A venous catheter complication adaptive identification and nursing recommendation system and method

Abstract

The application provides a venous catheter complication adaptive identification and nursing recommendation system and method, relates to the technical field of artificial intelligence, and comprises a multi-modal acquisition terminal layer module, which is used for synchronously collecting hyperspectral images, local skin temperature fields and catheter port exudate near-infrared spectrum data of a puncture site of a patient to obtain multi-modal data; an edge computing node layer, which is deployed locally in a ward and is used for receiving and processing the multi-modal data, performing multi-modal data fusion, complication type and grading identification and nursing measure recommendation, and generating a structured nursing operation instruction set; and a cloud platform layer, which is used for being in communication connection with the edge computing node layer, globally storing patient data, managing and dynamically updating a standard nursing knowledge graph and training and optimizing a complication identification model, and pushing the updated model and the knowledge graph to the edge computing node layer.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an adaptive identification and nursing recommendation system and method for intravenous catheter complications. Background Technology

[0002] Intravenous therapy is one of the most common clinical treatments, with hundreds of millions of patients worldwide receiving intravenous infusions, chemotherapy, and parenteral nutrition each year. However, the incidence of catheter-related complications remains high, seriously affecting patient safety and the quality of care. Traditional complication identification relies mainly on nurses' visual observation and experience, which suffers from high subjectivity, low early identification rates, and inconsistent assessment standards, often leading to delays in diagnosis and treatment. Therefore, there is an urgent need for an objective, intelligent, and real-time decision support tool to improve the safety and efficiency of nursing care.

[0003] Currently, some artificial intelligence technologies have been applied to the field of intravenous therapy and nursing. By collecting ordinary RGB images of the puncture site, convolutional neural networks are used to extract and classify image features to determine the presence or absence of phlebitis and classify it into grades I to IV. After identification, corresponding processing prompts are output to provide nurses with some reference information.

[0004] However, existing technologies rely solely on a single RGB image, which cannot capture changes in tissue oxygenation, abnormal temperature fields, and differences in the chemical composition of exudates caused by early inflammation. This results in low accuracy in identifying grade I phlebitis and early exudates, making timely warnings difficult. Furthermore, the output prompts lack standardized and executable operational instructions, are disconnected from clinical nursing procedures, and nurses still need to consult nursing guidelines on their own, limiting their practicality. Summary of the Invention

[0005] To address the shortcomings of existing technologies that rely solely on a single RGB image, failing to capture changes in tissue oxygenation, abnormal temperature fields, and differences in the chemical composition of exudates caused by early inflammation, resulting in low accuracy in identifying grade I phlebitis and early exudates and difficulty in providing timely warnings; and to address the lack of standardized and executable operational instructions in the output prompts, which are disconnected from clinical nursing procedures and require nurses to consult nursing guidelines independently, thus limiting practicality, this invention provides an adaptive identification and nursing recommendation system and method for intravenous catheter complications.

[0006] The technical solutions provided by the embodiments of the present invention are as follows: A first aspect of the present invention provides an adaptive identification and nursing recommendation system for intravenous catheter complications, comprising: The multimodal acquisition terminal layer is used to simultaneously acquire hyperspectral images of the patient's puncture site, local skin temperature field, and near-infrared spectral data of catheter orifice exudate to obtain multimodal data. The edge computing node layer, deployed locally in the ward, is used to receive and process multimodal data, perform multimodal data fusion, identify complication types and grades, recommend nursing interventions, and generate a set of structured nursing operation instructions. The cloud platform layer is used to communicate and connect with the edge computing node layer, globally store patient data, manage and dynamically update the standard nursing knowledge graph, train and optimize the complication identification model, and push the updated model and knowledge graph to the edge computing node layer.

[0007] A second aspect of the present invention provides an adaptive identification and nursing recommendation method for venous catheter complications, applied to the adaptive identification and nursing recommendation system for venous catheter complications in the first aspect, comprising: S1: Obtain basic patient information; S2: Initialize preprocessing parameters and inference rules for basic patient information; S3: Through the multimodal acquisition terminal layer, hyperspectral images of the puncture site, local skin temperature field, and near-infrared spectral data of catheter orifice exudate are simultaneously acquired to obtain multimodal data; S4: Through the edge computing node layer, multimodal data is synchronized in time, aligned in space, and preprocessed adaptively. S5: Through a gated fusion mechanism, the features of the three preprocessed modalities are fused to generate a unified fused feature vector; S6: Input the fused feature vectors into a multi-expert hybrid model or a Transformer-based multimodal fusion network to identify the types and severity levels of six complications; S7: Based on the type and severity classification of complications, combined with dynamically updated standard nursing knowledge graphs, complication identification results and patient characteristics, generate a personalized set of structured nursing operation instructions; S8: Based on the structured nursing operation instruction set, nurses execute recommended nursing measures, record the execution status, and upload the data to the cloud platform layer; S9: Through the cloud platform layer, the uploaded data is used regularly for model training and knowledge graph updates, and the updated model and knowledge graph are pushed to the edge computing node layer.

[0008] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the adaptive identification and nursing recommendation method for venous catheter complications as described in the second aspect.

[0009] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this embodiment of the invention, by simultaneously acquiring three modal data—hyperspectral imaging, temperature field, and near-infrared spectroscopy—it is possible to capture microscopic tissue changes and metabolic abnormalities in early complications, significantly improving the accuracy of identifying grade I phlebitis and early exudation, and achieving early warning of complications. Simultaneously, based on a dynamically updated standard nursing knowledge graph, this invention automatically infers and generates a set of executable structured nursing operation instructions containing operation names, operation frequencies, key points, precautions, and assessment criteria. The identification results are directly converted into standardized nursing plans, allowing nurses to execute the instructions without needing to consult additional guidelines. This truly achieves a fully intelligent closed-loop process from identification to intervention, improving clinical practicality and nursing efficiency. Attached Figure Description

[0010] 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 accompanying 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.

[0011] Figure 1 This is a schematic diagram of the structure of an adaptive identification and nursing recommendation system for intravenous catheter complications provided in an embodiment of the present invention.

[0012] Figure 2 This is a schematic diagram of the composition of an adaptive identification and nursing recommendation system for intravenous catheter complications provided in an embodiment of the present invention.

[0013] Figure 3 This is a schematic diagram illustrating the computational principle of an edge computing node provided in an embodiment of the present invention.

[0014] Figure 4 This is a schematic diagram of a multi-expert mixed complication identification process provided in an embodiment of the present invention.

[0015] Figure 5 This is a schematic diagram of a knowledge graph-based nursing intervention recommendation mechanism provided in an embodiment of the present invention.

[0016] Figure 6 This is a flowchart illustrating an adaptive identification and nursing recommendation method for intravenous catheter complications provided in an embodiment of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0019] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0020] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0021] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0022] Reference manual attached Figure 1 The diagram shows a structural schematic of an adaptive identification and nursing recommendation system for intravenous catheter complications provided by an embodiment of the present invention.

