Port-related complication warning system and method

Abstract

The application relates to the technical field of medical treatment, in particular to a port-related complication early warning system and method, wherein the system comprises a collection device and a computer device; the collection device is used for collecting at least one of thermal imaging data, pressure data, ultrasonic image data and visible light image data of a region where a user's port is located, and obtaining clinical data of the user; the computer device communicates with the collection device, generates modality data according to the data collected by the collection device, calls a pre-trained prediction model, inputs the modality data into the prediction model, outputs a comprehensive risk level of the user existing in a complication through the prediction model, and executes a complication early warning action according to the risk level, so that the monitoring and early warning of the port-related complication are realized. Therefore, the problem that related technologies lack monitoring of early micro changes of the port and lead to delayed discovery of the port-related complication is solved.

Description

Technical Field

[0001] This application relates to the field of medical technology, and in particular to an early warning system and method for complications related to infusion ports. Background Technology

[0002] Infusion ports, as long-term implantable venous access devices, are widely used in clinical settings such as chemotherapy for tumors, long-term parenteral nutrition, and the treatment of chronic diseases. Their advantages include reducing repeated punctures, improving patients' quality of life, and lowering the risk of catheter-related infections. However, infusion ports still face the threat of various complications during use, mainly including catheter-related infections, drug extravasation, catheter obstruction, clamping syndrome, port inversion, and abnormal blood return.

[0003] Currently, the clinical identification of these complications mainly relies on patient complaints (such as local redness, swelling, and pain) or obvious physical signs (such as fever, tissue necrosis, and inability to aspirate blood), often only being discovered when the pathological process has progressed to the middle or late stages. This symptom-driven passive monitoring model has significant lag, making early intervention difficult, increasing treatment difficulty and medical costs, and potentially endangering patient safety. Summary of the Invention

[0004] This application provides an early warning system and method for complications related to infusion ports, in order to solve the problem that the lack of monitoring of early minor changes in infusion ports in related technologies leads to delayed detection of complications related to infusion ports.

[0005] The first aspect of this application provides an early warning system for complications related to infusion ports, comprising: a data acquisition device for acquiring at least one of thermal imaging data, pressure data, ultrasound image data, and visible light image data of the area where the user's infusion port is located, and acquiring the user's clinical data; and a computer device communicating with the data acquisition device, wherein the computer device generates modal data based on at least one of the thermal imaging data, pressure data, ultrasound image data, visible light image data, and clinical data, calls a pre-trained prediction model, inputs the modal data into the prediction model, outputs a comprehensive risk level of complications for the user through the prediction model, and executes a complication early warning action based on the risk level. The prediction model includes multiple task heads and a comprehensive risk head, each task head predicts a risk score for corresponding complications related to the infusion port, and a comprehensive score for complications for the user is determined based on the comprehensive risk head and the risk scores predicted by each task head, and the risk level of complications for the user is determined based on the comprehensive score.

[0006] Optionally, each task head includes a fully connected layer, a nonlinear activation layer, a regularization layer, and an output layer. Each task head predicts the risk score of complications corresponding to the infusion port, including: each task head uses the fully connected layer to extract feature data of the corresponding complications from thermal imaging data, pressure data, ultrasound image data, visible light image data, and clinical data; the feature data is sequentially input into the nonlinear activation layer and the regularization layer; the feature data after passing through the nonlinear activation layer and the regularization layer is passed through the output layer, and the output layer maps the feature data to the risk probability of the corresponding complications, and assesses the risk score of the corresponding complications based on the risk probability.

[0007] Optionally, by combining the risk scores predicted by the risk head and each task head, a comprehensive score for the user's complications is determined, including: weighting and fusing the risk scores predicted by each task head to obtain a main risk score; calculating the associated risk scores based on the risk scores of different complications; and calculating the comprehensive score for the user's complications based on the main risk score and the associated risk scores.

[0008] Optionally, the main risk scoring formula is:

[0009]

[0010] in, For the first Complication risk score predicted by each task head ; For predefined importance coefficients The fusion weights obtained by normalization ,in, Let be the clinical importance coefficient of the j-th complication. For the first to the second The sum of the clinical importance coefficients of each individual is obtained by adding them together;

[0011] The formula for calculating the association risk score is:

[0012]

[0013] in, For the first Complication risk score predicted by each task head; The complication risk score predicted for the i-th task head; For the first The effective association strength between the complication risk score predicted by the i-th task head and the complication risk score predicted by the i-th task head;

[0014] The formula for calculating the overall score is:

[0015]

[0016] in, This is the correlation amplification factor obtained by training the prediction model.

[0017] Optionally, when the computer device detects that the risk level is greater than the risk threshold, it determines the target complication based on the risk score predicted by each task head, and generates early warning information based on the target complication, so as to locate the target complication related to the infusion port while issuing early warnings for complications related to the infusion port through the early warning information.

[0018] Optionally, the multiple task heads include multiple infection heads, clamping heads, extravasation heads, occlusion heads, reversal heads, and blood return heads, with the infection heads, clamping heads, extravasation heads, occlusion heads, reversal heads, and blood return heads determined based on complications associated with the infusion port.

[0019] Optionally, the computer device is also used to train the prediction model before calling the pre-trained prediction model. The training process includes: obtaining the training dataset of the prediction model, wherein the training dataset includes training samples and ground truth labels. The training samples are thermal imaging data, pressure data, ultrasound image data, visible light image data, and clinical data of the area where the user's port is located. The ground truth labels are the risk score of the user's port-related complications and the comprehensive score of the user having complications. The prediction model is trained using the training samples in the training dataset. During the training process, the training loss is calculated using the ground truth labels and the predicted risk score and predicted comprehensive score of the prediction model. The model parameters of the prediction model are updated according to the training loss.

[0020] Optionally, the computer device further includes a thermal imaging encoder, an ultrasound encoder, a pressure encoder, and a clinical data encoder. After generating modal data based on at least one of the thermal imaging data, pressure data, ultrasound image data, visible light image data, and clinical data, the device further includes: extracting thermal imaging data from the modal data using the thermal imaging encoder to obtain a feature vector of the thermal imaging data; extracting ultrasound image data from the modal data using the ultrasound encoder to obtain a feature vector of the ultrasound image data; extracting pressure data from the modal data using the pressure encoder to obtain a feature vector of the pressure data; and extracting clinical data from the modal data using the clinical data encoder to obtain a feature vector of the clinical data.

