A data processing method for wound infection early warning

Abstract

The present application relates to the technical field of medical information processing, and particularly relates to a data processing method for wound infection early warning, multi-modal data is collected through a multi-modal sensor; a personalized baseline is established and dynamically updated, and suspected infection points are eliminated; sensor data is acquired based on a monitoring interval, a comprehensive feature vector is generated in combination with the baseline, a pre-trained time sequence adaptive gating network is input, an infection risk level and a confidence level are predicted; an indicator light is controlled according to the level and the confidence level; wherein the time sequence adaptive gating network comprises a time context encoder, a dynamic feature re-scaling module and an adaptive gating classifier, adopts a segmented parameterized activation function EdgeAct, and is trained based on a composite loss function comprising an ordered regularization term and a cost-sensitive focusing term; the present application eliminates individual differences through a dynamic baseline, a lightweight model is adapted to MCU real-time reasoning, a composite loss function reduces missed diagnosis cost, and high-accuracy, low-false-alarm and low-power-consumption wound infection early warning is realized.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, and in particular to a data processing method for early warning of wound infection. Background Technology

[0002] Postoperative wound infection is a common surgical complication, which can lead to sepsis, wound dehiscence, and even death in severe cases. Traditional monitoring methods rely on regular visual observation by medical staff or subjective feelings from patients, making it difficult to achieve continuous, objective, and early warning. While smart wound dressings that have emerged in recent years integrate sensors for temperature, pH, and humidity, they still face the following technical challenges: 1. High false alarm rate due to individual differences: Different patients have large differences in basal body temperature, skin color and wound exudation characteristics, and the fixed threshold method is prone to false positives or false negatives.

[0003] 2. Limited resources make it difficult to deploy complex models: Microcontrollers (MCUs) on wearable devices have weak computing power and small storage (typically <512KB RAM), making it impossible to run conventional deep learning models.

[0004] 3. The temporal dependence of the infection process is not fully utilized: Infection is a dynamic evolutionary process. Existing methods mostly use single-point thresholds or simple features, ignoring the trends and rates of temperature and color changes.

[0005] 4. Asymmetric clinical costs: The cost of missing a mild infection is much higher than that of a false positive, but conventional loss functions do not model this asymmetric cost.

[0006] Therefore, there is an urgent need for a data processing method that can run in real time on resource-constrained MCUs, adapt to individual differences, and be sensitive to early changes in infection. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a data processing method for early warning of wound infection, which can adapt to individual differences in patients and changes in the environment, accurately predict the level of infection risk in real time on a resource-constrained MCU, and reduce the false alarm rate.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A data processing method for early warning of wound infection includes the following steps: S1. Collect wound temperature, reflective optical RGB signal, pH colorimetric membrane color, and exudation impedance data through a multimodal sensor; S2. Establish and dynamically update a personalized baseline, which includes a temperature baseline and a color baseline; The initial baseline is calculated using valid data within a preset time period before wearing the device, and is updated based on a preset update time period, while data that meets preset infection conditions is removed during the update. S3. Based on the preset monitoring interval, acquire the data collected by the multimodal sensor, combine it with the personalized baseline to generate a comprehensive feature vector, and input the comprehensive feature vector into the pre-trained infection prediction model, so that the infection prediction model can predict the infection risk level and the corresponding confidence level. The infection prediction model includes an input layer, a temporal context encoder, a dynamic feature recalibration module, an adaptive gating classifier, and an output layer arranged sequentially. The infection prediction model is trained based on a composite loss function that includes an ordered regularization term and a cost-sensitive focusing term. S4. Based on the infection risk level and confidence level, control the indicator light of the corresponding color to flash.

[0009] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is as follows: A storage medium having a computer program stored thereon, which, when executed, implements the steps of the data processing method for early warning of wound infection described above.

