A child infusion risk early warning method based on multi-modal data and artificial intelligence

Abstract

The application discloses a kind of children infusion risk early warning method based on multi-modal data and artificial intelligence, the puncture point state data of the collection children infusion process, the behavior data and basic data of children and the first flow rate of infusion pump are calculated after children to the behavior interference degree caused by infusion process, the pipeline pressure in puncture point state data and the flow rate of infusion pump are repaired in conjunction with Kalman filter, the abnormal information of infusion pump in infusion process is obtained by pipeline pressure and flow rate after repair;The feature extraction of thermal imaging data and pipeline pressure in puncture point state data is carried out by behavior interference degree, the extracted feature and children blood vessel diameter are input into the U-Net model of double branch, obtain swelling probability graph and swelling risk score;After physical verification succeeds to swelling probability graph corresponding swelling area, the swelling risk score is calibrated by preset blood vessel brittleness coefficient and behavior interference degree, and early warning information is constructed in conjunction with swelling area and the abnormal information of infusion pump.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for medical devices, and relates to, but is not limited to, a method for early warning of risks associated with intravenous infusion in children based on multimodal data and artificial intelligence. Background Technology

[0002] In pediatric intravenous infusions, children's crying, restlessness, fragile blood vessels, and limited ability to express themselves make it difficult for them to accurately describe early discomfort symptoms such as local swelling, pain, and burning. This makes it difficult to detect extravasation during the infusion process early. Furthermore, in pediatric wards with many patients, nurses spend a considerable amount of time conducting a full round, and each child is typically checked only a few times during an infusion. This infrequent manual observation makes it easy to miss extravasation during the infusion process. Often, by the time obvious swelling is visible, extravasation has already caused subcutaneous tissue damage, and may even lead to serious complications such as necrosis and infection. Therefore, early warning of potential extravasation risks and critical events such as infusion pump malfunctions during infusions has become an urgent priority to ensure the safety of children receiving intravenous infusions.

[0003] In related technologies, traditional methods such as flow rate monitoring, remaining volume reminders, and pressure alarms are used to optimize infusion management. Although these methods can provide some indication of the infusion endpoint or make a rough judgment of tubing blockage, most of them are more suitable for adults and have not been customized for children as a special group. Furthermore, they lack in-depth modeling of the clinical characteristics of children, such as non-cooperation, fragile blood vessels, and unclear expression. Therefore, it is difficult to achieve early identification and proactive warning of high-risk events such as leakage.

[0004] Therefore, how to efficiently and accurately provide early warnings of risks that may arise during intravenous infusion in children has become an urgent problem to be solved. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a method for early warning of risks in children's intravenous infusion based on multimodal data and artificial intelligence, which at least solves the problem that related technologies cannot efficiently and accurately warn of risks that occur during children's intravenous infusion.

[0006] According to a first aspect of the present invention, a method for early warning of intravenous infusion risks in children based on multimodal data and artificial intelligence is provided, comprising: Data on the puncture site status, the child's behavior, baseline data, and the first flow rate of the infusion pump were collected during the child's infusion process; and the degree of behavioral interference caused by the child during the infusion process was obtained based on the puncture site status data and the child's behavior data. The first tubing pressure and the first flow rate of the infusion pump in the puncture point status data are repaired based on behavioral interference degree and Kalman filter to obtain the repaired second tubing pressure and second flow rate. Based on behavioral interference, features are extracted from the thermal imaging data and the first tubing pressure in the puncture point status data to obtain initial stitching features; and based on the bi-branch U-Net model, the swelling area around the puncture point is predicted using the initial stitching features and the vessel diameter in the basic data to obtain the first swelling probability map and the first swelling risk score; the loss function of the bi-branch U-Net model is constructed based on the output of the trained PINN model. After the physical verification of the swelling region corresponding to the first swelling probability map based on the thermal diffusion equation is successful, the first swelling risk score is calibrated by preset vascular fragility coefficient and behavioral interference degree to obtain the second swelling risk score; Abnormal information of the infusion pump during the infusion process is obtained based on the second pipeline pressure and the second flow rate; and early warning information is constructed based on the second swelling risk score, swelling area and abnormal information of the infusion pump.

[0007] Optionally, the initial stitching features are obtained by extracting features from the thermal imaging data and the first tubing pressure in the puncture point state data based on behavioral interference degree, including: Substitute the behavioral interference degree into the mask coefficient formula to obtain the mask coefficient; and extract features from the thermal imaging data in the puncture point state data to obtain the first thermal imaging feature. The first thermal imaging feature is multiplied by the mask coefficient to obtain the second thermal imaging feature; the index features are calculated based on the first pipeline pressure; the index features include fluctuation variance, peak offset and number of consecutive drops; The index features, the second thermal imaging features, and the behavioral interference degree are stitched together to obtain the initial stitched features; The formula for the mask coefficient is as follows: In the above formula, Mask is the mask coefficient, and inteference_score is the behavioral interference degree; The U-Net model based on two branches predicts the swelling area around the puncture point using the initial splicing features and the vessel diameter in the basic data, resulting in a first swelling probability map and a first swelling risk score, including: The vessel diameter, age, and initial splicing features in the basic data are spliced ​​together to obtain the target splicing features. The bi-branch U-Net model is then used to predict the swelling area around the puncture point using the target splicing features, resulting in the first swelling probability map and the first swelling risk score.

[0008] Optionally, the step of calibrating the first swelling risk score by pre-setting a vascular fragility coefficient and behavioral interference degree to obtain a second swelling risk score includes: The second swelling risk score is obtained by substituting the first swelling risk score, behavioral interference level, and preset vascular fragility coefficient into the swelling risk scoring formula; the swelling risk scoring formula is as follows: In the above formula, S represents the second swelling risk score, and A represents the first swelling risk score. K represents the behavioral interference level, and K is the preset vascular fragility coefficient.

