An outpatient infusion exosmosis risk early warning method and system

Abstract

This application relates to a method and system for early warning of extravasation risk in outpatient infusions. The method includes: extracting corresponding visual frames, physiological signal data points, and medical event records from real-time acquired visual frame sequences, physiological signal sequences, and medical event streams; extracting and integrating features from these data to construct spatiotemporal sample units; inputting these spatiotemporal sample units into a gated spatiotemporal joint encoder to generate potential risk states characterizing the comprehensive risk status at the current moment; adding these potential risk states to a risk potential state sequence maintained for the patient to obtain an updated risk potential state sequence; performing risk analysis based on the updated risk potential state sequence to generate an evidence chain; and performing decision mapping based on the evidence chain to generate a decision report. This method, through multimodal spatiotemporal data fusion and a gated dynamic coding mechanism, achieves real-time and accurate perception and dynamic evolution modeling of extravasation risk, improving the intelligence level of outpatient infusion safety monitoring and the timeliness of risk intervention.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to a method and system for early warning of extravasation risk during outpatient infusion. Background Technology

[0002] Intravenous infusion is one of the most widely used methods of drug administration in outpatient clinical treatment. However, extravasation is a common adverse nursing event. The risks it causes, such as local tissue redness, swelling, blisters, necrosis, and even compartment syndrome, seriously threaten the patient's treatment safety and medical experience. In particular, the extravasation hazards of special solutions such as hypertonic solutions and vasoactive drugs are more significant.

[0003] Specifically, current clinical methods for early warning of intravenous extravasation still have many shortcomings. For example, traditional monitoring mainly relies on regular rounds by nursing staff and subjective feedback from patients. Limited by manpower, this can lead to gaps in rounds, and the identification of early atypical extravasation symptoms is highly dependent on individual experience, making real-time and accurate early warning difficult. Furthermore, while some studies use single sensors to monitor physiological signals or local pressure changes, they fail to integrate multi-source data such as visual puncture site status, medication information in the medical event flow, and infusion parameters, creating "data silos." This results in incomplete risk assessment dimensions and high false alarm and false negative rates. In addition, existing early warning methods are mostly based on fixed thresholds to judge risk, lacking the ability to analyze the dynamic evolution of risk during infusion, failing to capture key characteristics such as the speed of risk change and cumulative trends, and lacking clear risk attribution mechanisms and complete evidence chains after warning. Intervention measures mostly rely on general nursing guidelines, making it difficult to provide personalized decision support by combining individual patient risk characteristics and historical cases, leading to delayed emergency response and fragmented treatment procedures. Summary of the Invention

[0004] Therefore, it is necessary to provide a method and system for early warning of extravasation risk in outpatient infusion to address the above-mentioned technical problems, aiming to improve the real-time perception accuracy and early identification capability of extravasation risk.

[0005] Firstly, this application provides a method for early warning of extravasation risk during outpatient intravenous infusion, including:

[0006] Real-time acquisition of visual frame sequences, physiological signal sequences, and medical event streams; based on a preset sliding time window, extraction of corresponding visual frames, physiological signal data points, and medical event records from the visual frame sequences, physiological signal sequences, and medical event streams; feature extraction of visual frames, physiological signal data points, and medical event records respectively; integration of corresponding standardized image block sequences, comprehensive signal feature vectors, and event context vectors to construct spatiotemporal sample units;

[0007] The spatiotemporal sample unit is input into the gated spatiotemporal joint encoder for processing to generate a risk potential state that represents the comprehensive risk status at the current moment. The risk potential state is added to the risk potential state sequence maintained for the patient to obtain an updated risk potential state sequence.

[0008] Risk analysis is performed based on the updated sequence of potential risk states to generate a chain of evidence. Decision mapping is then performed based on the chain of evidence to generate a decision report.

[0009] In one embodiment, features are extracted from visual frames, physiological signal data points, and medical event records, respectively. The corresponding standardized image patch sequences, integrated signal feature vectors, and event context vectors are then integrated to construct a spatiotemporal sample unit, including:

[0010] Automatic puncture point localization is performed on the visual frame to obtain the puncture point localization result; based on the puncture point localization result, the visual frame is cropped to obtain cropped image blocks; illumination and color correction is performed on the cropped image blocks to obtain a standardized image block sequence.

[0011] The physiological signal data points are bandpass filtered to obtain the denoised physiological signal; the time domain features, frequency domain features and nonlinear features are calculated based on the denoised physiological signal to obtain multi-dimensional signal features; the multi-dimensional signal features are combined to obtain a comprehensive signal feature vector.

[0012] The medical event records are vectorized and embedded to obtain event embedding vectors; the event embedding vectors are weighted and aggregated based on a time attention mechanism to obtain event context vectors.

[0013] Standardized image patch sequences, integrated signal feature vectors, and event context vectors are packaged in an ordered manner to construct spatiotemporal sample units.

[0014] In one embodiment, the gated spatiotemporal co-encoder includes a visual branch, a non-visual branch, a gated fusion module, and an output module. Spatiotemporal sample units are input into the gated spatiotemporal co-encoder for processing to generate a potential risk state characterizing the overall risk state at the current moment, including:

[0015] Standardized image patch sequences are extracted from spatiotemporal sample units, and the standardized image patch sequences are input into the visual branch of the gated spatiotemporal co-encoder. Spatiotemporal features are extracted through a three-dimensional convolutional neural network to obtain visual spatiotemporal feature vectors.

[0016] The integrated signal feature vector and event context vector are extracted from the spatiotemporal sample unit. The integrated signal feature vector and event context vector are concatenated to obtain the non-visual joint feature vector. The non-visual joint feature vector is input into the non-visual branch of the gated spatiotemporal joint encoder and subjected to nonlinear transformation by a multilayer perceptron to obtain the non-visual feature vector.

[0017] Based on the gated fusion module, a gated vector is obtained by calculating the visual spatiotemporal feature vector and the non-visual feature vector using a preset gated weight generation function. The gated vector is then used to perform element-wise weighted fusion of the visual spatiotemporal feature vector and the non-visual feature vector to obtain the fused feature vector.

[0018] Based on the output module, the fused feature vector is mapped to generate potential risk states.

[0019] In one embodiment, risk analysis is performed based on the updated sequence of potential risk states to generate a chain of evidence, including:

[0020] Extract the continuous risk latent states arranged in chronological order from the updated risk latent state sequence, input the continuous risk latent states into a causal dilated convolutional network to obtain a discrete risk value sequence; perform spline interpolation on the discrete risk value sequence to construct a continuous risk trajectory function;

[0021] Based on the risk trajectory function, the risk value, risk change rate, and risk change acceleration at the current moment are calculated using the numerical differentiation method; preset absolute value thresholds, velocity thresholds, and acceleration thresholds are obtained, and the risk value, risk change rate, and risk change acceleration are compared with the corresponding absolute value thresholds, velocity thresholds, and acceleration thresholds to obtain the comparison results;

[0022] Based on the comparison result, a compound logic judgment is performed, and when the comparison result meets the preset triggering condition, an early warning mechanism is triggered.

[0023] When the early warning mechanism is triggered, the integral gradient method is used to calculate the contribution of the risk value to each basic feature in the spatiotemporal sample unit, and the preset quantitative feature with the highest contribution is selected as the key risk contribution factor; the sensitivity of the risk increment in the most recent preset period to the risk potential state at each historical moment in the updated risk potential state sequence is calculated by the time backpropagation method, and the moment with the highest sensitivity is determined as the starting point of the key risk accumulation period.

[0024] Integrate key risk contributing factors, key risk accumulation time periods, risk values, risk change rates, and risk change acceleration information to generate a chain of evidence.

[0025] In one embodiment, a decision report is generated based on a chain of evidence for decision mapping, including:

[0026] The evidence chain is analyzed to extract information such as risk value, risk change rate, risk change acceleration, key risk contribution factors, and key risk accumulation time period. Based on the preset risk level classification standard, the current risk level is determined by combining the risk value, risk change rate, and risk change acceleration.

[0027] Based on key risk contribution factors and key risk accumulation time periods, risk change trend information is determined by combining risk change rate and risk change acceleration; risk level, key risk contribution factors and risk change trend information are integrated to generate query conditions.