[0023] Reference manual attached Figure 2 The diagram shows a schematic representation of the composition of an adaptive identification and nursing recommendation system for intravenous catheter complications provided in an embodiment of the present invention.

[0024] This invention provides an adaptive identification and nursing recommendation system 30 for intravenous catheter complications, the system comprising: The multimodal acquisition terminal layer 301 is used to simultaneously acquire hyperspectral images of the patient's puncture site, local skin temperature field, and near-infrared spectral data of catheter orifice exudate to obtain multimodal data.

[0025] In one possible implementation, the multimodal acquisition terminal layer 301 specifically includes: The hyperspectral imaging module is used to acquire hyperspectral images in the 400-1000nm band.

[0026] Infrared thermal imaging module, used to collect local skin temperature field.

[0027] The near-infrared spectroscopy detection module is used to collect the near-infrared spectrum of exudates in the 900-1700nm wavelength range.

[0028] The terminal control and communication module is used to trigger the synchronous data collection of each module through hardware and send the collected data to the edge computing node layer to obtain multimodal data.

[0029] It should be noted that the core server of this invention is an Intel Xeon Platinum 8369B processor, 64GB of memory, and a 2TB SSD. One NVIDIA A10 GPU is used for model training. Five application servers are used to handle user requests. Three database servers are used for data storage. Load balancing technology is employed to achieve high availability, with a system availability rate of 99.9%.

[0030] Specifically, the multimodal acquisition terminal is a handheld integrated device consisting of the following four modules: a hyperspectral imaging module, an infrared thermal imaging module, a near-infrared spectral detection module, and a terminal control and communication module.

[0031] Furthermore, the hyperspectral imaging module is used to acquire hyperspectral images of the puncture site in the 400-1000nm band (the 400-1000nm band covers the main absorption peaks of hemoglobin, water and melanin, and can effectively distinguish between normal tissue and inflamed tissue) to capture microstructural changes in skin tissue.

[0032] Furthermore, in terms of specific implementation technology, pushbroom hyperspectral imaging technology is adopted, with a spectral resolution of 5nm and a spatial resolution of 0.1mm / pixel (the 5nm spectral resolution and 0.1mm spatial resolution balance imaging quality and speed), and the imaging time is less than 1 second. Reflectivity correction is performed using a built-in standard white board, with the standard formula as follows:

[0033] in, Indicates wavelength The reflectivity at that location. Indicates the sample at wavelength The original image intensity at that location. Indicates at wavelength Dark current noise at the location. Indicates the standard whiteboard at wavelength Image intensity at the puncture site. After scanning and processing the optical signals at the puncture site, the output is a hyperspectral data cube with dimensions (512, 512, 121).

[0034] Specifically, the infrared thermal imaging module employs an uncooled infrared focal plane array with a temperature resolution of 0.05℃ and a temperature measurement range of 10-50℃ (the 0.05℃ resolution can detect temperature increases of 0.1-0.5℃ caused by early inflammation). The imaging frequency is 30Hz (a 30Hz imaging frequency can capture dynamic changes in the temperature field). This module is used to acquire the local skin temperature field at the puncture site, reflecting the degree of tissue inflammation. Temperature measurement accuracy is improved through a two-point temperature correction method; the expression for two-point temperature correction is:

[0035] in, This indicates the actual temperature. This indicates the detector's output voltage. and These represent the correction coefficients, obtained by measuring two blackbody sources at known temperatures. The final output is a temperature field matrix with dimensions (256, 256).

[0036] Furthermore, the near-infrared spectroscopy detection module employs a miniature fiber optic spectrometer with a spectral range of 900-1700 nm (the 900-1700 nm band is sensitive to components such as water, protein, and glucose, and can effectively distinguish between normal exudate, blood, and pus), and a spectral resolution of 2 nm. The near-infrared spectroscopy detection module acquires the near-infrared spectrum of the exudate at the catheter orifice and analyzes its chemical composition. Multiple scattering correction (MSC) is used to eliminate the influence of scattering.

[0037] in, This indicates the corrected spectrum. Represents the original spectrum. This represents the average value of the spectra of all samples. This represents the standard deviation of the spectra of all samples.

[0038] Near-infrared spectroscopy detection technology is a well-known technology and will not be elaborated here.

[0039] Specifically, the terminal control and communication module controls the synchronous operation of each acquisition module, sending the acquired data to the edge computing node. The module controller can, for example, use an ARM Cortex-M4 processor with built-in Wi-Fi and Bluetooth 5.0 communication modules, supporting encrypted data transmission. Synchronous acquisition by the three acquisition modules is achieved through hardware triggering, with a synchronization accuracy of less than 1ms. For details, please refer to the composition and principles of existing IoT communication control terminals.

[0040] Furthermore, before data collection, nurses and other medical staff need to enter the patient's basic information on the edge computing node, including name, gender, age, hospital number, diagnosis, allergy history, underlying diseases, and intravenous treatment plan. The system automatically initializes preprocessing parameters and knowledge graph inference rules based on the patient information. Personalized system initialization can improve the accuracy of data preprocessing and the relevance of nursing intervention recommendations.

[0041] It should be noted that during data collection, the nurse holds the multimodal acquisition terminal 10-15cm away from the puncture site, keeping the lens perpendicular to the skin, and presses the acquisition button. The terminal synchronously activates the hyperspectral camera, infrared thermal imager, and near-infrared spectrometer via hardware triggering, completing the acquisition of all data within one second. After acquisition, the terminal automatically sends the data to the edge computing node. Synchronous acquisition ensures the temporal consistency of multimodal data, improving the accuracy of subsequent fusion and recognition.

[0042] In this embodiment of the invention, hardware-triggered synchronous acquisition of three modal data—hyperspectral imaging, temperature field, and near-infrared spectroscopy—enables multidimensional information capture within one second of tissue oxygenation changes, inflammatory heat distribution, and exudate chemical composition at the puncture site. This captures early microscopic tissue changes and metabolic abnormalities that are undetectable by conventional visual observation. Each module employs clinically adapted band and resolution designs, coupled with standardized correction algorithms, to ensure data quality and consistency. This provides a reliable data foundation for the accurate fusion and identification of subsequent edge computing nodes, serving as the primary guarantee for achieving early warning and intelligent nursing decision-making.

[0043] Reference manual attached Figure 3 The diagram illustrates the computational principle of the edge computing node provided in this embodiment of the invention.

[0044] Edge computing node layer 302, deployed locally in the ward, is used to receive and process multimodal data, perform multimodal data fusion, identify complication types and grades, recommend nursing interventions, and generate a structured nursing operation instruction set.

[0045] In one possible implementation, the edge computing node layer specifically includes: The data receiving and synchronization module is used to receive multimodal data and upsample the temperature field data to hyperspectral image resolution through timestamp synchronization algorithm and spatial alignment operation.

