Infection risk assessment method and system for burn nursing

Abstract

The invention discloses an infection risk assessment method and system for burn nursing, and the system comprises a long-term trend analysis module and a short-term anomaly detection module, the long-term trend analysis module employs an improved Transform-XL model, and captures the chronic cumulative effect of inflammation indexes within 72 hours through a relative position coding and memory multiplexing mechanism; the short-term anomaly module is based on a bidirectional LSTM reconstruction network, combines a dynamic threshold algorithm to identify sudden physiological anomaly in a window of 6 hours in real time, generates a final infection risk index through a confidence-driven adaptive fusion algorithm and dynamic weighted dual-channel output, triggers clinical response in combination with a hierarchical early warning mechanism, and introduces closed-loop feedback optimization, so as to achieve the early warning of the physiological anomaly. The actual intervention effect feedback model is iterated, the infection recognition time is shortened to 3.2 hours, the ROC-AUC reaches 0.94, and the reasonable utilization rate of antibiotics is increased to 89%.

Description

Technical Field

[0001] The present invention relates to an infection risk assessment method and system for burn care, belonging to the field of medical technology. Background Art

[0002] Burns are a common trauma, and managing infection risk is crucial during their care. Burn wounds provide an ideal environment for the growth of bacteria and other microorganisms. If improperly managed, they can easily lead to infection, which can cause a series of serious complications and even threaten the patient's life. Therefore, accurately and promptly assessing the infection risk of burn patients and implementing appropriate interventions have always been important issues in burn care.

[0003] Traditional burn infection risk assessment methods rely primarily on the clinical experience of healthcare professionals, determining infection risk through observation and analysis of indicators such as the patient's body temperature, white blood cell count, and wound condition. However, this method suffers from high subjectivity, long assessment cycles, and insufficient accuracy, making it difficult to meet the demands of modern burn care for efficient and precise infection risk management.

[0004] With the rapid development of medical information technology and artificial intelligence (AI), a growing number of researchers are exploring the use of big data and machine learning to improve burn infection risk assessment methods. For example, by collecting and analyzing multidimensional data such as patients' physiological indicators, medical history, and treatment records, predictive models are constructed to assess infection risk. However, existing machine learning-based infection risk assessment methods still have some limitations.

[0005] On the one hand, burn infection is a complex process involving physiological changes across multiple timescales. For example, the cumulative effects of the inflammatory response may gradually manifest over hours or even days, while certain sudden abnormalities (such as a sudden rise in body temperature) may occur in a short period of time. Existing assessment methods often struggle to simultaneously capture these physiological changes across timescales, resulting in inaccurate and ineffective assessment results.

[0006] On the other hand, the physiological data of burn patients is often affected by various factors, such as sensor noise and missing data. These factors can lead to a decline in data quality, thereby affecting the accuracy of assessment results. Existing assessment methods often lack dynamic data quality assessment and adjustment mechanisms, making them difficult to adapt to the complex and changing clinical environment.

[0007] Furthermore, existing infection risk assessment methods often remain at the early warning stage and lack close integration with clinical intervention measures. Even if infection risk can be accurately warned, the incidence of infection cannot be truly reduced without timely and effective intervention measures. Therefore, establishing a clinical pathway that enables real-time monitoring, early warning, intervention, and iterative optimization is crucial for improving the accuracy and effectiveness of burn infection risk assessment. Summary of the Invention

[0008] In order to overcome the problems raised in the above background technology, the present invention proposes an infection risk assessment method and system for burn care.

[0009] The technical solution of the present invention is: an infection risk assessment method for burn care, comprising the following steps: S11: Data collection, collecting data on burn patients’ wounds through high-definition cameras, sensors, and laboratory equipment; S12: Data standardization and processing: standardization and feature engineering of the collected data. Standardization includes linear interpolation of low-frequency data based on the highest frequency, filling short-term missing data with a moving average method, and marking long-term missing data. S13: Dynamic risk assessment model construction: establish a dynamic risk assessment model, input the processed data into the dynamic risk assessment model, and output an assessment report; S14: Real-time early warning and clinical response, based on the output of the dynamic risk assessment model, using a hierarchical early warning mechanism and closed-loop feedback management approach to respond; S15: Verification and continuous optimization: Use a combination of manual evaluation and system evaluation to verify the evaluation results and responses, and regularly input new case data for iterative optimization of the model.

