Pediatric infusion drop rate remote monitoring and risk correction system based on hierarchical diagnosis and treatment

Abstract

This invention discloses a remote monitoring and risk correction system for pediatric infusion drip rate based on hierarchical medical treatment, belonging to the field of monitoring and automated control technology. It includes: a data acquisition module that acquires image sequences of the infusion vessel, vibration sequences, patient's body acceleration sequences, drug solution environmental parameters, and medical order background data, and generates a patient profile vector; a fusion estimation module that outputs the original drip rate estimation sequence; a compensation module that generates body motion influencing factors and, combined with drug solution environmental parameters, generates accurate corrected drip rate data through fluid-structure interaction calculations; an inference module that constructs a medical risk graph structure and uses the body motion influencing factors to dynamically recalibrate edge weights, outputting a real-time risk score; and a decision module that inputs the risk score, accurate corrected drip rate, and body motion influencing factors into a pre-trained reinforcement learning model, retrieves the optimal adjustment step size, and outputs corrective control commands. This invention achieves precise monitoring and adaptive risk control of pediatric infusions.

Description

Technical Field

[0001] This invention relates to the field of monitoring and automated control technology, specifically to a remote monitoring and risk correction system for pediatric infusion drip rate based on hierarchical diagnosis and treatment. Background Technology

[0002] Pediatric infusion monitoring technology is an important component of clinical nursing and the hierarchical medical system. It aims to monitor the infusion rate of medications in real time through automated means, ensuring the safety and accuracy of medication administration to children. In recent years, with advancements in sensor technology and multimodal data fusion algorithms, infusion monitoring solutions based on image recognition or micro-vibration sensing have gradually become more widespread, achieving an initial transition from manual inspection to digital early warning.

[0003] Existing technologies mostly use photoelectric sensors or visual cameras installed on the side of the infusion pot to capture the instantaneous characteristics of droplet shedding to calculate the drip rate, and combine this with a common preset threshold to provide over-speed or drip stop alarms. These methods perform well in monitoring infusions in adults or in a static state, and have a certain real-time monitoring capability. However, for the special scenario of pediatric infusions, due to the young age of the children, their weak self-control and high pain sensitivity, they often have frequent rolling over, scratching, or large-amplitude limb swings and other non-steady-state body movements. This violent body movement not only brings a lot of motion artifacts and vibration noise to the visual sensors installed on the infusion pot, causing overlapping or missed detections in the counting logic, but also, at a deeper level, the displacement and acceleration changes of the limbs are transmitted to the drug-liquid interface through the infusion tubing, generating a fluid-structure interaction effect, which instantaneously changes the flow resistance characteristics of the drug-liquid and the pressure distribution inside the pot.

[0004] Current technologies fail to establish a deep compensation mechanism for the physiological and behavioral characteristics of pediatric patients. They often misinterpret artifacts caused by body movement as abnormal drip rates, or continue to use static sensing parameters even when the child's vigorous movement alters the flow resistance of the medication, resulting in a significant deviation between the acquired "apparent drip rate" and the "true flow rate" within the tubing. This data-level distortion is amplified step by step through the algorithm, causing subsequent risk assessment models to output incorrect risk scores due to a lack of dynamic correction based on the child's individual background information, ultimately triggering ineffective or even erroneous corrective instructions. This not only increases the false alarm rate in clinical nursing but also may lead to infusion reactions or drug penetration risks due to improper medication adjustments, limiting the reliable application of remote monitoring of pediatric infusions in complex clinical environments. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a remote monitoring and risk correction system for pediatric infusion drip rate based on hierarchical medical treatment.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] This invention discloses a remote monitoring and risk correction system for pediatric infusion drip rate based on hierarchical medical treatment, comprising:

[0008] The data acquisition module is used to simultaneously acquire the image sequence and vibration sequence of the infusion bottle, the body acceleration sequence of the child, the drug solution environment parameters and the background data of the child's medical orders, and generate a child profile vector based on the background data of the child's medical orders.

[0009] The fusion estimation module is used to extract multimodal features from the image sequence and the vibration sequence, and to fuse the multimodal features through a preset adaptive weight allocation network to output the original drop rate estimation sequence.

[0010] The compensation module is used to generate a body motion influence factor based on the correlation analysis between the child's body motion acceleration sequence and the original drip rate estimation sequence; if the body motion influence factor exceeds a preset perturbation threshold, it triggers interference stripping calculation to eliminate body motion artifacts in the original drip rate estimation sequence; at the same time, based on the coupling mapping between the drug solution environment parameters and the body motion influence factor, it generates true corrected drip rate data.

[0011] The inference module is used to construct a medical risk graph structure using the patient profile vector, and inject the real corrected drip rate data into the corresponding drip rate node of the medical risk graph structure; dynamically recalibrate the edge weights between nodes in the medical risk graph structure using the body motion influence factor, and output a real-time risk score; if the real-time risk score exceeds a preset safety threshold, a trigger signal is generated.

[0012] The decision module is used to receive the trigger signal, input the real-time risk score, the actual corrected drip rate data and the body motion influencing factor into the pre-trained reinforcement learning model, retrieve the optimal adjustment step size in the preset adjustment action space, and output the correction control command.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0014] 1. This invention achieves comprehensive perception of the multi-dimensional physical characteristics of the infusion process by constructing a multi-source heterogeneous data synchronous acquisition mechanism. The data acquisition module simultaneously acquires image sequences of the infusion vessel, vibration sequences, the child's body acceleration sequences, drug solution environmental parameters, and medical order background data, providing a rich and mutually corroborating data foundation for subsequent processing. The fusion estimation module performs multi-modal feature extraction and adaptive weight fusion on the images and vibration sequences, enabling stable original drip rate estimates to be obtained even under complex environments such as different lighting and occlusion conditions. This avoids the reliability issues of single sensing methods in pediatric infusion scenarios, thereby improving the basic accuracy of drip rate monitoring.

[0015] 2. This invention introduces the concept of body motion influencing factors through a compensation module. Based on the correlation analysis between the child's body motion acceleration sequence and the original drip rate estimation sequence, it quantifies the degree of interference of the child's involuntary body movements on drip rate monitoring. When the body motion influence exceeds a preset threshold, interference stripping calculation is triggered to eliminate body motion artifacts in the original drip rate. Simultaneously, the coupled mapping of drug solution environmental parameters and body motion influencing factors generates accurate corrected drip rate data. This design of stripping interference at the physical mechanism level ensures that the final drip rate data can truly reflect the physiological state during the infusion process, providing accurate data support for subsequent risk assessment.

[0016] 3. This invention constructs a medical risk graph structure using the child's profile vector and dynamically recalibrates the edge weights of nodes in the medical risk graph structure by injecting the actual corrected drip rate. This enables risk assessment to evolve from static thresholds to individualized dynamic features. By dynamically adjusting the risk transmission weights using body movement influencing factors, the risk score can accurately map the child's physiological tolerance and safety level under different body movement states. Combined with a pre-trained reinforcement learning model, the optimal adjustment step size is retrieved in the action space. This ensures that the corrective control command not only has the ability to adjust for drip rate deviations but also takes into account the execution stability under the child's body movement state. This achieves a closed-loop feedback from environmental perception to precise control, improving the accuracy of intelligent intervention in the context of hierarchical medical resource allocation. Attached Figure Description

[0017] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals refer to the same parts. Wherein:

[0018] Figure 1 This is a system module connection diagram of the present invention;

[0019] Figure 2 This is a flowchart of the workflow steps of the present invention. Detailed Implementation

[0020] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0021] In existing technologies, pediatric infusion monitoring largely relies on single photoelectric counting or image recognition algorithms to capture the droplet shedding process, making it difficult to balance robustness and data accuracy. Traditional methods are susceptible to motion artifacts and mechanical vibrations when children experience high-frequency, random limb movements, leading to overlapping or missed drip rate counts and consequently systemic monitoring biases. Existing equipment cannot simultaneously perceive the impact of changes in the child's physiological behavior on the drug dynamics within the tubing. Especially when vigorous movement causes nonlinear fluctuations in drug flow resistance, fixed drip rate discrimination models may logically fail, making it difficult to meet the needs of precise drug administration and risk warning in pediatric tiered healthcare systems.