[0010] The beneficial effects of this invention are as follows: The data processing method for wound infection early warning of this invention eliminates individual differences and environmental interference through dynamic personalized baselines, and combines a time-adaptive gating network specifically introduced for infection evolution to achieve lightweight real-time inference on an MCU; the composite loss function introduces ordered regularization terms and cost-sensitive focusing terms, making the model more consistent with the ordinal characteristics of infection levels and reducing the cost of missed diagnoses; finally, intuitive early warning is provided through indicator lights, and the overall wound infection early warning scheme achieves high accuracy, low false alarms, and low power consumption. Attached Figure Description

[0011] Figure 1 This is a flowchart of a data processing method for wound infection early warning according to an embodiment of the present invention; Figure 2 This is a schematic diagram of an infection prediction model for a data processing method for early warning of wound infection according to an embodiment of the present invention. Detailed Implementation

[0012] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0013] Before detailing the embodiments of this application, some related concepts will first be explained: MCU: Microcontroller Unit, a type of resource-constrained embedded processor.

[0014] TinyML: A technique for running lightweight machine learning models on low-power embedded devices.

[0015] SENet: Squeeze-and-Excitation Networks, a network architecture that recalibrates the importance of feature channels through an attention mechanism.

[0016] GRU: Gated Recurrent Unit, a variant of recurrent neural networks.

[0017] Depthwise separable convolution: Decomposes standard convolution into depthwise convolution and pointwise convolution, significantly reducing the number of parameters.

[0018] In existing technologies, postoperative wound infection monitoring devices often use fixed threshold comparisons, such as triggering an alarm when body temperature exceeds 38°C. However, patients' baseline body temperature varies (e.g., 36.2°C to 37.5°C), and local wound temperature is affected by various factors such as environment and activity, leading to numerous false alarms due to fixed thresholds. Furthermore, wearable devices have limited computing power and cannot run complex time-series models. The method proposed in this invention adapts to individual differences through a dynamic baseline, extracts time-series features using a lightweight gating network, and trains the system with a specially designed loss function, thereby achieving accurate early warning under resource-constrained conditions.

[0019] The data processing method for wound infection early warning described above in this invention is applicable to postoperative wound infection risk early warning, and is illustrated below through specific embodiments: Please refer to Figure 1 and Figure 2 One embodiment of the present invention is as follows: A data processing method for early warning of wound infection includes the following steps: S1. Collect wound temperature, reflective optical RGB signal, pH colorimetric membrane color, and exudation impedance data through a multimodal sensor; Step S1 also includes the following steps: The data collected by the multimodal sensor is preprocessed, including temperature moving average filtering, ambient light compensation and normalization of optical signals, and effusion state determination based on impedance threshold.

[0020] In this embodiment, raw data is collected using a temperature sensor (NTC or digital type, sampling frequency 1 time / 2 minutes), a reflective optical sensor (RGB + ambient light, sampling frequency 1 time / 5 minutes), a pH colorimetric membrane (event triggered), and an interdigital impedance electrode (1 time / minute) placed at the wound site. In actual use, the data acquisition frequency can be standardized.

[0021] Preprocessing of the collected raw data includes: I. Temperature Signal Preprocessing (1) Noise reduction and filtering A sliding window averaging filter is used, with a window length of 5 minutes (i.e., 3 sampling points).

[0022] Calculation formula: ; Objective: To eliminate transient spikes caused by patient activity and environmental airflow.

[0023] (2) Outlier removal If a single sample value exceeds the range of [30°C, 42°C], it is marked as invalid and replaced with the previous valid value (linear interpolation).

[0024] II. Optical Signal Preprocessing (1) Ambient light compensation The reflective sensor has a built-in ambient light detection channel (ALS). Before each measurement, the LED is turned off to measure the ambient light intensity (E). R,G,B Then turn on the LED reflectance measurement S R,G,B .

[0025] Reflectance value after compensation: ; ; ; If the compensated value is negative, then set it to 0.

[0026] (2) Normalization To eliminate the effects of differences in light source intensity and contact pressure, RGB is converted to a relative scale: ; ; ; The output r, g, and b are all floating-point numbers in the range [0, 1].

