requirements per unit time. The feedback calibration unit is configured to detect in real time the deviation between the sum of the actual flow rates of each channel and the commanded infusion rate; The closed-loop adjustment unit is configured to dynamically correct the driving parameters of the multi-channel execution unit based on the deviation value, so that the actual infusion volume per unit time does not deviate from the command within a preset range.

[0013] Preferably, it also includes an anomaly monitoring module, the anomaly monitoring module comprising: The threshold storage unit is configured to pre-store the safety range thresholds for physiological parameters of preterm infants and the maximum tolerated infusion volume threshold per unit time. The real-time comparison unit is configured to perform dual comparisons between biosensor data and physiological safety thresholds, and between actual infusion volume and maximum tolerance thresholds. The emergency response unit is configured to generate an alarm signal and trigger an infusion pause command when any threshold is exceeded.

[0014] Preferably, the anomaly monitoring module further includes a trend prediction unit, which is configured as follows: Synergistic trend analysis was performed on the physiological parameters and infusion volume per unit time for a preset number of consecutively collected cycles; Calculate the correlation coefficient between the rate of change of physiological indicators and the rate of adjustment of infusion volume; When the correlation coefficient exceeds the warning threshold, a dose and metabolic abnormality warning signal is generated in advance and pushed to the medical terminal.

[0015] Beneficial effects: The edge computing-based intelligent infusion control system of this invention can effectively reduce the dosage error caused by manual calculation in intravenous nutrition support for premature infants. By processing multi-source data in real time through an edge AI chip and combining reinforcement learning, it dynamically adjusts the infusion formula and infusion volume, quickly adapting to the individual metabolic differences of premature infants and avoiding the lag of traditional manual adjustments. At the same time, relying on closed-loop regulation and abnormal monitoring functions, it can identify risks and issue warnings in a timely manner, improve infusion safety, reduce the overall workload of medical staff, and ensure the accuracy and reliability of intravenous nutrition supply for premature infants. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a structural block diagram of an intelligent infusion control system based on edge computing, provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] In this document, the term "comprising" is intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0020] This embodiment aims to address the technical problems of traditional manual calculation methods in intravenous nutritional support for preterm infants, which are prone to dosage errors, adjustment lags, and difficulty in adapting to individual metabolic differences. It provides a method... Figure 1 The edge computing-based intelligent infusion control system shown includes: The intelligent infusion pump communicates with the data acquisition module; An edge AI chip is built into the smart infusion pump, and a reinforcement learning model runs in the edge AI chip. The reinforcement learning model performs feature extraction and fusion processing on electronic health record data and real-time biosensor data, and dynamically generates infusion formula adjustment instructions. The infusion formula adjustment instructions include at least the nutritional component ratio parameters of multiple nutrients and the infusion volume per unit time. The intelligent infusion pump automatically adjusts the nutrient composition ratio parameters of each component in the output intravenous nutrition solution and the total infusion volume per unit time according to the infusion formula adjustment instructions.

[0021] The data acquisition module communicates with the intelligent infusion pump via Bluetooth Low Energy (BLE 5.0) or medical WiFi (compliant with IEEE 802.11ax standard) to ensure data transmission latency ≤100ms. Real-time biosensor data, in addition to blood glucose data, may also include serum electrolyte (such as sodium, potassium, and calcium) concentration data, collected via a non-invasive electrochemical sensor at a sampling frequency of once per minute. Various nutrients, including glucose, amino acids, and fat emulsions, are included, with initial formulation parameters set according to the "Guidelines for Clinical Application of Neonatal Nutritional Support in China" (e.g., daily...). For premature infants aged 1-3 days, the initial glucose-to-amino acid ratio is 3:1; the edge AI chip uses a medical-grade low-power chip (such as the NVIDIA Jetson Nano 2GB medical version), and the reinforcement learning model running adopts the DQN (DeepQ-Network) model in deep reinforcement learning. The model input dimension is 128 dimensions (including 8-dimensional static features and 120-dimensional dynamic temporal features), and the output is a 3-dimensional vector (corresponding to the ratio adjustment coefficient of the three nutrients and the infusion volume per unit time); the adjustment response time of the intelligent infusion pump is ≤500ms.