[0022] To address the aforementioned issues, the study discovered a strong correlation between the child's body acceleration and drip rate monitoring artifacts. By establishing a body acceleration-drip rate correlation analysis model and introducing interference stripping calculations, accurate compensation for motion noise was achieved. Further findings revealed that the image mode exhibits high accuracy in capturing droplet morphology, while the vibration mode demonstrates good stability in physical stripping actions. Therefore, a method for dynamically fusing multimodal features through an adaptive weight allocation network was proposed. Further experimental verification involved introducing a weight recalibration mechanism between the child's medical order background-constructed profile vector and the medical risk graph structure into the risk reasoning process. A reinforcement learning model was then used to retrieve the optimal adjustment step size within the dynamic action space, forming a perceptual self-evolution and correction closed-loop feedback system.

[0023] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] Example:

[0025] like Figure 1 As shown, the pediatric infusion drip rate remote monitoring and risk correction system based on hierarchical medical treatment includes:

[0026] The data acquisition module is used to simultaneously acquire image and vibration sequences of the infusion bottle, body acceleration sequences of the child, drug solution environmental parameters, and background data of the child's medical orders, and generate a child profile vector based on the background data of the child's medical orders.

[0027] The fusion estimation module is used to extract multimodal features from image sequences and vibration sequences, and to fuse these features through a preset adaptive weight allocation network to output the original drop rate estimation sequence.

[0028] The compensation module is used to generate a body motion influence factor based on the correlation analysis between the child's body motion acceleration sequence and the original drip rate estimation sequence; if the body motion influence factor exceeds a preset perturbation threshold... This triggers interference stripping calculations to eliminate body motion artifacts in the original drip rate estimation sequence; simultaneously, based on the coupled mapping of drug solution environmental parameters and body motion influencing factors, true corrected drip rate data is generated.

[0029] The inference module is used to construct a medical risk graph structure using the patient's profile vector and inject the real corrected drip rate data into the corresponding drip rate node of the medical risk graph structure; it uses the body motion influence factor to dynamically recalibrate the edge weights between nodes in the medical risk graph structure and outputs a real-time risk score; if the real-time risk score exceeds the preset safety threshold, a trigger signal is generated.

[0030] The decision module receives trigger signals and inputs real-time risk scores, actual corrected drip rate data, and body motion influencing factors into a pre-trained reinforcement learning model. It then retrieves the optimal adjustment step size within a preset adjustment action space and outputs correction control commands.

[0031] like Figure 2 As shown, the system integrates a closed-loop architecture with multimodal sensing capabilities and nonlinear correction resilience into a medical IoT edge gateway, aiming to solve the problems of monitoring data artifacts and regulatory instability caused by random body movements of children during pediatric infusion.

[0032] The data acquisition module utilizes the embedded processor's multi-parallel peripheral interfaces (such as MIPI-CSI (Mobile Industry Processor Interface - Camera Serial Interface) for images, and I2C (Internal Integrated Circuit Bus) / SPI (Serial Peripheral Interface) for accelerometers) to simultaneously acquire image and vibration sequences from the infusion bottle and the child's body acceleration sequence. During the initialization phase, the system retrieves the child's medical order background data from the Hospital Information System (HIS) and generates a child profile vector using an embedded lightweight feature extraction operator. The child profile vector includes at least:

[0033] 1. Basic physiological tolerance threshold: The safe drip rate range is set based on the child's age, weight, and allergy history, usually between 20 gtt / min and 60 gtt / min;

[0034] 2. Activity rhythm factor: A dynamic weighted parameter generated based on the frequency of agitation in the child's previous nursing records, with an initial value of 0.4.

[0035] The fusion estimation module extracts multimodal features from the image and vibration sequence through an embedded neural network processor (NPU). Specifically, the system constructs an adaptive weight allocation network, the core logic of which is to monitor the signal-to-noise ratio of the visual signal in real time. When the infusion bottle shakes and causes the image to blur, the system dynamically reduces the weight of the image features and simultaneously increases the weight of the vibration sequence based on the piezoelectric sensor, and outputs the original drip rate estimation sequence through complementary fusion.

[0036] The adaptive weight allocation network deployed in the fusion estimation module is designed to dynamically adjust the contribution of multimodal features based on the quality of the real-time perceived environment, ensuring the atomic extraction accuracy of the original drip rate estimation sequence under all operating conditions. The input features of this network consist of high-dimensional representations of the image branch and the vibration branch. Specifically, for the synchronously acquired infusion bottle image sequence, the system first extracts spatial feature maps through the backbone of a lightweight convolutional neural network (such as MobileNetV3 or EfficientNet-Lite), and focuses on the spatiotemporal changes of droplet formation and falling within the ROI (Region of Interest), generating a 128-dimensional dynamic image feature vector. Simultaneously, the vibration sequence (usually acquired by a piezoelectric thin-film sensor attached to the wall of the infusion bottle) is subjected to a short-time Fourier transform (STFT) to convert it into a two-dimensional spectrum containing time-frequency domain information. Then, an encoder consisting of a two-layer temporal convolutional network (TCN) extracts the rhythmic features of the vibration events, outputting another 64-dimensional vibration feature vector. The two feature vectors are concatenated to form a 192-dimensional joint feature space, which serves as the original input to the adaptive weight allocation network.

[0037] At the network architecture level, this adaptive weight allocation network employs a three-layer fully connected structure for dynamic weight prediction: the first layer is a fully connected layer containing 128 neurons with a ReLU activation function, used for initial fusion of nonlinear correlations between multimodal features; the second layer is a compression layer containing 64 neurons, designed to extract core confidence representations in the current perceptual environment; the third layer is an output layer containing 2 neurons, normalized by a Softmax function, ultimately outputting a set of weight coefficients summing to 1, corresponding to the contribution ratios of the image mode and vibration mode in the current fusion. This network structure design balances inference efficiency in embedded environments with sensitivity to dynamic adjustment, with fewer than 50K parameters and a single inference time on the edge NPU controlled within 5ms, ensuring real-time response to drop rate changes.

[0038] The system employs a strategy combining offline pre-training and online fine-tuning. During offline training, a hybrid dataset of simulation and real-world measurements is constructed, containing over 100,000 sets of synchronized image-vibration-real drip rate labels. This dataset covers various interference scenarios that may be encountered in infusion monitoring, including but not limited to: gradual and abrupt changes in light intensity from 10 lux to 500 lux, random shaking of the infusion vessel within a frequency range of 2 Hz to 10 Hz simulated by a mechanical vibration table, and interference from different drug colors (such as transparent and milky white) on imaging. Real drip rate labels are synchronously acquired using a high-precision micro-flowmeter (accuracy ±0.1 gtt / min). During training, image and vibration features are input into the network, and the output weights are used to weightedly fuse the two features. The fused features are then used by a regression head to predict the final drip rate. The loss function uses mean squared error (MSE), with the deviation between the predicted drip rate and the actual flowmeter value serving as a supervisory signal, thereby backpropagating to optimize the parameters of the weight allocation network and the front-end feature extraction network. The optimizer used was Adam, with an initial learning rate of 0.001, 100 training epochs, and a batch size of 64. Early stopping was introduced to prevent overfitting. After offline training, the network weights were stored and burned into the memory of the edge computing unit.

[0039] The compensation module is used to physically reduce noise in the original sequences. The system calculates the cross-correlation coefficient between the child's body motion acceleration sequence and the drip rate sequence in real time, generating a body motion influence factor. If this factor exceeds a preset perturbation threshold... (Indicating that the child is in a state of vigorous movement such as scratching or rolling over), the system immediately triggers interference stripping calculations, using Kalman filtering or wavelet transform techniques to eliminate motion artifacts in the drip rate data. At the same time, the system retrieves drug solution environmental parameters such as the type of drug solution (e.g., hypertonic glucose, antibiotics) and real-time temperature, and generates accurate corrected drip rate data based on fluid-structure interaction empirical formulas, realizing the mapping from apparent monitoring to the actual physical state of the fluid.