[0009] Optionally, the step of obtaining the degree of behavioral interference caused by the child during the infusion process based on puncture site status data and child behavior data includes: Target features were extracted from the puncture site status data and the child's behavioral data, including acceleration features, audio features, and hand trajectory features. The probability of a child becoming agitated is predicted based on target features; and the probability of agitation is used as the degree of behavioral interference.

[0010] Optionally, the first flow rate in the first tubing pressure and the first flow rate of the infusion pump in the puncture point state data is repaired based on behavioral interference degree and Kalman filter to obtain the repaired second tubing pressure and second flow rate, including: The weights corresponding to the first pipeline pressure and the first flow velocity are determined in the preset weights based on the behavioral interference degree; and the observation equation is constructed through the first pipeline pressure and the first flow velocity, and the observation noise covariance matrix is ​​constructed through the weights. The observation equation and the observation noise covariance matrix are substituted into the Kalman filter for iterative repair to obtain the repaired second pipeline pressure and second flow velocity.

[0011] Optionally, before obtaining the second swelling risk score by calibrating the first swelling risk score using a preset vascular fragility coefficient and behavioral interference degree after successfully physically verifying the swelling region corresponding to the first swelling probability map based on the thermal diffusion equation, the method further includes: Based on the temperature field distribution in the thermal imaging data, the temperature change rate and spatial second derivative corresponding to the swelling region are calculated respectively; and when the product between the spatial second derivative and the preset thermal diffusivity is not less than the temperature change rate, the physical verification of the swelling region is successful; otherwise, the swelling probability map is shrunken until the physical verification of the swelling region is successful.

[0012] Optionally, the dual-branch U-Net model can be trained using the following steps: Multimodal data samples were collected during the infusion process in children. The multimodal data samples included puncture site status data samples, children's behavior data samples, baseline data samples, and the first flow rate sample of the infusion pump. The corresponding thermal diffusivity is determined using basic data samples; and the thermal diffusivity, thermal imaging samples from the puncture point state data samples, and the corresponding time are input into the PINN model to be trained to obtain the temperature anomaly field and the second swelling probability map. The first loss is calculated based on the second swelling probability map and the actual swelling probability map, and the second loss is calculated based on the temperature anomaly field and the thermal diffusion equation. The trained PINN model is obtained through the first loss and the physical constraint loss; and the predicted third swelling probability map is obtained based on the trained PINN model. Based on multimodal data samples and the bi-branch U-Net model to be trained, the predicted fourth swelling probability map and the second swelling risk score are obtained; and the third loss is calculated based on the third and fourth swelling probability maps. The fourth loss is calculated using the fourth swelling probability map and the actual swelling probability map, and the fifth loss is calculated using the second swelling risk score and the actual swelling risk score; the trained bi-branch U-Net model is obtained using the third, fourth, and fifth losses.

[0013] According to a second aspect of the present invention, an electronic device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; the memory is used to store at least one executable instruction, wherein the executable instruction causes the processor to perform an operation corresponding to the method described in the first aspect.

[0014] According to the scheme provided in the embodiments of the present invention, puncture point status data, child behavior data, basic data, and the first flow rate of the infusion pump are collected during the child's infusion process; and the degree of behavioral interference caused by the child to the infusion process is obtained based on the puncture point status data and the child's behavior data; the first tubing pressure and the first flow rate of the infusion pump in the puncture point status data are repaired based on the behavioral interference degree and a Kalman filter to obtain the repaired second tubing pressure and second flow rate; features are extracted from the thermal imaging data and the first tubing pressure in the puncture point status data based on the behavioral interference degree to obtain initial stitching features; and the initial stitching is performed based on a bi-branch U-Net model. Based on the vessel diameter in the feature and basic data, the swelling area around the puncture point is predicted to obtain a first swelling probability map and a first swelling risk score. The loss function of the bi-branch U-Net model is constructed based on the output of the trained PINN model. After the physical verification of the swelling area corresponding to the first swelling probability map is successful based on the heat diffusion equation, the first swelling risk score is calibrated by preset vessel fragility coefficient and behavioral interference degree to obtain a second swelling risk score. Abnormal information of the infusion pump during the infusion process is obtained based on the second tubing pressure and the second flow rate. And early warning information is constructed based on the second swelling risk score, the swelling area and the abnormal information of the infusion pump. In this process, behavioral interference is calculated using data from four modalities, accurately quantifying the degree of interference caused by the child's crying and limb movements during infusion, providing a basis for subsequent anti-interference measures. Based on behavioral interference and a Kalman filter, tubing pressure and flow rate are corrected, filtering out momentary jitter or spike noise caused by the child pulling on the infusion tubing, enabling more accurate acquisition of abnormal information from the infusion pump. Initial splicing features and vessel diameter are input into a bi-branch U-Net model, automatically enhancing sensitivity for children with small blood vessels, enabling individualized risk assessment. Physical verification of swollen areas and calibration of the first swelling risk score automatically filter out false positives that do not conform to physical laws and avoid underreporting of high-risk events. In summary, this allows for comprehensive judgment of events such as leakage and tubing blockage, providing efficient and accurate early warning of risks during pediatric infusion. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments 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, wherein: Figure 1 A flowchart illustrating a method for early warning of pediatric intravenous infusion risks based on multimodal data and artificial intelligence, provided in an embodiment of the present invention; Figure 2This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention. 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.

[0017] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0018] It should be noted that the terms "first, second, and third" used in the embodiments of the present invention are only used to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of the present invention described herein can be implemented in an order other than that illustrated or described herein.

[0019] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which these embodiments of the invention pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0020] Figure 1 This is a flowchart illustrating a method for early warning of pediatric infusion risks based on multimodal data and artificial intelligence, provided by an embodiment of the present invention. This method can be executed by an electronic device, such as a computer or server.

[0021] like Figure 1 As shown, a method for early warning of pediatric intravenous infusion risks based on multimodal data and artificial intelligence includes: S101. Collect puncture site status data, child behavior data, basic data, and the first flow rate of the infusion pump during the child's infusion process; and obtain the degree of behavioral interference caused by the child during the infusion process based on the puncture site status data and the child's behavior data.