[0022] Based on the above, the edge computing-based intelligent infusion control system of this embodiment, leveraging multi-source data fusion processing capabilities, can automatically integrate basic information from electronic health records with real-time physiological parameters collected by biosensors. This avoids subjective errors and cumbersome operations in manual data entry and calculation. Standardized processing by the data preprocessing unit further ensures the reliability of input data, laying a high-quality data foundation for subsequent formula optimization. Local deployment of the edge AI chip enables real-time data processing and analysis, significantly reducing data transmission latency compared to cloud computing. Combined with the hierarchical processing of static and dynamic features using reinforcement learning models and the enhancement of key metabolic data through attention mechanisms, it can quickly respond to the immediate metabolic changes of premature infants, dynamically generating nutritional component ratios and infusion volume adjustment schemes tailored to their individual needs. This solves the problems of lag in traditional manual adjustments and the inability of fixed templates to adapt to individual differences. Simultaneously, the multi-channel execution and closed-loop regulation design of the intelligent infusion pump ensures accurate execution of adjustment commands, while the dual threshold comparison and trend prediction functions of the abnormality monitoring module can promptly identify potential risks and trigger warnings, effectively reducing the risk of complications such as metabolic disorders. Overall, this system, through the deep integration of edge computing and artificial intelligence technologies, has achieved intelligent management of the entire process of intravenous nutrition for premature infants, from data collection and analysis to decision-making and execution control. It provides premature infants with more personalized nutritional support that is more in line with their physiological characteristics, which helps to improve treatment outcomes and long-term health prognosis.

[0023] In this embodiment, the data acquisition module includes: The recording interface unit is configured to extract basic health data of premature infants from the electronic health record system. The basic health data includes weight, gestational age and daily nutritional requirements baseline. The biosensor interface unit is configured to receive real-time physiological parameters acquired by the biosensor. The data integration unit is configured to aggregate basic health data and real-time physiological parameters into a unified input dataset, which includes at least metabolic indicators related to tolerance per unit time.

[0024] The recording interface unit uses the HL7FHIR (Fast Healthcare Interoperability Resources) protocol to interface with the hospital's electronic health record (EHR) system, supporting the automatic extraction of birth weight (accuracy ±5g), gestational age (accurate to the day), and daily nutritional baseline (e.g., energy requirement baseline set at 100kcal / (kg·d)) for premature infants, with a data extraction frequency of once per hour. The biosensor interface unit uses the I2C communication protocol, compatible with mainstream medical non-invasive sensors on the market (e.g., Abbott FreeStyle FreeStyle non-invasive blood glucose sensor, Medtronic electrolyte sensor), with a real-time physiological parameter reception delay ≤50ms. The data integration unit summarizes basic health data and real-time physiological parameters into an input dataset in JSON format. The metabolic indicators related to tolerance per unit time include hourly blood glucose changes and hourly serum albumin concentration changes. The data integration frequency is consistent with the real-time physiological parameter acquisition frequency (once per minute).

[0025] Based on the above, the edge computing-based intelligent infusion control system of this embodiment achieves seamless integration with the hospital's existing EHR system through the HL7FHIR protocol, enabling the acquisition of basic health data without modifying the original system, thus reducing hospital deployment costs. The design, which is compatible with mainstream biosensors, improves the system's hardware adaptability and avoids application limitations caused by sensor model restrictions. The JSON format input dataset standardizes data specifications, facilitating rapid parsing and subsequent computation by the edge AI chip, thereby improving data processing efficiency.

[0026] Furthermore, the data acquisition module also includes a data preprocessing unit, which is configured as follows: Imput missing values ​​in the input dataset using interpolation; Abnormal data that exceeds the physiologically reasonable range is marked and corrected; Real-time physiological parameters of different dimensions are standardized to a preset range, with the nutrient metabolism rate data per unit time in the real-time physiological parameters being normalized separately.