[0040] The inference module constructs a medical risk graph structure using a pre-set knowledge graph in non-volatile memory. The system injects the "actual corrected drip rate" as a dynamic variable into the drip rate nodes of the graph structure and dynamically recalibrates the edge weights connecting the child's condition and the medication risk using a body motion influencing factor. A real-time risk score is output through first-order neighborhood aggregation using a graph neural network (GNN). A trigger signal is generated only when the score exceeds a safety threshold dynamically adjusted based on the tiered healthcare system's capacity (e.g., the resource differences between primary healthcare centers and tertiary hospitals).

[0041] The decision module receives the trigger signal and activates a pre-trained reinforcement learning model (such as DQN or PPO algorithm). This model uses the risk score, actual drip rate, and body motion factor as state inputs, and performs a greedy search within a preset pump pressure adjustment action space to retrieve the optimal adjustment step size. The injection step size or weight adjustment step size for a single bias parameter is preferably set to 5% of the total. This parameter balances the system's evolution speed and stability. During the initial system deployment and debugging phase, it can be set to 8% to 12% to accelerate the environmental adaptation process; during stable system operation, it is recommended to reduce it to 2% to 5%, using a small step size iteration strategy to prevent oscillations in the perception logic caused by abnormal isolated samples. Finally, the system outputs a correction control command to the stepper motor actuator through the PWM (Pulse Width Modulation) interface to achieve precise physical reset of the drip rate.

[0042] like Figure 2 As shown, after the system completes the correction command, it updates the activity rhythm factor in the child's profile vector through a feedback loop, forming a closed-loop learning mechanism. This data-driven reconstruction of the monitoring logic ensures the atomic execution of key drug administration commands and avoids scheduling instability under high-frequency motion disturbances.

[0043] The system adds an execution-perception causal decoupling verification layer to the closed-loop circuit. This layer is applied after instruction execution. ( Within a cycle of (seconds), the system compares the encoder feedback of the stepper motor. (Calculated from the number of steps of the stepper motor issued by the controller, for example, the motor adjusts the rotation of the pressure roller) Theoretically, an increase of 5 gtt / min in the drip rate corresponds to... ) and displacement increments observed by visual modality (exist (The average drip rate increment calculated within the period).

[0044] like (Consistency threshold, base value set as) If a hardware-level fault is detected (such as a kinked pipe, a lost motor step, or blood returning to the needle), the system will lock the sensing module parameters and not adjust them, triggering a hardware alarm command. Only when the two show strong consistency can the deviation be determined to originate from the sensing accuracy, and the deviation can be fed back to the weighted allocation network for fine-tuning. This causal verification logic cuts off the coupling propagation of errors between the sensing layer and the execution layer, preventing the system from entering a vicious cycle of performance degradation.

[0045] This application achieves predictability and resilience in drip rate control within the complex pediatric medical environment through data-driven reconstruction of infusion monitoring logic. It effectively solves the problems of signal fragmentation and monitoring artifacts caused by high-frequency random body movements in children, ensuring the atomic extraction of the original drip rate perception under unsteady physical conditions, thus avoiding data distortion caused by vigorous movement interference in traditional monitoring methods. The system ensures that the risk score objectively reflects the child's true medication needs under different intensities of body movement, improving monitoring accuracy in environments with limited resources in tiered medical care. Through closed-loop decision-making by a reinforcement learning model and real-time retrieval of the optimal adjustment step size, the system possesses adaptive correction capabilities for complex clinical scenarios, achieving nonlinear smooth control of the infusion drip rate. This ensures timely response to correction commands and provides an automated, self-reflective guarantee solution for precise pediatric medication administration.

[0046] This application further proposes that the specific steps for generating body motion influencing factors based on the correlation analysis between the child's body motion acceleration sequence and the original drip rate estimation sequence include:

[0047] Within a preset sliding time window, the child's body acceleration sequence and the original drip rate estimation sequence are normalized.

[0048] Calculate the covariance of the normalized child's body acceleration sequence and the original drip rate estimation sequence under different time lags, and construct the cross-correlation matrix;

[0049] The largest eigenvalue in the cross-correlation matrix is ​​extracted as the body motion influence factor.

[0050] The system calculates the body motion influence factor based on the temporal coupling between the child's body motion acceleration sequence and the original drop rate estimation sequence, using a pre-defined correlation analysis algorithm. This factor characterizes the intensity of the nonlinear perturbation caused by limb movement during droplet shedding. Specifically, the system uses the largest eigenvalue of the cross-correlation coefficient matrix for quantification, and the calculation model is as follows:

[0051] ;

[0052] in, This represents the calculated body motion influence factor, and its value is usually set between 0 and 1;

[0053] The normalized sequence of body motion accelerations of the child;

[0054] For those with time lag The normalized original drop rate estimation sequence;

[0055] m is the preset sliding time window number of steps, which is usually set between 10 and 50 sampling points based on the common limb tremor frequency in pediatric clinical practice, preferably 20.

[0056] After determining the factors influencing body motion, the system maps the statistical distribution characteristics of the cross-correlation matrix to the associated interference stripping gain through a mapping matrix. This step decouples the question from "how much body motion is involved" to "how much data should be compensated." To ensure the accuracy of the mapping, the correlation compensation model can be trained offline using a multilayer perceptron (MLP). The MLP structure consists of a 128-node input layer (corresponding to the expanded dimension of the cross-correlation matrix), a 64-node hidden layer, and a 1-node output layer (output compensation coefficients). The loss function used is mean squared error (MSE). The training conditions are set as follows: number of iterations N = 2000, learning rate... The weight matrix generated through this training process is stored in the memory of the edge computing unit, enabling the system to perform calculations based on the feature values ​​generated in real time. With preset disturbance threshold (Based on data from 100 pediatric patients, the range was set to 0.65-0.80, with 0.70 being the typical value) to perform rapid interference identification.

[0057] Through the above technical solutions, this application achieves the standardized transformation of physical motion data into sensory correction parameters, ensuring the timeliness and accuracy of the system's response to the non-steady-state behavior of the child. By calculating the covariance of multiple hysteresis and extracting eigenvalues, the system can eliminate randomly occurring minor tremor noise and generate influencing factors only for intense physical motion events with strong coupling characteristics, significantly reducing the computational load of the correction module under normal infusion conditions. Simultaneously, the identification mapping mechanism decouples the motion feature source from the fluid compensation algorithm, enhancing the universality of the monitoring framework across different age groups and activity levels of children, providing a highly confident physical basis for subsequent real drip rate correction.

[0058] This application further proposes that the specific steps for generating accurate corrected drip rate data based on the coupled mapping of drug solution environmental parameters and body motion influencing factors include:

[0059] Based on the drug type identifier and its real-time temperature data in the drug environment parameters, the corresponding flow resistance compensation coefficient is matched from the preset flow resistance database.

[0060] The volumetric influence factor and the flow resistance compensation coefficient are input into a preset fluid-structure interaction empirical formula for coupling calculation, and the actual corrected drip rate data is output.

[0061] Among them, the fluid-structure interaction empirical formula is pre-constructed based on the principles of fluid dynamics and is used to simulate the dynamic influence of the drug liquid sloshing caused by the child's body movement on the drug liquid flow resistance characteristics.

[0062] Specifically, the system uses an empirical formula for fluid-structure interaction for correction, as follows:

[0063] ;

[0064] in, The actual corrected drip rate data obtained from the calculation;

[0065] The original drop rate estimation sequence output by the fusion estimation module;

[0066] The flow resistance compensation coefficient is obtained by searching based on the drug type ID (such as electrolyte, macromolecular colloid) and real-time temperature T. Its value usually fluctuates between 0.85 and 1.15.

[0067] It is a fluid-structure interaction operator; As a factor affecting body movement, Let be the dynamic viscosity of the liquid drug. This fluid-structure interaction operator... This is used to characterize the velocity distortion of a drug solution in an infusion line under the influence of centripetal and inertial forces generated by limb movement. Specifically, it involves fluid-structure interaction empirical operators. An exponential expression based on fluid dynamics principles is used, which is a parameterized engineering approximation of the fluid-structure interaction effect of an infusion system under forced oscillation (the child's body movement). Its core physical idea is that the oscillation of the infusion pot and tubing generates an additional inertial force field, which can be equivalent to a dynamic pressure gradient, thereby changing the effective driving force of the medication flowing through the drip chamber and tubing. The formula is as follows:

[0068] ;

[0069] in, As a factor influencing body movement;

[0070] The dynamic viscosity of the liquid (unit: mPa·s) can be found in the physical property database based on the type of liquid and the real-time temperature.