[0046] Specifically, this module receives multimodal data sent by the acquisition terminal and performs time synchronization and data alignment. A timestamp-based synchronization algorithm is used to align the hyperspectral image, temperature field, and near-infrared spectral data to the same time reference. Bilinear interpolation is used to upsample the temperature field data to the same spatial resolution as the hyperspectral image, achieving pixel-level alignment. The data receiving module receives data packets sent by the acquisition terminal every 2 seconds, extracts the timestamp of each data point, calculates the time deviation, and corrects it. Then, the temperature field data is upsampled from (256, 256) to (512, 512) and pixel-level aligned with the hyperspectral image.

[0047] The adaptive data preprocessing module performs wavelet denoising, adaptive histogram equalization, and Savitzky-Golay smoothing filtering on hyperspectral images, temperature fields, and near-infrared spectra, respectively, and performs adaptive normalization based on patient historical data in the multimodal data to obtain normalized data.

[0048] Specifically, denoising, enhancing, and normalizing multimodal data can improve data quality. The following are some preprocessing methods: Hyperspectral image denoising uses a wavelet thresholding algorithm, selecting the db4 wavelet basis (suitable for medical images), with a decomposition level of 3 (3-level decomposition balances denoising effect and computational complexity), and the threshold formula is:

[0049] in, Indicates the threshold. Indicates the standard deviation of noise. Represents logarithmic operations. Indicates the signal length.

[0050] Furthermore, the temperature field is enhanced by employing an adaptive histogram equalization algorithm to improve its contrast and highlight abnormal temperature regions. Near-infrared spectral preprocessing utilizes Savitzky-Golay smoothing filtering with an 11-point window and a polynomial degree of 2 to remove spectral noise (the 11-point window and second-order Savitzky-Golay filtering effectively smooth the spectrum while preserving characteristic peaks).

[0051] Specifically, adaptive normalization automatically adjusts the normalization parameters based on the patient's skin color and basal body temperature.

[0052] in, This represents the average of the patient's historical data. This represents the standard deviation of the patient's historical data. This represents the normalization parameter.

[0053] Furthermore, the data receiving and synchronization module performs time synchronization and spatial alignment on the received data, upsampling the temperature field data to the same resolution as the hyperspectral image. Then, the adaptive data preprocessing module performs wavelet denoising, histogram equalization, and / or Savitzky-Golay smoothing on the data. Finally, the multimodal gated fusion module dynamically calculates the weights of each modality to generate the fused feature vector. Data preprocessing removes noise and interference, improving data quality. The gated fusion mechanism can fully utilize the complementary information of different modalities, enhancing the feature representation capability.

[0054] The multimodal gating fusion module is used to dynamically calculate the modal weights of normalized data and generate a unified fusion feature vector.

[0055] In one possible implementation, the multimodal gating fusion module employs a gating fusion mechanism, which specifically includes: The hyperspectral image features are multiplied by the first weight, the temperature field features are multiplied by the second weight, and the near-infrared spectral features are multiplied by the third weight to obtain the fused feature vector.

[0056] The first weight, the second weight, and the third weight are the weights of the three modes, and the sum of the three is 1, which is learned by the gating network.

[0057] Specifically, this module fuses features from three modalities: hyperspectral imagery, temperature field, and near-infrared spectroscopy, to generate a unified feature representation. The module employs a gated fusion mechanism to dynamically adjust the weights of different modalities.

[0058] in, , as well as These represent the characteristics of the hyperspectral, temperature field, and near-infrared spectra, respectively. , as well as These represent the weights of each mode. The gating network structure is learned from the gating network and consists of: input layer (3D) → hidden layer (16D, ReLU activation) → output layer (3D, Softmax activation), which will not be elaborated here.

[0059] It should be noted that the gated fusion mechanism can automatically adjust the weights of each modality based on the characteristics of different complications. For example, for phlebitis, the weight of the temperature field will be automatically increased. For exudation, the weights of hyperspectral and near-infrared spectra will be automatically increased. The final output is a fused feature vector (256 dimensions).

[0060] The multi-expert hybrid complication identification module is used to identify six complications and their corresponding severity levels in parallel based on fused feature vectors: phlebitis, exudation, thrombosis, infection, catheter occlusion, and catheter dislodgement.

[0061] In one possible implementation, the multi-expert hybrid complication identification module is based on a multi-expert hybrid model (MoE) for symptom identification, and the module specifically includes: There are 6 expert networks, each corresponding to a complication. The structure consists of an input layer, a first hidden layer, a second hidden layer, and an output layer. The output layer contains 5 neurons, corresponding to no complication, grade I, grade II, grade III, and grade IV, respectively.

[0062] A gated network is used to calculate the weights of six expert networks.

[0063] The multi-expert hybrid complication identification module is trained using a multi-task cross-entropy loss function. The calculation method for the multi-task cross-entropy loss function is as follows: For each complication, the corresponding clinical importance weight is multiplied by the sum of the products of the true labels and predicted logarithms of the five complication grades to obtain the loss value for that complication. The loss values ​​for the six complications are summed and negative to obtain the final loss value.

[0064] The module's final output is a weighted sum of the outputs from each expert network, enabling simultaneous identification and grading of multiple complications.

[0065] In one possible implementation, the multi-expert hybrid complication identification module performs symptom identification based on a Transformer multimodal fusion network, which specifically includes: The multimodal feature extraction branch uses a 3D convolutional neural network to extract spatial-spectral joint features, a 2D convolutional neural network to extract temperature distribution features, and a 1D convolutional neural network to extract spectral features.

[0066] The cross-modal attention fusion module adopts a multi-head self-attention mechanism to realize the interaction and fusion between features of different modalities and outputs a fused feature vector.

[0067] The complication classification and grading module uses a fully connected network to perform multi-task classification and grading of six complications, and outputs a probability matrix with a dimension of 6×5.

[0068] Specifically, the multi-expert hybrid complication identification module inputs the fused feature vector into six expert networks, with each expert network outputting the probability of complication classification. A gating network automatically calculates the weights of each expert network based on the input features, and the weighted sum yields the final identification result. The system displays the identification results on the screen, using different colors to mark complications of varying severity. The multi-expert hybrid model can simultaneously identify six common complications and perform accurate classification, achieving an accuracy rate of over 92%.

[0069] Reference manual attached Figure 4 The diagram illustrates a multi-expert mixed complication identification process provided by an embodiment of the present invention.

[0070] In this application, a multi-expert hybrid model (MoE) is used for complication identification. The fused features are analyzed to identify complication types and perform grading. The MoE model consists of 6 expert networks and 1 gated network, with each expert network corresponding to a complication type. The structure of each expert network is: Input layer (256 dimensions) → Hidden layer 1 (128 dimensions, ReLU) → Hidden layer 2 (64 dimensions, ReLU) → Output layer (5 dimensions, Softmax). The 5 neurons in the output layer correspond to: no complications, Grade I, Grade II, Grade III, and Grade IV. Model construction parameters: Training dataset: Contains multimodal data from 10,000 patients, including 4,000 normal samples and 6,000 complication samples. Optimizer: Adam, learning rate = 0.001, batch size = 32. Loss function: Multi-task cross-entropy loss, formula:

[0071] in, This represents the weight of the i-th complication, set according to clinical importance, for example: phlebitis 1.2, exudation 1.1, thrombosis 1.3, infection 1.4, catheter blockage 1.0, and catheter dislodgement 1.0. This indicates the actual label. This represents the model's predicted value.