[0027] Specifically, missing value interpolation is performed using the existing linear interpolation method, which calculates missing values ​​from adjacent valid data points. If a missing value is found in more than three consecutive data points, the historical mean of the parameter is used for supplementation. Anomaly detection is based on the 3σ principle, where data values ​​exceeding the historical mean ± three standard deviations of the parameter are considered anomalies. Anomaly correction uses the moving average of five adjacent valid data points. Standardization of real-time physiological parameters of different dimensions uses the existing min-max standardization method, unifying the data to the [0,1] interval (e.g., the blood glucose standardization interval corresponds to the safe blood glucose range of 2.6-7.0 mmol / L for preterm infants). Nutritional metabolic rate data per unit time (e.g., hourly glucose metabolism rate) is normalized using the existing Z-score method to highlight its crucial impact on formulation adjustments.

[0028] Based on the above, the edge computing-based intelligent infusion control system of this embodiment effectively solves the problems of missing and abnormal data through linear interpolation and moving average correction, avoiding model calculation deviations caused by erroneous data and ensuring the reliability of input data; min-max standardization eliminates dimensional differences, allowing different physiological parameters (such as blood glucose and weight) to participate in model calculations in the same dimension; and the separate normalization of nutrient metabolic rate data strengthens the contribution of this type of key data to model decision-making and improves the accuracy of prescription adjustment.

[0029] Secondly, the data preprocessing unit is also configured as follows: Time series smoothing processing is performed on continuously collected blood glucose data; Calculate the correlation coefficient between the fluctuation range of blood glucose and the amount of nutrition infusion and metabolic indicators per unit time. The correlation coefficient is used as one of the input features for the reinforcement learning model to evaluate the rationality of the infusion volume.

[0030] The time-series smoothing of continuous blood glucose data employs the existing 5-point moving average method to filter out short-term fluctuations caused by sensor noise. The blood glucose fluctuation amplitude within a unit of time (set to 1 hour) is obtained by calculating the difference between the maximum and minimum values ​​of the smoothed blood glucose data within that time period; if the difference > 2.0 mmol / L, it is marked as abnormal blood glucose fluctuation. The correlation coefficient between nutrient infusion and metabolic indicators is calculated using the Pearson correlation coefficient, with the formula: In the formula This refers to the amount of nutrition infused per unit time (e.g., glucose infusion, in ml / kg·h). For the corresponding metabolic indicators (such as blood glucose level, unit mmol / L). The sample size (60 data points within one hour) and the correlation coefficient. The value range of is [-1, 1]. When The coefficient is used as an input feature of the reinforcement learning model to evaluate the matching degree between the current infusion volume and the metabolic state.

[0031] Based on the above, the edge computing-based intelligent infusion control system of this embodiment reduces the interference of sensor noise on blood glucose data through 5-point moving average processing, making the blood glucose change trend clearer and facilitating the model to capture the real metabolic state; the calculation of blood glucose fluctuation amplitude can quickly identify metabolic instability tendencies, providing a clear direction for model adjustment; the correlation coefficient, as a model input feature, enables the model to quantify the correlation between infusion volume and metabolic indicators, avoiding blind adjustments and improving the rationality of infusion volume setting.

[0032] In this embodiment, the edge AI chip includes: The data receiving unit is configured to receive the input dataset after it has been processed by the data acquisition module. The model computation unit is configured to load the reinforcement learning model and perform computation, and simultaneously calculate the co-optimization parameters of nutrient composition ratio parameters and unit time infusion volume. The instruction generation unit is configured to convert the model calculation results into executable infusion formula adjustment instructions, which include the nutritional component ratio parameters of each component and the corresponding infusion volume per unit time parameters.