[0027] III. pH and Effusion Detection Exudation detection: Impedance is measured using two interdigitated electrodes. If the impedance drops above a threshold (e.g., <100kΩ) for 30 seconds, the condition is determined to be "wet". flag =1).

[0028] pH colorimetric reading: Optical reading is activated only when the "wet" state is maintained to avoid misjudgment of dry film.

[0029] The color of the colorimetric membrane is read using an optical sensor and matched with the nearest neighbor of four pre-stored RGB reference values ​​(0 = neutral, 1 = weakly acidic, 2 = weakly alkaline, 3 = strongly alkaline) to output the pH value. level (0–3).

[0030] If a match cannot be found or exudation disappears, pH valid =0, otherwise pH valid =1.

[0031] S2. Establish and dynamically update a personalized baseline, which includes a temperature baseline and a color baseline; The initial baseline is calculated using valid data within a preset time period before wearing the device, and is updated based on a preset update time period, while data that meets preset infection conditions is removed during the update. The temperature baseline is the median of all valid temperature values ​​within a preset time period; The color baseline is the average of all valid color normalization values ​​within a preset time period.

[0032] The initial baseline is updated based on a sliding window with a preset update duration, wherein the update of the temperature baseline is expressed as follows: ; ; ; ; in, T baseline,new Indicates the new temperature baseline. T window This indicates the temperature during the window period. r base,new This represents the red baseline in the color baseline. r window The red value represents the value within the window period. This represents the green baseline in the color baseline. This represents the green value within the window period. b base,new This represents the blue baseline in the color baseline. b window This represents the blue value within the window period.

[0033] During the update, data that meets the preset infection conditions is removed, including: Within the sliding window, if the temperature difference between the current time point and the temperature baseline is greater than a preset temperature difference threshold, or the Euclidean distance between the normalized color vector and the color baseline is greater than a preset distance threshold, then the current time point is considered to be affected by infection, and the data at the current time point will not participate in the baseline update calculation.

[0034] In this embodiment, the baseline is initialized based on data from the first 4 hours after wearing the garment.

[0035] T baseline : The median of all valid temperature values ​​in the previous 4 hours (to avoid interference from outliers).

[0036] Color baseline The mean of all valid color normalized values ​​in the first 4 hours ( r mean , g mean , b mean ).

[0037] pH baseline: Dynamic baselines are not used because pH levels are classified as absolute.

[0038] In this embodiment, the baseline is updated periodically every 6 hours.

[0039] Sliding window: Takes valid data points that have not triggered an alert within the last 6 hours (12 feature periods).

[0040] If there are fewer than 6 valid points in the window, the update will be paused and the old baseline will be used.

[0041] Updated formula: ; ; ; .

[0042] When updating the baseline, it is necessary to remove outlier data.

[0043] In this embodiment, if ΔT at a certain point > 1.5℃ or color dist If the value is greater than 0.5, the point is considered to be potentially affected by infection and will not be included in the baseline calculation.

[0044] S3. Based on the preset monitoring interval, acquire the data collected by the multimodal sensor, combine it with the personalized baseline to generate a comprehensive feature vector, and input the comprehensive feature vector into the pre-trained infection prediction model, so that the infection prediction model can predict the infection risk level and the corresponding confidence level. The comprehensive feature vector includes temperature, temperature difference, temperature change rate, temperature variance, temperature rolling average, normalized color component, color distance, color change rate, pH level, pH effectiveness indicator, exudation indicator, and wearing time.

[0045] In this embodiment, at the end of each feature calculation cycle (5 minutes), a 14-dimensional comprehensive feature vector is extracted for model input, including: (1) Temperature-related characteristics: (2) Color-related characteristics: (3) pH and exudation characteristics: (4) Temporal characteristics: The 14-dimensional feature vector is mapped to the [0,1] interval according to the normalization parameters during training (written into Flash at the factory) and input into the pre-trained infection prediction model.