[0028] The similarity calculation is performed between the query conditions and the condition patterns in the predefined structured intervention decision knowledge base, and intervention rules with a matching degree exceeding the preset threshold are retrieved.

[0029] The intervention measures corresponding to the merged intervention rules are calculated, and an initial basic score is calculated for each intervention measure based on the matching degree and inherent priority of each intervention rule. The initial basic scores are then used to sort and generate a preliminary set of intervention measures.

[0030] The trajectory segment with a preset duration centered on the warning time is extracted from the risk trajectory function. The similarity between the trajectory segment and the case trajectory segment in the historical case database is calculated based on the dynamic time warping algorithm. The case with the highest preset number of similarities is selected as similar cases. The frequency of occurrence and average effect score of the intervention measures actually taken and marked as effective in the similar cases are statistically analyzed to obtain the case measure statistical data set.

[0031] For each intervention in the initial intervention set, the corresponding frequency of occurrence and average effect score are extracted from the case intervention statistics set, and the correlation coefficient is calculated. The initial base score is weighted using the correlation coefficient to obtain the calibrated priority score for each intervention in the initial intervention set.

[0032] Interventions with both frequency of occurrence and average effect score higher than the corresponding preset threshold are selected from the statistical data set of case measures. Interventions already included in the preliminary intervention measure set are removed to obtain supplementary measures. Based on the frequency of occurrence and average effect score of the supplementary measures in the statistical data set of case measures, an initial score for the supplementary measures is calculated by weighted summation. The supplementary measures are assigned the corresponding initial score and added to the preliminary intervention measure set to obtain the candidate intervention measure set.

[0033] Extract each candidate intervention from the candidate intervention set, sort them in descending order according to the scores of each candidate intervention, check the logical dependencies between each candidate intervention, and generate a sorted list of personalized interventions.

[0034] Each intervention measure in the personalized intervention measure list is analyzed to generate intervention instructions; information on risk level, key risk contribution factors, risk change trend, risk value, risk change rate, risk change acceleration, and key risk accumulation period is integrated to obtain early warning basis;

[0035] The early warning criteria are linked to each intervention instruction to generate a decision report.

[0036] In one embodiment, the mathematical expression of the preset gating weight generation function is:

[0037]

[0038]

[0039] in, This is the gate vector; Use the Sigmoid activation function; and All are learnable weight matrices; Norm(·) is the L2 normalization operation; It is a visual spatiotemporal feature vector; Non-visual feature vectors; Conf represents the joint confidence level of the visual spatiotemporal feature vector and the non-visual feature vector, with values ​​[0,1]. This represents the batch mean of the visual spatiotemporal feature vector; The batch mean of the non-visual feature vectors; It is an L2 norm; This is the bias vector.

[0040] In one embodiment, the formula for calculating the average performance score is:

[0041]

[0042] in, For the first The average effectiveness score of the interventions that were actually taken and marked as effective; For the use of the historical case library The total number of effective cases for which an actual intervention was taken and marked as effective; For the first The first one adopts the first The original effect scores recorded in similar cases of actual interventions that were marked as effective, with a range of values. ; This is the time decay coefficient; For the first The time interval between a similar case and the current warning time; For the first In the 1st similar case The duration of the effective effect of an actual intervention measure that is marked as effective; For the first In the 1st similar case The total duration of implementation of each intervention measure that was actually taken and marked as effective; For the first Sample weights for similar cases.

[0043] Secondly, an outpatient intravenous infusion extravasation risk early warning system includes:

[0044] The spatiotemporal sampling and feature fusion module is used to acquire visual frame sequences, physiological signal sequences, and medical event streams in real time. Based on a preset sliding time window, it extracts the corresponding visual frames, physiological signal data points, and medical event records from the visual frame sequences, physiological signal sequences, and medical event streams. It then performs feature extraction on the visual frames, physiological signal data points, and medical event records respectively, and integrates the corresponding standardized image patch sequences, comprehensive signal feature vectors, and event context vectors to construct spatiotemporal sample units.

[0045] The gated dynamic risk coding module is used to input spatiotemporal sample units into the gated spatiotemporal joint encoder for processing, generate risk potential states to characterize the comprehensive risk status at the current moment, add the risk potential states to the risk potential state sequence maintained for the patient, and obtain the updated risk potential state sequence.

[0046] The evidence-driven risk decision-making module is used to perform risk analysis based on the updated risk potential state sequence, generate an evidence chain, perform decision mapping based on the evidence chain, and generate a decision report.

[0047] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the first aspect.

[0048] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the first aspect.

[0049] The aforementioned outpatient infusion extravasation risk early warning method and system firstly collects multi-source data in real time and constructs spatiotemporal sample units based on a preset sliding time window, solving the problem of difficult unified representation of multi-source heterogeneous medical data and achieving effective integration of visual, physiological, and medical event data. Secondly, the spatiotemporal sample units are input into a gated spatiotemporal joint encoder to generate risk latent states and update sequences, achieving adaptive fusion of multimodal risk features and improving the accuracy of comprehensive risk state representation. Furthermore, risk analysis is performed based on the updated risk latent state sequences to generate evidence chains, realizing dynamic risk assessment and attribution, and reducing the probability of false alarms and missed alarms. Finally, decision mapping is completed based on the evidence chains, and a decision report is generated, achieving a closed-loop connection from early warning to intervention and improving the targeting and efficiency of clinical intervention. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A flowchart of an outpatient infusion extravasation risk early warning method is provided as an exemplary embodiment of the present invention;

[0052] Figure 2 A flowchart of a method for generating potential risk states is provided as an exemplary embodiment of the present invention;

[0053] Figure 3 This is a schematic diagram of an outpatient infusion extravasation risk early warning system provided as an exemplary embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0055] In one embodiment, such as Figure 1 As shown, an outpatient infusion extravasation risk warning method is provided. This embodiment illustrates the method's application to a terminal. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0056] S101: Real-time acquisition of visual frame sequences, physiological signal sequences, and medical event streams; Based on a preset sliding time window, extracting corresponding visual frames, physiological signal data points, and medical event records from the visual frame sequences, physiological signal sequences, and medical event streams; Performing feature extraction on the visual frames, physiological signal data points, and medical event records respectively, and integrating the corresponding standardized image block sequences, comprehensive signal feature vectors, and event context vectors to construct spatiotemporal sample units.

[0057] Specifically, visual information can reflect local tissue changes at the puncture site, such as obvious signs like redness, swelling, and effusion. Physiological signals can reflect the patient's stress response to infusion, such as changes in local blood flow and fluctuations in skin temperature. The medical event stream can include key background information such as medication type, infusion rate, and the patient's basic medical history. Real-time acquisition of this information allows for a comprehensive analysis of the risk of extravasation, facilitating timely intervention. Illustratively, visual frame sequences can be captured by cameras deployed near the infusion puncture site. Physiological signal sequences can be acquired through multi-parameter sensors attached to the skin surrounding the puncture site, including key physiological indicators such as skin temperature, local blood flow, and tissue impedance. The medical event stream can be synchronized in real-time through the interface of the hospital's existing information system, including structured data such as basic patient information, medication orders (drug type, concentration, dosage), infusion start time, and puncture operation records.

[0058] Furthermore, to avoid the interference of the randomness of a single data sampling on risk assessment, and to balance the real-time nature and temporal correlation of data processing, a preset sliding time window can be used to extract data from multiple sources. The length and step size of the sliding time window can be preset according to the evolutionary pattern of clinical extravasation risk; for example, the window sliding step size can be set to 1-2 seconds. Based on this sliding time window, all visual frames falling within the window's time range, continuous physiological signal data points, and complete medical event records can be extracted from the cached visual frame sequences, physiological signal sequences, and medical event streams, achieving alignment of multi-source data in the temporal dimension.