[0021] Optionally, the computer device is also used to: convert the feature vector of each modality into a query vector, a key vector, and a value vector through linear projection; calculate attention-weighted feature vectors from other modalities based on the query vector, key vector, and value vector of the current modality and the query vector, key vector, and value vector of other modalities; add the feature vector of the current modality to the attention-weighted feature vectors from other modalities and concatenate them in the channel dimension to obtain the fused modal data.

[0022] The second aspect of this application provides a method for early warning of port-related complications. The method is applied to the computer equipment of the port-related complications early warning system of any one of the first aspects. The method includes the following steps: generating modal data from at least one of thermal imaging data, pressure data, ultrasound image data, visible light image data, and clinical data; inputting the modal data into a prediction model, and outputting a comprehensive risk level of complications for the user through the prediction model; and executing a complication early warning action according to the risk level.

[0023] Therefore, this application has the following beneficial effects:

[0024] This application embodiment can collect thermal imaging data, pressure data, ultrasound image data, visible light image data, and the user's clinical data of the area where the infusion port is located using a data acquisition device. A computer communicates with the acquisition device to generate modal data based on the collected data. A pre-trained prediction model is then invoked, and the modal data is input into the prediction model. The prediction model outputs a comprehensive risk level for complications in the user and executes complication warning actions based on the risk level. By integrating multimodal information such as thermal imaging, pressure sensing, ultrasound, visible light images, and clinical data, and using a pre-trained intelligent prediction model to perform fusion analysis and collaborative reasoning on early subtle abnormalities in the infusion port area, it can accurately identify early signs of common complications such as infection, obstruction, and extravasation, and output a clinically interpretable comprehensive risk level, thereby achieving early detection and early warning of complications. Compared with the passive monitoring methods in related technologies that rely on obvious symptoms, this significantly improves the sensitivity and timeliness of the warning. Therefore, it solves the problem of delayed detection of infusion port-related complications due to the lack of monitoring of early subtle changes in the infusion port in related technologies.

[0025] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0026] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0027] Figure 1 This is a schematic diagram of an early warning system for port-of-care complications provided according to an embodiment of this application;

[0028] Figure 2 This is a flowchart illustrating the method for early warning of complications related to infusion ports provided in accordance with the embodiments of this application. Detailed Implementation

[0029] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0030] The following description, with reference to the accompanying drawings, illustrates an embodiment of an infusion port-related complication early warning system and method. To address the issue mentioned in the background that related technologies lack monitoring of early, subtle changes in infusion ports, leading to delayed detection of port-related complications, this application provides an early warning system for port-related complications. In this system, thermal imaging data, pressure data, ultrasound image data, visible light image data, and the user's clinical data are collected by a data acquisition device. A computer communicates with the acquisition device to generate modal data based on the collected data. A pre-trained prediction model is then invoked, and the modal data is input into the prediction model. The prediction model outputs a comprehensive risk level for the user's potential complications and executes early warning actions based on this risk level. By integrating multimodal information such as thermal imaging, pressure sensing, ultrasound, visible light images, and clinical data, and utilizing a pre-trained intelligent prediction model to perform fusion analysis and collaborative reasoning on early, subtle abnormalities in the port area, the system can accurately identify early signs of common complications such as infection, obstruction, and extravasation, and output a clinically interpretable comprehensive risk level, thereby achieving early detection and early warning of complications. Compared to traditional passive monitoring methods that rely on obvious symptoms, this system significantly improves the sensitivity and timeliness of early warning. This solves the problem that the lack of monitoring for early, subtle changes in infusion ports in related technologies leads to delayed detection of infusion port-related complications.

[0031] Specifically, Figure 1 This is a schematic diagram of an early warning system for complications related to infusion ports provided in an embodiment of this application.

[0032] like Figure 1 As shown, the infusion port-related complication early warning system includes: data acquisition device 101 and computer device 102.

[0033] The acquisition device 101 is used to acquire at least one of the thermal imaging data, pressure data, ultrasound image data, and visible light image data of the area where the user's port is located, and to acquire the user's clinical data. The computer device 102 communicates with the acquisition device 101. The computer device 102 generates modal data based on at least one of the thermal imaging data, pressure data, ultrasound image data, visible light image data, and clinical data. It calls a pre-trained prediction model, inputs the modal data into the prediction model, and outputs the comprehensive risk level of the user's complications through the prediction model. It then executes complication warning actions based on the risk level. The prediction model includes multiple task heads and a comprehensive risk head. Each task head predicts the risk score of the corresponding complications related to the port. The comprehensive risk head and the risk scores predicted by each task head determine the comprehensive score of the user's complications, and the risk level of the user's complications is determined based on the comprehensive score.

[0034] Modal data refers to structured feature representations formed after preprocessing information from different data sources, i.e., different modalities. For example, thermal imaging data is converted into temperature distribution features, ultrasound images are extracted into morphological features, and pressure signals are converted into temporal fluctuation features. Each data type constitutes a modality. Clinical data includes user case data and laboratory data. Medical record data includes body temperature, pulse, respiratory rate, blood pressure, blood oxygen saturation, pain score, and blood glucose, while laboratory data includes blood cell data, liver and kidney function data, infection marker detection data (C-reactive protein, procalcitonin, etc.), and pathogen detection data (blood, urine, stool, sputum culture, etc.). The predictive model is an artificial intelligence model pre-trained on a large number of historical cases to predict the risk of complications from multimodal inputs. The training process will be described in detail below and will not be repeated here.

[0035] It is understood that the embodiments of this application include a data acquisition device 101 and a computer device 102. The data acquisition device 101 is used to acquire at least one of the following data from the user's port-a-cath area: thermal imaging data, pressure data, ultrasound image data, and visible light image data. It also acquires the user's clinical information simultaneously and synchronizes the above data to the computer device 102. The computer device 102 fuses the above multi-source data to generate multimodal data and inputs it into a pre-trained prediction model. The prediction model includes multiple task heads and a comprehensive risk head. Each task head assesses the risk score of a specific complication (such as infection, obstruction, extravasation, etc.). The comprehensive risk head calculates the user's comprehensive risk score based on these scores and their medical relevance, and determines the overall risk level of the complication accordingly. Finally, it automatically executes the corresponding early warning action based on the level, thereby achieving early, accurate, and intelligent monitoring and early warning of port-a-cath related complications.

[0036] In this embodiment, the acquisition device 101 includes an infrared thermal imaging device, a pressure sensor device, an ultrasound acquisition device, a visible light camera device, and a data interface. It acquires thermal imaging data, pressure data, ultrasound image data, and visible light image data from the infusion port and obtains the user's clinical data through the data interface.