[0072] Furthermore, the gated network structure is as follows: input layer (256 dimensions) → hidden layer (64 dimensions, ReLU) → output layer (6 dimensions, Softmax). The 6 neurons in the output layer correspond to the weights of the 6 expert networks. The model output expression is as follows:

[0073] in, This represents the weight of the i-th expert network. This represents the output of the i-th expert network.

[0074] Furthermore, the six expert networks were optimized for the characteristics of different complications, improving the recognition accuracy. The gating network can automatically select the most suitable expert network for the current sample. The multi-task loss function takes into account the clinical importance of different complications, improving the recognition rate of serious complications.

[0075] Furthermore, by fusing feature vectors (256 dimensions), a probability matrix (6×5) of the types and grades of six complications is obtained. In addition, the system automatically switches nursing alarm measures based on the complication identification results: when the probability of all complications is <0.3, the system is in normal mode, displaying a "No abnormality" message and suggesting that nursing care be performed according to routine procedures. When the probability of a grade I complication is >0.7, the system enters warning mode, displaying a yellow warning message, prompting nurses to strengthen observation, and recommending preventative nursing measures. When the probability of a grade II or higher complication is >0.7, or multiple complications occur concurrently, the system enters intervention mode, displaying a red warning message, automatically generating a detailed set of nursing operation instructions, and reminding the head nurse to review them.

[0076] It should be noted that the complication identification in this case can also employ a Transformer-based multimodal fusion network to replace the original multi-expert hybrid model. The model input consists of raw data from hyperspectral imagery, temperature fields, and near-infrared spectra, and end-to-end identification is achieved through the following three key modules: Furthermore, for multimodal feature extraction, independent feature extraction branches are designed for each modality: The hyperspectral branch uses a 3D convolutional neural network (3D-CNN) with an input size of (512, 512, 256), extracting spatial-spectral joint features through three 3D convolutional layers, resulting in a 256-dimensional output feature. The temperature field branch uses a 2D convolutional neural network (2D-CNN) with an input size of (512, 512, 1), extracting temperature distribution features through three 2D convolutional layers, resulting in a 128-dimensional output feature. The near-infrared spectral branch uses a 1D convolutional neural network (1D-CNN) with an input size of (1, 1024), extracting spectral features through three 1D convolutional layers, resulting in a 128-dimensional output feature.

[0077] Furthermore, cross-modal attention fusion is achieved by introducing a multi-head self-attention mechanism to enable interaction and fusion between features from different modalities. The attention network structure is as follows: input layer (512 dimensions) → multi-head attention layer (8 heads, output 512 dimensions) → feedforward network layer (output 512 dimensions) → output layer (256 dimensions).

[0078] Furthermore, for complication classification and grading, a fully connected network is used to achieve multi-task classification and grading of 6 types of complications. The output layer uses the Softmax activation function, and the output dimension is 6×5 (6 types of complications, 5 grades for each type).

[0079] Furthermore, regarding model training and optimization strategies, the dataset was built upon the original 10,000 cases, adding 5,000 cases of complex complications, covering scenarios involving multiple concurrent and rare complications, bringing the total dataset size to 15,000 cases. The training strategy employed transfer learning, fine-tuning the pre-trained model on Image Net. The optimizer used was Adam W, with an initial learning rate of 0.0001, employing a cosine annealing learning rate scheduler. The batch size was set to 16, and the training run consisted of 50 epochs. The loss function used was multi-task cross-entropy loss, with weights assigned based on the clinical importance of different complications: phlebitis 1.2, exudation 1.1, thrombosis 1.3, infection 1.4, catheter occlusion 1.0, and catheter dislodgement 1.0. Focal Loss was also introduced to address class imbalance, with α set to 0.25 and γ set to 2.0.

[0080] Specifically, the complication probability matrix output by the deep learning identification model is directly input into the nursing intervention recommendation module, replacing the original multi-expert hybrid model output. The recommendation module, based on the original knowledge graph reasoning, adds the following optimization mechanisms: personalized recommendation based on reinforcement learning, constructing a reinforcement learning agent that dynamically adjusts the nursing intervention recommendation strategy using patient prognosis as the reward signal. The agent's state consists of patient characteristics, complication identification results, and current nursing interventions; the action space includes all available nursing interventions; and the reward function is defined as the weighted sum of the patient recovery time reduction rate and the complication recurrence rate reduction rate.

[0081] Furthermore, a real-time feedback closed-loop optimization mechanism is established to provide real-time feedback on the effectiveness of nursing interventions. This mechanism feeds actual patient recovery data (such as symptom relief time and complication outcomes) back into the deep learning model, enabling continuous model optimization. Every 1000 new feedback data points automatically triggers a model fine-tuning, ensuring that model performance continuously improves with clinical practice.

[0082] Furthermore, the deep learning module adopts a modular design, allowing for seamless integration with existing edge computing nodes. In scenarios with ample edge computing resources, complete deep learning models can be deployed directly. In resource-constrained scenarios, model compression techniques (such as knowledge distillation and quantization) can be used to compress the model size to 30% of its original size, ensuring real-time operation on low-computing-power devices.

[0083] Furthermore, through internal testing, the deep learning model demonstrated an 8-12% improvement in accuracy for identifying complex complications compared to the original multi-expert hybrid model, while reducing the average identification time by 30%. In terms of nursing intervention recommendations, reinforcement learning-based personalized recommendations shortened the average recovery time for patients by 15-20% and reduced the recurrence rate of complications by 10-15%.

[0084] The nursing intervention recommendation module is used to generate a set of structured nursing operation instructions based on the severity of complications and the corresponding severity levels, combined with a standard nursing knowledge graph.

[0085] In one possible implementation, the nursing intervention recommendation module makes inferences based on a dynamically updated knowledge graph of standard intravenous therapy nursing care.

[0086] The knowledge graph specifically includes three types of entities: complication entities, nursing intervention entities, and patient characteristic entities, as well as five types of relationships: cause, applicability, contraindication, adjustment, and substitution.

[0087] The reasoning process combines rule-based reasoning with similarity-based reasoning to generate a structured nursing operation instruction set that includes operation name, operation frequency, operation points, precautions, and evaluation criteria.

[0088] Specifically, this module is used to infer personalized nursing interventions from a standard nursing knowledge graph based on complication identification results. Its recommendation mechanism is as follows: First, a knowledge graph is constructed based on the "Specifications for Intravenous Therapy Nursing Techniques" (WS / T433-2013) and the latest clinical guidelines, containing three types of entities: complication entities (6 types), nursing intervention entities (42 types), and patient characteristic entities (15 types). Five types of relationships are established: causing, applicable, contraindicated, adjustment, and substitution. A method for constructing this knowledge graph is provided below.