[0010] Preferably, when collecting data, the collected data includes: A11: Wound surface data: Regularly photograph the wound surface with a high-definition camera to record the wound surface color, exudate volume, and edge redness and swelling. A12: Physiological indicators: sensors monitor the patient's body temperature, heart rate, and blood oxygen saturation in real time, and simultaneously record fever events; A13: Laboratory data: The patient's wound is analyzed using laboratory equipment. The analyzed data include the bacterial load and inflammatory factors in the wound exudate; A14: Environmental data, monitoring the humidity and temperature around the wound.

[0011] Preferably, when constructing a dynamic risk assessment model, the structure of the constructed dynamic risk assessment model is: A21: Parallel processing of two channels, including a long-term trend channel and a short-term anomaly channel. The long-term trend channel is used to capture the evolution of slow variables across time windows, and the short-term anomaly channel is used to identify sudden abnormal signals. A22: Information fusion and decision-making layer, used to fuse the data results of long-term trend channels and short-term abnormal channels and make decisions.

[0012] As a preference, when constructing a dynamic risk assessment model, the long-term trend channel structure is as follows: A31: The input layer uses multivariate time series normalization and time window partitioning to divide 72 hours of data into 18 time points, retaining the timestamps of key medical events and standardizing to eliminate dimensional differences, so that the model treats indicators of different magnitudes equally; A32: Embedding layer, which uses a fully connected layer + ReLU activation + residual connection technology to map low-dimensional clinical indicators to a high-dimensional semantic space, capture nonlinear relationships, and retain the original numerical features through residual connections to prevent gradient disappearance; A33: Position encoding layer, which uses relative position encoding technology to replace traditional absolute position encoding, enhances the model's understanding of the timeliness of medical events, and allows attention calculation across time windows; A34: Transformer-XL layer, which uses multi-head self-attention, memory reuse, and layer normalization techniques to identify long-term pathological evolution patterns through attention across time windows. It also retains key historical states in a memory bank to assist in determining whether the current abnormality is a recurrence. A35: The output layer uses time pooling and Sigmoid activation technology. Time pooling technology is used to select the most significant risk signals to avoid dilution of key events by average pooling. Sigmoid activation technology is used to compress the risk score to the range of 0-1 to intuitively reflect the probability of infection.

[0013] As a preference, when constructing a dynamic risk assessment model, the working principles of the long-term trend channel include: S21: Input processing: divide the 72-hour data into 18 time points, each of which contains 5-dimensional features, including body temperature, CRP, PCT, wound moisture, and pain score; S22: Embedding and position encoding, mapping the 5-dimensional input to 128 dimensions through the fully connected layer, introducing nonlinear relationships, and calculating for any two time points i and j ,in, represents the query vector at position i, represents the key vector at position j, represents the embedding vector with time difference ji, represents query-key interaction; S23: Memory reuse mechanism: Each Transformer layer retains the hidden state of the previous three time windows, controls the information strength through weight decay, and calculates the attention weight by combining the query vector of the current window and the key vector of the historical window to capture long-term dependencies; S24: Risk probability calculation, the hidden state of the last layer at the 18th time point outputs the baseline infection probability through the Sigmoid function.

[0014] As a preference, when constructing a dynamic risk assessment model, the structure of the short-term abnormal channel is: A41: Input layer, which uses a sliding window mechanism to maintain a rolling transmission of 6 hours, sampling every 5 minutes, and is used to receive time series data of original physiological parameters as the input basis of the model; A42: Sliding window normalization, which is used to calculate the mean and standard deviation of each feature within each 6-hour window to eliminate dimensional differences between features; A43: Bidirectional LSTM layer, with a 2-layer stacked 128-unit structure, consisting of a first bidirectional LSTM layer and a second bidirectional LSTM layer. The forward LSTM of the first bidirectional LSTM layer processes the sequence from time steps 1 to 72 to capture gradual trends. The reverse LSTM of the first bidirectional LSTM layer processes the sequence from time steps 72 to 1 in reverse order to locate the starting point of the mutation. Each direction of the first bidirectional LSTM layer outputs a 72×127 hidden state, which is concatenated to obtain a 72×256 sequence. The output of the second bidirectional LSTM layer is the 72×256 sequence output by the first bidirectional LSTM layer. The structure of the second bidirectional LSTM layer is the same as that of the first bidirectional LSTM layer. The second bidirectional LSTM layer is used to further extract high-order temporal patterns. A44: reconstruction layer, which uses a fully connected neural network to restore the physiological parameter values in normal mode from the hidden state; A45: Anomaly scorer, which uses dynamic thresholds and exponential moving average technology to quantify the degree of anomaly at each time point, avoiding false positives and false negatives caused by fixed thresholds.