[0033] The data receiving unit uses the MQTT (Message Queuing Telemetry Transport) protocol to receive the input dataset, with a QoS level of 2 (ensuring messages are delivered at least once and without duplication). The data receiving buffer is set to 1024KB to prevent data overflow. The reinforcement learning model loaded by the model computing unit is pre-trained on a cloud server (using an NVIDIA Tesla V100 GPU). The training dataset contains intravenous nutrition treatment data from 1000 premature infants. After training, the model parameters (approximately 50MB) are burned to the Flash storage unit of the edge AI chip via USB. The power consumption during model computing is controlled below 5W to meet the low power consumption requirements of medical equipment. The co-optimization parameters of nutrient composition ratio and infusion volume per unit time are calculated with the objective function of maximizing metabolic stability and maximizing infusion safety. The gradient descent method is used to solve for the optimal solution. The objective function expression is: In the formula To comprehensively optimize the index (value range [0,1], the closer to 1 the better), Metabolic stability (calculated based on blood glucose fluctuations), 4.4 represents the maximum reasonable fluctuation range of blood glucose in premature infants. For infusion safety (calculated based on the ratio of infusion volume to maximum tolerated volume), , (Maximum tolerable infusion volume per unit time) , The weighting coefficients (all set to 0.5) are calculated iteratively to make... The value is at its maximum (approaching 1), and the corresponding parameter is the co-optimized parameter. The instruction generation unit converts the optimized parameter into an instruction that conforms to the communication protocol of the intelligent infusion pump (such as RS485 protocol). The instruction format is component ID-ratio parameter-infusion volume per unit time, such as glucose-0.4-3.5ml / kg·h, to ensure that the intelligent infusion pump can directly parse and execute it.

[0034] Based on the above, the edge computing-based intelligent infusion control system of this embodiment ensures that the high reliability of the MQTT protocol guarantees no loss of input datasets, providing complete data support for model calculations; the local computation of the edge AI chip avoids cloud transmission delays, meeting the real-time requirements for intravenous nutrition adjustment in premature infants; the improved objective function takes into account both metabolic stability and infusion safety, avoiding the risks caused by optimizing a single indicator (such as over-infusion in pursuit of metabolic stability); and the standardized instruction format ensures that the intelligent infusion pump can be directly parsed and executed, reducing instruction conversion time.

[0035] Furthermore, the steps of loading the reinforcement learning model and performing computations by the model computation unit include: The input dataset is segmented into features, with weight and gestational age from electronic health record data used as static features, and metabolic indicators from real-time biosensor data used as dynamic features. The weighting of data related to metabolic load per unit time (including rate of change of blood glucose and rate of nutrient consumption per unit time) in dynamic features is enhanced by attention mechanisms. The fused feature vector is iteratively processed by a reinforcement learning model to output the optimal nutrient composition ratio parameters and the corresponding safe infusion volume per unit time in the current state.

[0036] In the feature layering process, static features (weight, gestational age) are normalized and used as the static input layer of the model (dimension 2), while dynamic features (blood glucose, electrolyte concentration) are processed by the existing LSTM layer to obtain 60 time-series data points within one hour and used as the dynamic input layer (dimension 64). The attention mechanism adopts the existing multi-head attention mechanism, setting four attention heads and using the existing softmax function to assign higher weights (weight ratio 0.7) to data related to metabolic load per unit time (blood glucose change rate per unit time, nutrient consumption rate). The reinforcement learning model has 150 iterations and a learning rate of 0.001. The existing mean squared error (MSE) is used as the loss function, and the iteration is considered to have converged when the loss function value is less than 0.01, at which point the optimal parameters are output.

[0037] Based on the above, the edge computing-based intelligent infusion control system of this embodiment processes static and dynamic features in a hierarchical manner, enabling the model to capture both the basic differences of preterm infants (such as initial nutritional needs determined by weight and gestational age) and to respond to metabolic changes (such as blood glucose fluctuations) in real time, thereby improving parameter adaptability. The multi-head attention mechanism strengthens the weight of key metabolic data and reduces interference from irrelevant data, making the model's decisions more focused on the core factors affecting infusion safety. Iterative calculations until the loss function converges ensure that the output nutrient composition ratio and infusion volume are the optimal solutions under the current state, avoiding adjustment deviations caused by insufficient calculations.