[0071] For reference viscosity, the dynamic viscosity of water at 20℃ is taken. ), used for dimensionless processing;

[0072] k is a preset coupling strength coefficient, which reflects the gain of volumetric strength on flow resistance disturbance. It is calibrated through mechanical simulation experiments in typical infusion scenarios, and its value range is usually 0.2~0.3.

[0073] n is the viscosity influence index, which is determined based on the Reynolds number and disturbance propagation law in fluid mechanics. Its value ranges from 0.4 to 0.6. For liquids with high viscosity (such as fat emulsions), n can be taken at the upper limit to enhance the correction range.

[0074] By introducing a nonlinear combination of body movement factors and drug viscosity, the additional inertial force and pressure fluctuations generated when a child's limb movements are transmitted through the infusion tube to the liquid surface in the vessel can be effectively simulated, thereby calibrating the apparent drip rate to a value closer to the actual flow rate. In practical applications, the specific values ​​of k and n can be fine-tuned according to the characteristics of different age groups of children. For example, for active children with thin blood vessels, the value of k can be appropriately increased to improve compensation sensitivity. Fluid-structure interaction empirical operator The calculation process is completed in real time in the edge computing unit, ensuring that the delay of the corrected drip rate is less than 50ms, which meets the real-time requirements of closed-loop control.

[0075] After determining the flow resistance compensation coefficient, the system maps the acquired physical environment characteristics to associated fluid dynamics correction gains using a mapping matrix. Through this technical solution, this application achieves standardized transformation of environmental sensing data into precise medical-grade flow indicators, ensuring the timeliness and accuracy of drug administration rate monitoring in dynamic infusion scenarios. By matching the flow resistance database and calculating using fluid-structure interaction empirical formulas, the system can eliminate instantaneous flow velocity artifacts caused by changes in drug viscosity or slight limb movement, generating correction data only for significant flow regime shifts affecting clinical safety, reducing the redundant computational load on the backend inference module under changing environmental conditions. Simultaneously, the parameter mapping mechanism decouples the sensor's underlying data from high-level pharmacological risk assessment, enhancing the monitoring framework's adaptability to drugs of different viscosities and complex ward temperature control environments, providing high-confidence state feedback for subsequent optimal adjustment step size retrieval.

[0076] Furthermore, the edge weights between nodes in the medical risk graph structure are adaptively adjusted based on the child's risk sensitivity data. This mechanism dynamically corrects the "sensitivity" of risk transmission by monitoring the child's physiological homeostasis under different body movement disturbances in real time, ensuring that the system maximizes the reduction of false alarm interference while maintaining monitoring sensitivity. When dynamically recalibrating the edge weights between nodes in the medical risk graph structure using body movement influence factors, the following steps are performed:

[0077] Extract the initial weight vector connecting the drip rate node and the adjacent risk node in the medical risk graph structure;

[0078] Construct an exponential mapping function or a linear gain operator using the body motion influence factor;

[0079] By multiplying the initial weight vector using an exponential mapping function or a linear gain operator, the risk weights of abnormal drip rates in a specific patient context can be nonlinearly amplified or attenuated.

[0080] Specifically, in another preferred embodiment of the present invention, the system extracts the initial weight vector connecting the drip rate node and the adjacent risk node in the medical risk graph structure. Simultaneously, the system monitors the residual volatility of the drip rate within the most recent time window (e.g., 10s to 60s), i.e., the dispersion between the actual corrected drip rate and the target drip rate ordered by the physician within that time window, to characterize the physical stability of the current drug infusion. The system calculates the recalibration gain value according to a preset risk sensitivity assessment formula. The formula is as follows:

[0081] ;

[0082] in, As a factor influencing body movement;

[0083] The residual volatility is the drop rate.

[0084] The sensitivity weight is preset, and its value is determined based on the clinical classification and monitoring priority of the child. For example, in monitoring scenarios for high-risk children (such as extremely low birth weight infants requiring extremely low-rate, precise drug delivery), a sensitivity weight can be set. It can suppress interference noise by giving higher weight to body motion; in the case of monitoring in a general ward, it can be configured in reverse to improve the response speed to flow velocity deviation.

[0085] To achieve smooth weight adjustments, dynamic recalibration of edge weights typically employs incremental saturation control logic, setting the step size for each adjustment to be within the total weight range. To prevent drastic fluctuations in risk scores, the parameters of this adjustment model can be calibrated on a clinical simulation platform. Training conditions include: simulating a drip rate increase from 10 gtt / min to 100 gtt / min, superimposed with mechanical vibration interference of different frequencies, and determining the optimal preset sensitivity weights through offline learning of no fewer than 2000 sets of clinical risk evolution sequences. The combination of .

[0086] Furthermore, an exponential mapping function is constructed using the recalibration gain to apply the initial weight vector connecting the drip rate node and its adjacent risk nodes. A nonlinear adjustment is performed to obtain the recalibrated weight vector.

[0087] ;

[0088] in: This is the initial weight vector connecting the drip rate node and adjacent risk nodes (such as drug extravasation risk, cardiopulmonary overload risk, etc.) in the medical risk graph structure.

[0089] v is a preset scaling parameter (usually taken as 0.3~0.7), used to control the sensitivity of weight adjustment;

[0090] It is an exponential function with the natural constant e as its base, which enables nonlinear amplification or decay of the weights.

[0091] When the body movement influencing factors When the child is older (during strenuous exercise), As it increases, the initial weights are adjusted through exponential mapping. The system makes corresponding adjustments to dynamically correct the transmission path of abnormal drip rates in risk assessment: if vigorous body movement leads to a decrease in the reliability of the original drip rate, the system reduces the transmission weight from the drip rate node to the risk node through an exponential decay mechanism, establishing a protective barrier against "motion spurious errors"; conversely, if the child is at rest and the drip rate fluctuates significantly, the system increases the risk weight through an exponential amplification mechanism to ensure the system's sensitive response to real drip rate abnormalities.

[0092] In another preferred embodiment, a linear gain operator can also be used instead of an exponential mapping, i.e. ,in This represents the linear gain coefficient. Both implementations aim to achieve dynamic adaptive adjustment of the edge weights according to the body's dynamic state; the specific choice depends on the clinical scenario's requirements for the degree of nonlinearity.

[0093] By introducing body movement factors as feedback variables, this application transforms static risk assessment indicators into dynamic mapping parameters with "behavioral awareness," establishing a dynamic game balance between risk monitoring sensitivity and anti-interference resilience. This adaptive recalibration based on physiological behavioral feedback enhances the operational stability of the monitoring system in dynamic infusion environments, ensuring that medical monitoring resources can always respond optimally according to the real-time threat intensity of the child.

[0094] Furthermore, the system utilizes a graph neural network (GNN) feature propagation mechanism to perform deep feature evolution on the medical risk graph structure, outputting a real-time risk score. This mechanism dynamically reflects the transmission and superposition effects of risk states within the topological space by aggregating heterogeneous features from multiple source nodes in real time, ensuring that the system can capture weak risk precursors and suppress transient monitoring noise.

[0095] Specifically, the system will accurately correct the drip rate data. Real-time environmental parameters and patient profile vectors This is mapped to the corresponding node space of the medical risk graph structure. For each node, feature information of neighboring drop rate nodes is collected along the recalibrated edge weight path. Each node not only carries a single physical indicator, but is also transformed into a high-dimensional initial node feature matrix through the embedding layer. The system monitors the feature consistency entropy value within the most recent period, i.e., the similarity of feature distributions among neighboring nodes, to characterize the clustering quality of the system's current risk features. The system calculates the node update gain according to a preset aggregation game formula:

[0096] ;

[0097] in, This is the updated drip rate node state vector;

[0098] These are the edge weights after recalibration using the body motion influence factor;

[0099] and Preset feature retention and feature absorption weights. For example, in scenarios involving viscous drug solutions with complex infusion dynamics models, the following settings can be configured. To maintain the stability of the node's own state and prevent score oscillations caused by noise interference; in ICU monitoring scenarios where drug concentration is extremely sensitive, weights can be configured in reverse to enhance the sensitivity of capturing changes in neighboring risk nodes.