[0059] Subsequently, feature extraction can be performed on the three types of extracted data, transforming the original unstructured / semi-structured data into computer-processable quantitative features. For example, for visual frames, image preprocessing techniques can be used to remove background interference and enhance the features of the puncture site region, transforming them into a standardized image patch sequence to ensure that visual data from different patients and puncture sites have a unified representation dimension. For physiological signal data points, quantitative indicators reflecting the essential characteristics of the signal can be extracted after signal preprocessing (such as denoising and baseline correction), forming a comprehensive signal feature vector, achieving dimensionality reduction of physiological signals while preserving key information. For medical event records, structured encoding can be used to transform textual and numerical event information into vector-based event context vectors, ensuring that medical background information can be collaboratively computed with visual and physiological features. Finally, the three types of features are systematically integrated to construct a spatiotemporal sample unit. This unit contains risk features of various dimensions while preserving the temporal correlation of the data, providing standardized input for subsequent multimodal feature fusion and risk state modeling.

[0060] S102: Input the spatiotemporal sample unit into the gated spatiotemporal joint encoder for processing to generate a risk potential state that represents the comprehensive risk status at the current moment. Add the risk potential state to the risk potential state sequence maintained for the patient to obtain an updated risk potential state sequence.

[0061] Specifically, the visual image patch sequences, comprehensive signal feature vectors, and event context vectors in the spatiotemporal sample units belong to different modalities, with significant differences in feature distribution and representation dimensions. Ordinary feature concatenation or simple weighted fusion cannot fully explore the complementary relationships between features of different modalities, and may even lead to feature distortion due to modal conflicts. The core advantage of the gated spatiotemporal co-encoder lies in its modal adaptive fusion capability, which can dynamically adjust the fusion weights according to the contribution of different modal features to risk judgment. Therefore, the constructed spatiotemporal sample units can be input into the gated spatiotemporal co-encoder, and different modal features can be processed through different feature extraction branches. For example, for visual image patch sequences, the spatiotemporal correlation features can be extracted, while for comprehensive signal feature vectors and event context vectors, the numerical and semantic correlation features can be extracted. Through the gating mechanism inside the encoder, the features extracted from each branch can be dynamically weighted and fused to generate a fused feature vector that can comprehensively reflect all current risk factors. For example, when obvious redness and swelling signals appear in the visual features, the fusion weight of the visual modality can be automatically increased. When the medical event stream shows that the patient is using hypertonic drugs, the weight of the medical event modality can be automatically increased, thereby improving the targeting and accuracy of the fused features.

[0062] Furthermore, based on the fused feature vector, a risk potential state can be generated to characterize the overall risk status at the current moment. This risk potential state is a high-dimensional quantification vector, with each dimension corresponding to a core risk-related indicator. It can comprehensively and objectively quantify the current risk level of extravasation, avoiding the one-sidedness of traditional single-indicator judgments. In addition, considering that extravasation risk is a dynamic and cumulative process, a single-moment risk potential state cannot reflect the evolution trend of the risk, such as whether the risk continues to rise or the rate of rise. Therefore, the risk potential state generated at the current moment can be added to a risk potential state sequence specifically maintained for that patient, forming an updated risk potential state sequence. This sequence, with time as its axis, completely records all risk status changes from the start of infusion to the current moment, providing time-series data support for subsequent calculations of the rate and acceleration of risk changes and analysis of the risk evolution trajectory. Simultaneously, by maintaining a separate sequence for each patient, the individualized nature of the risk analysis is ensured, avoiding feature interference between different patients.

[0063] S103: Perform risk analysis based on the updated risk potential state sequence, generate a chain of evidence, perform decision mapping based on the chain of evidence, and generate a decision report.

[0064] Specifically, by analyzing the risk potential sequence, we can not only determine the current risk level but also capture the trend of risk changes, generating key dynamic indicators such as the rate of increase and acceleration of risk values. These indicators can effectively distinguish between sudden high risks and slowly accumulating high risks, providing a more accurate basis for clinical intervention decisions. When the risk assessment meets the warning conditions, an evidence chain can be generated. This evidence chain can be a structured data set integrating core information such as risk level, risk evolution trend, and key influencing factors, thereby clarifying the cause of the risk and providing clear risk attribution basis for clinical medical staff, avoiding blind intervention. Subsequently, decision mapping can be performed based on the evidence chain to establish a scientific correspondence between risk characteristics and intervention measures. The information such as risk level and evolution trend contained in the evidence chain can accurately match the priority and direction of clinical intervention. Through the decision mapping process, abstract risk data can be transformed into specific and actionable clinical intervention strategies. Ultimately, a decision report can be generated, which is a systematic presentation of risk analysis results and intervention strategies. It can include both early warning evidence (i.e., core information in the chain of evidence) and specific intervention instructions, such as adjusting the infusion rate, changing the puncture site, and local nursing measures, thus realizing a closed loop from risk warning to scientific intervention.

[0065] The aforementioned method first achieves unified representation and standardized integration of visual, physiological signals, and medical events through multimodal spatiotemporal sampling and feature fusion. Second, it overcomes the limitation of traditional fixed-threshold judgments in capturing the temporal characteristics of risk evolution by dynamically encoding and maintaining the sequence of risk latent states using a gated spatiotemporal co-encoder, thus improving the dynamism and accuracy of risk modeling. Finally, through evidence chain generation and decision mapping, it addresses the lack of attribution mechanisms and personalized support for early warning results, enhancing the interpretability and relevance of intervention decisions.

[0066] In one embodiment, features are extracted from visual frames, physiological signal data points, and medical event records respectively. The corresponding standardized image patch sequences, integrated signal feature vectors, and event context vectors are then integrated to construct a spatiotemporal sample unit, including:

[0067] Automatic puncture point localization is performed on the visual frame to obtain the puncture point localization result; based on the puncture point localization result, the visual frame is cropped to obtain cropped image blocks; illumination and color correction is performed on the cropped image blocks to obtain a standardized image block sequence.

[0068] The physiological signal data points are bandpass filtered to obtain the denoised physiological signal; the time domain features, frequency domain features and nonlinear features are calculated based on the denoised physiological signal to obtain multi-dimensional signal features; the multi-dimensional signal features are combined to obtain a comprehensive signal feature vector.

[0069] The medical event records are vectorized and embedded to obtain event embedding vectors; the event embedding vectors are weighted and aggregated based on a time attention mechanism to obtain event context vectors.

[0070] Standardized image patch sequences, integrated signal feature vectors, and event context vectors are packaged in an ordered manner to construct spatiotemporal sample units.

[0071] Specifically, image feature enhancement techniques can be used to strengthen the texture and grayscale differences between the puncture point area and the surrounding skin. Then, a multi-scale feature fusion network can be employed to extract key puncture-related features from the visual frame, such as the edge contour of the puncture needle and local tissue texture changes at the puncture site. For example, the anatomical patterns of clinical puncture procedures can be combined to define a candidate region for the puncture point. The final puncture point localization result can be determined by calculating the feature matching degree of the candidate regions. This result can represent the center position of the puncture point and the boundary range of the surrounding key areas in the form of a coordinate set. Based on this puncture point localization result, the visual frame can be cropped. That is, based on the boundary coordinates in the localization result, the puncture point and the surrounding tissue area within a predetermined range are retained, while irrelevant background pixels are removed. This reduces the impact of background interference on subsequent feature extraction, focusing on the core area where extravasation risk is most likely to occur. Subsequently, illumination and color correction is performed on the cropped image blocks to eliminate image feature distortion caused by differences in light intensity and color temperature of light sources under different infusion environments. For example, the brightness distribution and color channel response of the image can be adjusted by an adaptive illumination equalization algorithm to ensure that the image blocks obtained under different conditions maintain consistency in brightness, contrast, and color space, and finally form a standardized image block sequence to ensure that the sequence has uniform size specifications and feature distribution characteristics.

[0072] Specifically, bandpass filtering of physiological signal data points can further filter out noise components such as power frequency interference and baseline drift, retaining effective signal frequency bands related to extravasation risk. For example, an adaptive filter can be used. By analyzing the frequency distribution characteristics of the physiological signal, the passband range of the filter is determined, allowing only signal components within this range to pass through, thus obtaining a denoised physiological signal that more realistically reflects the physiological changes at the puncture site. Based on the denoised physiological signal, time-domain features, frequency-domain features, and nonlinear features are calculated. Time-domain features focus on the temporal distribution and dynamic changes of the signal, including the amplitude range, peak density, mean, and variance, directly reflecting the instantaneous changes and stable states of physiological indicators. Frequency-domain features can be obtained by converting the time-domain signal to the frequency domain using Fourier transform, extracting the dominant frequency component and frequency distribution entropy, characterizing the frequency distribution pattern of the physiological signal. Nonlinear features, based on chaos theory and fractal geometry, can extract approximate entropy and sample entropy of signals to reflect the nonlinear dynamic characteristics of physiological systems. These three types of features complement each other from different dimensions, representing the essential attributes of physiological signals and together constituting multidimensional signal features. Subsequently, the above features can be normalized to eliminate dimensional differences between different feature dimensions. Then, the normalized features are concatenated in a predetermined order to form a comprehensive signal feature vector. This vector can comprehensively integrate the multidimensional information of physiological signals, providing quantitative support for risk assessment.