[0037] Among them, infrared thermal imaging equipment is a non-contact imaging device that generates thermal images reflecting local temperature distribution by detecting infrared radiation on the surface of human skin, used to identify abnormal temperature rises around the infusion port caused by inflammation or infection; pressure sensor equipment is a sensing device installed at the connection of the infusion line or port, which monitors fluid pressure changes in real time during the infusion process, used to capture pressure fluctuations or continuous increases caused by obstruction, clamping, or abnormal blood return; ultrasound acquisition equipment refers to portable or fixed ultrasound probes and supporting systems, used to acquire high-frequency sound wave reflection images of the infusion port and surrounding tissues, which can observe the port position, catheter morphology, presence of thrombi, extravasation, or overturning and other structural abnormalities; visible light imaging equipment is a common optical camera used to capture visible light images of the infusion port puncture site, assisting in the identification of visible signs such as redness, swelling, exudation, and skin damage; data interface refers to the software or hardware communication module that interfaces with hospital information systems (such as electronic medical record systems and laboratory information systems), used to automatically acquire the patient's clinical data.

[0038] It is understood that the acquisition device 101 in this embodiment of the application consists of an infrared thermal imaging device, a pressure sensor device, an ultrasonic acquisition device, a visible light camera device, and a data interface. The infrared thermal imaging device is used to acquire temperature distribution data of the infusion port area, the pressure sensor device is used to monitor pressure changes in the infusion pathway, the ultrasonic acquisition device is used to acquire ultrasound images of the port and surrounding tissues, the visible light camera device is used to record the appearance of the puncture site, and the data interface is used to automatically acquire the patient's clinical data from the hospital information system, thereby providing comprehensive multimodal input for subsequent intelligent analysis.

[0039] It should be noted that the data acquisition device 101 in this embodiment of the application acquires data according to a set frequency. For example, if the user is a normal patient, the thermal imaging data acquisition frequency can be preset to once a day, and if the user is a high-risk patient, the thermal imaging data acquisition frequency can be preset to once every 8 hours. The data acquisition frequency of other pressure sensor devices, ultrasound acquisition devices, etc. is similar and can be specifically set according to actual needs, and is not specifically limited here. At the same time, there can be multiple sets of the data acquisition device 101 in this embodiment of the application to realize data acquisition of infusion port-related complications of multiple users.

[0040] In this embodiment, each task head includes a fully connected layer, a nonlinear activation layer, a regularization layer, and an output layer. Each task head predicts the risk score of complications corresponding to the infusion port, including: each task head uses the fully connected layer to extract feature data of the corresponding complications from thermal imaging data, pressure data, ultrasound image data, visible light image data, and clinical data; the feature data is sequentially input into the nonlinear activation layer and the regularization layer; the feature data after passing through the nonlinear activation layer and the regularization layer is passed through the output layer, and the output layer maps the feature data to the risk probability of the corresponding complications, and assesses the risk score of the corresponding complications based on the risk probability.

[0041] The task head is a sub-network module in the prediction model specifically designed to identify a particular complication (such as infection, obstruction, or extravasation). Multiple task heads work in parallel to achieve multi-task learning. The fully connected layer is a basic layer in a neural network, where each input neuron is connected to each output neuron. It is used to linearly combine and map input features in a high dimension, thereby extracting discriminative features relevant to the target task. Nonlinear activation layers typically use activation functions such as ReLU (Rectified Linear Unit) to introduce nonlinearity, enabling the model to fit complex data relationships and preventing the network from degenerating into a linear model. Regularization layers are often used to randomly disable some neurons during training to prevent overfitting and improve generalization ability. The output layer is the last layer of the task head, usually containing one neuron and using the Sigmoid activation function (a classic nonlinear activation function, its formula is...). The output of the previous layer is compressed to the (0,1) interval, which represents the probability of the complication occurring.

[0042] It is understood that each task head in this embodiment consists of a fully connected layer, a nonlinear activation layer, a regularization layer, and an output layer, specifically designed to predict a risk score for a specific complication related to an infusion port. Specifically, the task head first extracts the most relevant features to the complication from the fused modal data (including thermal imaging data, pressure data, ultrasound image data, visible light image data, and clinical data) through a fully connected layer. Subsequently, the extracted features are passed through a nonlinear activation layer to enhance the model's expressive power, and a regularization layer is used to suppress overfitting. Finally, the output layer maps the processed features to a risk probability between 0 and 1, and generates a risk score for the complication based on this risk probability. This score is used to measure the likelihood of the complication's occurrence and support subsequent comprehensive risk assessment and early warning decisions. By designing independent task heads for each infusion port complication, the prediction model can specifically learn the multimodal discriminative features unique to that complication.

[0043] In this embodiment of the application, a comprehensive score for determining whether a user has complications is determined by combining the risk head and the risk score predicted by each task head, including: weighting and fusing the risk scores predicted by each task head to obtain a main risk score; calculating the associated risk score based on the risk scores of different complications; and calculating the comprehensive score for determining whether a user has complications based on the main risk score and the associated risk score.

[0044] The main risk score is a basic composite risk value obtained by summing the clinically important risk scores of each complication, reflecting the different medical hazard weights of different complications (e.g., infection has a higher weight than port-wine inversion). The associated risk score measures the synergistic amplification effect when multiple complications are at high risk simultaneously. Based on the co-occurrence patterns of complications learned from historical data (e.g., "obstruction is often accompanied by no blood return"), when the risks of two strongly related complications are both high, the associated risk score adds an extra layer to the total risk. The composite score is the final overall risk quantification indicator, which is a linear combination of the main risk score and the associated risk score, used to uniformly measure the overall risk of a patient currently experiencing port-related complications.

[0045] It is understood that this application embodiment fuses the complication risk scores output by each task head through a comprehensive risk head to determine the comprehensive score of the user's complication status. Specifically, firstly, different weights are assigned to the risk scores of each complication based on its clinical importance, and the weighted sum is used to obtain the main risk score. Secondly, based on the medical correlation between complications, the synergistic enhancement effect generated when the risks of multiple related complications increase simultaneously is calculated to form a correlation risk score. Finally, the main risk score and the correlation risk score are superimposed proportionally to generate the final comprehensive score. By introducing clinical importance weights, the prediction model can prioritize complications that are more harmful and require urgent treatment, avoiding the dilution of key risks caused by egalitarianism. At the same time, by modeling the medical correlation between complications, the model effectively captures the situation of multi-factor synergistic deterioration, enabling the model to actively amplify the overall danger signal when multiple risks coexist, avoiding missed warnings. The final comprehensive score respects medical priorities and reflects pathological interaction logic, thus more realistically and comprehensively reflecting the patient's actual risk status and providing a scientific, reliable, and interpretable decision-making basis for graded early warning. This score comprehensively reflects the patient's current complication risk level and provides a basis for subsequent early warning level classification.