[0089] Reference manual attached Figure 5 This diagram illustrates a knowledge graph-based nursing intervention recommendation mechanism provided by an embodiment of the present invention. An ontology-based construction method is employed. Ontology, as the core of the knowledge graph, can accurately define concepts, attributes, and relationships within the domain of intravenous therapy nursing. Like medical ontology, it provides a solid framework for the entire knowledge graph, clarifying the categories and interrelationships of various entities such as complications, nursing interventions, and patient characteristics. The construction process is as follows: Specifically, data collection and preprocessing involved gathering information from a wide range of data sources, including the "Standardized Operating Procedures for Intravenous Therapy Nursing" (WS / T433-2013), the latest clinical guidelines, relevant medical literature, and actual case data. The collected data was then cleaned and formatted to remove duplicates, errors, and irrelevant information, ensuring accuracy and consistency. For example, the descriptions of the same complication in different literature were standardized, laying a solid foundation for subsequent knowledge extraction.

[0090] Furthermore, entity recognition and linking, leveraging natural language processing technology, accurately identifies key entities from the preprocessed text, such as the six complication entities, 42 nursing intervention entities, and 15 patient characteristic entities mentioned earlier. These identified entities are then linked to entities in the existing knowledge base to ensure their uniqueness and accuracy. For example, the identified entity "phlebitis" is associated with existing definitions and related information in the medical knowledge base, enriching the entity's meaning.

[0091] Furthermore, relation extraction and knowledge fusion involve in-depth analysis of relationships between entities, such as five types of relationships: "cause," "applicable," "contraindicated," "adjust," and "substitute." These relationships are automatically extracted from text using machine learning and deep learning algorithms. Simultaneously, the extracted knowledge is integrated with existing knowledge graphs to eliminate conflicts and contradictions, enhancing the graph's coherence and accuracy. For example, after extracting the relationship "a certain nursing measure is applicable to a certain type of patient characteristics," it is integrated into the existing knowledge graph, improving the relationship network between entities.

[0092] Specifically, the knowledge graph validation and improvement process involves inviting experts in the medical field to validate the constructed knowledge graph, checking the accuracy of entities, the rationality of relationships, and the completeness of knowledge. Based on the experts' feedback, the knowledge graph is further refined and optimized. For example, experts might point out that the definitions of certain entities are not accurate enough, or the descriptions of certain relationships are not clear enough, requiring us to modify and adjust these parts to improve the quality of the knowledge graph.

[0093] It is worth noting that a well-constructed knowledge graph for intravenous therapy nursing has broad application prospects. In clinical decision support, doctors and nurses can quickly obtain the best nursing interventions for different patient characteristics and complications by consulting the knowledge graph, providing patients with more precise and personalized treatment plans. In intelligent nursing training, the knowledge graph can supplement training materials, helping nursing staff to learn intravenous therapy nursing knowledge more systematically and improve their professional skills. Furthermore, the knowledge graph can also be applied to nursing quality assessment. By analyzing the data in the knowledge graph, the effectiveness of nursing interventions can be evaluated, providing a basis for continuous improvement of nursing quality.

[0094] Specifically, the reasoning rules employ a combination of rule-based reasoning and similarity-based reasoning: rule-based reasoning matches corresponding standard nursing interventions based on the type and grade of complications, while similarity-based reasoning calculates the similarity between the current patient and historical patients and recommends effective nursing interventions for similar patients.

[0095] Furthermore, the nursing interventions are represented in a structured manner, which is a set of structured instructions containing operation name, operation frequency, operation points, precautions and assessment criteria. The nursing intervention recommendation module receives the complication identification results and the patient's basic information. First, it matches the corresponding standard nursing interventions in the knowledge graph, and then makes personalized adjustments based on the patient's age, allergy history, underlying diseases and other characteristics. Finally, it generates a set of structured nursing operation instructions.

[0096] Furthermore, during operation, the nursing intervention recommendation module receives complication identification results and patient characteristics. First, it matches corresponding standard nursing interventions in the knowledge graph, and then makes personalized adjustments based on the patient's age, allergy history, underlying diseases, and other characteristics. For example, for elderly patients, the dressing change interval is appropriately extended. For patients with a history of drug allergies, the use of iodine-containing disinfectants is avoided. Finally, a structured set of nursing operation instructions is generated, including the operation name, operation frequency, key points, precautions, and assessment criteria. Knowledge graph-based nursing recommendations ensure the standardization and normalization of nursing interventions while considering individual patient differences, thus improving nursing outcomes.

[0097] The local storage and display module is used to store local data and provide a visual operation interface based on the structured nursing operation instruction set.

[0098] Specifically, the local storage and display module includes a local storage module and a display module, used for locally storing patient examination data and nursing records, respectively, and providing nurses with a visual operating interface. The local storage module uses an SQLite database to store local data, with a storage capacity of 128GB, capable of storing complete data for 10,000 patients. The display module uses a 10.1-inch touchscreen display, and the interface includes four functional modules: data acquisition, recognition results, nursing recommendations, and historical records.

[0099] In this embodiment of the invention, the weights of each modality are dynamically adjusted through a gating fusion mechanism to adapt to individual patient differences. Combined with a multi-expert parallel recognition architecture and a dynamically updated nursing knowledge graph, six complications and their severity levels can be accurately identified simultaneously, and structured, executable, personalized nursing operation instructions can be generated. This achieves a seamless closed loop of identification and intervention, significantly improving the efficiency of clinical decision-making and the standardization of nursing care.

[0100] The cloud platform layer 303 is used to communicate and connect with the edge computing node layer, globally store patient data, manage and dynamically update the standard nursing knowledge graph, train and optimize the complication identification model, and push the updated model and knowledge graph to the edge computing node layer.

[0101] It should be noted that the cloud platform is deployed on a mainstream public cloud platform, adopts a distributed architecture, and can elastically scale computing and storage resources. It consists of the following four modules: In one possible implementation, the cloud platform layer 303 specifically includes: The data storage module adopts a hybrid architecture of object storage, relational database, and graph database.

[0102] Specifically, this module stores multimodal data, identification results, nursing records, and system logs for all patients. The data storage module employs a hybrid storage architecture: object storage for unstructured data such as hyperspectral images and infrared thermograms; a relational database for structured data such as patient information, identification results, and nursing records; and a graph database for storing a knowledge graph of standard intravenous therapy nursing care. The data storage period is 15 years for patients, meeting medical data preservation requirements.

[0103] The knowledge graph management module supports visual editing by nursing experts and automatically retrieves the latest clinical guidelines and literature to update the knowledge graph.

[0104] Specifically, the knowledge graph management module provides a visual editing interface for the knowledge graph, allowing nursing experts to add, modify, and delete knowledge entities and relationships. It also automatically retrieves the latest clinical guidelines and research literature regularly, extracting new nursing knowledge and updating it into the knowledge graph.