[0073] Furthermore, vectorized embedding encoding of medical event records can transform different types of medical event information into low-dimensional dense vectors that can be processed by computers. Medical event records can include textual information such as drug names and operation types, and numerical information such as infusion rate and drug dosage. For textual information, a semantic encoding model can be used to transform it into a vector representation with semantic relevance. For numerical information, it can be mapped to a vector space of the same dimension after standardization. Ultimately, all types of medical event records are transformed into event embedding vectors with uniform dimension, which retain the core semantic and numerical features of the medical event. Weighted aggregation of event embedding vectors based on a time attention mechanism can dynamically allocate aggregation weights for different event embedding vectors according to the time interval between the occurrence time of the medical event and the current processing time, and the importance of the event type to risk assessment. For example, the corresponding attention weights can be calculated using the following formula:

[0074]

[0075] in, Indicates the first Attention weights for each event embedding vector. Indicates the first The similarity score of each event embedding vector is calculated by combining the matching degree between the event embedding vector and the feature query vector at the current time, and the time decay coefficient between the event occurrence time and the current time. This represents the total number of event embedding vectors participating in the aggregation. After calculating the attention weight of each event embedding vector using this formula, the event embedding vectors are weighted and summed with their corresponding attention weights to obtain the event context vector. This vector not only integrates information from all medical events but also reflects the differences in the contribution of different events to the current risk state through attention weights, thus strengthening the influence weight of key medical events.

[0076] Finally, the standardized image patch sequence, integrated signal feature vector, and event context vector can be ordered and packaged according to a preset modal order, clarifying the storage location and index relationship of various features in the spatiotemporal sample unit. This ensures that the features of the corresponding modality can be extracted quickly and accurately in subsequent processing. The integrated spatiotemporal sample unit simultaneously contains feature information and temporal correlation characteristics of three modalities: visual, physiological, and medical events, forming a comprehensive quantitative representation of the current state of the infusion process. This provides standardized input data for subsequent multimodal feature fusion and risk state modeling.

[0077] In one embodiment, the gated spatiotemporal joint encoder includes a visual branch, a non-visual branch, a gated fusion module, and an output module. For example... Figure 2As shown, the spatiotemporal sample units are input into a gated spatiotemporal joint encoder for processing to generate a potential risk state characterizing the overall risk state at the current moment, including:

[0078] S201: Extract standardized image patch sequences from spatiotemporal sample units, input the standardized image patch sequences into the visual branch of the gated spatiotemporal co-encoder, and extract spatiotemporal features through a three-dimensional convolutional neural network to obtain visual spatiotemporal feature vectors;

[0079] S202: Extract the integrated signal feature vector and event context vector from the spatiotemporal sample unit, concatenate the integrated signal feature vector and event context vector to obtain the non-visual joint feature vector; input the non-visual joint feature vector into the non-visual branch of the gated spatiotemporal joint encoder, and perform nonlinear transformation through a multilayer perceptron to obtain the non-visual feature vector.

[0080] S203: Based on the gated fusion module, the gated vector is calculated using a preset gated weight generation function according to the visual spatiotemporal feature vector and the non-visual feature vector; the gated vector is then used to perform element-wise weighted fusion processing on the visual spatiotemporal feature vector and the non-visual feature vector to obtain the fused feature vector.

[0081] S204: Based on the output module, the fused feature vector is mapped to generate the potential risk state.

[0082] Specifically, after extracting standardized image patch sequences from spatiotemporal sample units, these sequences can be input into the visual branch of a gated spatiotemporal co-encoder. This visual branch can employ a 3D convolutional neural network as its core feature extraction component. The 3D convolutional kernels in this network can simultaneously perform convolution operations on both the spatial dimension of the image and the temporal dimension of the frame sequence, thereby synchronously capturing the spatial morphological features and temporal evolution features of the puncture site region. For example, the 3D convolutional neural network can be composed of multiple layers of 3D convolutional layers, batch normalization layers, and nonlinear activation layers stacked together. After the standardized image patch sequences are input into the network in a temporal order, different levels of spatiotemporal features are extracted through the sliding operation of the 3D convolutional kernels. Shallow convolutional layers capture basic spatial features such as the edges and textures of the image patches, as well as minute displacement features between frames. Deep convolutional layers fuse shallow features to extract high-level spatiotemporal correlation features such as the spread of redness and swelling at the puncture site, and tissue swelling and deformation. Finally, a global average pooling layer transforms the high-dimensional spatiotemporal feature map into a fixed-dimensional vector, resulting in a visual spatiotemporal feature vector. This vector fully encodes the core spatiotemporal features related to the risk of extravasation in the visual modality.

[0083] Specifically, after extracting the comprehensive signal feature vector and event context vector from the spatiotemporal sample units, the elements of the comprehensive signal feature vector and the event context vector can be concatenated according to their feature dimensions to form a non-visual joint feature vector with a dimension equal to the sum of the dimensions of the two types of vectors. This provides a standardized input for the unified processing of the non-visual branch. The non-visual joint feature vector can then be input into the non-visual branch of the gated spatiotemporal co-encoder. This non-visual branch can be constructed using a multilayer perceptron, which includes an input layer, at least one hidden layer, and an output layer. The hidden layers use non-linear activation functions to achieve non-linear transformation of features. Through linear transformation and non-linear activation of multilayer weight matrices, it can uncover the potential correlation between physiological signal features and medical event features, such as the correlation between hyperosmolar drug input and tissue impedance changes. After the non-visual joint feature vector enters the multilayer perceptron through the input layer, it sequentially passes through each hidden layer to complete the non-linear mapping and dimensionality compression of the features. Finally, the output layer can output a non-visual feature vector of fixed dimensions. This vector encodes the core association features related to the risk of extravasation in non-visual modalities, and its dimension is similar to that of the visual spatiotemporal feature vector. Maintaining consistency lays the foundation for subsequent fusion processing.

[0084] Indicatively, the gated fusion module can calculate a gate vector based on visual spatiotemporal feature vectors and non-visual feature vectors using a preset gate weight generation function. Then, dynamic weighted fusion of the two types of features is achieved using this gate vector. The mathematical expression of the preset gate weight generation function can be:

[0085]

[0086] in, The gating vector has the same dimension as the visual spatiotemporal feature vector. Non-visual feature vectors The dimensions are completely consistent, and the value range of each element in the vector is [0,1]. The value of a single element represents the weight distribution ratio of the fusion of visual features and non-visual features in the corresponding dimension. The Sigmoid activation function maps the linear calculation result to the [0,1] interval to accommodate the weight values. It is a learnable weight matrix for visual features, with row dimensions consistent with the gating vector dimension and column dimensions consistent with the visual spatiotemporal feature vector dimension, used to perform linear transformation on the normalized visual spatiotemporal feature vector. The learnable weight matrix for non-visual features has the same row dimension as the gating vector dimension and the same column dimension as the non-visual feature vector dimension. It is used to perform linear transformation on the normalized non-visual feature vector. The learnable weight matrix for the joint confidence score has the same row dimension as the gate vector dimension and a column dimension of 1, used for linear transformation of the joint confidence score; Norm(·) is the L2 normalization operation, which eliminates the dimensional difference between the visual spatiotemporal feature vector and the non-visual feature vector, ensuring that the linear transformation results of the two types of features are on the same order of magnitude; Conf The joint confidence score of the visual spatiotemporal feature vector and the non-visual feature vector is used to quantify the effectiveness of the two types of features, and its mathematical expression is:

[0087]

[0088] In the above expression, The batch mean of the visual spatiotemporal feature vector represents the average level of the visual spatiotemporal feature vectors during the training or inference process of the same batch. It is used to measure the degree of deviation of the current visual spatiotemporal feature vector from the batch average level. The batch mean of non-visual feature vectors represents the average level of non-visual feature vectors during the training or inference process of the same batch, and is used to measure the degree of deviation of the current non-visual feature vector from the batch average level. The L2 norm is used to calculate the Euclidean distance between an eigenvector and its batch mean, which reflects the specificity and effectiveness of the eigenvector. It represents the maximum L2 norm of all visual spatiotemporal feature vectors in the same batch and their batch mean. The maximum L2 norm of all non-visual feature vectors in the same batch and their batch mean is used as a normalization factor to map the feature deviation to the [0,1] interval. The joint confidence score (Conf) can be obtained by averaging the deviations of visual and non-visual features. Its value ranges from [0,1]. The larger the value, the higher the effectiveness of the two types of features and the greater their contribution to risk assessment. This is the bias vector of the gated fusion module, whose dimension is the same as that of the gated vector, and is used to adjust the calculation basis of the gated weight generation function.