[0022] In embodiments of the present invention, during intravenous infusion in children, data on the puncture site status, the child's behavior, the child's baseline data, and the first flow rate of the infusion pump are collected. The puncture site status data includes thermal imaging data and tubing pressure data. The child's behavior data includes limb movements, sounds, and facial expressions. An index value is calculated based on the puncture site status data and the child's behavior data to measure the degree of behavioral interference caused by the child's limb movements such as crying, shaking, and patting during the infusion process. The first flow rate of the infusion pump can directly monitor whether the infusion rate is within a safe range, preventing heart failure and pulmonary edema caused by excessively fast flow, or affecting drug efficacy due to excessively slow or stagnant flow. Simultaneously, dynamic changes in the first flow rate can identify potential problems such as tubing blockage and venous leakage in advance. The behavioral interference level is a continuous value between 0 and 1, used to quantify the intensity of interference caused to sensor data by the child's limb movements such as crying, shaking, and patting during the infusion process. It can subsequently dynamically adjust the reliability weight of the sensor data and drive Kalman filtering to repair distorted data, thereby effectively suppressing motion artifacts from interfering with the monitoring of the infusion status. Specifically, during intravenous infusion in children, thermal imaging data can be collected by attaching a flexible thermal imaging film to the child's skin. A miniature pressure sensor can be integrated into the infusion clamp to collect pressure waveforms in the infusion tubing to sense the puncture site status and tubing flow, thus obtaining tubing pressure data. A three-axis accelerometer can be worn on the child's wrist, and microphones and cameras can be installed around it to collect behavioral data, including limb movements, voice, and facial expressions. The infusion pump's built-in infrared drop count sensor, ultrasonic flow sensor, or an interface adapter can be used to collect the initial flow rate of the infusion pump. The child's basic data, including blood vessel diameter and age, can be obtained through clinical records.

[0023] In some embodiments of the present invention, a temporal Markov transition matrix and a byte distribution histogram are extracted from the first flow rate of the infusion pump, i.e., the raw flow rate data. In the temporal Markov transition matrix, frequently occurring (high-probability) state transition paths in actual observations are identified and retained to obtain an initial Markov transition matrix. The initial Markov transition matrix and the byte distribution histogram are then concatenated to obtain concatenated features. Dimensionality reduction, pooling, and nonlinear transformations are performed on the concatenated features to obtain processed target features. The similarity between the target features and the features corresponding to preset infusion pump models is estimated, and a preset number of initial models are obtained from the preset infusion pump models based on the similarity. The Euclidean distance between the processed target features and the features corresponding to the initial models is calculated, and the initial model corresponding to the shortest distance is taken as the infusion pump model. After obtaining the infusion pump model, the corresponding protocol configuration can be found in a preset library. Then, the first flow rate is parsed using the protocol configuration to obtain the final timestamp. The timestamp is used to perform linear interpolation on other data in the multimodal data to align the time, and then subsequent data processing is performed.

[0024] S102. Based on behavioral interference and Kalman filter, the first tubing pressure and the first flow rate of the infusion pump in the puncture point status data are repaired to obtain the repaired second tubing pressure and second flow rate.

[0025] In embodiments of the present invention, during intravenous infusion in children, the collected tubing pressure and flow rate may be inaccurate due to interference from children and inherent sensor limitations, further hindering the effective and accurate detection of abnormal information from the infusion pump. To enable the Kalman filter to obtain dynamically adjusted observation noise covariance parameters, thereby achieving optimal state estimation in complex environments with children's behavioral interference, the first tubing pressure and first flow rate can be weighted by the degree of behavioral interference. Then, the Kalman filter iteratively repairs the weighted tubing pressure and flow rate until the repaired tubing pressure and flow rate are obtained.

[0026] S103. Based on the behavioral interference degree, feature extraction is performed on the thermal imaging data and the first tubing pressure in the puncture point state data to obtain the initial stitching features; and based on the two-branch U-Net model, the swelling area around the puncture point is predicted by the initial stitching features and the blood vessel diameter in the basic data to obtain the first swelling probability map and the first swelling risk score; the loss function of the two-branch U-Net model is constructed based on the output of the trained PINN model.

[0027] In an embodiment of the invention, the dual-branch U-Net model, based on the standard U-Net architecture, retains the original decoder output branch for generating the swelling probability map, and adds a parallel branch derived from the encoder bottleneck layer for outputting the swelling risk score, forming a dual-output head multi-task learning architecture. The output of the trained PINN model is used as the label for the dual-branch U-Net model during training. A loss is calculated based on the label, and the calculated loss, combined with other losses, is used to construct the total loss. A mask coefficient is calculated based on the behavioral interference level. Features are extracted from thermal imaging data and the first pipeline pressure, and initial stitching features are obtained based on the extracted features and the mask coefficient. Then, based on the initial stitching features and the vessel diameter, the dual-branch U-Net model is used for prediction to obtain the first swelling probability map and the first swelling risk score.

[0028] The first swelling probability map can be presented as a heatmap to reflect the spatial distribution probability of venous leakage around the infusion puncture site. The heatmap uses color gradients, such as blue representing low probability and red representing high probability, to visually display the spatial distribution of swelling probability, clearly locating the central area and diffusion boundary of leakage. The first swelling risk score is a comprehensive quantitative indicator, presented as low / medium / high risk levels, used to assess the overall risk of venous leakage during intravenous infusion in children.

[0029] Specifically, during the dual-branch output, the high-resolution feature map of the last layer of the standard U-Net model decoder is input into the... The convolutional layer maps the number of channels to the number of target categories, such as binary classification of leakage / non-leakage. Then, it is normalized pixel-by-pixel to the 0-1 interval using a Sigmoid activation function, generating a first swelling probability map of the same size as the input image. This first swelling probability map visually presents the spatial probability distribution of venous leakage in the tissue surrounding the puncture point in the form of a heatmap. The spatial dimension is compressed into a feature vector, which is then nonlinearly transformed through multiple fully connected layers. Finally, it is mapped to the 0-1 interval using a Sigmoid activation function, outputting a scalar value as the first swelling risk score.