[0038] Secondly, the steps for the model computation unit to calculate the co-optimization parameters of nutrient component ratio parameters and unit time infusion volume include: The metabolically stable range (e.g., 2.6-7.0 mmol / L for preterm infants) in historical infusion data and the maximum tolerated infusion volume per unit time (3-5 ml / kg·h, calculated based on weight and applicable to preterm infants) are used as the dual-dimensional reward benchmarks. The system correlates the deviation values ​​of metabolic indicators with the deviation values ​​of infusion volume in real time. When either deviation value exceeds the preset threshold (metabolic deviation ±10%, infusion volume deviation ±8%), the optimization parameters are corrected through a penalty function. Based on the corrected parameters, a coordinated adjustment scheme is generated, which includes the infusion rate gradient of various nutrients (e.g., adjusted by 0.5 ml / kg·h every 30 minutes) and the total dose matching relationship per unit time.

[0039] Firstly, the dual-dimensional reward criteria for preterm infants have been revised: the metabolic stability range has been adjusted to the safe blood glucose range of 2.6-7.0 mmol / L for preterm infants, and the maximum tolerated infusion volume per unit time has been set according to age (as mentioned above). ); Metabolic index deviation value This refers to the relative deviation of the current metabolic indicator from the median of the stable range (e.g., when the median blood glucose level is 4.8 mmol / L, and the current blood glucose level is 5.3 mmol / L). Infusion volume deviation value The relative deviation between the current infusion volume and the commanded infusion volume; the penalty function adopts the improved L2 penalty function, the expression of which is: In the formula This is the penalty coefficient (ranging from [0, 0.8]). As the penalty weight (taken as 0.8), when or At that time, through Parameter correction; infusion rate gradient dynamically adjusted based on metabolic deviation value: Adjust the volume by 0.2 ml / kg·h every 30 minutes. Adjust the volume by 0.3 ml / kg·h every 30 minutes. Adjust by 0.5 ml / kg·h every 30 minutes, and ensure that the total dose per unit time matches the total dose of each component, which is equal to the total dose per unit time.

[0040] Based on the above, the edge computing-based intelligent infusion control system of this embodiment adapts the reward benchmark for preterm infants to avoid parameter deviations caused by using adult or full-term infant standards, thereby improving the clinical adaptability of the protocol; the improved penalty function can dynamically adjust the penalty intensity according to the degree of deviation, which avoids excessive correction of minor deviations affecting stability, and effectively constrains deviations exceeding the threshold to ensure infusion safety; the dynamic infusion rate gradient matches the adjustment range with the degree of metabolic deviation, avoiding response lag caused by too small an adjustment range or new metabolic fluctuations caused by too large an adjustment range.

[0041] In this embodiment, the intelligent infusion pump includes: The multi-channel execution unit is configured to independently control the infusion flow rate of multiple nutrient solution components and synchronously match the nutrient component ratio parameters with the infusion volume requirements per unit time. The feedback calibration unit is configured to detect in real time the deviation between the sum of the actual flow rates of each channel and the commanded infusion rate; The closed-loop control unit is configured to dynamically correct the drive parameters of the multi-channel execution unit based on the deviation value, so that the actual infusion volume per unit time does not deviate from the command within a preset range.

[0042] The multi-channel execution unit has three independent channels (corresponding to glucose, amino acids, and fat emulsions, respectively). Each channel uses a micro stepper motor to drive the infusion pump head, with a step angle of 1.8°, a subdivision coefficient of 16, and a flow control accuracy of ±0.05 ml / h. The feedback calibration unit uses a micro electromagnetic flow sensor (accuracy ±0.1 ml / h) to detect the actual flow rate of each channel, collecting flow data every 10 seconds and calculating the sum, comparing it with the commanded infusion volume to obtain the deviation value. The closed-loop regulation unit uses the existing PID (proportional-integral-derivative) control algorithm to correct the drive parameters. The PID parameters are determined by the existing Ziegler-Nichols tuning method (proportional coefficient Kp=2.0, integral coefficient Ki=0.5, derivative coefficient Kd=0.1), ensuring that the deviation between the actual infusion volume and the command is stably controlled within ±5%.