[0100] During the aggregation process, the system employs a weighted summation aggregation function combined with a nonlinear activation function (such as LeakyReLU or ELU) to perform a nonlinear transformation on the state vector. If the calculated characteristic consistency entropy value exceeds the preset steady-state reference range (usually set at...), the system will proceed as follows: Within the specified range, this indicates a drastic change in characteristics among the current risk nodes, potentially teetering on the brink of drip rate runaway or sensor failure. In this situation, the system automatically increases the slope parameter of the nonlinear activation function, amplifying feature differences through more sensitive mapping logic, and updates the state vector of each drip rate node, ensuring the global state vector can quickly respond to sudden local risks.

[0101] To ensure the smoothness and reliability of the risk score output, the mapping from the global state vector to the real-time risk score employs a moving average recalibration logic, setting the amplitude of a single score change to be limited by the historical average. The parameter calibration of this mapping model can be performed on a highly realistic pediatric infusion risk database. Training conditions include: covering no fewer than 50 different drug flow rate curves, simulating 100 typical medical risk evolution events (such as needle displacement, tubing kinking, and osmosis), and determining the optimal feature retention weights through offline learning of no fewer than 5000 sets of spatiotemporal graph sequences. With feature absorption weights The combination of .

[0102] In this embodiment, the constructed medical risk graph structure G=(V,E) is defined as follows: the node set V includes drip rate nodes, patient physiological state nodes (such as heart rate, body temperature), and risk event nodes (such as drug extravasation, phlebitis). The initial weights of the edge set E are set based on prior probabilities in the clinical knowledge base.

[0103] The construction of the clinical knowledge base integrates the following two types of publicly available data sources:

[0104] 1. Structured clinical guidelines and expert consensus: such as evidence-based guidelines for the diagnosis and management of acute fever of unknown etiology in children aged 0 to 5 years in China, and standards for intravenous infusion therapy. From these, qualitative correlations between risk events (such as drug extravasation, phlebitis, and febrile reactions) and relevant physiological parameters (abnormal drip rate, heart rate, and body temperature) are extracted and transformed into nodes and edges in a graph structure.

[0105] 2. Publicly available medical event statistics databases: Statistical data from publicly available clinical databases such as MIMIC-III or published retrospective studies of adverse events following pediatric intravenous infusions are used to initialize edge weights (i.e., prior probabilities). For example, statistics from a pediatric ward might show that when the infusion rate exceeds the set value by 50%, accompanied by pediatric agitation, the statistical probability of drug extravasation is approximately 0.3-0.5. This probability range can serve as a reference for assigning initial edge weights connecting the "excessive infusion rate" node and the "risk of drug extravasation" node.

[0106] In the implementation of graph neural network inference, a two-layer GraphSAGE (Graph Sample and Aggregation) framework is adopted. The first graph convolutional layer maps the input node features (64-dimensional) to a 128-dimensional hidden layer, using the LeakyReLU activation function (with a negative slope of 0.2). The second layer maps the 128-dimensional hidden layer to the final 32-dimensional state vector. For neighborhood aggregation, a mean aggregator is used, which averages the features of neighboring nodes and then concatenates them with the features of the current node.

[0107] To prevent overfitting, a Dropout layer was added between the two layers, with a dropout rate of 0.3. The entire graph neural network was trained offline using the Adam optimizer with a learning rate of 0.001 for 200 epochs on a dataset containing 5000 simulated infusion risk events, employing a cross-entropy loss function to optimize the accuracy of risk scoring.

[0108] By introducing neighborhood feature aggregation as a feedback variable, this invention transforms isolated drip rate monitoring indicators into global scoring indicators with "spatiotemporal correlation," establishing a dynamic game balance between local physical fluctuations and overall medical risk: in high-noise environments, a monitoring barrier is established by shrinking feature absorption weights; in high-risk environments, early warning computing power is released by expanding weights. This adaptive scoring based on graph structure evolution enhances the diagnostic consistency of the system in heterogeneous hierarchical medical environments, ensuring that limited corrective decision-making resources can always be optimally allocated according to real-time risk levels.

[0109] In another preferred embodiment of the present invention, the system implements real-time self-auditing of the trigger signal execution effect through a closed-loop consistency verification module. This mechanism dynamically corrects the "biased cognition" of the perception and inference layers by monitoring the degree of deviation between the feedback results of the physical execution end and the logical expectations, ensuring that the system can maintain the closed-loop robustness of regulation even under hardware aging or extreme environmental interference.

[0110] Specifically, the system acquires the feedback drip rate sequence collected in real time by sensors after the execution of the correction control command. Simultaneously, the system calculates the feedback sequence and the target adjustment amount in the corrective control command. instantaneous deviation between To characterize the collaborative quality during execution, the system monitors the execution robustness factor within the most recent control cycle (e.g., 2s to 5s). This refers to the ratio of the variance in the feedback drip rate to the adjustment amount. The system calculates the execution confidence index based on a preset consistency evaluation function.

[0111] ;

[0112] in, The calculated execution confidence index is set to a value between 0 and 1.

[0113] This is the instantaneous deviation value;

[0114] To implement the robustness factor;

[0115] A preset consistency weight is used. For example, in micro-drug administration scenarios that emphasize high-precision infusion, the system is set... This enhances sensitivity to the absolute value of deviations; however, in mobile transport scenarios with severe environmental interference, adjustments can be made appropriately. Weights are added to allow for a wider range of physical oscillations.

[0116] The calculated execution confidence index is compared in real time with a preset consistency threshold range within the controller, which is typically fixed between 0.70 and 0.90. For high-precision micro-infusion pumps, it is recommended to set it within the range of 0.85 to 0.95 to achieve rigorous accuracy self-verification; for ordinary gravity infusion monitoring scenarios, it can be set within the range of 0.65 to 0.80. If the execution confidence index is lower than the lower limit of this threshold, it indicates that there has been a significant logical drift in the current perception layer output or inference layer decision. At this time, the system triggers a compensation mechanism correction instruction, synchronously feeding back the instantaneous deviation value as a bias parameter to the fusion estimation module (used to correct weight allocation bias) and the inference module (used to recalculate the risk graph node state).

[0117] By introducing execution confidence as a feedback variable, this invention transforms unidirectional instruction issuance into a bidirectional verification logic with "self-reflection capabilities," establishing a dynamic game balance between instruction expectation and physical response: when execution deviates, a perceptual defense is established through feedback bias; when precise synchronization is achieved, algorithmic computing power is released by maintaining parameters. This adaptive correction based on consistency verification enhances the anti-decay capability of the embedded monitoring system during long-term operation, ensuring that correction decisions are always based on a physical closed loop with high confidence.

[0118] In another preferred embodiment of the invention, the fusion estimation module achieves self-evolving gain adjustment of the perception layer by receiving correction instructions from a compensation mechanism. This mechanism dynamically corrects the bias in the multimodal fusion process by analyzing the distribution pattern between physical execution deviation and perception layer output, ensuring that the system can autonomously optimize perception accuracy when environmental features drift.

[0119] Specifically, the system extracts the instantaneous deviation value from the compensation mechanism correction instruction. And use a sliding window to extract the original drop rate estimation sequence within the most recent period. The system determines the correlation between the instantaneous deviation value and the distribution of the sensing sequence by calculating the probability distribution divergence of the two (such as KL divergence or JS divergence). To quantify the evolutionary dynamics of the sensing layer, the system monitors the modal confidence entropy in the current operating state, i.e., the balance between the contributions of image and vibration features. The system calculates the modal weight adjustment gain according to a preset self-evolutionary evaluation formula.

[0120] ;

[0121] in, Adjust the gain to match the calculated modal weights;

[0122] This is a measure of distribution correlation.

[0123] Modal confidence entropy;

[0124] To preset self-evolution weights. For example, in scenarios where drastic fluctuations in lighting conditions (such as lights being turned on in a hospital ward at night) cause image modality distortion, the system sets... This accelerates the response speed to deviation distributions; and in scenarios where sensor hardware performance is highly stable, the adjustable... The ratio is adjusted to maintain the long-term stability of the perceived weight.