[0030] S104. After the physical verification of the swelling region corresponding to the first swelling probability map based on the thermal diffusion equation is successful, the first swelling risk score is calibrated by the preset vascular fragility coefficient and behavioral interference degree to obtain the second swelling risk score.

[0031] In an embodiment of the present invention, the thermal diffusion equation is as follows: In the above formula, T represents temperature and t represents time. For the Laplace operator of temperature, Where is the thermal diffusivity, This represents the rate of temperature change.

[0032] The thermal diffusivity is a physical parameter describing the speed at which heat spreads through tissue; a higher value indicates faster temperature changes diffuse within the tissue. In monitoring intravenous leakage in children, it constrains the swelling range predicted by the U-Net model, ensuring that the rate of temperature change does not exceed a physical limit and preventing invalid predictions that violate thermodynamic laws. The rate of temperature change over time is obtained by calculating the temperature changes of pixels within the swelling region. The Laplace operator for the temperature at each pixel is calculated using thermal imaging data. Then, the rate of temperature change over time, the Laplace operator, and the thermal diffusivity are substituted into the thermal diffusivity equation to determine the physical consistency of the currently verified pixel. If the absolute value of the difference between the left and right sides of the equation is greater than a preset threshold, the pixel is deemed to violate physical laws and is removed from the swelling region; otherwise, it is retained. This process is repeated until all remaining pixels within the region satisfy the physical constraints of the thermal diffusivity equation, thus completing the physical verification of the swelling region.

[0033] Furthermore, after completing the physical verification of the swollen area, individual physiological differences were calibrated by the vascular fragility coefficient, and the influence of motion artifacts was suppressed by the behavioral interference degree. This ensured that the final swelling risk score could not only truly reflect the leakage risk level of the child, but also effectively avoid false alarms caused by crying and shaking, thus obtaining the second swelling risk score.

[0034] S105. Obtain abnormal information of the infusion pump during the infusion process based on the second pipeline pressure and the second flow rate; and construct early warning information based on the second swelling risk score, swelling area and abnormal information of the infusion pump.

[0035] In an embodiment of the present invention, a sliding window method is used to determine the pressure of the repaired second pipeline, employing the average pressure value within a 5-second sliding time window, where the average is μ and the variance is σ. 2When the variance is not less than a certain multiple of the mean, the window is considered to have abnormal fluctuations. This multiple can be set, for example, to 3 times the mean. This value is an empirical threshold set by engineers based on historical data, experimental tests, or industry practices to distinguish between normal and abnormal fluctuations. If multiple consecutive sliding windows meet this condition, the infusion pump is considered to be abnormal. Further, the abnormality type is determined by the repaired second flow rate. For example, an increase in second tubing pressure accompanied by a decrease in flow rate indicates tubing blockage, ultimately obtaining the abnormal information of the infusion pump. After obtaining the second swelling risk score, the second swelling risk score, i.e., the risk level corresponding to the calibrated risk score, can be further determined. When the variance is less than a certain multiple of the mean, it indicates that the current working state of the infusion pump is normal, and the abnormal information of the infusion pump in the warning message is displayed as none.

[0036] This can be achieved by obtaining multiple publicly available and valid clinical datasets. These datasets contain anonymized statistical data, including only calibrated risk scores and quantile information, without any patient privacy data. Calibrated risk scores are calculated at the initial quantiles of 20%, 40%, 60%, and 80%, and denoted as follows: , , and Based on the second swelling risk score, the swelling risk level is determined as follows: Safety level: Risk score after calibration < ; Note the level: ≤ Post-calibration risk score < ; Mild warning: ≤ Post-calibration risk score < ; Moderate alert: ≤ Post-calibration risk score < ; Severe warning: Risk score ≥ 1 after calibration ; Among them, the safety level indicates no abnormal risk and no intervention is required; the attention level indicates potential minor abnormalities and regular inspections are recommended; the mild warning level indicates possible early leakage and timely inspection is recommended; the moderate warning level indicates a high probability of leakage and immediate treatment is required; and the severe warning level indicates severe swelling or tissue damage and emergency intervention is required.

[0037] It is understood that, in the embodiments of the present invention, puncture point status data, child behavior data and basic data, and the first flow rate of the infusion pump are collected during the child's infusion process; and the degree of behavioral interference caused by the child to the infusion process is obtained based on the puncture point status data and the child's behavior data; the first tubing pressure and the first flow rate of the infusion pump in the puncture point status data are repaired based on the behavioral interference degree and a Kalman filter to obtain the repaired second tubing pressure and second flow rate; features are extracted from the thermal imaging data and the first tubing pressure in the puncture point status data based on the behavioral interference degree to obtain initial stitching features; and the initial stitching features are obtained based on a two-branch U-Net model. The swelling area around the puncture point is predicted by splicing features and the blood vessel diameter in the basic data, resulting in a first swelling probability map and a first swelling risk score. The loss function of the bi-branch U-Net model is constructed based on the output of the trained PINN model. After successful physical verification of the swelling area corresponding to the first swelling probability map based on the heat diffusion equation, the first swelling risk score is calibrated by preset blood vessel fragility coefficient and behavioral interference degree to obtain a second swelling risk score. Abnormal information of the infusion pump during infusion is obtained based on the second tubing pressure and the second flow rate. Early warning information is constructed based on the second swelling risk score, the swelling area, and the abnormal information of the infusion pump. In this process, behavioral interference is calculated from the collected data, accurately quantifying the degree of interference caused by the child's crying and limb movements during the infusion process, providing a basis for subsequent anti-interference measures. Based on behavioral interference and a Kalman filter, tubing pressure and flow rate are corrected, filtering out momentary jitter or spike noise caused by the child pulling on the infusion tubing, allowing for more accurate acquisition of abnormal information from the infusion pump. Initial splicing features and age are input into a bi-branch U-Net model, automatically enhancing sensitivity for children with fine blood vessels, enabling individualized risk assessment. Physical verification of swollen areas and calibration of the first swelling risk score automatically filter out false positives that do not conform to physical laws and avoid underreporting of high-risk events. In summary, this allows for comprehensive judgment of events such as leakage and tubing blockage, enabling efficient and accurate early warning of risks during pediatric infusions.