[0046] In this embodiment of the application, the main risk scoring formula is:

[0047]

[0048] in, For the first Complication risk score predicted by each task head ; For predefined importance coefficients The fusion weights obtained by normalization ,in, The clinical importance coefficient for the j-th complication is a positive number predefined by clinical experts based on medical experience, reflecting the severity, urgency, and treatment priority of the complication. For the first to the second The sum of the clinical importance coefficients of each individual is obtained by adding them together;

[0049] The formula for calculating the association risk score is:

[0050]

[0051] in, For the first Complication risk score predicted by each task head; This indicates iterating through all different complication pairs; The complication risk score predicted for the i-th task head; For the first The effective association strength between the complication risk score predicted by the i-th task head and the complication risk score predicted by the i-th task head. It can be calculated from historical cases and Pearson correlation coefficient Specifically, we can set a correlation threshold. , ;

[0052] The formula for calculating the overall score is:

[0053]

[0054] in, The correlation amplification coefficient is obtained by training the prediction model and is used to control the amplification of correlation risk in the total risk.

[0055] It is understood that the calculation of the comprehensive score in this application embodiment is divided into three steps. First, the complication risk scores output by each task head are weighted and fused according to their clinical importance to obtain the main risk score, where the weights are obtained by normalization of the predefined clinical importance coefficients. Second, the correlation risk score is calculated by iterating through all different complication pairs, multiplying their risk scores by the corresponding effective correlation strength, and then summing them to capture the synergistic deterioration effect when multiple complications are at high risk simultaneously. Finally, the main risk score and the correlation risk score are linearly combined to obtain the final comprehensive score, where the correlation amplification coefficient is automatically learned through model training to balance the contribution of independent risk and correlation enhancement, thereby comprehensively and accurately quantifying the overall risk level of patients developing port-of-care complications.

[0056] In this embodiment of the application, when the computer device 102 detects that the risk level is greater than the risk threshold, it determines the target complication based on the risk score predicted by each task head, and generates early warning information based on the target complication, so as to locate the target complication related to the infusion port while issuing early warning information for complications related to the infusion port.

[0057] The risk level is a discrete warning level derived from a comprehensive score mapping. For example, risk levels can be set as: Normal, Level 1 Warning, Level 2 Warning, Level 3 Warning, and Level 4 Warning. The risk threshold characterizes the severity of the patient's overall risk. The risk threshold is a preset critical value used to determine whether a warning needs to be activated and to identify target complications. For example, setting the risk threshold to Level 2 Warning means that a warning message is generated when the risk level is greater than Level 2, and the target complication is located. The risk threshold can be a fixed value or dynamically adjusted based on individual patient factors, set according to actual needs. On the warning interface, different colors can be used to highlight the risk level; for example, blue is used for Level 1 Warning to indicate mild abnormalities. Normal, observation is recommended. Level 2 warning uses yellow, indicating moderate abnormality; Level 3 warning uses orange, indicating high abnormality; Level 4 warning uses red, indicating emergency requiring immediate action. The risk score predicted by the task head is a quantitative risk value output by each task head for a specific complication (such as infection, obstruction, etc.), reflecting the degree of abnormality of the complication. Based on the risk score, the most likely complication for the user can be determined, i.e., the target complication. The warning information is a structured message containing the target complication type, risk level, key abnormal indicators (such as "local temperature increase + white blood cell increase → indicating infection"), and recommended treatment measures, which is used to push to medical staff terminals or integrate into the electronic medical record system.

[0058] It is understood that when the computer device 102 in this application embodiment does not detect that the patient's overall risk level exceeds the preset risk threshold, it only uses different colors to highlight the risk level. When it detects that the patient's overall risk level exceeds the preset risk threshold, it further analyzes the complication risk scores output by each task head, identifies one or more target complications with the highest risk, and sends warning information containing specific types, risk basis, and clinical recommendations to the medical staff terminal. This can achieve overall warning of infusion port-related complications, accurately locate the most likely specific complication type, and provide a clear direction for rapid clinical intervention.

[0059] In the embodiments of this application, the multiple task heads include multiple infection heads, clamping heads, extravasation heads, occlusion heads, flipping heads, and blood return heads. The infection heads, clamping heads, extravasation heads, occlusion heads, flipping heads, and blood return heads are determined according to the complications related to the infusion port.

[0060] Among them, the infection head is a task head specifically designed to predict the risk of catheter infection complications, using early detection through local temperature elevation, abnormal heat diffusion patterns, and systemic reactions (such as elevated temperature in the lymph node area); the clamping head is a task head specifically designed to predict the risk of clamping syndrome, using data such as catheter compression, changes in body position, or ultrasound imaging to determine whether the catheter is damaged; the extravasation head is a task head specifically designed to predict the risk of drug extravasation complications, assessing the amount of extravasation and the risk of tissue damage based on heat diffusion patterns, tissue swelling, and drug characteristics; the occlusion head is a task head specifically designed to predict catheter occlusion complications, using infusion pressure, aspiration resistance, and ultrasound blood flow signals to determine catheter patency and to identify the type (thrombotic / non-thrombotic) and location of the occlusion; the flipping head is a task head specifically designed to predict the risk of port infusion seat flipping complications, using ultrasound or visible light imaging to analyze whether the port's position and orientation are abnormal; and the blood return head is a task head specifically designed to predict the risk of complications from no blood return during aspiration, combining aspiration pressure data analysis, catheter tip position assessment, and blood flow signal detection to determine the cause of blood return obstruction.

[0061] It should be noted that since some complications can have more detailed outputs, such as extravasation which may require outputs of extravasation volume and severity, the output of the prediction model may also be designed according to the specific needs of each complication. For example, the task head design for predicting the risk of drug extravasation complications may include an output layer with multiple outputs, which can simultaneously output whether extravasation has occurred, the severity level, and the estimated extravasation volume.