[0125] After determining the adjustment gain, the system dynamically adjusts the weight ratios for the vibration sequence in the adaptive weight allocation network using a linear mapping operator. If the distribution correlation shows that the deviation value oscillates synchronously with the drop rate estimation, it indicates that the current vibration mode is less affected by limb interference and has higher residual explanatory power. In this case, the system actively increases the weight ratio of the vibration sequence to dynamically strengthen the logic of the perception layer. Conversely, if the correlation is weak, the system reduces the contribution of this mode by suppressing the operator to prevent the execution deviation from penetrating back into the perception layer.

[0126] To achieve stability through perceptual self-evolution, the dynamic adjustment of weight ratios employs an adaptive learning rate decay logic, setting the step size for each weight adjustment to be within a certain range of the total weights. To prevent oscillating evolution of the perception layer due to occasional execution jitter, the parameters of this evolutionary model can be calibrated on a multi-scene perception database. Training conditions include: simulated image noise from... Upgraded to To simulate data drift caused by sensor temperature drift, the optimal preset self-evolving weights are determined through offline learning of no fewer than 3000 sets of "bias-weight" mapping sequences. combination.

[0127] By introducing distribution correlation as a feedback variable, this application transforms the static feature fusion logic into a self-evolving framework with "perception and awakening" capabilities, establishing a dynamic game balance between execution layer errors and perception layer weights: modal defense is established through distribution determination when perception fails, and algorithm computing power is released through weight expansion when accuracy regresses. This bias-guided self-evolving adjustment enhances the embedded monitoring system's adaptive adjustment capability to heterogeneous infusion environments, ensuring that the perception layer can always reconstruct optimal features based on the real feedback from the execution end.

[0128] This application further proposes that the inference module also introduces the weight of hierarchical medical resources to make sensitivity correction to the real-time risk score. This mechanism quantifies the guarantee capacity of medical institutions as a regulating factor for risk transmission, dynamically corrects the alarm priority of the system for abnormal states, and ensures that in the environment of limited medical resources, the lack of intensity of manual monitoring can be compensated by increasing risk sensitivity.

[0129] Specifically, the system parses the current medical institution's grade code from the patient's medical order background data. And retrieve the corresponding medical resource guarantee coefficient from the preset database. This coefficient reflects the institution's comprehensive capacity in dimensions such as medical staff ratio, emergency response, and specialized equipment. Simultaneously, the system monitors the rate of risk evolution under the current monitoring pathway, i.e., the instantaneous change in risk score per unit time. The system calculates the final weighted risk score based on a preset sensitivity correction formula, as follows:

[0130] ;

[0131] in, This is the revised real-time risk score;

[0132] The original risk score output by the medical risk map structure;

[0133] This is a correction increment calculated based on the medical resource security coefficient (medical resource security coefficient and level code). Proportional, correcting the increment With grade coding Inversely proportional, meaning the lower the level, the greater the correction increment. (The higher)

[0134] Preset sensitivity weights. For example, in scenarios with low medical resource availability, such as primary healthcare centers, the system is set to... By positively increasing the correction increment to amplify the original risk score, even slight fluctuations in drip rate can trigger high-level warnings. In scenarios with well-established support systems, such as top-tier hospitals, the weight can be appropriately reduced to minimize invalid interference alarms.

[0135] After determining the weighted risk score, the system recalibrates the score using nonlinear mapping logic. If the calculated risk evolution rate exceeds the preset steady-state benchmark range (usually set at...), the system will take action. (Within the range of minutes / seconds), this indicates that the child's physiological state is in a period of rapid fluctuation. At this time, the system, combined with the constraint of a low protection coefficient, proactively lowers the preset safety threshold. This involves triggering corrective actions through more stringent access control logic. This step decouples the decision-making process from "physical risk assessment" to "resource matching assessment," ensuring adaptive alignment of monitoring strategies across different organizational levels. Preset security thresholds are also included. The optimal score is 65, which is dynamically set based on the current protection level of medical institutions: in resource-rich scenarios such as tertiary hospitals, the score can be increased within the range of 70 to 80 to reduce the interference of low-risk alarms on medical staff; in primary health centers or at the end of the hierarchical medical system, the score can be decreased within the range of 55 to 65.

[0136] By introducing the weights of tiered healthcare resources as feedback variables, this invention transforms static clinical indicator assessment into a dynamic response framework with "resource awareness" capabilities, establishing a dynamic game balance between system assessment accuracy and institutional support capabilities: at lower-level institutions, weight expansion builds stronger risk barriers, while at higher-level institutions, weight contraction releases collaborative computing power for medical staff. This sensitivity correction based on resource awareness enhances the deployment consistency of the embedded monitoring system within regional medical consortia, ensuring that medical intervention resources are always optimally allocated according to real-time risk levels and institutional support limits.

[0137] In another preferred embodiment of the invention, the decision-making module optimizes the reinforcement learning model online by constructing a multi-objective joint reward function. This mechanism dynamically corrects the smoothness and intensity of corrective actions by adjusting the weight of different dimensions of reward evaluation in real time, ensuring that the system can quickly eliminate drip rate risks and guarantee the physical safety of the infusion tubing in complex dynamic environments.

[0138] Specifically, the system acquires infusion tubing pressure data in real time through a data acquisition module. This is then transformed into a core indicator characterizing physical stability. Simultaneously, the system monitors the action cost consistency factor within the current adjustment cycle, i.e., the ratio of the pressure pulse caused by the corrective action to the step size adjustment. The system calculates the total reward value according to a preset multi-objective joint reward formula, as follows:

[0139] ;

[0140] in, To reinforce the feedback reward value of the learning model;

[0141] The first reward item is the reduction in risk level after executing the corrective action;

[0142] The second reward item is the stability of pressure fluctuation in the infusion line;

[0143] For dynamically adjusted weighting ratios, As a weight for the magnitude of the risk level decrease, Weighting is applied to stability fluctuations. For example, when the body movement factor indicates that the child is in a state of strenuous exercise, the system automatically increases the weighting. The weight (usually its proportion is increased to) Within a certain range, the system sacrifices some reset time in exchange for a more stable pressure output, preventing instantaneous high-pressure shocks from occurring during violent shaking; conversely, in scenarios where the child is at rest and the risk score increases sharply, the system increases pressure in the opposite direction. The weighting of the data accelerates the execution of corrective actions.

[0144] After determining the reward weighting, the system updates its parameters using the policy gradient of the reinforcement learning model. If the calculated pipeline pressure stability fluctuation falls below a preset steady-state reference range (usually set at...), the system will update its parameters accordingly. If the pressure fluctuation is within the rated pressure range, it indicates that the current corrective action may have triggered a risk of backflow or leakage in the pipeline. In this case, the system introduces a penalty term. (Where A is the action vector) The search space for actively constrained adjustment step size achieves decision decoupling from "aggressive correction" to "robust control." This step ensures that the trigger signal meets clinical indicators without exceeding the physical tolerance limit of the tubing.

[0145] The reinforcement learning model employs a Double DQN (Deep Double Q Network) architecture to improve training stability. Its network structure is a three-layer fully connected neural network: the input layer receives the state vector (dimension 10), the first hidden layer contains 128 neurons, and the second hidden layer contains 64 neurons, both using the ReLU activation function; the output layer's dimension is equal to the size of the action space (5 in this embodiment), corresponding to the Q-values ​​of each adjustment step.

[0146] The model training employs an experience replay mechanism, with an experience pool capacity of 10,000 experience tuples (s, a, r, s'). Each training iteration randomly samples 256 experiences for mini-batch updates. The target network is updated using a soft update method with a soft update coefficient τ set to 0.01. An ε-greedy exploration strategy is used, with an initial exploration rate ε set to 0.9, gradually decreasing to 0.05 with each training iteration. A discount factor γ is set to 0.95 to balance immediate and future rewards.

[0147] In the cloud server, the reinforcement learning model is periodically retrained using structured sample pairs uploaded from multiple edge devices to extract feature adjustment curves for specific diseases (such as pneumonia and diarrhea), and the updated network parameters are then distributed to each edge node.