[0038] In some embodiments of the present invention, the initial stitching features obtained by extracting features from the thermal imaging data and the first pipeline pressure in the puncture point state data based on behavioral interference degree in S103 can be achieved through the following steps: substituting the behavioral interference degree into the mask coefficient formula to obtain the mask coefficient; extracting features from the thermal imaging data in the puncture point state data to obtain the first thermal imaging feature; multiplying the first thermal imaging feature and the mask coefficient to obtain the second thermal imaging feature; calculating the index feature based on the first pipeline pressure; the index feature includes fluctuation variance, peak offset, and number of consecutive drops; stitching the index feature, the second thermal imaging feature, and the behavioral interference degree to obtain the initial stitching features. The mask coefficient formula is as follows: In the above formula, Mask is the mask coefficient, and inteference_score is the behavioral interference degree.

[0039] Specifically, the masking coefficient is a weighting factor dynamically calculated from the behavioral interference level, used to filter out thermal imaging artifacts caused by non-leakage factors such as limb friction and sweat evaporation at the feature level. The closer the masking coefficient is to 1, the more completely the thermal imaging features are preserved, in order to accurately capture the real temperature changes caused by venous leakage. In the thermal imaging data, for the same frame, the temperature difference between adjacent pixels is calculated, and then the ratio of the temperature difference to the spatial distance between adjacent pixels is calculated to obtain the spatial gradient value. Simultaneously, in consecutive frames, the temperature difference at the same pixel location is calculated, and then the temperature difference at the same pixel location is divided by the time interval between frames to obtain the temporal gradient value. The spatial gradient value and the temporal gradient value are used as features to construct the first thermal imaging feature. This process is existing technology and will not be described in detail here. Then, the first thermal imaging feature is multiplied by the calculated masking coefficient to obtain the second thermal imaging feature. The fluctuation variance, peak offset, and number of consecutive drops are calculated from the first pipeline pressure. The fluctuation variance can be further calculated by averaging all pressure values ​​within a set sliding window and then squared the difference between each pressure value and the average. The peak offset can be used as the difference between the detected local maxima and the baseline pressure by setting a baseline pressure. The number of consecutive drops can be further counted by comparing adjacent pressure values ​​frame by frame in the pipeline pressure data. After obtaining the index features and the second thermal imaging features, these are stitched together with the behavioral interference to obtain the initial stitched features. Here, 0.8 represents the optimal parameter value selected from a series of data through pre-conducted experiments by experts in the optimization of 2000 clinical cases.

[0040] It is understood that, in some embodiments of the present invention, the mask coefficients calculated by behavioral interference degree are used to weight the first thermal imaging feature, which can effectively suppress thermal artifacts caused by non-leakage factors such as limb friction and sweat evaporation, and significantly reduce the false alarm rate.

[0041] In some embodiments of the present invention, the U-Net-based model in S103 predicts the swelling area around the puncture point based on the initial splicing features and the blood vessel diameter in the basic data to obtain a swelling probability map. This can be achieved through the following steps: splicing the blood vessel diameter, age, and initial splicing features in the basic data to obtain the target splicing features; and using a two-branch U-Net model to predict the swelling area around the puncture point based on the target splicing features to obtain a first swelling probability map and a first swelling risk score.

[0042] Specifically, the basic data includes vessel diameter and age. Age determines vessel fragility, subcutaneous tissue thickness, and thermal diffusivity; that is, the younger the child, the thinner the vessel wall and the faster the exudate diffuses. The same amount of leakage in young children will result in a larger area of ​​swelling probability map and a higher risk score. Vessel diameter directly affects the dynamics of leakage; thinner vessels, such as 1.5mm, are more prone to leakage under the same pressure, and are more sensitive to changes in local temperature gradients after leakage. The response weights to temperature changes need to be dynamically adjusted based on vessel diameter. By combining the child's vessel diameter and age as input, the U-Net model's output can be adapted to individual physiological differences in children, making the prediction results more consistent with pediatric clinical practice and improving the accuracy and clinical reliability of the predictions. The initial spliced ​​features are then spliced ​​again, and the spliced ​​target spliced ​​features are input into a two-branch U-Net model. The two branches output the first swelling probability map and the first swelling risk score around the puncture point, respectively.

[0043] It is understood that, in some embodiments of the present invention, age is embedded in the input of the U-Net model, enabling the model to automatically enhance its sensitivity for young children and high-risk patients with fine blood vessels, thereby achieving individualized risk assessment.

[0044] In some embodiments of the present invention, the calibration of the first swelling risk score in S104 using a preset vascular fragility coefficient and behavioral interference degree to obtain a second swelling risk score can be described by the following steps: substituting the first swelling risk score, behavioral interference degree, and preset vascular fragility coefficient into the swelling risk scoring formula to obtain the second swelling risk score; the swelling risk scoring formula is as follows: In the above formula, S represents the second swelling risk score, and A represents the first swelling risk score. K represents the behavioral interference level, and K is the preset vascular fragility coefficient.

[0045] Specifically, the preset vascular fragility coefficient varies depending on the vessel diameter and age in the base data. To obtain the preset vascular fragility coefficient, the vessel diameter and age are first obtained, and then matched against a preset database to find the corresponding preset vascular fragility coefficient. Finally, the preset vascular fragility coefficient and behavioral interference level are used to calibrate the first swelling risk score to obtain the second swelling risk score.