[0043] Based on the above, the edge computing-based intelligent infusion control system of this embodiment achieves precise infusion of different nutrients through multi-channel independent control, meeting the requirements of the ratio parameters and avoiding component mixing errors caused by single-channel infusion; real-time flow feedback and PID closed-loop regulation form dynamic control, quickly correcting flow deviations, ensuring that the actual infusion volume is consistent with the command, and reducing the risk of dosage error; the deviation range of ±5% is much higher than the accuracy of clinical manual adjustment, improving the accuracy of intravenous nutrition supply.

[0044] In this embodiment, the system further includes an anomaly monitoring module, which includes: The threshold storage unit is configured to pre-store the safety range thresholds for physiological parameters of preterm infants and the maximum tolerated infusion volume threshold per unit time. The real-time comparison unit is configured to perform dual comparisons between biosensor data and physiological safety thresholds, and between actual infusion volume and maximum tolerance thresholds. The emergency response unit is configured to generate an alarm signal and trigger an infusion pause command when any threshold is exceeded.

[0045] The threshold storage unit uses existing EEPROM (Electrically Erasable Programmable Read-Only Memory) to store threshold data, which can be updated via a medical terminal APP. Pre-stored safe range thresholds for premature infant physiological parameters include: blood glucose 2.0-7.5 mmol / L (below 2.0 mmol / L is hypoglycemia, above 7.5 mmol / L is hyperglycemia), body temperature 36.0-38.0℃, blood oxygen saturation 88%-98%, and the maximum tolerated infusion volume per unit time threshold is the same. The real-time comparison unit acquires biosensor data and actual infusion volume data every 5 seconds and compares them with the corresponding threshold. The alarm signals generated by the emergency response unit include audible and visual alarms (buzzer frequency 2000Hz, red LED flashing once per second) and wireless alarms (pushed to the medical station terminal via the hospital intranet). At the same time, it triggers an infusion pause command, causing the multi-channel execution unit to immediately stop infusion until medical staff confirm the reset.

[0046] Based on the above, the edge computing-based intelligent infusion control system of this embodiment supports flexible updates of the EEPROM storage threshold, which can be adapted to the needs of premature infants of different ages and conditions, improving the system's versatility; real-time comparison every 5 seconds quickly detects threshold breaches, reducing risk response delays; multi-dimensional reminders of sound, light, and wireless alarms ensure that medical staff can detect abnormalities in a timely manner, and the infusion stop command can immediately block risks and avoid serious complications caused by hypoglycemia, hyperglycemia, or over-infusion.

[0047] Furthermore, the anomaly monitoring module also includes a trend prediction unit, which is configured as follows: Synergistic trend analysis was performed on the physiological parameters and infusion volume per unit time for a preset number of consecutively collected cycles; Calculate the correlation coefficient between the rate of change of physiological indicators and the rate of adjustment of infusion volume; When the correlation coefficient exceeds the warning threshold, a dose and metabolic abnormality warning signal is generated in advance and pushed to the medical terminal.

[0048] The preset number of cycles is set to 3 cycles (5 minutes per cycle, 15 minutes in total). The trend prediction unit analyzes physiological parameters (blood glucose, electrolyte concentration) and infusion volume per unit time within 15 consecutive minutes. The collaborative trend analysis uses linear regression, a method already in use, to fit trend lines to the parameters and calculate the rate of change (e.g., blood glucose change rate, infusion volume adjustment rate). The correlation coefficient between the rate of change of physiological indicators and the infusion volume adjustment rate is calculated using Pearson correlation coefficient, a method already in use. The warning threshold is set as follows: ,when When the infusion volume increases, it indicates a strong positive correlation between the two (e.g., an increase in infusion volume leads to a synchronous increase in blood glucose). This indicates a strong negative correlation between the two (e.g., increased infusion volume leads to decreased blood glucose); the warning signal is pushed to the mobile terminal of medical staff (e.g., nurse station PAD) via the WebSocket protocol, with a push delay of ≤1 second.