[0015] As a preference, when constructing a dynamic risk assessment model, the working principles of the short-term abnormal channel include: S31: Input processing, maintaining a 6-hour rolling window with a time point every 5 minutes, including 5-dimensional features, and calculating the mean and standard deviation of each feature separately to eliminate dimensional differences; S32: Bidirectional LSTM propagation, including: Forward LSTM, from the 1st time point to the 72nd time point, learns the normal physiological pattern; Reverse LSTM, traces back from the 72nd time point to the 1st time point to locate the abnormal starting point; State concatenation: concatenate the forward hidden state and the reverse hidden state into a 256-dimensional vector; S33: Anomaly detection, first perform reconstruction prediction, reconstruct the input vector from the 256-dimensional hidden state through the fully connected layer , and then perform error calculation, and set the threshold based on the exponential moving average of historical error data within a 24-hour range. The error calculation formula is: ,in, is the original input vector, is the reconstructed input vector; The principle formula for threshold setting is: ; in, is the dynamic threshold at time t, is the mean error at time t, represents the standard deviation of the error at time t, where , , is the EMA mean at the previous time point t−1, is the standard deviation of the previous time point t−1; S34: Alarm is triggered if the error of 3 consecutive time points Greater than threshold , it is marked as a red alert.

[0016] As a preference, the information fusion and decision-making layer, when fusing the data results of the long-term trend channel and the short-term abnormal channel and making a decision, specifically includes: By formula: , calculate the final risk index; in, is the output of the long-term trend channel, i.e., the long-term baseline risk, is the output of the short-term anomaly channel, i.e., the short-term anomaly intensity, is the threshold, where ;in, is the confidence of the Transformer-XL model, is the confidence of the LSTM model, and , , is the entropy of the predicted probability, and p is the original probability value output by the model.

[0017] Preferably, when responding to the output of the dynamic risk assessment model using a hierarchical early warning mechanism and closed-loop feedback management approach, the following steps are specifically included: S41: Dynamic threshold adjustment, which adjusts the threshold according to the medication status and the patient's immune system status; S42: Grading warnings: Classify warnings into medium-risk warnings and high-risk warnings, use corresponding warning notification methods to issue warnings, and automatically generate disposal suggestions for high-risk warnings; S43: Closed-loop feedback management, comparing indicator changes 24 hours after the warning, evaluating the effectiveness of interventions, and recording high-risk cases without infection for iterative model optimization.

[0018] Preferably, when automatically generating disposal suggestions for high-risk warnings, the following are specifically included: S51: Diagnostic label mapping: input the warning triggering reason, and map the warning triggering reason to a specific diagnostic symptom according to the mapping rules in the diagnostic database; S52: Knowledge graph query, searching in the structured diagnosis and treatment rule base according to diagnostic symptoms and performing dynamic matching; S53: Contraindication verification, first filtering according to patient characteristics, and then adjusting the dose according to the patient's physiological parameters; S54: Multi-model collaborative selection, using drug selection models to perform specific drug selection.

[0019] An infection risk assessment system for burn care includes: Data acquisition module, used to obtain full-dimensional clinical data of burn patients in real time; Data processing module, used to convert raw data into standardized features that can be analyzed; Transformer model module, used to integrate temporal dynamics and spatial distribution features for infection risk assessment; The output and warning module is used to convert the prediction results into executable clinical instructions.