[0046] Based on clinical data from 2000 pediatric intravenous infusion cases across three tertiary hospitals, Spearman correlation analysis revealed a strong negative correlation between age and vascular fragility coefficient. Further utilizing this data, a linear regression model was employed, with the clinical vascular fragility score as the dependent variable and age as the independent variable, applicable to children aged 0-12 years. The resulting formula for calculating the fragility coefficient is as follows: Vascular fragility coefficient = 1.0 - age × 0.05 Specifically, when the calculated vascular fragility coefficient for children is lower than 0.2, 0.2 is directly taken as the final value of the coefficient, and the calculated lower value is no longer used. A preset library for storing vascular fragility coefficients is constructed using the above formula, with 1.0 as the intercept and 0.05 as the regression coefficient, which are obtained by fitting linear regression models.

[0047] It is understood that, in some embodiments of the present invention, behavioral interference degree and vascular fragility coefficient are introduced to calibrate the first swelling risk score, so that even if the early swelling signal is weak, the warning level will be improved due to vascular fragility, avoiding missed reports of high-risk events and improving the warning sensitivity and clinical reliability of high-risk children.

[0048] In some embodiments of the present invention, obtaining the degree of behavioral interference caused by the child to the infusion process based on the puncture point status data and the child's behavior data in S101 can be achieved by the following steps: extracting target features from the puncture point status data and the child's behavior data respectively, the target features including acceleration features, audio features and hand trajectory features; predicting the probability value of the child's irritability in an irritable state through the target features; and using the irritability probability value as the degree of behavioral interference.

[0049] Specifically, in children's behavioral data, audio data can be... The sampling rate was used to acquire 40-Vimel frequency cepstral coefficients (MFCCs), which were then downsampled to... This preserves key information that distinguishes between crying and quiet states; simultaneously, the three-axis acceleration data, after mean pooling, is also synchronized to... Then, acceleration features reflecting limb activity are extracted. Puncture point status data can include the child's hand trajectory data during intravenous infusion. Hand trajectory features can be obtained by calculating the displacement, velocity, approach distance, and movement pattern of key hand points relative to the puncture point.

[0050] Furthermore, the extracted target features are concatenated and then input into a prediction model, which can be three stacked LSTM layers, each containing approximately 64 hidden units and including gating mechanisms such as input gates, forget gates, and output gates. Each LSTM output is then fed into an independent Dropout layer with a dropout rate of 0.2. The last LSTM layer can connect two cascaded fully connected layers, and the last fully connected layer is connected to a Softmax activation function, outputting a probability value (0~1) of the child being in a volatile state. This value is directly used as behavioral interference.

[0051] It is understood that, in some embodiments of the present invention, acceleration features, audio features, and hand trajectory features are used to predict the probability value of a child's irritability in an irritable state through target features; and the irritability probability value is used as the behavioral interference degree, which can dynamically suppress artifacts to reduce false alarms, compensate for signal distortion to prevent missed alarms, and significantly improve the robustness and reliability of leakage identification.

[0052] In some embodiments of the present invention, S102 can be implemented by the following steps: determining the weights corresponding to the first pipeline pressure and the first flow velocity in a preset weight based on the behavioral interference degree; constructing an observation equation through the first pipeline pressure and the first flow velocity; and constructing an observation noise covariance matrix through the weights. The observation equation and the observation noise covariance matrix are substituted into the Kalman filter for iterative repair to obtain the repaired second pipeline pressure and second flow velocity.

[0053] Specifically, to dynamically adjust the reliability of sensor data in the Kalman filter based on the child's real-time movement intensity, thereby suppressing motion artifacts from contaminating pressure and flow velocity data during periods of agitation and avoiding false alarms, and to preserve rapid responses to real physiological changes such as leakage or blockage during periods of calm to ensure no missed alarms, the weights corresponding to the first pipeline pressure and first flow velocity are determined from an existing weight library based on the value of behavioral interference. The observation noise covariance matrix is ​​set using these weights, and the observation equation is constructed using the first pipeline pressure and first flow velocity. Other parameters, such as the state equation being an identity matrix and the process noise following an N(0,0.005I) distribution (where I is the identity matrix and 0.005I is a manually set value), are used. Under these conditions, the Kalman filter iteratively repairs the data until the repair conditions are met, yielding the repaired second pipeline pressure and second flow velocity. The weight library can be constructed by combining historical experimental data statistics with expert experience calibration.

[0054] In some embodiments of the present invention, the following steps are included before S104: calculating the temperature change rate and spatial second derivative corresponding to the swelling region based on the temperature field distribution in the thermal imaging data; and when the product between the spatial second derivative and the preset thermal diffusivity is not less than the temperature change rate, the physical verification of the swelling region is successful; otherwise, the swelling probability map is shrunk at the boundary until the physical verification of the swelling region is successful.

[0055] Specifically, the rate of temperature change over time is obtained by calculating the temperature changes of pixels within the swelling region. The Laplace operator for temperature is calculated using thermal imaging data. Then, the rate of temperature change over time, the Laplace operator, and the thermal diffusivity are substituted into the thermal diffusivity equation to determine the physical consistency of the currently verified pixel. If the absolute value of the difference between the left and right sides of the equation is greater than a preset threshold, the pixel is deemed to violate physical laws and is removed from the swelling region; otherwise, it is retained. This process is repeated until all remaining pixels within the region satisfy the physical constraints of the thermal diffusivity equation, at which point the physical verification of the swelling region is complete.