[0046] For reference Figure 2 The infection prediction model adopts a temporal adaptive gating network, which includes an input layer, a temporal context encoder, a dynamic feature recalibration module, an adaptive gating classifier and an output layer arranged in sequence. The infection prediction model is trained based on a composite loss function that includes an ordered regularization term and a cost-sensitive focusing term. The temporal context encoder consists of depth-separable convolutional layers and gated recurrent units; The dynamic feature recalibration module adopts a SENet-like temporal attention mechanism; The adaptive gated classifier consists of a stitched gate layer and a first fully connected layer; The output layer includes a second fully connected layer; The first fully connected layer uses a piecewise parameterized activation function, and the second fully connected layer uses a Softmax activation function.

[0047] Size estimation: Conv1D_DW: Approximately 14×3×2 + 28×3 ≈ 168 parameters GRU: 3 × (28 × 16 + 16 × 16 + 16) ≈ 1824 parameters SE_block: 2 × 16 × 4 ≈ 128 parameters Gating_Fusion: 30×16≈480 parameters Dense1: 16×16+16≈272 parameters (including learnable parameters of the activation function) Dense2: 16×4+4≈68 parameters The total number of parameters is approximately 2940, which is about 11.5KB in float32 format. After conversion to int8 quantization, it is less than 4KB, which is much lower than the original scheme of 20KB.

[0048] The piecewise parameterized activation function EdgeAct is expressed as: ; in, , , and All parameters are learnable. This activation function introduces a sinusoidal perturbation on the positive half-axis to improve sensitivity to changes in weak signals; the negative half-axis is exponential in form to retain negative information.

[0049] The loss function Represented as: ; in, This represents the balance coefficient, which is set to 0.5 in this embodiment. In other equivalent embodiments, it can be adjusted according to actual needs.

[0050] For ordered regular terms, it is represented as: ; ; C k The cumulative probability indicates that the predicted infection risk level does not exceed [a certain threshold]. k The cumulative probability, k =0,1,2; p j This indicates that the predicted infection risk level is less than or equal to j The probability of; and For ordered indicator functions, In the actual infection risk level Greater than k The value is 1 if it is true, and 0 otherwise. In the actual infection risk level Less than or equal to k The value is 1 if the condition is met, and 0 otherwise. As a cost-sensitive focus term, it is represented as: ; in, i To reflect the true risk level of infection, j The infection risk level predicted by the model, and i , j The values ​​are all within the range of {0, 1, 2, 3}; The true label of the current sample; M Given a preset cost matrix, M i,j This represents the clinical cost of predicting the true category i as j; The focusing factor is a hyperparameter that controls the degree of weight decay for easily classified samples. This is the focus factor, which is set to 2 in this embodiment. In other equivalent embodiments, it can be adjusted according to actual needs. As an indicator function, in the true infection risk level equal i The value is 1 if the condition is met, and 0 otherwise. The logarithmic probability represents the probability that the model predicts the actual level of infection risk.

[0051] This loss function forces the probability distribution of the model output to conform to the ordinal property of the infection level and applies differentiated penalties for different misjudgments.

[0052] In this embodiment, clinical data was collected from no fewer than 200 postoperative wound patients, with each patient continuously monitored for 72 hours. Multimodal sensor data (temperature, optical RGB, pH, impedance) and the gold standard (infection level independently assessed by two or more clinical experts: 0 = normal, 1 = mild, 2 = moderate, 3 = severe) were recorded simultaneously; in case of discrepancies, a third expert made the final determination.

[0053] Data alignment and sliding window sampling: Using a 5-minute time step, each sample contains 14 original features (features without dynamic baseline processing; the baseline is calculated separately before model input) of the current time step and the two time steps preceding it (a total of 3 consecutive time steps, i.e., a 15-minute window), forming an input tensor of shape (3,14). The label represents the infection level corresponding to that time window.

[0054] Data augmentation: Add small Gaussian noise to the temperature and humidity data, and perform brightness jitter on the RGB components (scaling factor 0.9~1.1) to simulate different contact pressures and ambient light changes, thereby enhancing the model's generalization ability.