[0105] The model training and update module uses federated learning technology to train the model using data uploaded by each edge node and periodically pushes model updates to the edge computing node layer.

[0106] Specifically, the module is used to train and optimize the complication identification model, and periodically pushes model updates to the edge computing nodes. It employs federated learning technology to train the model using data from each edge node while protecting patient privacy. Incremental model updates are performed monthly, and full updates are performed quarterly.

[0107] The system management module is used for user permission management, device status monitoring, and system configuration.

[0108] Specifically, this module manages user permissions, device status, and system configuration, providing a web-based management interface that supports user registration, permission allocation, device monitoring, and system configuration. It also generates system operation reports and nursing quality analysis reports.

[0109] The management functions of the cloud platform can be understood by referring to the existing medical and nursing management backend.

[0110] It should be noted that nurses perform the recommended nursing procedures according to the system, and record the execution time, personnel, and details in the system after completion. The system automatically links the nursing records with the patient's identification results and stores them locally and on the cloud platform. Electronic nursing records improve work efficiency and facilitate nursing quality traceability and data analysis.

[0111] Specifically, the cloud platform regularly collects data and nursing records uploaded by each edge node and uses federated learning technology to incrementally update the complication identification model. Simultaneously, the knowledge graph management module regularly retrieves the latest clinical guidelines and research literature, extracts new nursing knowledge, and updates the knowledge graph. The updated model and knowledge graph are automatically pushed to all edge computing nodes. Continuous model and knowledge updates enable the system to adapt to new clinical situations, maintaining high identification accuracy and the quality of nursing recommendations.

[0112] In one possible implementation, the system supports three operating modes: normal mode, early warning mode, and intervention mode.

[0113] In the normal mode, when the probability of all complications is less than 0.3, the system displays "no abnormalities" and recommends routine nursing procedures.

[0114] The early warning mode displays a yellow warning message and recommends preventive nursing measures when the probability of a certain complication being grade I is greater than 0.7.

[0115] The intervention mode involves the system displaying a red warning message when the probability of a certain complication being grade II or above is greater than 0.7, or when multiple complications occur concurrently. The system will automatically generate a detailed set of nursing operation instructions and remind the head nurse to review them.

[0116] Specifically, the system's application control logic is as follows: Nurses use a multimodal acquisition terminal to scan the patient's puncture site, simultaneously acquiring hyperspectral images, temperature fields, and near-infrared spectral data. The acquisition terminal sends the multimodal data to the data receiving and synchronization module of the edge computing node via Wi-Fi. The data receiving and synchronization module performs time synchronization and spatial alignment on the data before sending it to the adaptive data preprocessing module. The adaptive data preprocessing module performs denoising, enhancement, and normalization on the data before sending it to the multimodal gating fusion module. The multimodal gating fusion module fuses the features of the three modalities to generate a unified feature representation, which is then sent to the multi-expert hybrid complication identification module. The multi-expert hybrid complication identification module identifies the complication type and performs a classification judgment, sending the result to the nursing intervention recommendation module. Based on the identification results and patient characteristics, the nursing intervention recommendation module infers personalized nursing interventions from the knowledge graph and displays them on the screen of the edge node. The edge computing node uploads all data and processing results to the cloud platform for storage. The cloud platform periodically uses the uploaded data for model training and knowledge graph updates, pushing the updated model and knowledge to the edge computing nodes.

[0117] In this embodiment of the invention, a distributed architecture and hybrid storage technology are employed to achieve secure storage and 15-year long-term preservation of patient multimodal data, nursing records, and knowledge graphs. The knowledge graph management module supports visual editing and can automatically retrieve the latest clinical guidelines for dynamic updates. The model training module, based on federated learning, continuously optimizes the recognition model while preserving patient privacy within the hospital and periodically pushes it to edge nodes. Combined with a three-tiered operating mode of normal, early warning, and intervention, a closed-loop management system is achieved, enabling the system to self-evolve and continuously ensure the long-term stability of recognition accuracy and nursing recommendation quality.

[0118] Reference manual attached Figure 6 The diagram illustrates a flowchart of an adaptive identification and nursing recommendation method for intravenous catheter complications provided by an embodiment of the present invention.

[0119] This invention also provides an adaptive identification and nursing recommendation method for venous catheter complications, applied to the aforementioned adaptive identification and nursing recommendation system for venous catheter complications. This method can be implemented by an adaptive identification and nursing recommendation device for venous catheter complications, which can be a terminal or a server. The processing flow of the adaptive identification and nursing recommendation method for venous catheter complications may include the following steps: S1: Obtain basic patient information.

[0120] S2: Initialize preprocessing parameters and inference rules for basic patient information.

[0121] S3: Through the multimodal acquisition terminal layer, hyperspectral images of the puncture site, local skin temperature field, and near-infrared spectral data of catheter orifice exudate are simultaneously acquired to obtain multimodal data.

[0122] S4: Through the edge computing node layer, multimodal data is synchronized in time, aligned in space, and preprocessed adaptively.

[0123] S5: Through a gating fusion mechanism, the features of the three preprocessed modalities are fused to generate a unified fused feature vector.

[0124] S6: Input the fused feature vectors into a multi-expert hybrid model or a Transformer-based multimodal fusion network to identify the types and severity levels of six complications.

[0125] S7: Based on the type and severity classification of complications, combined with dynamically updated standard nursing knowledge graphs, complication identification results, and patient characteristics, generate a personalized set of structured nursing operation instructions.

[0126] S8: Based on the structured nursing operation instruction set, nurses perform recommended nursing measures, record the execution status, and upload the data to the cloud platform layer.

[0127] S9: Through the cloud platform layer, the uploaded data is used regularly for model training and knowledge graph updates, and the updated model and knowledge graph are pushed to the edge computing node layer.

[0128] For example, taking the general surgery ward of a tertiary hospital as the application target, the adaptive identification and nursing measure recommendation system for intravenous catheter complications of this invention is deployed. The deployment scheme is as follows: Specifically, eight VT-MC100 handheld multimodal acquisition terminals are deployed, with one terminal assigned to each responsible nurse. One EC-N100 edge computing node is deployed at the nurse station. The cloud platform is deployed on Alibaba Cloud ECS, configured with two core servers, one NVIDIA A10 GPU, five application servers, and three database servers.

[0129] Furthermore, the implementation process of this system is recorded as follows: Taking patient Wang (male, 65 years old, post-gastric cancer surgery, receiving intravenous chemotherapy) as an example: First, patient information entry: At 8:00 AM on October 1, 2025, the responsible nurse, Li, entered patient Wang's basic information on the edge computing node: Name: Wang; Gender: Male; Age: 65 years old; Hospitalization number: 123456; Diagnosis: Post-gastric cancer surgery; Allergy history: penicillin allergy; Underlying disease: hypertension; Intravenous treatment plan: paclitaxel + cisplatin chemotherapy; Indwelling PICC catheter. The system automatically initializes preprocessing parameters and knowledge graph reasoning rules.