[0049] Based on the above, the edge computing-based intelligent infusion control system of this embodiment can detect potential abnormalities in advance through 15-minute trend analysis (such as persistently elevated blood glucose but not reaching the threshold), avoiding intervention only after the threshold is exceeded, and reserving 5-10 minutes of intervention time for medical staff; the correlation coefficient quantifies the relationship between physiological changes and infusion adjustments, accurately identifying abnormal correlations (such as increased infusion volume accompanied by decreased blood glucose, indicating metabolic abnormalities); early warning reduces the probability of complications.

[0050] In summary, the edge computing-based intelligent infusion control system of this embodiment provides a high-quality data foundation for intelligent decision-making through multi-source data integration and standardized processing by the data acquisition module. Relying on local reinforcement learning computation of the edge AI chip, it achieves real-time dynamic optimization of the intravenous nutrition solution formula and infusion volume for premature infants, adapting to individual metabolic differences while avoiding cloud transmission delays. The multi-channel execution and closed-loop adjustment mechanism of the intelligent infusion pump ensures the accurate implementation of adjustment commands. The dual threshold comparison and trend prediction functions of the anomaly monitoring module construct a multi-layered safety defense, enabling early identification and prevention of potential risks. Overall, through the deep integration of edge computing and artificial intelligence technologies, this system completely changes the limitations of traditional manual calculation schemes, which are prone to dosage errors and adjustment lags. It achieves intelligent management of the entire process of intravenous nutrition for premature infants, from data acquisition, analysis, decision-making to execution control, improving the accuracy, real-time nature, and safety of nutritional supply. It provides more personalized nutritional support tailored to the physiological characteristics of premature infants, effectively reducing the risk of complications such as metabolic disorders, and ultimately contributing to improved treatment outcomes and long-term health prognosis.

[0051] In the embodiments provided by this invention, it should be understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, code, or any suitable combination thereof. For hardware implementation, the processor can be implemented in one or more of the following: application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, other electronic units designed to implement the functions described herein, or combinations thereof. For software implementation, some or all of the processes of the embodiments can be performed by a computer program instructing the associated hardware. During implementation, the program can be stored in a computer-readable storage medium or transmitted as one or more instructions or code on a computer-readable storage medium. Computer-readable storage media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of a computer program from one place to another. Storage media can be any available medium accessible to a computer. Computer-readable storage media can include, but are not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code having the form of instructions or data structures and accessible to a computer.

[0052] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An edge computing-based intelligent infusion control system, characterized in that, include: The data acquisition module is used to acquire electronic health record data and real-time biosensor data of premature infants, wherein the real-time biosensor data includes at least blood glucose data. The intelligent infusion pump is communicatively connected to the data acquisition module; An edge AI chip is built into the intelligent infusion pump, and a reinforcement learning model runs in the edge AI chip; wherein, the reinforcement learning model performs feature extraction and fusion processing on the electronic health record data and real-time biosensor data, and dynamically generates infusion formula adjustment instructions, the infusion formula adjustment instructions including at least the nutrient composition ratio parameters of multiple nutrients and the infusion volume per unit time; The intelligent infusion pump automatically adjusts the nutrient composition ratio parameters of each component in the output intravenous nutrition solution and the total infusion volume per unit time according to the infusion formula adjustment command.

2. The edge computing based intelligent infusion control system according to claim 1, wherein, The data acquisition module includes: The recording interface unit is configured to extract basic health data of the premature infant from the electronic health record system, the basic health data including weight, gestational age and daily nutritional requirements baseline. The biosensor interface unit is configured to receive real-time physiological parameters acquired by the biosensor. The data integration unit is configured to aggregate the basic health data and real-time physiological parameters into an input dataset in a unified format, wherein the input dataset includes at least metabolic indicators related to tolerance per unit time. 3.The edge computing based intelligent infusion control system according to claim 2, characterized in that, The data acquisition module further includes a data preprocessing unit, which is configured to: Imput missing values ​​in the input dataset using interpolation; Abnormal data that exceeds the physiologically reasonable range is marked and corrected; Real-time physiological parameters of different dimensions are standardized to a preset range, with the nutrient metabolism rate data per unit time in the real-time physiological parameters being normalized separately.