[0148] Reinforcement learning reward function for decision-making module A 'clinical safety first' strategy was adopted. This was implemented after detecting body movement factors. In high-risk scenarios, the system automatically and dynamically increases the negative reward weight of 'dosage deviation'. ), and simultaneously lowered the weight of 'pressure stability' ( This ensures that the system corrects drip rate deviations as quickly as possible, rather than blindly pursuing stable pressure and maintaining an incorrect dosing rate.

[0149] Furthermore, the motion space is adjusted to an adaptive quantization step size based on the image vector: for newborn images (weight <5kg), the system activates a micro-adjustment mode, with the step size range set to... to For portraits of older children, the original style was restored. to gtt / min. This design achieves logical adaptation between the adjustment step size and the clinical infusion baseline value, ensuring precise individualization of correction actions.

[0150] By introducing a joint reward function regulated by body movement factors, this invention transforms the deviation adjustment of a single objective into a dynamic game framework with "anti-jitter execution," establishing a dynamic game balance between correction timeliness and pipeline safety: under intense body movement, pressure defense is established through weight transfer; under stable conditions, reset computing power is released through weight expansion. This adaptive decision-making based on multi-objective reward guidance enhances the continuity of the embedded monitoring system's actions in extreme motion scenarios, ensuring that the correction decision can always achieve optimal path retrieval while safeguarding the child's physiological safety.

[0151] In another preferred embodiment of the present invention, the system achieves deep data interaction between the edge monitoring terminal and the cloud management platform through a global state synchronization module. This mechanism transforms massive local discrete samples into global common knowledge, dynamically optimizes the generalization ability of the edge-side model, and ensures that the system can achieve "precise navigation" based on prior disease knowledge for different groups of children.

[0152] Specifically, the system encapsulates real-time corrected drip rate data, real-time risk scores, and corresponding corrective control commands into structured sample pairs with spatiotemporal attributes. Simultaneously, the system monitors the data transmission integrity factor under the current communication link, i.e., the ratio of successfully received acknowledgment messages to the total number of uploaded samples. The system calculates the knowledge update gain based on a preset cloud-edge synchronization evaluation formula.

[0153] ;

[0154] in, The calculated knowledge update gain;

[0155] For structured sample pairs;

[0156] This refers to the disease categories (such as pneumonia, dehydration, etc.) analyzed from the background data of medical orders.

[0157] The Euclidean distance between the parameters sent from the cloud and the current parameters on the edge;

[0158] α and β are preset synchronization weights. For example, in infusion monitoring scenarios involving rare diseases or complex complications, the system sets α=0.85 and prioritizes the use of feature adjustment curves extracted from large-scale cluster analysis in the cloud as constraints to compensate for insufficient sample size on the edge side; while in routine infusion scenarios, the β weight is increased to maintain the execution stability of the local model.

[0159] Upon receiving the feature adjustment curve from the cloud, the system transforms it into initialization parameters or state-space constraints for the reinforcement learning model. If the calculated knowledge update gain exceeds the preset steady-state baseline range (typically set within the range of [0.60, 0.85]), it indicates that the cloud model has captured a better rate control path for the specific disease. At this point, the system triggers the network parameter hot update logic, injecting the global feature curve into the local decision operator through a nonlinear mapping function, thus realizing the logical evolution from "individual experience learning" to "cluster wisdom sharing." This step ensures that the correction command, during execution, can both meet real-time physical feedback and align with clinical best practices.

[0160] By introducing feature adjustment curves as external feedback variables, this invention transforms isolated edge devices into collaborative units with "collective evolution perception" capabilities, establishing a dynamic game balance between the real-time nature of local perception and the global nature of cloud-based decision-making: in complex pathological scenarios, it establishes professional technical barriers by absorbing cloud constraints, while in routine scenarios, it releases local computing power through parameter fine-tuning. This adaptive update based on global state synchronization enhances the diagnostic consistency of the system in heterogeneous hierarchical diagnosis and treatment networks, ensuring that medical correction strategies can always evolve optimally based on the latest clinical clustering characteristics.

[0161] The following is a specific example of a remote monitoring and risk correction system for pediatric infusion drip rate based on hierarchical medical treatment:

[0162] A 5-year-old child with pneumonia (weighing 18 kg) was admitted to the hospital due to high fever and restlessness, and received intravenous ceftriaxone treatment at a local health center. A system using a MEMS accelerometer (sampling rate 100 Hz) attached to the child's mid-thigh on the non-injection side monitored the child's body movement in real time. When the child exhibited sustained leg kicking due to pain from the infusion (lasting 8 seconds, with a peak acceleration of 2.3 g), the system simultaneously acquired a sequence of images (30 fps) from a miniature camera on the infusion vessel and a piezoelectric film vibration signal (500 Hz). The adaptive weight allocation network in the fusion estimation module dynamically adjusted the modal weights based on the image signal-to-noise ratio (currently SNR=18 dB), reducing the image weight to 0.4 and increasing the vibration weight to 0.6. The fused output showed a raw drip rate estimation sequence, indicating a dramatic fluctuation in drip rate between 35 gtt / min and 52 gtt / min.

[0163] The compensation module performs correlation analysis on the original drip rate sequence and the body acceleration sequence, calculates the cross-correlation matrix within a sliding time window (m=20), and extracts the largest eigenvalue. As a factor influencing body movement, this value exceeds the preset perturbation threshold ( 0.65), triggering interference stripping calculation. The system retrieves the environmental parameters of the drug solution: the drug solution type is identified as ceftriaxone sodium (ID=CTRX), the real-time temperature is 26℃, and the flow resistance compensation coefficient α=1.08 is matched from the preset flow resistance database. The volumetric influence factor Γ=0.78 and the dynamic viscosity of the drug solution μ=1.2mPa·s are input into the fluid-structure interaction empirical formula. The calculated value is Φ = 1.18. (Original drip rate) Revised Formula Calculations showed that the actual corrected drip rate was calibrated to 42gtt / min ± 3gtt / min, which is basically consistent with the infusion pump setting of 40gtt / min, effectively eliminating the monitoring deviation caused by body motion artifacts.

[0164] The inference module extracts a profile vector from the child's medical order background data: age 5 years, weight 18 kg, negative allergy history, and clinical diagnosis of bronchopneumonia. Based on this profile vector, a medical risk graph structure is constructed, and the actual corrected drip rate of 42 gtt / min is injected into the corresponding drip rate node. An initial weight vector connecting the drip rate node to adjacent nodes such as "drug extravasation risk" and "cardiopulmonary load risk" is applied using a body motion influence factor Γ=0.78. Dynamic recalibration is performed, constructing the exponential mapping function f(Ω) = exp(0.5 × Ω), where Ω = κ1·Γ + κ2·σ² (κ1 = 0.7, κ2 = 0.3, σ² is the drop rate residual volatility of 0.15). The calculated Ω = 0.59, and the side weights become... After normalization, it is [0.60, 0.40]. The actual corrected drip rate and the image vector are mapped to the node space to generate the initial feature matrix. Neighborhood features are aggregated along the recalibrated edge weights. After first-order neighborhood aggregation and nonlinear activation of the graph neural network, the global state vector is mapped to a real-time risk score of 72 points (out of 100), which exceeds the preset safety threshold of the primary health care center. =65 points), the system generates a trigger signal.

[0165] After receiving the trigger signal, the decision-making module inputs the real-time risk score of 72, the actual corrected drip rate of 42 gtt / min, and the body motion influence factor of 0.78 into the pre-trained reinforcement learning model (based on the DQN algorithm). The model uses the current state vector as input and performs a greedy search within a preset adjustment action space (step size adjustment levels: -5, -3, 0, +3, +5 gtt / min), based on the reward function. (Currently experiencing intense physical activity) =0.7, =0.3) Retrieve the optimal action. Calculations show that selecting an adjustment step size of -3gtt / min maximizes the expected reward (both reducing the risk score and avoiding drastic fluctuations in pipeline pressure). The system outputs a correction control command to the stepper motor via the PWM interface, smoothly reducing the drip rate from 42gtt / min to 39gtt / min.

[0166] The closed-loop consistency verification module obtains the feedback drip rate sequence after instruction execution and calculates the instantaneous deviation Δv = 1.2gtt / min from the target adjustment (40gtt / min). Δv is then compared with the execution robustness factor. Input consistency evaluation function If the value is below the consistency threshold of 0.70, a compensation mechanism correction instruction is triggered, and the instantaneous deviation value Δv=1.2 is fed back to the fusion estimation module as a bias parameter. The weight ratio of the vibration sequence in the adaptive weight allocation network is dynamically adjusted (increased by 2%) to enhance the robustness of subsequent body motion perception.