[0020] Beneficial effects of the present invention: 1. Divide and Conquer - Synergistic Mechanism: Transformer-XL captures the cumulative effects of inflammation up to 72 hours, combined with bidirectional LSTM to detect sudden abnormalities (such as a sudden rise in body temperature) within 6 hours, achieving risk perception across time scales. 2. Entropy-driven adaptive fusion: Utilizes model-predicted entropy to dynamically calculate dual-channel weights. When the quality of a channel's data degrades (e.g., due to sensor noise), its decision weight is automatically reduced. 3. Closed-loop clinical pathway: The intervention effects (such as antibiotic response and changes in inflammatory indicators) within 24 hours after the warning are fed back to the model to form an iterative loop of "monitoring-warning-intervention-optimization". BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1Shown is a flow chart of the infection risk assessment method for burn care of the present invention; Figure 2 Shown is a schematic diagram of the structure of a dynamic risk assessment model in the infection risk assessment method for burn care of the present invention; Figure 3 Shown is a schematic diagram of the structure of the infection risk assessment system for burn care of the present invention. DETAILED DESCRIPTION

[0022] The present invention will be further described below with reference to the accompanying drawings and examples.

[0023] Example 1 See also Figure 1-2 The present invention provides an embodiment: an infection risk assessment method for burn care, comprising the following steps: 1. Multimodal Data Collection and Standardization 1.1 Data Collection The system deploys a four-layer sensor network: Wound surface: A medical-grade 4K embedded camera (resolution 4096×2160) was used to photograph the wound surface every 2 hours. Computer vision algorithms were used to quantify the color entropy (standard deviation of the red channel in HSV space), the proportion of exudate area (based on U-Net image segmentation), and the edge swelling index (contour curvature change rate). Physiological layer: A flexible electronic skin patch (0.3mm thickness) continuously monitors the wound microenvironment temperature (accuracy ±0.1°C) and humidity (capacitive sensor, accuracy ±2%RH), while simultaneously integrating a wrist device to collect heart rate variability (HRV) and blood oxygen saturation (SpO2); Laboratory level: Bedside microfluidic chips analyze wound exudate every six hours, quantitatively detecting bacterial load (16SrRNA copies / μL) and fungal β-D-glucan concentrations through qPCR, and electrochemical sensors detecting inflammatory factors such as IL-6 and TNF-α; Environmental layer: A distributed temperature and humidity sensor network in the ward (accuracy ±0.5°C) monitors the environment around the wound in real time, and triggers the dehumidifier when the humidity is >70%.

[0024] 1.2 Data Preprocessing To address the uneven sampling and missing characteristics of medical data, a three-stage processing pipeline is designed: 1. Frequency alignment: Using the highest sampling frequency (1 Hz for physiological parameters) as a benchmark, perform cubic spline interpolation on low-frequency data (such as CRP values every 6 hours) to ensure timestamp alignment; 2. Missing data processing: For short-term missing data (less than 3 consecutive points), adaptive Kalman filtering is used, and the state equation model is:

[0025] Among them, the state matrix A is dynamically adjusted according to the covariance of historical data; For long-term missing data (greater than 1 hour), a missing marker vector is introduced and attention masking is performed at the model input layer; 3. Feature Engineering: Constructing an Infection Risk Index: ; The texture features of the wound image are then extracted, and the energy responses in 8 directions and 4 scales are calculated using a Gabor filter bank to construct a 32-dimensional feature vector.

[0026] 2. Dynamic Risk Assessment Model Architecture 2.1 Long-term trend channel (modeling of chronic pathological evolution) Input layer design: Time window division: 72 hours of data are divided into 18 time points (every 4 hours is a node), each node contains 5-dimensional core features (body temperature, CRP, PCT, wound moisture, pain score); Medical event embedding: Key events (e.g., debridement surgery, antibiotic administration) are annotated with timestamps and fed into the model via learnable positional encoding vectors.