[0130] Further, multimodal data acquisition: At 8:05 AM, nurse Li held the multimodal acquisition terminal 12cm away from the PICC puncture site on the patient's right upper limb, keeping the lens perpendicular to the skin, and pressed the acquisition button. The terminal completed the acquisition of all data within one second: Hyperspectral image: dimensions (512, 512, 121), 400-1000nm band. Temperature field: dimensions (256, 256), temperature range 32.5-34.2℃. Near-infrared spectrum: dimensions (401,), 900-1700nm band. After acquisition, the terminal automatically sent the data to the edge computing node.

[0131] Further, data preprocessing and fusion are performed. The edge computing node receives the data at 8:05:02. First, time synchronization and spatial alignment are performed, upsampling the temperature field data to (512, 512). Then, data preprocessing is performed: Hyperspectral imagery: 3-level decomposition using the db4 wavelet basis, followed by threshold denoising. Temperature field: Adaptive histogram equalization is used to enhance contrast. Near-infrared spectrum: 11-point window, second-order polynomial Savitzky-Golay filtering is used. After preprocessing, the multimodal gated fusion module calculates the weights of each modality. , , Generate the fused feature vector (256 dimensions).

[0132] Specifically, regarding complication identification and grading: at 8:05:04, the multi-expert hybrid complication identification module analyzes the fused features, outputting the following results: Phlebitis: Grade I probability 0.92, Grade II probability 0.05, others 0.03. Exudation: Grade I probability 0.15, others 0.85. Thrombosis: probability 0.02. Infection: probability 0.01. Catheter occlusion: probability 0.03. Catheter dislodgement: probability 0.01. The system displays a yellow warning message, indicating "Grade I phlebitis".

[0133] Specifically, the nursing intervention recommendation module (8:05:05) infers personalized nursing interventions from the knowledge graph based on the identification results and patient characteristics: Operation Name: Moist Heat Compress. Operation Frequency: 3 times daily, 20 minutes each time. Operation Points: Use 50% magnesium sulfate solution at a temperature of 40-42℃ to avoid burns. Precautions: Observe changes in skin color and temperature; stop immediately if any abnormalities occur. Assessment Criteria: Assess the phlebitis grade daily; if no improvement is observed after 3 days, change the puncture site. Adjustment Suggestions: Due to the patient's history of penicillin allergy, avoid using topical medications containing penicillin. Due to the patient's advanced age, appropriately lower the heat compress temperature.

[0134] Furthermore, regarding nursing execution and recording, nurse Li performed a warm compress on the patient according to the system's recommended nursing interventions. After completing the procedure, she recorded the execution time (8:15), the person performing the procedure (Li), and the execution status (successfully completed, patient experienced no discomfort) in the system. The system automatically associated the nursing record with the patient's identification results and stored it locally and on the cloud platform.

[0135] Further follow-up was conducted, with nurse Li using the system to re-examine the patient daily at 8:00 AM. The re-examination results on October 4th showed that the probability of grade I phlebitis had decreased to 0.12, and the system returned to normal operation.

[0136] In this embodiment of the invention, the system successfully identified early-stage Grade I phlebitis and recommended personalized nursing interventions. The patient's phlebitis completely subsided after 3 days without developing into more serious complications. During a 3-month trial period, the system tested 500 patients undergoing intravenous therapy, identifying 126 cases of complications, achieving an accuracy rate of 93%. The nursing intervention recommendation compliance rate reached 95%, the complication rate decreased from the original 25%-35% to 11%, and patient satisfaction increased from the original 85% to 96%.

[0137] For example, during a three-month trial period, this system monitored 500 patients receiving intravenous therapy and identified 126 cases of complications. Compared to traditional methods relying on human experience, this system demonstrated significant advantages in several core indicators: the identification accuracy increased from approximately 70%-80% to 93%; the compliance rate of nursing intervention recommendations increased from approximately 75%-85% to 95%; the complication rate decreased from approximately 25%-35% to 11%; and patient satisfaction also increased from approximately 75%-85% to 96%. The data from the traditional methods were derived from the annual work report of the nursing director of the hospital's accreditation office, the year-end summary of intravenous therapy quality control, research on the impact of proactive nursing on intravenous infusion complications and patient satisfaction, and related practical application data.

[0138] Specifically, this invention offers significant technical advantages over traditional methods. First, in terms of accurate identification, this system utilizes advanced algorithms and big data analysis to more accurately identify complications from intravenous therapy, greatly improving the accuracy rate. Compared to traditional methods that rely on human experience, this reduces misdiagnosis and missed diagnosis. Second, the intelligent recommendation function provides personalized nursing recommendations based on the patient's specific condition and the type of complication, resulting in a higher compliance rate and making nursing care more scientific and effective. Third, the real-time monitoring function can promptly detect abnormal patient conditions, providing timely warnings to medical staff, facilitating early intervention and reducing the incidence of complications. Finally, this system also possesses continuous improvement capabilities. By continuously collecting and analyzing clinical data, optimizing algorithms and recommendation models, it further enhances the system's performance and effectiveness, providing patients with higher-quality nursing services.

[0139] This invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the adaptive identification and nursing recommendation method for intravenous catheter complications as described in the method embodiments.

[0140] The present invention provides a computer-readable storage medium that can realize the steps and effects of the adaptive identification and nursing recommendation method for intravenous catheter complications in the above-described method embodiments. To avoid repetition, the present invention will not repeat the details.

[0141] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0142] The following points need to be explained: (1) The accompanying drawings of the embodiments of the present invention only involve the structures involved in the embodiments of the present invention. Other structures can refer to the general design.

[0143] (2) For clarity, the thickness of layers or regions is enlarged or reduced in the drawings used to describe embodiments of the invention, i.e., these drawings are not drawn to scale. It is understood that when an element such as a layer, film, region or substrate is referred to as being “above” or “below” another element, the element may be “directly” located “above” or “below” the other element or there may be intermediate elements.

[0144] (3) Where there is no conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other to obtain new embodiments.

[0145] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An adaptive identification and nursing recommendation system for intravenous catheter complications, characterized in that, include: The multimodal acquisition terminal layer is used to simultaneously acquire hyperspectral images of the patient's puncture site, local skin temperature field, and near-infrared spectral data of catheter orifice exudate to obtain multimodal data. The edge computing node layer, deployed locally in the ward, is used to receive and process the multimodal data, perform multimodal data fusion, identify complication types and grades, recommend nursing interventions, and generate a structured nursing operation instruction set. The cloud platform layer is used to communicate with the edge computing node layer, globally store patient data, manage and dynamically update the standard nursing knowledge graph, train and optimize the complication identification model, and push the updated model and knowledge graph to the edge computing node layer.