[0056] In an embodiment of the present invention, the dual-branch U-Net model can be trained through the following steps: collecting multimodal data samples of the child's infusion process, including puncture point state data samples, child behavior data samples, baseline data samples, and the first flow rate sample of the infusion pump; determining the corresponding thermal diffusivity using the baseline data samples; and inputting the thermal diffusivity, thermal imaging samples from the puncture point state data samples, and the corresponding time into the PINN model to be trained to obtain the temperature anomaly field and the second swelling probability map; calculating the first loss based on the second swelling probability map and the actual swelling probability map, and calculating the loss based on the temperature anomaly field and the thermal diffusivity. The second loss is calculated using the dispersion equation; the trained PINN model is obtained using the first loss and the physical constraint loss; and the predicted third swelling probability map is obtained based on the trained PINN model; the predicted fourth swelling probability map and the second swelling risk score are obtained based on the multimodal data samples and the bi-branch U-Net model to be trained; and the third loss is calculated based on the third and fourth swelling probability maps; the fourth loss is calculated using the fourth swelling probability map and the actual swelling probability map, and the fourth loss is calculated using the second swelling risk score and the actual swelling risk score; and the trained bi-branch U-Net model is obtained using the third and fourth losses.

[0057] Specifically, samples of 2,000 exudation cases from multiple hospitals can be collected. These exudation case samples can consist of original samples and expanded samples, covering children aged 2-12 years. The exudation case samples include multimodal data samples of the children's infusion process, including puncture point status data samples, children's behavioral data samples, basic data samples, and the first flow rate sample of the infusion pump. Then, the training set, validation set, and test set are divided in a 7:2:1 ratio. First, the PINN model is trained. The thermal diffusivity coefficient is determined from a pre-defined thermal diffusivity coefficient library using the children's age and blood vessel diameter from the basic data samples. Then, the thermal diffusivity coefficient, thermal imaging samples from the puncture point state data samples, and the corresponding time are input into the PINN model to be trained, outputting a temperature anomaly field. A second swelling probability map is then calculated from the temperature anomaly field. The corresponding label, i.e., the actual swelling probability map, is obtained. A loss function, such as the mean squared error loss function or the binary cross-entropy loss function, is set. The second swelling probability map and the actual swelling probability map are substituted into the loss function to calculate the first loss. The rate of change of temperature over time and the Laplace operator of temperature are calculated from the temperature anomaly field. Then, the rate of change of temperature over time, the Laplace operator of temperature, and the thermal diffusivity coefficient are substituted into the thermal diffusivity equation to calculate the physical constraint loss as the second loss. The parameters of the PINN model to be trained are adjusted by summing the first and second losses until the training conditions are met, resulting in the trained PINN model. Finally, the training set and validation set can be substituted into the trained PINN model to obtain the temperature anomaly field and calculate the third swelling probability map. The thermal diffusivity database is created by integrating historical experimental measurement data, using statistical regression methods to establish a mapping relationship between children's age, blood vessel diameter, and thermal diffusivity, and then calibrating and correcting it based on expert experience.

[0058] Furthermore, after processing the training samples through steps S101 to S103, the U-Net model to be trained outputs a fourth swelling probability map and a second swelling risk score. Both the third and actual swelling probability maps are used as labels. The fourth and actual swelling probability maps are substituted into a loss function such as the mean squared error to obtain the fourth loss. Then, the third and fourth swelling probability maps can be substituted into the KL divergence loss function to obtain the KL divergence. This allows the thermodynamic physical laws inherent in the PINN model (i.e., the teacher model) to be transferred to the two-branch U-Net (student model) through knowledge distillation, enabling U-Net to maintain both high inference speed and physical consistency. Thus, while maintaining real-time inference speed, it outputs a swelling probability map that is more consistent with thermodynamic principles and more robust. Then, the second swelling risk score and the actual swelling risk score are substituted into the mean squared error loss function to obtain the fifth loss. The third, fourth and fifth losses are then weighted and summed to obtain the total loss. The two-branch U-Net model is then trained using the total loss until the trained two-branch U-Net model is obtained.

[0059] It is understood that, in some embodiments of the present invention, further training of the U-Net model through the dual-branch PINN model can embed physical laws into the data-driven deep learning process, enabling the dual-branch U-Net to not only fit clinical annotations but also follow real biothermodynamic behavior. On the one hand, this improves the physical consistency and generalization ability of the predicted swelling probability map, and on the other hand, it can provide reliable soft label guidance, ultimately enhancing the sensitivity and credibility of the dual-branch U-Net model to early and weak leakage infusion signals.

[0060] Reference Figure 2 The diagram shows a structural schematic of an electronic device according to an embodiment of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the electronic device.

[0061] like Figure 2 As shown, the electronic device may include: a processor 502, a communications interface 504, a memory 506, and a communications bus 508.

[0062] in: The processor 502, communication interface 504, and memory 506 communicate with each other via communication bus 508.

[0063] Communication interface 504 is used to communicate with other electronic devices or servers.

[0064] The processor 502 is used to execute program 510, specifically the relevant steps in the above method embodiments.

[0065] Specifically, program 510 may include program code that includes computer operation instructions.

[0066] Processor 502 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The smart device may include one or more processors of the same type, such as one or more CPUs; or it may include processors of different types, such as one or more CPUs and one or more ASICs.

[0067] Memory 506 is used to store program 510. Memory 506 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0068] Specifically, program 510 can be used to cause processor 502 to perform the operations corresponding to the methods described in the above method embodiments.

[0069] The specific implementation of each step in program 510 can be found in the corresponding descriptions of the steps and units in the above method embodiments, and will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the devices and modules described above can be referred to the corresponding process descriptions in the foregoing method embodiments, and will not be repeated here.

[0070] It should be noted that, depending on the implementation needs, the various components / steps described in the embodiments of the present invention can be broken down into more components / steps, or two or more components / steps or parts of the operation of components / steps can be combined into new components / steps to achieve the purpose of the embodiments of the present invention.

[0071] The methods described above according to embodiments of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for executing the methods shown herein.

[0072] Those skilled in the art will recognize that the units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the embodiments of the present invention.

[0073] The above embodiments are only used to illustrate the embodiments of the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the embodiments of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the embodiments of the present invention, and the patent protection scope of the embodiments of the present invention should be defined by the claims.