[0027] Improved Transformer-XL structure: 1. Semantic Embedding Layer: The fully connected network maps the 5-dimensional input to 128 dimensions: ; Introducing residual connections: ; 2. Relative position encoding: When calculating the attention between the query vector and the key vector, the embedding of the time difference Δ=i−j is introduced: ; in, is a learnable relative position vector used to enhance the model’s understanding of the timeliness of medical events; 3. Memory reuse mechanism: Each Transformer layer retains the hidden state of the previous three time windows, and historical information participates in the current calculation through the decay weight: ; In self-attention computation, the current query vector can access key states in the historical memory bank across time windows; Output layer optimization: Temporal attention pooling is used instead of traditional average pooling to automatically focus on high-risk periods: ; Final infection probability calculation: ; in, Sigmoid function, output range [0,1]; 2.2 Short-term abnormal channels (acute event detection) Sliding window processing: The window length is 6 hours, sliding every 5 minutes, with 5 input dimensions (body temperature, heart rate, SpO2, wound temperature, exudate flow rate), and the data within the window is normalized by Z-score: Bidirectional LSTM reconstruction network: 1. Encoder: Two layers of stacked bidirectional LSTM (128 units per layer). The forward LSTM learns the normal evolution pattern of physiological parameters, and the reverse LSTM locates the starting point of abnormalities. 2. Decoder: The fully connected network reconstructs the 256-dimensional hidden state into the original input dimension, minimizing the reconstruction error: ; 3. Dynamic threshold algorithm: Calculate the anomaly score for each time point: ; Set the threshold based on the exponential moving average (EMA) of the error over the previous 24 hours: ; ; ; Satisfied for 3 consecutive time points Greater than threshold A red alert is triggered when 2.3 Confidence Adaptive Fusion 1. Entropy calculation: For the original output probability p of each model, calculate the confidence: , ; Dynamic weight adjustment: ; Final Risk Index: ; 3. Clinical Response and System Verification Graded early warning mechanism: Yellow alert (0.4≤P_final<0.6): A pop-up window appears on the nurse's terminal, requiring wound swab culture and manual assessment to be completed within 4 hours; augmented reality (AR) glasses mark suspicious areas on the wound to assist in locating the sampling point.

[0028] Red alert (P_final ≥ 0.6 or persistent abnormality): The ward alarm light flashes, and the chief physician receives a push notification on his mobile phone (including a PDF of treatment recommendations); an antibiotic regimen is automatically generated: for example, the vancomycin dose is dynamically calculated based on creatinine clearance (CrCl), specifically: Diagnostic label mapping: input the warning trigger reason and map the warning trigger reason to a specific diagnostic symptom according to the mapping rules in the diagnostic database; Knowledge graph query, searching in the structured diagnosis and treatment rule base based on diagnostic symptoms and performing dynamic matching; Contraindication verification first filters according to patient characteristics, and then adjusts the dose according to the patient's physiological parameters. The dose adjustment algorithm is: defvancomycin_dose(weight,CrCl): ifCrCl>50:return "1gq12h" elif30 <CrCl<=50:return"1gq24h"; Multi-model collaborative selection, using drug selection models for specific drug selection: 3.2 Closed-loop feedback optimization Effectiveness evaluation: Compare key indicators 24 hours after the warning (such as the decrease in SOFA score and the rate of reduction of inflammatory factors) and calculate the intervention effectiveness: Model iteration: When the actual infection result deviates from the prediction by more than 20%, online learning is triggered: ; Among them, the learning rate , loss function is the weighted cross entropy.

[0029] Example 2 See also Figure 3 The present invention provides an embodiment of an infection risk assessment system for burn care, comprising: Data acquisition module, used to obtain full-dimensional clinical data of burn patients in real time; Data processing module, used to convert raw data into standardized features that can be analyzed; Transformer model module, used to integrate temporal dynamics and spatial distribution features for infection risk assessment; The output and warning module is used to convert the prediction results into executable clinical instructions.

[0030] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the scope of knowledge of those skilled in the art without departing from the spirit of the present invention.

Claims

1. An infection risk assessment method for burn care, characterized by: The following steps are included: S11: Data collection, collecting data on burn patients’ wounds through high-definition cameras, sensors, and laboratory equipment; S12: Data standardization and processing: standardization and feature engineering of the collected data. Standardization includes linear interpolation of low-frequency data based on the highest frequency, filling short-term missing data with a moving average method, and marking long-term missing data. S13: Dynamic risk assessment model construction: establish a dynamic risk assessment model, input the processed data into the dynamic risk assessment model, and output an assessment report; S14: Real-time early warning and clinical response, based on the output of the dynamic risk assessment model, using a hierarchical early warning mechanism and closed-loop feedback management approach to respond; S15: Verification and continuous optimization: Use a combination of manual evaluation and system evaluation to verify the evaluation results and responses, and regularly input new case data for iterative optimization of the model.