[0062] It is understood that the multiple task heads in the embodiments of this application are named infection head, clamping head, extravasation head, occlusion head, inversion head, and blood return head, respectively. Each task head models a typical port-of-care complication. The infection head is used to assess the risk of infection, the clamping head is used to identify clamping syndrome, the extravasation head is used to detect drug extravasation, the occlusion head is used to determine whether the catheter is blocked, the inversion head is used to detect abnormal inversion of the port injection seat, and the blood return head is used to predict the possibility of no blood return during aspiration. These task heads together constitute a multi-task prediction architecture, which realizes high-precision, high-efficiency, and clinically interpretable parallel risk assessment, ensuring accurate, parallel, and interpretable risk assessment for various complications.

[0063] In this embodiment of the application, the computer device 102 is further configured to train the prediction model before calling the pre-trained prediction model. The training process includes: obtaining a training dataset for the prediction model, wherein the training dataset includes training samples and ground truth labels. The training samples are thermal imaging data, pressure data, ultrasound image data, visible light image data, and clinical data of the area where the user's port is located. The ground truth labels are the risk score of the user's port-related complications and the comprehensive score of the user having complications. The prediction model is trained using the training samples in the training dataset. During the training process, the training loss is calculated using the ground truth labels and the predicted risk score and predicted comprehensive score of the prediction model. The model parameters of the prediction model are updated according to the training loss.

[0064] The training dataset is a collection of data used to train the prediction model, consisting of a large number of historical patient samples. Training samples are the input to each training data point, containing multimodal data collected from the same patient at a specific time point, including thermal imaging data, pressure data, ultrasound image data, visible light image data, and clinical data. Ground truth labels are the true target values ​​corresponding to each training sample, labeled by clinical experts or subsequent diagnostic results, including ground truth values ​​for risk scores and comprehensive scores for each complication. Predicted risk scores are the risk score values ​​output by each task head in the prediction model for a single complication. Predicted comprehensive scores are the comprehensive risk level values ​​output by the comprehensive risk head in the prediction model. Training loss is a quantitative indicator measuring the difference between the model's predicted values ​​and the ground truth labels, including the losses of each task head and the comprehensive risk head. The losses of each task head and the comprehensive risk head can be calculated using mean squared error, and the total loss is the weighted sum of the losses of each task head and the comprehensive risk head. Model parameter updates can be performed using the backpropagation algorithm, optimizing the model weights through gradient descent based on the loss function, gradually bringing the prediction results closer to the ground truth until the preset requirements are met, indicating that the prediction model training is complete.

[0065] It is understood that the computer device 102 in this application embodiment needs to train the prediction model before using it. Specifically, a training dataset containing a large number of historical cases is first obtained. Each training sample consists of thermal imaging data, pressure data, ultrasound image data, visible light image data, and clinical data of the patient's port area. The corresponding ground truth labels include the patient's true risk score for each complication and the overall risk level. During training, the prediction model takes these multimodal data as input and outputs the predicted risk score and predicted overall risk level for each complication. By calculating the training loss between the predicted value and the ground truth label, the model's internal parameters are continuously adjusted using backpropagation and optimization algorithms until the model can accurately learn the early risk patterns of complications from the input data, thus laying the foundation for subsequent real-time early warning. Through multimodal joint training on a large number of training samples, the prediction model can deeply learn and integrate complex correlation patterns in thermal imaging, pressure, ultrasound, visible light, and clinical data, accurately capture the early signs of various port complications, and has high clinical adaptability and robustness. This provides a solid technical foundation for subsequent real-time and reliable complication early warning, significantly improving the possibility of early intervention and patient safety.

[0066] In this embodiment, the computer device 102 further includes a thermal imaging encoder, an ultrasonic encoder, a pressure encoder, and a clinical data encoder. After generating modal data based on at least one of the thermal imaging data, pressure data, ultrasonic image data, visible light image data, and clinical data, the device further includes: extracting thermal imaging data from the modal data using the thermal imaging encoder to obtain a feature vector of the thermal imaging data; extracting ultrasonic image data from the modal data using the ultrasonic encoder to obtain a feature vector of the ultrasonic image data; extracting pressure data from the modal data using the pressure encoder to obtain a feature vector of the pressure data; and extracting clinical data from the modal data using the clinical data encoder to obtain a feature vector of the clinical data.

[0067] The thermal imaging encoder uses a residual network, such as ResNet-50 (a 50-layer residual network, a deep convolutional neural network that solves the gradient vanishing problem in deep networks through residual connections; in this embodiment, it is used for fine-tuning to adapt to single-channel infrared thermal imaging). By inputting an infrared thermal imaging image to the encoder, the encoder progressively extracts multi-level features through residual convolutional blocks, uses global average pooling, and averages the high-dimensional feature map output from the last convolutional layer across the spatial dimension, compressing it into a fixed-length vector to obtain the feature vector of the thermal imaging data. The ultrasound encoder can be the encoder part of U-Net (an encoder-decoder network commonly used for medical image segmentation; here, only its encoder (downsampling path) is used to extract features). By inputting ultrasound image data, the encoder alternately uses convolutional and pooling layers to progressively downsample and extract multi-scale morphological features to obtain the feature vector of the ultrasound image data. The pressure encoder can be a 1DCNN (One-Dimensional Convolutional Neural Network). The network takes stress data as input, uses a one-dimensional convolutional kernel to slide along the time axis to automatically detect local patterns, and gradually reduces dimensionality through pooling layers to retain key temporal features. Finally, it uses a fully connected layer to summarize all temporal information and generate a feature vector of stress data. The clinical data encoder can be a Transformer (a deep learning model based on self-attention mechanism). It takes structured clinical data as input, converts each feature into an embedding vector, adds positional encoding, and uses a multi-head self-attention mechanism to allow the encoder to automatically learn the dependencies between different clinical variables, and finally outputs a feature vector of clinical data.

[0068] It is understood that the computer device 102 in this embodiment of the application also includes multiple dedicated encoder modules, including a thermal imaging encoder, an ultrasound encoder, a pressure encoder, and a clinical data encoder. After receiving multimodal input (i.e., at least one of thermal imaging data, ultrasound images, pressure signals, visible light images, and clinical data) provided by the acquisition device 101 and organizing it into modal data, the corresponding encoders are called to perform feature extraction: the thermal imaging encoder extracts feature vectors reflecting local temperature abnormalities from the thermal image, the ultrasound encoder extracts feature vectors related to the port body and duct structure from the ultrasound image, the pressure encoder extracts feature vectors of dynamic patterns such as obstruction or wall adhesion from the time-series pressure data, and the clinical data encoder transforms the patient's laboratory indicators and medical history into clinical semantic feature vectors. These feature vectors are then fused and input into the task head and comprehensive risk head of the prediction model, providing a high-quality multimodal representation basis for accurate complication early warning.