2. The adaptive identification and nursing recommendation system for intravenous catheter complications according to claim 1, characterized in that, The multimodal acquisition terminal layer specifically includes: The hyperspectral imaging module is used to acquire hyperspectral images in the 400-1000nm wavelength band; Infrared thermal imaging module, used to acquire local skin temperature field; The near-infrared spectroscopy detection module is used to collect the near-infrared spectrum of exudates in the 900-1700nm wavelength range; The terminal control and communication module is used to trigger the synchronous data collection of each module through hardware and send the collected data to the edge computing node layer to obtain the multimodal data.

3. The adaptive identification and nursing recommendation system for intravenous catheter complications according to claim 1, characterized in that, The edge computing node layer specifically includes: The data receiving and synchronization module is used to receive the multimodal data and upsample the temperature field data to hyperspectral image resolution through a timestamp synchronization algorithm and spatial alignment operation. An adaptive data preprocessing module is used to perform wavelet denoising, adaptive histogram equalization, and Savitzky-Golay smoothing filtering on the hyperspectral image, temperature field, and near-infrared spectrum, respectively, and to perform adaptive normalization based on the patient's historical data in the multimodal data to obtain normalized data. A multimodal gating fusion module is used to dynamically calculate the modal weights of the normalized data and generate a unified fusion feature vector. A multi-expert hybrid complication identification module is used to identify six complications and their corresponding severity levels in parallel based on the fused feature vector: phlebitis, exudation, thrombosis, infection, catheter occlusion, and catheter dislodgement. The nursing intervention recommendation module is used to infer and generate the structured nursing operation instruction set based on the complications and their corresponding severity levels, combined with the standard nursing knowledge graph. The local storage and display module is used to store local data and provide a visual operation interface based on the structured nursing operation instruction set.

4. The adaptive identification and nursing recommendation system for intravenous catheter complications according to claim 3, characterized in that, The multimodal gating fusion module employs a gating fusion mechanism, which specifically includes: The fused feature vector is obtained by multiplying the hyperspectral image features by a first weight, adding the temperature field features by a second weight, and adding the near-infrared spectral features by a third weight. The first weight, the second weight, and the third weight are the weights of the three modes, respectively, and the sum of the three is 1, which is learned by the gating network.

5. The adaptive identification and nursing recommendation system for intravenous catheter complications according to claim 3, characterized in that, The multi-expert hybrid complication identification module is based on the multi-expert hybrid model (MoE) for symptom identification. The module specifically includes: Six expert networks, each corresponding to a complication, are structured as an input layer, a first hidden layer, a second hidden layer, and an output layer. The output layer contains five neurons, corresponding to no complication, grade I, grade II, grade III, and grade IV, respectively. The gated network is used to calculate the weights of the six expert networks; The multi-expert hybrid complication identification module is trained using a multi-task cross-entropy loss function, which is specifically calculated as follows: For each complication, the loss value corresponding to the complication is obtained by multiplying the clinical importance weight of the complication by the sum of the products of the logarithms of the true labels and predicted values ​​of the five grades under the complication; the loss values ​​of the six complications are summed and negative to obtain the final loss value. The final output of the module is a weighted sum of the outputs of each of the expert networks, enabling simultaneous identification and hierarchical judgment of multiple complications.

6. The adaptive identification and nursing recommendation system for intravenous catheter complications according to claim 3, characterized in that, The multi-expert hybrid complication identification module identifies symptoms based on a Transformer multimodal fusion network, which specifically includes: The multimodal feature extraction branch uses a 3D convolutional neural network to extract spatial-spectral joint features, a 2D convolutional neural network to extract temperature distribution features, and a 1D convolutional neural network to extract spectral features. The cross-modal attention fusion module employs a multi-head self-attention mechanism to achieve interaction and fusion between features from different modalities, and outputs a fused feature vector. The complication classification and grading module uses a fully connected network to perform multi-task classification and grading of six complications, and outputs a probability matrix with a dimension of 6×5.

7. The adaptive identification and nursing recommendation system for intravenous catheter complications according to claim 3, characterized in that, The nursing intervention recommendation module performs reasoning based on a dynamically updated knowledge graph of standard intravenous therapy nursing care. The knowledge graph specifically includes three types of entities: complication entities, nursing intervention entities, and patient characteristic entities, as well as five types of relationships: cause, applicable, contraindicated, adjustment, and substitution. The reasoning process combines rule-based reasoning with similarity-based reasoning to ultimately generate a structured nursing operation instruction set that includes operation name, operation frequency, operation points, precautions, and evaluation criteria.

8. The adaptive identification and nursing recommendation system for intravenous catheter complications according to claim 1, characterized in that, The cloud platform layer specifically includes: The data storage module adopts a hybrid architecture of object storage, relational database, and graph database; The knowledge graph management module supports visual editing by nursing experts and automatically retrieves the latest clinical guidelines and literature to update the knowledge graph; The model training and update module uses federated learning technology to train the model using data uploaded by each edge node and pushes model updates to the edge computing node layer regularly. The system management module is used for user permission management, device status monitoring, and system configuration.

9. The adaptive identification and nursing recommendation system for intravenous catheter complications according to claim 1, characterized in that, The system supports three operating modes: normal mode, early warning mode, and intervention mode. The normal mode is defined as follows: when the probability of all complications is less than 0.3, the system displays "no abnormalities" and recommends routine nursing procedures. The warning mode is such that when the probability of a certain complication being grade I is greater than 0.7, the system displays a yellow warning message and recommends preventive nursing measures. The intervention mode involves the system displaying a red warning message and automatically generating a detailed set of nursing operation instructions when the probability of a certain complication being grade II or above is greater than 0.7, or when multiple complications occur concurrently. The head nurse is then reminded to review the instructions.

10. A method for adaptive identification and nursing recommendations for intravenous catheter complications, characterized in that, The adaptive identification and care recommendation system for intravenous catheter complications applied to any one of claims 1 to 9 includes: S1: Obtain basic patient information; S2: Initialize the preprocessing parameters and inference rules for the patient's basic information; S3: Through the multimodal acquisition terminal layer, hyperspectral images of the puncture site, local skin temperature field, and near-infrared spectral data of catheter orifice exudate are simultaneously acquired to obtain multimodal data; S4: The multimodal data is time-synchronized, spatially aligned, and adaptively preprocessed through the edge computing node layer; S5: Through a gated fusion mechanism, the features of the three preprocessed modalities are fused to generate a unified fused feature vector; S6: Input the fused feature vector into a multi-expert hybrid model or a Transformer-based multimodal fusion network to identify the types and severity levels of the six complications; S7: Based on the type and severity classification of the aforementioned complications, combined with the dynamically updated standard nursing knowledge graph, complication identification results, and patient characteristics, generate a personalized set of structured nursing operation instructions; S8: Based on the structured nursing operation instruction set, the nurse performs the recommended nursing measures, records the execution status, and uploads the data to the cloud platform layer; S9: Through the cloud platform layer, the uploaded data is used periodically for model training and knowledge graph updates, and the updated model and knowledge graph are pushed to the edge computing node layer.

Images