2. The infection risk assessment method for burn care according to claim 1, characterized in that: During data collection, the collected data includes: A11: Wound surface data: Regularly photograph the wound surface with a high-definition camera to record the wound surface color, exudate volume, and degree of redness and swelling at the edge; A12: Physiological indicators: sensors monitor the patient's body temperature, heart rate, and blood oxygen saturation in real time, and simultaneously record fever events; A13: Laboratory data: The patient's wound is analyzed using laboratory equipment. The analyzed data include the bacterial load and inflammatory factors in the wound exudate; A14: Environmental data, monitoring the humidity and temperature around the wound.

3. The infection risk assessment method for burn care according to claim 2, characterized in that: When constructing a dynamic risk assessment model, the structure of the constructed dynamic risk assessment model is: A21: Parallel processing of two channels, including a long-term trend channel and a short-term anomaly channel. The long-term trend channel is used to capture the evolution of slow variables across time windows, and the short-term anomaly channel is used to identify sudden abnormal signals. A22: Information fusion and decision-making layer, used to fuse the data results of long-term trend channels and short-term abnormal channels and make decisions.

4. The infection risk assessment method for burn care according to claim 3, characterized in that: When constructing a dynamic risk assessment model, the long-term trend channel structure is: A31: The input layer uses multivariate time series normalization and time window partitioning to divide 72 hours of data into 18 time points, retaining the timestamps of key medical events and standardizing to eliminate dimensional differences, so that the model treats indicators of different magnitudes equally; A32: Embedding layer, which uses a fully connected layer + ReLU activation + residual connection technology to map low-dimensional clinical indicators to a high-dimensional semantic space, capture nonlinear relationships, and retain the original numerical features through residual connections to prevent gradient disappearance; A33: Position encoding layer, which uses relative position encoding technology to replace traditional absolute position encoding, enhances the model's understanding of the timeliness of medical events, and allows attention calculation across time windows; A34: Transformer-XL layer, which uses multi-head self-attention, memory reuse, and layer normalization techniques to identify long-term pathological evolution patterns through attention across time windows. It also retains key historical states in a memory bank to assist in determining whether the current abnormality is a recurrence. A35: The output layer uses time pooling and Sigmoid activation technology. Time pooling technology is used to select the most significant risk signals to avoid dilution of key events by average pooling. Sigmoid activation technology is used to compress the risk score to the range of 0-1 to intuitively reflect the probability of infection.

5. The infection risk assessment method for burn care according to claim 4, characterized in that: When building a dynamic risk assessment model, the working principles of the long-term trend channel include: S21: Input processing: divide the 72-hour data into 18 time points, each of which contains 5-dimensional features, including body temperature, CRP, PCT, wound moisture, and pain score; S22: Embedding and position encoding, mapping the 5-dimensional input to 128 dimensions through the fully connected layer, introducing nonlinear relationships, and calculating for any two time points i and j ,in, represents the query vector at position i, represents the key vector at position j, represents the embedding vector with time difference ji, represents query-key interaction; S23: Memory reuse mechanism: Each Transformer layer retains the hidden state of the previous three time windows, controls the information strength through weight decay, and calculates the attention weight by combining the query vector of the current window and the key vector of the historical window to capture long-term dependencies; S24: Risk probability calculation, the hidden state of the last layer at the 18th time point outputs the baseline infection probability through the Sigmoid function.