4. The edge computing based intelligent infusion control system according to claim 3, wherein, The data preprocessing unit is further configured to: Time series smoothing processing is performed on continuously collected blood glucose data; Calculate the correlation coefficient between the fluctuation range of blood glucose and the amount of nutrition infusion and metabolic indicators per unit time. The correlation coefficient is used as one of the input features for the reinforcement learning model to evaluate the rationality of the infusion volume.

5. The edge computing based intelligent infusion control system of claim 1, wherein, The edge AI chip includes: The data receiving unit is configured to receive the input dataset after it has been processed by the data acquisition module. The model computation unit is configured to load the reinforcement learning model and perform computation, and simultaneously calculate the co-optimization parameters of nutrient composition ratio parameters and unit time infusion volume. The instruction generation unit is configured to convert the model calculation results into executable infusion formula adjustment instructions, which include the nutrient component ratio parameters of each component and the corresponding infusion volume per unit time parameters. 6.The edge computing based intelligent infusion control system according to claim 5, wherein, The steps of loading the reinforcement learning model and performing operations in the model computation unit include: The input dataset is segmented into features, with weight and gestational age from electronic health record data used as static features, and metabolic indicators from real-time biosensor data used as dynamic features. The weighting of data related to metabolic load per unit time in dynamic features is enhanced through attention mechanisms. The reinforced learning model is used to iteratively operate on the fused feature vectors, and output the optimal nutrient component ratio parameters and the corresponding safe infusion volume per unit time in the current state.

7. The edge computing based intelligent infusion control system according to claim 6, wherein, The model operation unit calculates the synergistic optimization parameters of the nutrient component ratio parameters and the infusion volume per unit time, including: The metabolic stability interval and the maximum tolerable infusion volume per unit time in the historical infusion data are used as the two-dimensional reward benchmarks; The metabolic index deviation value and the infusion volume deviation value are associated in real time, and when any deviation value exceeds the preset threshold, the optimization parameters are corrected by a penalty function; Based on the corrected parameters, a synergistic adjustment scheme is generated, which includes the infusion rate gradient of multiple nutrient components and the total dose matching relationship per unit time.

8. The edge computing based intelligent infusion control system of claim 1, wherein, The intelligent infusion pump includes: A multi-channel execution unit configured to independently control the infusion flow of multiple nutrient liquid components, and to synchronize the matching of the nutrient component ratio parameters and the infusion volume per unit time requirements; A feedback calibration unit configured to detect the deviation between the actual flow sum of each channel and the instructed infusion volume in real time; A closed-loop adjustment unit configured to dynamically correct the driving parameters of the multi-channel execution unit according to the deviation value, so that the actual infusion volume per unit time does not deviate from the instruction by more than a preset range. 9.The edge computing based intelligent infusion control system according to claim 1, wherein, It also includes an abnormality monitoring module, which includes: A threshold storage unit configured to pre-store the physiological parameter safety range threshold of premature infants and the maximum tolerable infusion volume per unit time threshold; A real-time comparison unit configured to perform double comparison between the biosensor data and the physiological safety threshold, and between the actual infusion volume and the maximum tolerable threshold; An emergency response unit configured to generate an alarm signal and trigger an infusion pause instruction when any threshold is broken. 10.The edge computing based intelligent infusion control system according to claim 9, wherein, The abnormality monitoring module further includes a trend prediction unit configured to: Perform synergistic trend analysis on the physiological parameters and the infusion volume per unit time collected continuously for a preset number of periods; Calculate the correlation coefficient of the physiological index change rate and the infusion volume adjustment rate; When the correlation coefficient exceeds the early warning threshold, generate an early warning signal for abnormal dose and metabolism and push it to the medical terminal.

Images