[0089] Specifically, after calculating the gating vector Then, the gating vector is used to analyze the visual spatiotemporal feature vector. Non-visual feature vectors Element-wise weighted fusion is performed, following the rule of element-wise weighted summation, to fuse the feature vectors. The calculation logic is as follows Where © represents element-wise multiplication, and 1 represents a vector of all 1s consistent with the dimension of the gate vector. This fusion method allows for the dynamic allocation of weights between visual and non-visual features based on the element values ​​of the gate vector in each dimension of the features. For example, when the effectiveness of visual features decreases due to camera occlusion, the joint confidence decreases, the element values ​​of the corresponding dimension of the gate vector decrease, and the weight of non-visual features increases accordingly. Conversely, the weight of visual features increases, thereby achieving adaptive fusion of multimodal features to obtain a fused feature vector that accurately reflects the current risk state. .

[0090] Specifically, the output module can use a fully connected layer as the core mapping component. The input dimension of the fully connected layer is consistent with the dimension of the fused feature vector, and the output dimension is a preset risk potential state dimension. This dimension can be determined based on the number of core characterizing indicators of infusion extravasation risk. After the fused feature vector is input into this fully connected layer, it can undergo a linear transformation through a learnable weight matrix, and then be processed by a nonlinear activation function to map the linear transformation result to a preset numerical range, ultimately generating a risk potential state. This risk potential state is a high-dimensional quantification vector with a fixed dimension. Each dimension corresponds to a core characterizing indicator related to infusion extravasation risk, including the severity of the risk, the type of risk cause, and the risk development trend. It can comprehensively and objectively quantify the overall risk status at the current moment, providing standardized feature input for subsequent risk time series analysis.

[0091] In one embodiment, risk analysis is performed based on the updated sequence of potential risk states to generate a chain of evidence, including:

[0092] Extract the continuous risk latent states arranged in chronological order from the updated risk latent state sequence, input the continuous risk latent states into a causal dilated convolutional network to obtain a discrete risk value sequence; perform spline interpolation on the discrete risk value sequence to construct a continuous risk trajectory function;

[0093] Based on the risk trajectory function, the risk value, risk change rate, and risk change acceleration at the current moment are calculated using the numerical differentiation method; preset absolute value thresholds, velocity thresholds, and acceleration thresholds are obtained, and the risk value, risk change rate, and risk change acceleration are compared with the corresponding absolute value thresholds, velocity thresholds, and acceleration thresholds to obtain the comparison results;

[0094] Based on the comparison result, a compound logic judgment is performed, and when the comparison result meets the preset triggering condition, an early warning mechanism is triggered.

[0095] When the early warning mechanism is triggered, the integral gradient method is used to calculate the contribution of the risk value to each basic feature in the spatiotemporal sample unit, and the preset quantitative feature with the highest contribution is selected as the key risk contribution factor; the sensitivity of the risk increment in the most recent preset period to the risk potential state at each historical moment in the updated risk potential state sequence is calculated by the time backpropagation method, and the moment with the highest sensitivity is determined as the starting point of the key risk accumulation period.

[0096] Integrate key risk contributing factors, key risk accumulation time periods, risk values, risk change rates, and risk change acceleration information to generate a chain of evidence.

[0097] Specifically, we can first extract the continuous risk latent states arranged in chronological order from the updated risk latent state sequence. This continuous risk latent state sequence covers all risk latent states within a preset time period tracing back from the current moment, forming an ordered temporal feature matrix. Then, this temporal feature matrix is ​​input into a causal dilated convolutional network. This causal dilated convolutional network is constructed based on causal constraints and a dilated convolutional structure. The causal constraints ensure that the network, when calculating the output at the current moment, relies only on the risk latent states at historical moments, meeting the real-time requirements of outpatient infusion risk warning. The dilated convolution, through an exponentially increased kernel dilation rate, significantly expands the network's receptive field without increasing computational load, effectively capturing long-term risk evolution correlation features in the risk latent state sequence, such as the delayed correlation between early drug-induced physiological signal changes and the later manifestation of extravasation risk. Schematic, after continuous risk latent states are processed through multiple layers of convolution, activation, and normalization in a causal dilated convolutional network, the output layer can map them into a one-dimensional discrete numerical sequence, i.e., a discrete risk value sequence. Each element in this sequence corresponds to a quantized risk value at a given time, thus realizing the mapping from high-dimensional features to single-dimensional risk quantification. Furthermore, a continuous risk trajectory function can be constructed using a cubic spline interpolation algorithm to perform spline interpolation on the discrete risk value sequence. This algorithm can construct a piecewise cubic polynomial function, making the interpolated function continuous and second-order differentiable throughout its domain. This transforms the discrete risk quantification result into a continuously differentiable mathematical function, providing a mathematical foundation for subsequent calculation of dynamic risk change indicators through differentiation. The resulting risk trajectory function can accurately characterize the continuous evolution of risk values ​​over time during infusion.

[0098] Specifically, based on the constructed continuous risk trajectory function, the risk value, risk change rate, and risk change acceleration at the current moment are calculated using numerical differentiation methods. The risk value at the current moment is directly determined by the function value of the risk trajectory function at the current time point. The risk change rate is the first derivative of the risk trajectory function, and the risk change acceleration is the second derivative of the risk trajectory function. Illustratively, the forward difference method can be used for numerical differentiation, with the core calculation formula being the first-order difference expression for the risk change rate:

[0099]

[0100] in, For the current moment The speed of risk change For the risk trajectory function at the current moment The function value is the current risk value. The risk trajectory function is calculated at the previous step size at the current time. The function value at time t. The time sampling interval is defined as the risk potential state sequence. The risk change acceleration is further calculated based on the forward difference result of the risk change rate. The risk change rate and acceleration obtained in this way can quantify the dynamic evolution trend of risk, overcoming the deficiency of traditional static threshold early warning systems in capturing the patterns of risk change. Preset absolute value thresholds, velocity thresholds, and acceleration thresholds are obtained. These thresholds are set based on the evolution data of clinical infusion extravasation risk and expert consensus, corresponding to the critical values ​​for risk severity, rapid risk escalation, and accelerated risk deterioration, respectively. The current risk value is compared with the absolute value threshold, the risk change rate is compared with the velocity threshold, and the risk change acceleration is compared with the acceleration threshold, resulting in three independent comparison results. These comparison results are represented in Boolean form to indicate whether each indicator has reached a critical state.