[0167] This embodiment fully presents the application process of the system in a primary care pediatric infusion scenario. Through the quantification and comprehensive application of body movement influencing factors, precise compensation for drip rate monitoring in agitated children is achieved. Dynamic correction of risk thresholds based on the weights of tiered healthcare resources enables primary healthcare institutions to capture potential risks with greater sensitivity. The reinforcement learning model smoothly retrieves the adjustment step size under body movement disturbances, avoiding overshoot oscillations in traditional PID control under severe disturbances. The closed-loop consistency verification mechanism feeds back execution deviations to the perception layer, forming a complete intelligent closed loop of "monitoring-compensation-decision-verification-evolution," effectively improving the reliability and adaptability of pediatric infusion monitoring within the tiered healthcare system.

[0168] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A pediatric infusion drip rate remote monitoring and risk correction system based on hierarchical medical treatment, characterized in that, include: The data acquisition module is used to simultaneously acquire the image sequence and vibration sequence of the infusion bottle, the body acceleration sequence of the child, the drug solution environment parameters and the background data of the child's medical orders, and generate a child profile vector based on the background data of the child's medical orders. The fusion estimation module is used to extract multimodal features from the image sequence and the vibration sequence, and to fuse the multimodal features through a preset adaptive weight allocation network to output the original drop rate estimation sequence. The compensation module is used to generate a body motion influence factor based on the correlation analysis between the child's body motion acceleration sequence and the original drip rate estimation sequence; if the body motion influence factor exceeds a preset perturbation threshold, it triggers interference stripping calculation to eliminate body motion artifacts in the original drip rate estimation sequence; at the same time, based on the coupling mapping between the drug solution environment parameters and the body motion influence factor, it generates true corrected drip rate data. The inference module is used to construct a medical risk graph structure through the patient profile vector, and inject the real corrected drip rate data into the corresponding drip rate node of the medical risk graph structure; and to dynamically recalibrate the edge weights between nodes in the medical risk graph structure using the body motion influence factor, and output a real-time risk score. If the real-time risk score exceeds a preset safety threshold, a trigger signal is generated; The decision module is used to receive the trigger signal, input the real-time risk score, the actual corrected drip rate data and the body motion influencing factor into the pre-trained reinforcement learning model, retrieve the optimal adjustment step size in the preset adjustment action space, and output the correction control command.

2. The pediatric infusion drip rate remote monitoring and risk correction system based on hierarchical diagnosis and treatment as described in claim 1, characterized in that, The specific steps for generating the body motion influencing factor based on the correlation analysis between the child's body motion acceleration sequence and the original drip rate estimation sequence include: Within a preset sliding time window, the child's body acceleration sequence and the original drip rate estimation sequence are normalized. Calculate the covariance of the normalized child's body acceleration sequence and the original drip rate estimation sequence under different time lags, and construct a cross-correlation matrix; The largest eigenvalue in the cross-correlation coefficient matrix is ​​extracted as the body motion influence factor.

3. The pediatric infusion drip rate remote monitoring and risk correction system based on hierarchical diagnosis and treatment as described in claim 1, characterized in that, The specific steps for generating accurate corrected drip rate data based on the coupled mapping between the drug solution environmental parameters and the body motion influencing factors include: Based on the drug type identifier and its real-time temperature data in the drug environment parameters, the corresponding flow resistance compensation coefficient is matched from the preset flow resistance database. The body dynamics influence factor and the flow resistance compensation coefficient are input into a preset fluid-structure interaction empirical formula for coupling calculation, and the actual corrected drip rate data is output. The fluid-structure interaction empirical formula is pre-constructed based on the principles of fluid dynamics and is used to simulate the dynamic influence of drug sloshing caused by the child's body movement on the drug's flow resistance characteristics.

4. The pediatric infusion drip rate remote monitoring and risk correction system based on hierarchical diagnosis and treatment as described in claim 1, characterized in that, When dynamically recalibrating the edge weights between nodes in the medical risk graph structure using the aforementioned body motion influence factor, the following steps are performed: Extract the initial weight vector connecting the drip rate node and adjacent risk nodes in the medical risk graph structure; Construct an exponential mapping function or a linear gain operator using the aforementioned body motion influence factors; The initial weight vector is multiplied by the exponential mapping function or the linear gain operator to achieve nonlinear amplification or attenuation of the risk weight of abnormal drip rate in a specific child context.

5. The pediatric infusion drip rate remote monitoring and risk correction system based on hierarchical diagnosis and treatment as described in claim 1, characterized in that, The specific steps for generating a real-time risk score include: The actual corrected drip rate data and the patient profile vector are mapped to the corresponding node space of the medical risk graph structure to generate an initial node feature matrix. For each node, collect feature information of neighboring drop rate nodes along the recalibrated edge weight path; The collected feature information is weighted and summed using an aggregation function, and the state vector of each drop rate node is updated, thus mapping the global state vector to the real-time risk score.

6. The pediatric infusion drip rate remote monitoring and risk correction system based on hierarchical diagnosis and treatment as described in claim 1, characterized in that, The system also includes a closed-loop consistency verification module, which performs the following steps: Obtain the feedback drip rate sequence generated after the execution of the correction control command, and calculate the instantaneous deviation between the feedback drip rate sequence and the target adjustment amount in the correction control command; The instantaneous deviation value is input into a preset consistency evaluation function to calculate the execution confidence index; If the execution confidence index is lower than the preset consistency threshold, a compensation mechanism correction instruction is triggered, and the instantaneous deviation value is synchronously fed back to the fusion estimation module and the inference module as a bias parameter.

7. The pediatric infusion drip rate remote monitoring and risk correction system based on hierarchical diagnosis and treatment as described in claim 6, characterized in that, The fusion estimation module utilizes the compensation mechanism to correct instructions and achieve self-evolution of the perception layer. The specific steps are as follows: Extract the instantaneous deviation value from the compensation mechanism correction instruction, and determine the correlation between the instantaneous deviation value and the distribution of the original drip rate estimation sequence; Based on the distribution correlation, the weight ratio for the vibration sequence in the adaptive weight allocation network is dynamically adjusted.

8. The pediatric infusion drip rate remote monitoring and risk correction system based on hierarchical diagnosis and treatment as described in claim 1, characterized in that, The inference module also performs sensitivity correction on the real-time risk score by introducing weights from tiered medical resources. Specific steps include: The level code of the current medical institution is parsed from the background data of the child's medical orders, and the corresponding medical resource guarantee coefficient is retrieved. The original risk scores output by the medical risk map structure are weighted and summed using the medical resource guarantee coefficient. The medical resource guarantee coefficient is directly proportional to the grade code.

9. The pediatric infusion drip rate remote monitoring and risk correction system based on hierarchical diagnosis and treatment as described in claim 1, characterized in that, The decision-making module optimizes the reinforcement learning model by constructing a multi-objective joint reward function, the specific logic of which is as follows: The pressure of the infusion line is obtained through the data acquisition module; The weight ratio of the first reward item and the second reward item is dynamically adjusted by the body motion influence factor; wherein, the first reward item is defined as the degree of risk level reduction after executing the corrective control command, and the second reward item is defined as the degree of stability fluctuation of the infusion pipeline pressure; When the physical activity influencing factor indicates that the child is in a state of vigorous exercise, the weight of the second reward item is increased.

10. The pediatric infusion drip rate remote monitoring and risk correction system based on hierarchical diagnosis and treatment as described in claim 1, characterized in that, The system also includes a global state synchronization module for enabling data interaction between the edge and the cloud, with specific steps including: The actual corrected drip rate data, the real-time risk score, and the corresponding corrective control instructions are encapsulated into structured sample pairs; The structured sample pairs are asynchronously uploaded to a cloud server, and feature adjustment curves for specific disease categories are extracted through cluster analysis. The system receives the feature adjustment curve sent by the cloud server and uses it as a constraint or initialization parameter to update the network parameters of the reinforcement learning model, and outputs the correction control command.

Images