Claims

1. A method for early warning of pediatric intravenous infusion risks based on multimodal data and artificial intelligence, characterized in that, include: Collect data on the puncture site status, the child's behavior, baseline data, and the initial flow rate of the infusion pump during the child's intravenous infusion process; And based on puncture site status data and child behavior data, the degree of behavioral interference caused by children during the infusion process was obtained; The first tubing pressure and the first flow rate of the infusion pump in the puncture point status data are repaired based on behavioral interference degree and Kalman filter to obtain the repaired second tubing pressure and second flow rate. Based on behavioral interference, feature extraction is performed on thermal imaging data and first tubing pressure in puncture point status data to obtain initial stitching features; Based on the dual-branch U-Net model, the swelling area around the puncture point is predicted using the initial splicing features and the vessel diameter in the basic data, resulting in the first swelling probability map and the first swelling risk score; the loss function of the dual-branch U-Net model is constructed based on the output of the trained PINN model. After the physical verification of the swelling region corresponding to the first swelling probability map based on the thermal diffusion equation is successful, the first swelling risk score is calibrated by preset vascular fragility coefficient and behavioral interference degree to obtain the second swelling risk score; Abnormal information of the infusion pump during the infusion process is obtained based on the pressure and flow rate of the second pipeline; Early warning information is constructed based on the second swelling risk score, swelling area, and abnormal information of the infusion pump.

2. The method according to claim 1, characterized in that, The initial stitching features are obtained by extracting features from the thermal imaging data and the first tubing pressure in the puncture point state data based on behavioral interference, including: Substitute the behavioral interference degree into the mask coefficient formula to obtain the mask coefficient; and extract features from the thermal imaging data in the puncture point state data to obtain the first thermal imaging feature. The first thermal imaging feature is multiplied by the mask coefficient to obtain the second thermal imaging feature; the index features are calculated based on the first pipeline pressure; the index features include fluctuation variance, peak offset and number of consecutive drops; The index features, the second thermal imaging features, and the behavioral interference degree are stitched together to obtain the initial stitched features; The formula for the mask coefficient is as follows: In the above formula, Mask is the mask coefficient, and inteference_score is the behavioral interference degree; The U-Net model based on two branches predicts the swelling area around the puncture point using the initial splicing features and the vessel diameter in the basic data, resulting in a first swelling probability map and a first swelling risk score, including: The vessel diameter, age, and initial splicing features in the basic data are spliced ​​together to obtain the target splicing features. The bi-branch U-Net model is then used to predict the swelling area around the puncture point using the target splicing features, resulting in the first swelling probability map and the first swelling risk score.

3. The method according to claim 1, characterized in that, The process of calibrating the first swelling risk score using a preset vascular fragility coefficient and behavioral interference level to obtain a second swelling risk score includes: Substituting the first swelling risk score, behavioral interference level, and preset vascular fragility coefficient into the swelling risk scoring formula, the second swelling risk score is obtained; the swelling risk scoring formula is as follows: In the above formula, S represents the second swelling risk score, and A represents the first swelling risk score. K represents the behavioral interference level, and K is the preset vascular fragility coefficient.

4. The method according to claim 1, characterized in that, The method of obtaining the degree of behavioral interference caused by children during the infusion process based on puncture site status data and children's behavioral data includes: Target features were extracted from the puncture point status data and the child's behavioral data, including acceleration features, audio features, and hand trajectory features. The probability of a child becoming agitated is predicted based on target features; and the probability of agitation is used as the degree of behavioral interference.

5. The method according to claim 1, characterized in that, The method of repairing the first tubing pressure and the first flow rate of the infusion pump in the puncture point state data based on behavioral interference degree and Kalman filter to obtain the repaired second tubing pressure and second flow rate includes: The weights corresponding to the first pipeline pressure and the first flow velocity are determined in the preset weights based on the behavioral interference degree; and the observation equation is constructed through the first pipeline pressure and the first flow velocity, and the observation noise covariance matrix is ​​constructed through the weights. The observation equation and the observation noise covariance matrix are substituted into the Kalman filter for iterative repair to obtain the repaired second pipeline pressure and second flow velocity.

6. The method according to claim 1, characterized in that, After the physical verification of the swelling region corresponding to the first swelling probability map based on the thermal diffusion equation is successful, before calibrating the first swelling risk score by setting a preset vascular fragility coefficient and behavioral interference degree to obtain the second swelling risk score, the method further includes: Based on the temperature field distribution in the thermal imaging data, the temperature change rate and spatial second derivative corresponding to the swelling region are calculated respectively; and when the product between the spatial second derivative and the preset thermal diffusivity is not less than the temperature change rate, the physical verification of the swelling region is successful; otherwise, the swelling probability map is shrunken until the physical verification of the swelling region is successful.

7. The method according to claim 1, characterized in that, The dual-branch U-Net model can be trained through the following steps: Multimodal data samples were collected during the infusion process in children. The multimodal data samples included puncture site status data samples, children's behavior data samples, basic data samples, and the first flow rate sample of the infusion pump. The corresponding thermal diffusivity is determined using basic data samples; and the thermal diffusivity, thermal imaging samples from the puncture point state data samples, and the corresponding time are input into the PINN model to be trained to obtain the temperature anomaly field and the second swelling probability map. The first loss is calculated based on the second swelling probability map and the actual swelling probability map, and the second loss is calculated based on the temperature anomaly field and the thermal diffusion equation. The trained PINN model is obtained through the first loss and the physical constraint loss; and the predicted third swelling probability map is obtained based on the trained PINN model. The predicted fourth swelling probability map and second swelling risk score are obtained based on multimodal data samples and the two-branch U-Net model to be trained; The third loss is calculated based on the third and fourth swelling probability maps; The fourth loss is calculated using the fourth swelling probability map and the actual swelling probability map, and the fifth loss is calculated using the second swelling risk score and the actual swelling risk score; the trained bi-branch U-Net model is obtained using the third, fourth, and fifth losses.

Images