[0101] Based on the above three comparison results, a composite logical judgment is performed. The preset triggering conditions can be "risk value greater than absolute value threshold" and "risk change rate greater than speed threshold or risk change acceleration greater than acceleration threshold". This takes into account both the severity of the risk and its dynamic evolution trend, thus avoiding false warnings caused by short-term fluctuations at low risk levels and ineffective warnings in stable states at high risk levels. The warning mechanism is only triggered when the risk has reached a critical level and is in a rapid deterioration phase. When the comparison results simultaneously meet the preset triggering conditions, the warning mechanism can be activated and enter the risk attribution analysis process. For example, when the warning mechanism is triggered, the contribution of the current risk value to each basic feature in the spatiotemporal sample unit can be calculated first using the integral gradient method. That is, by calculating the gradient integral of the risk value relative to each basic feature, the influence of each basic feature on the risk warning result can be quantified, avoiding the one-sidedness of a single gradient calculation. The basic features in the spatiotemporal sample unit can include pixel features of visual image block sequences, features of each dimension of the comprehensive signal feature vector, and features of each dimension of the event context vector. The integral gradient value is calculated for each basic feature, and the absolute value of the gradient value is used as the feature contribution. Then, the contribution of all features is normalized, and the preset number of features with the highest contribution can be selected as key risk contribution factors. These factors directly point to the core causes of extravasation risk, such as tissue impedance characteristics, the proportion of pixels in the pale area of ​​the puncture site, and hyperosmolar drug event characteristics.

[0102] Subsequently, the sensitivity of the risk increment within the most recent preset time period to the risk latent states at each historical moment in the updated risk latent state sequence can be calculated using the time backpropagation method. The risk increment is the difference between the risk value at the current moment and the risk value at the start of the preset time period. The time backpropagation method uses this risk increment as the loss value and backpropagates it to each historical moment in the risk latent state sequence, calculating the gradient of the risk latent state at each historical moment to this loss value. This gradient value is the sensitivity; the higher the sensitivity, the greater the contribution of the risk latent state at the corresponding historical moment to the current risk increment. The moment with the highest sensitivity is determined as the starting point of the critical risk accumulation period. Combined with the current moment, a complete critical risk accumulation period can be formed. This period accurately locates the core period when risk begins to accumulate rapidly, providing a temporal basis for clinically tracing the risk evolution process.

[0103] By structurally integrating key risk contribution factors, key risk accumulation time periods, current risk values, and information on the rate and acceleration of risk change according to a pre-defined structured data format, an evidence chain can be generated. This evidence chain provides a comprehensive traceability and quantitative explanation of the warning result, clearly identifying the core characteristic factors that triggered the risk, pinpointing the key time range for risk accumulation, and including the current quantitative level and dynamic trend of the risk. It provides clinical medical staff with clear and verifiable evidence for risk assessment. For example, key risk contribution factors can be stored as key-value pairs of "feature name - contribution level," key risk accumulation time periods can be stored as time intervals, and risk values, risk change rates, and risk change accelerations can be stored as quantitative values. Corresponding labels and descriptions can be added to each type of information to ensure the evidence chain is clear and the data is complete.

[0104] In one embodiment, a decision report is generated based on a chain of evidence for decision mapping, including:

[0105] The evidence chain is analyzed to extract information such as risk value, risk change rate, risk change acceleration, key risk contribution factors, and key risk accumulation time period. Based on the preset risk level classification standard, the current risk level is determined by combining the risk value, risk change rate, and risk change acceleration.

[0106] Based on key risk contribution factors and key risk accumulation time periods, risk change trend information is determined by combining risk change rate and risk change acceleration; risk level, key risk contribution factors and risk change trend information are integrated to generate query conditions.

[0107] The similarity calculation is performed between the query conditions and the condition patterns in the predefined structured intervention decision knowledge base, and intervention rules with a matching degree exceeding the preset threshold are retrieved.

[0108] The intervention measures corresponding to the merged intervention rules are calculated, and an initial basic score is calculated for each intervention measure based on the matching degree and inherent priority of each intervention rule. The initial basic scores are then used to sort and generate a preliminary set of intervention measures.

[0109] The trajectory segment with a preset duration centered on the warning time is extracted from the risk trajectory function. The similarity between the trajectory segment and the case trajectory segment in the historical case database is calculated based on the dynamic time warping algorithm. The case with the highest preset number of similarities is selected as similar cases. The frequency of occurrence and average effect score of the intervention measures actually taken and marked as effective in the similar cases are statistically analyzed to obtain the case measure statistical data set.

[0110] For each intervention in the initial intervention set, the corresponding frequency of occurrence and average effect score are extracted from the case intervention statistics set, and the correlation coefficient is calculated. The initial base score is weighted using the correlation coefficient to obtain the calibrated priority score for each intervention in the initial intervention set.

[0111] Interventions with both frequency of occurrence and average effect score higher than the corresponding preset threshold are selected from the statistical data set of case measures. Interventions already included in the preliminary intervention measure set are removed to obtain supplementary measures. Based on the frequency of occurrence and average effect score of the supplementary measures in the statistical data set of case measures, an initial score for the supplementary measures is calculated by weighted summation. The supplementary measures are assigned the corresponding initial score and added to the preliminary intervention measure set to obtain the candidate intervention measure set.

[0112] Extract each candidate intervention from the candidate intervention set, sort them in descending order according to the scores of each candidate intervention, check the logical dependencies between each candidate intervention, and generate a sorted list of personalized interventions.

[0113] Each intervention measure in the personalized intervention measure list is analyzed to generate intervention instructions; information on risk level, key risk contribution factors, risk change trend, risk value, risk change rate, risk change acceleration, and key risk accumulation period is integrated to obtain early warning basis;

[0114] The early warning criteria are linked to each intervention instruction to generate a decision report.

[0115] Specifically, based on preset risk level classification standards, the current risk level can be determined by combining risk value, risk change rate, and risk change acceleration. The preset risk level classification standards can be set according to the severity and evolution speed of clinical infusion extravasation risk. The risk level can be obtained by matching the values ​​of the three current indicators with corresponding intervals. Based on key risk contributing factors and key risk accumulation time periods, risk change trend information can be determined by combining the risk change rate and risk change acceleration. That is, the sign and magnitude of the risk change rate and acceleration can be used to determine the risk. For example, when both the risk change rate and acceleration are positive and exceed the corresponding threshold, it can be determined as a rapidly deteriorating risk trend; when the risk change rate is positive and the acceleration is negative, it can be determined as a slowly rising risk trend. Integrating risk level, key risk contributing factors, and risk change trend information, query conditions can be generated. These conditions are represented in structured vector form, covering elements such as risk level labels, key risk contributing factor types, and risk change trend labels.

[0116] Subsequently, a cosine similarity algorithm can be used to calculate the similarity between the query conditions and the condition patterns in a predefined structured intervention decision knowledge base. This structured intervention decision knowledge base is constructed by clinical experts based on treatment guidelines and practical experience. Each intervention rule includes a condition pattern, a set of intervention measures, and an inherent priority. The condition pattern can be represented by a structured vector with the same dimension as the query conditions, while the set of intervention measures can include recommended clinical procedures under that rule. The inherent priority can be set based on the urgency and level of evidence-based support. Illustrated, the matching degree can be obtained by calculating the cosine of the angle between the query condition vector and the condition pattern vectors of each rule. A higher matching degree indicates a stronger fit between the rule and the current risk scenario. After retrieving intervention rules with matching degrees exceeding a preset threshold, all intervention measures corresponding to these rules can be merged. An initial base score is calculated for each intervention measure based on its matching degree and inherent priority. This initial base score is equal to the weighted sum of the rule's matching degree and its inherent priority, with the weights set according to clinical expert consensus. By sorting interventions in descending order based on the initial baseline scores, a preliminary set of interventions can be generated, which provides basic intervention options for subsequent case calibration.