[0069] In this embodiment of the application, the computer device 102 is further configured to: convert the feature vector of each modality into a query vector, a key vector, and a value vector through linear projection; calculate attention-weighted feature vectors from other modalities based on the query vector, key vector, and value vector of the current modality and the query vector, key vector, and value vector of other modalities; add the feature vector of the current modality to the attention-weighted feature vectors from other modalities, and concatenate them in the channel dimension to obtain fused modal data.

[0070] The query vector represents what the current modality wants to know. For example, when thermal imaging features are used as the query vector, the intention is to ask which information from other modalities is related to the current abnormal temperature rise of the thermal imaging features. The key vector represents what searchable content other modalities can provide. The value vector represents the useful information actually carried by other modalities. The attention-weighted feature vector from other modalities is calculated based on the query vector, key vector, and value vector of the current modality and the query vector, key vector, and value vector of other modalities, using the following formula:

[0071]

[0072] in, For attention-weighted feature vectors from other modalities, For query vector, For key vectors, For value vectors, This is the transpose of the key vector. The vector dimension is used to scale the dot product to prevent gradient saturation. The normalized attention weights are used to convert the original attention scores into a probability distribution.

[0073] The feature vector of the current modality is added to the attention-weighted feature vectors from other modalities and concatenated along the channel dimension to obtain the fused modal data. Specifically, this includes adding the features of each modality itself to the attention-weighted features from other modalities to obtain cross-modal enhanced features, and concatenating all cross-modal enhanced feature vectors along the feature dimension to form a high-dimensional joint representation, i.e., the fused modal data.

[0074] It is understood that the computer device 102 in this application embodiment is also used for multimodal feature fusion. First, the feature vectors obtained by encoding thermal imaging data, pressure data, ultrasound image data, and clinical data are respectively converted into query vectors, key vectors, and value vectors through learnable linear projection. Then, the similarity between the query vector of a certain modality (such as thermal imaging) and the key vectors of other modalities is calculated, and the value vectors of other modalities are weighted and summed accordingly to obtain attention-weighted feature vectors from other modalities. Then, the original features of the modality itself are added to the cross-modal attention information received to generate cross-modal enhanced features. This computational method calculates the cross-modal enhancement features for each modality. Finally, the enhancement features of all modalities are concatenated along the channel dimension to form a unified high-dimensional fusion vector, which serves as the input for the subsequent complication task head and comprehensive risk head. By introducing a multimodal feature fusion strategy based on an attention mechanism, it is possible to dynamically mine complementary and correlated information between different data sources. The resulting high-dimensional fusion vector is not only richer in information and more comprehensive in representation, but also significantly improves the model's ability to understand complex clinical scenarios, making complication identification more accurate and risk assessment more context-aware. This provides a multimodal perception foundation for highly reliable intelligent early warning decisions.

[0075] The port-of-care port-related complication early warning system proposed in this application collects thermal imaging data, pressure data, ultrasound image data, visible light image data, and the user's clinical data from the area where the port is located using a data acquisition device. A computer communicates with the acquisition device to generate modal data based on the collected data. A pre-trained prediction model is then invoked, and the modal data is input into the prediction model. The prediction model outputs a comprehensive risk level for the user's potential complications. Based on this risk level, a complication early warning action is executed. By integrating multimodal information such as thermal imaging, pressure sensing, ultrasound, visible light images, and clinical data, and utilizing a pre-trained intelligent prediction model to perform fusion analysis and collaborative reasoning on early minor abnormalities in the port area, the system can accurately identify early signs of common complications such as infection, obstruction, and extravasation, and output a comprehensive risk level with clinical interpretability. This enables early detection and early warning of complications. Compared to traditional passive monitoring methods that rely on obvious symptoms, this significantly improves the sensitivity and timeliness of the early warning system.

[0076] The following is a specific embodiment of the early warning system for port-of-care complications. The overall early warning process of the port-of-care complications early warning system in this embodiment is as follows:

[0077] After the port-of-care-related complication early warning system in this embodiment is started, it first executes the initialization process, which includes loading the pre-trained prediction model, connecting to the clinical database through the data interface, and starting the real-time communication service. After the data acquisition module and computer equipment are ready, it enters the standby state.

[0078] The system in this embodiment supports two operating modes: In single-patient monitoring mode, it sequentially completes the acquisition of multi-source data for individual patients (such as at least one of thermal imaging data, pressure data, ultrasound image data, and visible light image data), data preprocessing, risk analysis of complications by a predictive model, comprehensive risk assessment, and generation of personalized early warnings. The results are then stored and pushed to the medical staff's terminal. In multi-patient centralized monitoring mode, the system obtains a list of all patients, collects data from each patient synchronously through a parallel mechanism, performs batch processing and concurrent predictive model analysis, conducts a comprehensive assessment of the group risk, and finally updates and aggregates all early warning information in real time on the centralized monitoring interface, realizing efficient, intelligent, and scalable full-process monitoring and early warning of port-of-care complications.

[0079] Specifically, the data acquisition process in this embodiment is as follows: The connection status of the acquisition device at the user's location is checked. If the acquisition device is online, multimodal sensors such as infrared thermal imaging, pressure sensors, ultrasound acquisition devices, and visible light cameras are activated to simultaneously acquire multi-source physiological and imaging data of the infusion port area. If the acquisition device is offline, its offline status is recorded and the current process ends. The acquired data is then uniformly verified for quality. If any modal data is unqualified, a retry mechanism is triggered. If the retry limit is exceeded (e.g., setting the retry limit to three times), the data is marked as abnormal and the process terminates. If the data quality is qualified, all modal data are timestamped, packaged, and transmitted to the computer device, awaiting confirmation. Upon successful confirmation, the patient's monitoring status is updated, completing this round of high-quality, synchronized, safe, and reliable data acquisition closed loop.

[0080] This embodiment acquires multimodal data of the patient's infusion port area through acquisition equipment, including at least one of thermal imaging data, pressure data, ultrasound image data, and visible light image data, as well as clinical data from the hospital information system. This raw data is then fed into a computer, where dedicated encoders extract features: a thermal imaging encoder (based on ResNet-50) extracts semantic features of temperature distribution; an ultrasound encoder (based on U-Net encoder) extracts structural features of the port and catheter; a pressure encoder (based on 1D CNN) extracts temporal pressure dynamic patterns; and a clinical data encoder (based on Transformer) fuses the interaction between medical history and laboratory indicators to generate high-dimensional feature vectors for each modality. Subsequently, a cross-modal attention mechanism is employed to convert the features of each modality into query vectors, key vectors, and value vectors through linear projection. Using the current modality as the query, attention weights are calculated with the keys and values ​​of other modalities to generate attention-weighted features from other modalities, which are then added to the current modality's own features to form cross-modal enhanced features. Finally, all enhanced features are concatenated along the channel dimension to obtain unified fused modal data.