6. The infection risk assessment method for burn care according to claim 5, characterized in that: When constructing a dynamic risk assessment model, the structure of the short-term abnormal channel is: A41: Input layer, which uses a sliding window mechanism to maintain a rolling transmission of 6 hours, sampling every 5 minutes, and is used to receive time series data of original physiological parameters as the input basis of the model; A42: Sliding window normalization, which is used to calculate the mean and standard deviation of each feature within each 6-hour window to eliminate dimensional differences between features; A43: Bidirectional LSTM layer, with a 2-layer stacked 128-unit structure, consisting of a first bidirectional LSTM layer and a second bidirectional LSTM layer. The forward LSTM of the first bidirectional LSTM layer processes the sequence from time steps 1 to 72 to capture gradual trends. The reverse LSTM of the first bidirectional LSTM layer processes the sequence from time steps 72 to 1 in reverse order to locate the starting point of the mutation. Each direction of the first bidirectional LSTM layer outputs a 72×127 hidden state, which is concatenated to obtain a 72×256 sequence. The output of the second bidirectional LSTM layer is the 72×256 sequence output by the first bidirectional LSTM layer. The structure of the second bidirectional LSTM layer is the same as that of the first bidirectional LSTM layer. The second bidirectional LSTM layer is used to further extract high-order temporal patterns. A44: reconstruction layer, which uses a fully connected neural network to restore the physiological parameter values in normal mode from the hidden state; A45: Anomaly scorer, which uses dynamic thresholds and exponential moving average technology to quantify the degree of anomaly at each time point, avoiding false positives and false negatives caused by fixed thresholds.

7. The infection risk assessment method for burn care according to claim 6, characterized in that: When building a dynamic risk assessment model, the working principles of the short-term abnormal channel include: S31: Input processing, maintaining a 6-hour rolling window with a time point every 5 minutes, including 5-dimensional features, and calculating the mean and standard deviation of each feature separately to eliminate dimensional differences; S32: Bidirectional LSTM propagation, including: Forward LSTM, from the 1st time point to the 72nd time point, learns the normal physiological pattern; Reverse LSTM, traces back from the 72nd time point to the 1st time point to locate the abnormal starting point; State concatenation: concatenate the forward hidden state and the reverse hidden state into a 256-dimensional vector; S33: Anomaly detection, first perform reconstruction prediction, reconstruct the input vector from the 256-dimensional hidden state through the fully connected layer , and then perform error calculation, and set the threshold based on the exponential moving average of historical error data within a 24-hour range. The error calculation formula is: ,in, is the original input vector, is the reconstructed input vector; The principle formula for threshold setting is: ; in, is the dynamic threshold at time t, is the mean error at time t, represents the standard deviation of the error at time t, where , , is the EMA mean at the previous time point t−1, is the standard deviation of the previous time point t−1; S34: Alarm is triggered if the error of 3 consecutive time points Greater than threshold , it is marked as a red alert.

8. The infection risk assessment method for burn care according to claim 7, characterized in that: When the information fusion and decision-making layer fuses the data results of the long-term trend channel and the short-term abnormal channel and makes decisions, it specifically includes: By formula: , calculate the final risk index; in, is the output of the long-term trend channel, i.e., the long-term baseline risk, is the output of the short-term anomaly channel, i.e., the short-term anomaly intensity, is the threshold, where ;in, is the confidence of the Transformer-XL model, is the confidence of the LSTM model, and , , is the entropy of the predicted probability, and p is the original probability value output by the model.

9. The infection risk assessment method for burn care according to claim 8, characterized in that: When responding based on the output of the dynamic risk assessment model, a hierarchical early warning mechanism and closed-loop feedback management approach are adopted, specifically including: S41: Dynamic threshold adjustment, which adjusts the threshold according to the medication status and the patient's immune system status; S42: Grading warnings: Classify warnings into medium-risk warnings and high-risk warnings, use corresponding warning notification methods to issue warnings, and automatically generate disposal suggestions for high-risk warnings; S43: Closed-loop feedback management, comparing indicator changes 24 hours after the warning, evaluating the effectiveness of interventions, and recording high-risk cases without infection for iterative model optimization.

10. An infection risk assessment system for burn care, used in the infection risk assessment method for burn care according to any one of claims 1 to 9, characterized in that: Includes: Data acquisition module, used to obtain full-dimensional clinical data of burn patients in real time; Data processing module, used to convert raw data into standardized features that can be analyzed; Transformer model module, used to integrate temporal dynamics and spatial distribution features for infection risk assessment; The output and warning module is used to convert the prediction results into executable clinical instructions.

Images