[0117] Furthermore, a trajectory segment of a preset duration centered on the warning time can be extracted from the risk trajectory function. This trajectory segment covers the risk evolution process before and after the warning, accurately representing the dynamic characteristics of the current risk event. Based on the dynamic time warping algorithm, the similarity between the trajectory segment and case trajectory segments in the historical case database can be calculated. This algorithm, through non-linear alignment of the two trajectories, can eliminate the influence of trajectory length and local deformation. The calculated similarity value can objectively reflect the evolutionary similarity between the current risk event and historical cases. Subsequently, a preset number of cases with the highest similarity can be selected as similar cases. Similar cases represent historical events that are closest to the evolutionary pattern of the current risk scenario, and their intervention experience has high reference value for current decision-making. By statistically analyzing the frequency and average effect score of the intervention measures actually taken and marked as effective in the similar cases, a statistical data set of case measures can be obtained. The formula for calculating the average effect score is:

[0118]

[0119] in, For the first The average effectiveness score of the interventions that were actually taken and marked as effective was used to quantify the overall effectiveness of the intervention in similar historical cases. For the use of the historical case library The total number of effective cases in which an actual intervention measure was taken and marked as effective. An effective case refers to a case in which the risk was effectively controlled after the implementation of the measure. For the first The first one adopts the first The original effect scores are recorded in similar cases where the intervention was actually implemented and marked as effective. These original effect scores were assigned by clinicians after the case period based on the degree of risk mitigation, and the range of values ​​is [value missing]. ; This is the time decay coefficient, with a value greater than 0. It is used to characterize the timeliness of cases. More recent cases have a higher weight in the scoring calculation. The default value is determined based on iterative optimization of clinical data. For the first The time interval between similar cases and the current warning time, in days. The longer the time interval, the lower the weight of the case's timeliness. For the first In the 1st similar case The duration of the effective effect of an actual intervention measure that is marked as effective, in minutes, reflecting the duration of the effect after the measure takes effect; For the first In the 1st similar case The total execution time of each intervention measure that was actually taken and marked as effective, in minutes, is used to measure the efficiency of the measure's execution. For the first The sample weights of each similar case are assigned based on the match between the case patient's age, underlying disease, infusion type, and the current patient's condition, with values ​​ranging from [value range missing]. The higher the matching degree, the greater the sample weight. This average effect score, by comprehensively considering the timeliness of the case, the effectiveness of the measures, and the matching degree of the sample, avoids the bias caused by simple averaging and can more accurately reflect the expected effect of the intervention measures in the current risk scenario. The frequency of occurrence is [number missing]. The ratio of the number of times an actual intervention measure that was marked as effective was used in similar cases to the total number of similar cases is used to measure the universality of the measure in similar scenarios.

[0120] Specifically, for each intervention in the initial intervention set, the corresponding frequency of occurrence and average effect score can be extracted from the case intervention statistics set. A correlation coefficient is obtained by calculating the weighted sum of the frequency of occurrence and the average effect score, where the weights are set according to clinical needs. The frequency of occurrence represents universality, and the average effect score represents effectiveness; both together determine the relevance of the intervention. Using the correlation coefficient to weight the initial baseline score, a calibrated priority score can be obtained for each intervention in the initial intervention set. The calibrated priority score integrates the evidence-based nature of rule matching with the specificity of case experience, more accurately reflecting the priority of the intervention in the current scenario. Interventions with both frequency of occurrence and average effect score higher than the corresponding preset threshold are selected from the case intervention statistics set, and those already included in the initial intervention set are removed, resulting in supplementary interventions. These supplementary interventions are high-frequency, effective intervention options in similar cases that are not covered by the rule base, further enriching the diversity of intervention programs. Furthermore, based on the frequency of occurrence and average effect score of the supplementary interventions in the case intervention statistics set, an initial score for the supplementary interventions can be calculated by weighted summation, with the weights consistent with those in the correlation coefficient calculation. By assigning initial scores to supplementary measures, these measures, carrying their initial scores, can be added to the preliminary intervention set, resulting in a candidate intervention set. This candidate intervention set integrates rule recommendations and case experience-based intervention options, providing a foundation for the subsequent generation of personalized intervention lists.

[0121] Specifically, each candidate intervention is extracted from the candidate intervention set and sorted in descending order according to its calibrated priority score or initial score. During the sorting process, logical dependencies between candidate interventions can be checked simultaneously. For example, if pausing intravenous infusion is a prerequisite for local wet compresses, the priority of the prerequisite intervention must be higher than that of the subsequent intervention. This results in a sorted personalized intervention list, arranged from highest to lowest priority, clearly defining the execution order of the interventions. Parsing each intervention in the personalized intervention list generates intervention instructions. These instructions not only transform abstract interventions into concrete, executable clinical operation descriptions but also clarify key parameters and precautions. Integrating information on risk level, key risk contributing factors, risk change trends, risk values, risk change rates, risk change accelerations, and key risk accumulation periods yields early warning criteria. These criteria are presented in structured text, clearly explaining the severity of the current risk, its core causes, evolutionary trends, and key accumulation periods, providing clinicians with comprehensive risk background information. By associating the early warning criteria with each intervention instruction, a corresponding execution rationale can be matched for each instruction. This rationale is generated based on the correspondence between risk characteristics and intervention measures. Finally, combining the early warning criteria, intervention instructions, and execution rationales in a preset format generates a decision report. This report is output as a standardized document, covering core content such as risk descriptions, intervention priority lists, and operational procedures, providing decision support for clinical emergency response.

[0122] Based on the same inventive concept, this application also provides an outpatient infusion extravasation risk warning system for implementing the aforementioned outpatient infusion extravasation risk warning method. The solution provided by this system is similar to the solution described in the above method; therefore, the specific limitations of one or more embodiments of the outpatient infusion extravasation risk warning system provided below can be found in the above-described limitations of the outpatient infusion extravasation risk warning method, and will not be repeated here.

[0123] In one exemplary embodiment, such as Figure 3 As shown, an outpatient infusion extravasation risk early warning system 300 is provided, including:

[0124] The spatiotemporal sampling and feature fusion module 301 is used to acquire visual frame sequences, physiological signal sequences, and medical event streams in real time. Based on a preset sliding time window, it extracts corresponding visual frames, physiological signal data points, and medical event records from the visual frame sequences, physiological signal sequences, and medical event streams. It performs feature extraction on the visual frames, physiological signal data points, and medical event records respectively, and integrates the corresponding standardized image block sequences, comprehensive signal feature vectors, and event context vectors to construct spatiotemporal sample units.

[0125] The gated dynamic risk coding module 302 is used to input the spatiotemporal sample unit into the gated spatiotemporal joint encoder for processing, generate a risk potential state to characterize the comprehensive risk state at the current moment, add the risk potential state to the risk potential state sequence maintained for the patient, and obtain an updated risk potential state sequence.

[0126] The evidence-driven risk decision-making module 303 is used to perform risk analysis based on the updated risk potential state sequence, generate an evidence chain, perform decision mapping based on the evidence chain, and generate a decision report.

[0127] In one exemplary embodiment, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the outpatient infusion extravasation risk warning method of this application. A multi-core processor is preferred to improve the system's parallel processing capability. The memory provides sufficient temporary storage space to support program execution and data processing. The memory capacity should be large enough to accommodate large amounts of data and computational tasks.

[0128] In one exemplary embodiment, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the outpatient infusion extravasation risk warning method of this application. The computer-readable storage medium may include: a read-only memory, a random access memory, a solid-state drive, or an optical disk, etc.

[0129] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for early warning of extravasation risk during outpatient intravenous infusion, characterized in that, The method includes: Real-time acquisition of visual frame sequences, physiological signal sequences, and medical event streams; based on a preset sliding time window, extraction of corresponding visual frames, physiological signal data points, and medical event records from the visual frame sequences, physiological signal sequences, and medical event streams; feature extraction of the visual frames, physiological signal data points, and medical event records respectively; integration of the corresponding standardized image patch sequences, comprehensive signal feature vectors, and event context vectors to construct spatiotemporal sample units; The spatiotemporal sample unit is input into the gated spatiotemporal joint encoder for processing to generate a risk potential state that characterizes the comprehensive risk status at the current moment. The risk potential state is added to the risk potential state sequence maintained for the patient to obtain an updated risk potential state sequence. Risk analysis is performed based on the updated potential risk state sequence to generate a chain of evidence. Decision mapping is then performed based on the chain of evidence to generate a decision report.

2. The method according to claim 1, characterized in that, The process involves extracting features from the visual frame, the physiological signal data points, and the medical event record, respectively, and then integrating the corresponding standardized image patch sequence, comprehensive signal feature vector, and event context vector to construct a spatiotemporal sample unit, including: Automatic puncture point localization is performed on the visual frame to obtain puncture point localization results; based on the puncture point localization results, the visual frame is cropped to obtain cropped image blocks; illumination and color correction is performed on the cropped image blocks to obtain the standardized image block sequence. The physiological signal data points are subjected to bandpass filtering to obtain denoised physiological signals; time-domain features, frequency-domain features, and nonlinear features are calculated based on the denoised physiological signals to obtain multi-dimensional signal features; the multi-dimensional signal features are combined to obtain the comprehensive signal feature vector. The medical event record is vectorized and embedded to obtain an event embedding vector; the event embedding vector is weighted and aggregated based on a time attention mechanism to obtain the event context vector. The standardized image patch sequence, the integrated signal feature vector, and the event context vector are packaged in an ordered manner to construct the spatiotemporal sample unit.