[0081] The overall prediction process of this embodiment is as follows: The fused modal data is input into a pre-trained prediction model. The model contains six task heads (infection head, clamping head, extravasation head, obstruction head, inversion head, and blood return head). Each task head outputs the risk probability of the corresponding complication as a risk score through a fully connected layer, ReLU activation, regularization layer, and output layer. The comprehensive risk head calculates the main risk score based on these scores, on the one hand, according to the weights normalized by the clinical importance coefficient, and on the other hand, calculates the co-risk score using the effective association strength learned in the past. Combined with a learnable amplification coefficient, a comprehensive score is obtained, and the risk level is mapped according to this comprehensive score.

[0082] In this embodiment, after obtaining the risk level output by the predictive model, if the risk level is lower than a preset threshold, it is marked as low risk and recorded. If it is higher than the threshold, it enters the detailed evaluation stage, that is, each complication is evaluated separately, the patient's historical monitoring data is automatically retrieved, the risk change trend is analyzed, and the warning level is determined accordingly. Based on this, structured warning information and targeted treatment suggestions are generated, and the warnings are prioritized according to the urgency of the condition, further constructing a personalized clinical decision tree. If it is determined that immediate intervention is required, an emergency notification is triggered and a to-do item is created and pushed to the medical staff workstation; otherwise, it is included in routine monitoring and the subsequent monitoring frequency is dynamically adjusted. Medical staff can confirm after receiving the notification. If they do not respond in time, the reminder is automatically escalated. Once the processing is initiated, this embodiment will continuously track the processing progress until the task is completed and the results are recorded.

[0083] It should be noted that when this embodiment is used to monitor port-of-care complications for multiple users, the alert interface can be set up as follows: obtain a list of all users to be monitored and arrange the interface according to the selected layout mode (grid, list, or grouped by alert level); create an independent monitoring card for each user and asynchronously load their multimodal data, sequentially render the thermal imaging view, overlay the predictive model analysis results, and display the corresponding alert information; the interface continuously maintains data freshness in two ways: prioritizing real-time messages pushed by the server (such as heatmap updates, analysis results, device status, or system notifications), and if real-time data is not enabled... The connection is degraded to a timed polling mechanism. When new data is received, the interface automatically parses the message type and updates the image, warning status, or connection identifier accordingly. At the same time, this embodiment also supports rich interactions for medical staff, including clicking on user cards to view detailed information, clicking on warnings to view handling suggestions, and clicking on heatmaps to zoom in for observation. All interactions will trigger a partial refresh of the interface. During continuous monitoring, data updates and interaction responses are executed cyclically until the medical staff actively exits. At this time, the connection is closed, the current interface state is saved, and resources are released and the interface is destroyed, thereby achieving an efficient, dynamic, and interactive multi-user centralized monitoring experience.

[0084] Next, with reference to the accompanying drawings, a method for early warning of complications related to infusion ports proposed according to embodiments of this application is described.

[0085] Figure 2 This is a flowchart illustrating the method for early warning of port-related complications proposed in this application. The method is applied to the computer equipment of the aforementioned early warning system for port-related complications, such as... Figure 2 As shown, the method includes the following steps:

[0086] In step S201, at least one modal data of thermal imaging data, pressure data, ultrasound image data, visible light image data, and clinical data is generated.

[0087] It is understood that the embodiments of this application collect thermal imaging data, pressure data, ultrasound image data, and visible light image data of the area where the user's infusion port is located, as well as the user's clinical information. These data are then processed into modal data, that is, feature vectors of each modality are extracted by multiple encoders, and information from different sources is fused using a cross-modal attention mechanism and spliced ​​and fused into processed modal data, laying the foundation for subsequent feature extraction and risk prediction.

[0088] In step S202, modal data is input into the prediction model, and the prediction model outputs the comprehensive risk level of the user's complications.

[0089] It is understood that in this embodiment of the application, modal data is input into a pre-trained prediction model, and the fused representation is analyzed in parallel by multiple task heads to output individual risk scores for six types of complications, namely infection, clamping, extravasation, obstruction, port inversion, and abnormal blood return. Based on this, the comprehensive risk head calculates an overall comprehensive risk score by combining the clinical importance and interrelationship of each complication, and maps it to a discrete risk level (such as normal, level 1 warning, level 2 warning, etc.). Finally, the comprehensive risk level is output as the core basis for judging whether a patient has port-related complications and their severity, and is used to trigger subsequent warning and clinical intervention processes.

[0090] In step S203, a complication warning action is performed based on the risk level.

[0091] It is understood that, based on the comprehensive risk level output by the prediction model, the embodiments of this application automatically execute graded early warning actions. The specific early warning information and reminder methods can be set according to actual needs and are not specifically limited here. For example, if the risk threshold is set to a level 2 early warning, and the risk level is lower than level 2 (such as normal or level 1 early warning), it will only be displayed in blue or green on the monitoring interface, the data will be recorded and the regular monitoring frequency will be maintained. If the risk level is greater than or equal to level 2 early warning, the early warning and the judgment of the target complication will be initiated. For example, when level 2 early warning is reached, a yellow warning will be highlighted on the interface, the early warning information will be pushed to the workstation of medical staff, and it will be recommended to strengthen observation. If the risk level is level 3 or level 4 early warning, a red alarm will be triggered immediately, and a structured early warning information containing the type of target complication, key abnormal indicators and preliminary treatment suggestions will be generated. This information will be actively pushed through pop-ups, sound prompts and mobile notifications, and to-do items will be automatically created, the subsequent monitoring frequency will be increased, and the electronic medical record system will be linked to recommend emergency assessment when necessary. All early warning actions support medical staff confirmation, processing tracking and feedback closed loop to ensure that high-risk patients receive timely and accurate clinical intervention.