3. The method according to claim 2, characterized in that, The gated spatiotemporal co-encoder includes a visual branch, a non-visual branch, a gated fusion module, and an output module. The process of inputting the spatiotemporal sample unit into the gated spatiotemporal co-encoder for processing generates a potential risk state characterizing the comprehensive risk state at the current moment, including: The standardized image patch sequence is extracted from the spatiotemporal sample unit, and the standardized image patch sequence is input into the visual branch of the gated spatiotemporal co-encoder. Spatiotemporal features are extracted through a three-dimensional convolutional neural network to obtain a visual spatiotemporal feature vector. The integrated signal feature vector and the event context vector are extracted from the spatiotemporal sample unit. The integrated signal feature vector and the event context vector are concatenated to obtain a non-visual joint feature vector. The non-visual joint feature vector is input into the non-visual branch of the gated spatiotemporal joint encoder and subjected to nonlinear transformation by a multilayer perceptron to obtain a non-visual feature vector. Based on the gated fusion module, a gated vector is obtained by calculating the visual spatiotemporal feature vector and the non-visual feature vector using a preset gated weight generation function; the gated vector is then used to perform element-wise weighted fusion processing on the visual spatiotemporal feature vector and the non-visual feature vector to obtain a fused feature vector. Based on the output module, the fused feature vector is mapped to generate the potential risk state.

4. The method according to claim 1, characterized in that, The risk analysis based on the updated potential risk sequence, generating a chain of evidence, includes: Extract the continuous risk latent states arranged in chronological order from the updated risk latent state sequence, input the continuous risk latent states into a causal dilated convolutional network to obtain a discrete risk value sequence; perform spline interpolation on the discrete risk value sequence to construct a continuous risk trajectory function; Based on the risk trajectory function, the risk value, risk change rate, and risk change acceleration at the current moment are calculated using a numerical differentiation method; preset absolute value thresholds, velocity thresholds, and acceleration thresholds are obtained, and the risk value, risk change rate, and risk change acceleration are compared with the corresponding absolute value thresholds, velocity thresholds, and acceleration thresholds to obtain comparison results; Based on the comparison result, a compound logic judgment is performed, and when the comparison result meets the preset triggering condition, an early warning mechanism is triggered; When the early warning mechanism is triggered, the integral gradient method is used to calculate the contribution of the risk value to each basic feature in the spatiotemporal sample unit, and the preset quantitative feature with the highest contribution is selected as the key risk contribution factor; the time backpropagation method is used to calculate the sensitivity of the risk increment in the most recent preset period to the risk potential state at each historical moment in the updated risk potential state sequence, and the moment with the highest sensitivity is determined as the starting point of the key risk accumulation period. The evidence chain is generated by integrating the key risk contribution factors, the key risk accumulation period, the risk value, the risk change rate, and the risk change acceleration information.

5. The method according to claim 4, characterized in that, The step of performing decision mapping based on the evidence chain and generating a decision report includes: The evidence chain is analyzed to extract the risk value, the rate of risk change, the acceleration of risk change, the key risk contribution factor, and the key risk accumulation time period. Based on a preset risk level classification standard, the risk value, the rate of risk change, and the acceleration of risk change are combined to determine the level of the current risk and obtain the risk level. Based on the key risk contribution factors and the key risk accumulation time period information, combined with the risk change rate and the risk change acceleration, risk change trend information is determined; the risk level, the key risk contribution factors, and the risk change trend information are integrated to generate query conditions; The query conditions are matched with the condition patterns in the predefined structured intervention decision knowledge base to calculate similarity, and intervention rules with a matching degree exceeding a preset threshold are retrieved. The intervention measures corresponding to the intervention rules are merged, and an initial basic score is calculated for each intervention measure based on the matching degree and inherent priority of each intervention rule. A preliminary set of intervention measures is generated by sorting the intervention measures based on the initial basic scores. From the risk trajectory function, a trajectory segment of a preset duration centered on the warning time is extracted. The similarity between the trajectory segment and the case trajectory segments in the historical case database is calculated based on the dynamic time warping algorithm. The preset number of cases with the highest similarity are selected as similar cases. The frequency of occurrence and average effect score of the intervention measures actually taken and marked as effective in the similar cases are statistically analyzed to obtain a case measure statistical data set. For each intervention measure in the preliminary intervention measure set, the corresponding occurrence frequency and average effect score are extracted from the case measure statistical data set, and a correlation coefficient is calculated; the initial base score is weighted using the correlation coefficient to obtain the calibrated priority score corresponding to each intervention measure in the preliminary intervention measure set; Interventions with occurrence frequencies and average effect scores higher than the corresponding preset thresholds are selected from the case measure statistics set, and interventions already included in the preliminary intervention measure set are removed to obtain supplementary measures; based on the occurrence frequency and average effect score of the supplementary measures in the case measure statistics set, an initial score for the supplementary measures is calculated by weighted summation, and the supplementary measures are assigned the corresponding initial scores. The supplementary measures carrying the initial scores are added to the preliminary intervention measure set to obtain a candidate intervention measure set; Extract each candidate intervention measure from the candidate intervention measure set, sort them in descending order according to the scores corresponding to each candidate intervention measure, check the logical dependencies between each candidate intervention measure, and generate a sorted personalized intervention measure list. Each intervention measure in the personalized intervention measure list is analyzed to generate an intervention instruction; the risk level, the key risk contribution factor, the risk change trend, the risk value, the risk change rate, the risk change acceleration, and the key risk accumulation time period information are integrated to obtain the early warning basis; The warning criteria are associated with each of the intervention instructions to generate the decision report.

6. The method according to claim 3, characterized in that, The mathematical expression for the preset gate weight generation function is: in, Let be the gate vector; Use the Sigmoid activation function; and All are learnable weight matrices; Norm(·) is the L2 normalization operation; The visual spatiotemporal feature vector; The non-visual feature vector; Conf The joint confidence score of the visual spatiotemporal feature vector and the non-visual feature vector takes the value [0,1]. The batch mean of the visual spatiotemporal feature vector; The batch mean of the non-visual feature vector; It is an L2 norm; This is the bias vector.

7. The method according to claim 5, characterized in that, The formula for calculating the average performance score is as follows: in, For the first The average effectiveness score of the interventions actually taken and marked as effective as described in the item; The historical case library adopts the first The total number of valid cases for which the aforementioned interventions were actually implemented and marked as effective; For the first The first one adopts the first The original effect scores recorded in similar cases of the actual interventions taken and marked as effective, with a range of values. ; This is the time decay coefficient; For the first The time interval between a similar case and the current warning time; For the first In the 1st similar case The duration of the effective effect produced by the intervention measures actually taken and marked as effective as described in the item; For the first In the 1st similar case The total duration of the intervention measures actually taken and marked as effective as described in the item; For the first Sample weights for similar cases.

8. An outpatient intravenous infusion extravasation risk early warning system, characterized in that, The system includes: The spatiotemporal sampling and feature fusion module is used to acquire visual frame sequences, physiological signal sequences, and medical event streams in real time. Based on a preset sliding time window, it extracts corresponding visual frames, physiological signal data points, and medical event records from the visual frame sequences, physiological signal sequences, and medical event streams. It then performs feature extraction on the visual frames, physiological signal data points, and medical event records respectively, and integrates the corresponding standardized image patch sequences, comprehensive signal feature vectors, and event context vectors to construct spatiotemporal sample units. The gated dynamic risk coding module is used to input the spatiotemporal sample unit into the gated spatiotemporal joint encoder for processing, generate a risk potential state to characterize the comprehensive risk state at the current moment, and add the risk potential state to the risk potential state sequence maintained for the patient to obtain an updated risk potential state sequence. The evidence-driven risk decision-making module is used to perform risk analysis based on the updated risk potential state sequence, generate an evidence chain, perform decision mapping based on the evidence chain, and generate a decision report.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images