[0092] The method for early warning of complications related to infusion ports proposed in this application can collect thermal imaging data, pressure data, ultrasound image data, visible light image data, and clinical data of the user's infusion port area using a data acquisition device. A computer communicates with the data acquisition device to generate modal data based on the collected data. A pre-trained prediction model is then invoked, and the modal data is input into the prediction model. The prediction model outputs a comprehensive risk level of complications for the user and executes early warning actions based on the risk level. By integrating multimodal information such as thermal imaging, pressure sensing, ultrasound, visible light images, and clinical data, and using a pre-trained intelligent prediction model to perform fusion analysis and collaborative reasoning on early minor abnormalities in the infusion port area, it can accurately identify early signs of common complications such as infection, obstruction, and extravasation, and output a comprehensive risk level with clinical interpretability, thereby achieving early detection and early warning of complications. Compared with traditional passive monitoring methods that rely on obvious symptoms, this method significantly improves the sensitivity and timeliness of early warning.

[0093] It should be noted that the foregoing explanation of the embodiment of the port-related complication early warning system also applies to the port-related complication early warning method of this embodiment, and will not be repeated here.

[0094] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0095] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0096] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0097] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any of the following techniques known in the art, or a combination thereof: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0098] Those skilled in the art will understand that all or part of the steps of the methods implementing the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0099] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A port-of-care port-related complication early warning system, characterized in that, include: The data acquisition device is used to acquire at least one of the following: thermal imaging data, pressure data, ultrasound image data, and visible light image data of the area where the user's infusion port is located, and to obtain the user's clinical data. A computer device communicates with the acquisition device. Based on at least one generated modal data from the thermal imaging data, pressure data, ultrasound image data, visible light image data, and clinical data, the computer device calls a pre-trained prediction model, inputs the modal data into the prediction model, outputs a comprehensive risk level for the user's complications through the prediction model, and executes a complication warning action based on the risk level. The prediction model includes multiple task heads and a comprehensive risk head. Each task head predicts a risk score for a corresponding complication related to the infusion port. A comprehensive score for the user's complications is determined using the comprehensive risk head and the risk scores predicted by each task head. The risk level of the user's complications is determined based on the comprehensive risk head and the risk scores predicted by each task head. Determining the comprehensive score for the user's complications using the comprehensive risk head and the risk scores predicted by each task head includes weighted fusion of the risk scores predicted by each task head to obtain a master risk score, calculating a correlation risk score based on the risk scores of different complications, and calculating the comprehensive score for the user's complications based on the master risk score and the correlation risk scores. The formula for the master risk score is: in, For the first Complication risk score predicted by each task head ; For predefined importance coefficients The fusion weights obtained by normalization ,in, Let be the clinical importance coefficient of the j-th complication. For the first to the second The sum of the clinical importance coefficients of each individual is obtained by adding them together; The formula for calculating the correlation risk score is as follows: in, For the first Complication risk score predicted by each task head; The complication risk score predicted for the i-th task head; For the first The effective association strength between the complication risk score predicted by the i-th task head and the complication risk score predicted by the i-th task head; The formula for calculating the comprehensive score is as follows: in, This is the correlation amplification factor obtained by training the prediction model.

2. The early warning system for port-of-care complications according to claim 1, characterized in that, Each task head includes a fully connected layer, a nonlinear activation layer, a regularization layer, and an output layer. Each task head predicts the risk score of related complications associated with the infusion port, including: Each task head utilizes the fully connected layer to extract feature data of corresponding complications from the thermal imaging data, the pressure data, the ultrasound image data, the visible light image data, and the clinical data; The feature data is sequentially input into the nonlinear activation layer and the regularization layer; The feature data that has passed through the nonlinear activation layer and the regularization layer is passed through the output layer, which maps the feature data to the risk probability of related complications, and evaluates the risk score of related complications based on the risk probability.

3. The early warning system for port-of-care complications according to claim 1, characterized in that, When the computer device detects that the risk level is greater than the risk threshold, it determines the target complication based on the risk score predicted by each task head, and generates early warning information based on the target complication, so as to provide early warning of infusion port-related complications and locate the target complication related to the infusion port at the same time.

4. The early warning system for port-of-care complications according to claim 1, characterized in that, The plurality of task heads includes multiple infection heads, clamping heads, extravasation heads, occlusion heads, flipping heads, and blood return heads, wherein the infection head, clamping head, extravasation head, occlusion head, flipping head, and blood return head are determined based on complications associated with the infusion port.

5. The early warning system for port-of-care complications according to claim 1, characterized in that, The computer device is also used to train the prediction model before invoking a pre-trained prediction model, the training process including: Obtain the training dataset for the prediction model, wherein the training dataset includes training samples and ground truth labels. The training samples are thermal imaging data, pressure data, ultrasound image data, visible light image data and clinical data of the area where the user's port is located. The ground truth labels are the risk score of the user's port-related complications and the comprehensive score of the user's complication status. The prediction model is trained using training samples from the training dataset. During the training process, the training loss is calculated using the ground truth labels and the prediction risk score and prediction comprehensive score of the prediction model. The model parameters of the prediction model are then updated based on the training loss.

6. The early warning system for port-of-care complications according to claim 1, characterized in that, The computer device further includes a thermal imaging encoder, an ultrasound encoder, a pressure encoder, and a clinical data encoder. After generating modal data based on at least one of the thermal imaging data, the pressure data, the ultrasound image data, the visible light image data, and the clinical data, it further includes: The thermal imaging data is extracted from the modal data using the thermal imaging encoder to obtain the feature vector of the thermal imaging data; The ultrasonic encoder is used to extract the ultrasonic image data from the modal data to obtain the feature vector of the ultrasonic image data; The pressure data in the modal data is extracted using the pressure encoder to obtain the feature vector of the pressure data; The clinical data is extracted from the modal data using the clinical data encoder to obtain the feature vector of the clinical data.

7. The early warning system for port-of-care complications according to claim 6, characterized in that, The computer equipment is also used for: The feature vector of each modality is transformed into a query vector, key vector, and value vector through linear projection. Calculate attention-weighted feature vectors from other modalities based on the query vector, key vector, and value vector of the current modality and the query vector, key vector, and value vector of other modalities; The feature vector of the current modality is added to the attention-weighted feature vectors from other modalities, and then concatenated along the channel dimension to obtain the fused modal data.

8. A method for early warning of complications related to infusion ports, characterized in that, The method is applied to the computer equipment of the port-of-care complication early warning system according to any one of claims 1-7, and the method includes the following steps: At least one modal data of the thermal imaging data, the pressure data, the ultrasound image data, the visible light image data, and the clinical data is generated; The modal data is input into the prediction model, and the prediction model outputs the comprehensive risk level of the user's complications. Implement complication warning actions based on risk levels.

Images