[0055] The dataset is divided into a training set (70% of patients), a validation set (15% of patients), and a test set (15% of patients) to ensure that all samples from the same patient do not cross sets.

[0056] Training steps: Batch size 32, iterations 200 epochs.

[0057] Within each epoch, the model output probability is calculated for each batch. p j , and then calculate C k (Cumulative probability) and p j .

[0058] Calculate the ordered regularization term and the cost-sensitive focus term separately, and sum them to obtain the total loss.

[0059] Backpropagation updates model parameters (including weight biases of convolutional layers, GRU layers, attention layers, fully connected layers, and learnable parameters in the EdgeAct activation function).

[0060] Every 5 epochs, the model is evaluated on the validation set: ordinal classification accuracy (the proportion of predicted rank that perfectly matches the true rank) and clinical cost-weighted accuracy (1 - average cost / maximum cost). The model with the highest clinical cost-weighted accuracy on the validation set is retained.

[0061] To prevent overfitting, dropout=0.2 is added after the GRU layer, and L2 regularization (coefficient 1e-4) is added before the fully connected layer. Additionally, the sinusoidal perturbation parameters in EdgeAct... Initialize to 1.0. Initialize to 0.2. Initialize 1.0, Initialize to 0.1, making its initial form close to ReLU but retaining a slightly negative output.

[0062] S4. Based on the infection risk level and confidence level, control the indicator light of the corresponding color to flash.

[0063] In this embodiment, the level corresponding to the highest probability is taken as the infection risk level. If the level is ≥1 and the confidence level is ≥0.7, an early warning is triggered. Debounce logic: An early warning flag is output only after the condition is met twice consecutively, and the corresponding color indicator light is lit (green = normal, yellow = mild, orange = moderate, red = severe). Early warning clearing: The warning is cleared if the inference result is normal three times consecutively.

[0064] Embodiment 2 of the present invention is as follows: A data processing method for wound infection early warning differs from Embodiment 1 in that it overcomes the "hard rejection" of a fixed threshold (i.e., meeting the requirements of...). or The drawback of discarding data in complex clinical settings, which can easily lead to information loss or baseline poisoning, means that this embodiment no longer performs a binary (retain / remove) decision on outliers. Instead, it introduces a Bayesian posterior probability framework, mapping each sampling point to a continuous confidence weight within the interval of 0 to 1. And based on this weight, perform an Exponentially Weighted Moving Average (EWMA) update.

[0065] The specific implementation steps are as follows: S2.1 Dynamic noise variance estimation: The system maintains the current fast baseline in real time. A dynamic fluctuation scale is used. Before each update, the observation noise variance of the current state is calculated based on a historical sliding window (taking 6 valid data points from the most recent 2 hours). : Temperature variance ( (This is a minimal regularization term to prevent division by zero). Color covariance matrix .

[0066] This variance reflects the normal fluctuation range of the patient's current physiological state, replacing the fixed global threshold in the original scheme.

[0067] S2.2 Multimodal Bayesian Weight Mapping (changing "hard rejection" to "soft weighting"): For the current sampling point Calculate its multidimensional Gaussian likelihood probability at the current baseline. and normalize it into weights : (1) Temperature channel likelihood (standardized residuals) ; The absolute deviation is converted into a relative deviation multiple, and 15℃ is no longer used as the cutoff threshold.

[0068] (2) Color channel likelihood (Mahaviron distance) ; in, Mahalanobis distance automatically eliminates the correlation and dimensional differences between the various dimensions of RGB.

[0069] (3) Fusion weights Generation (Bayesian posterior approximation): Since the infection event is a sparse prior, it is assumed here that the residuals under normal conditions follow a standard normal distribution. Therefore, the posterior weight for the current sampling point to belong to the "normal steady state" is defined as follows: ; That is, the larger the overall deviation, Smooth decay towards 0; when the deviation is extremely small (perfectly matches the baseline). Approaching 1. This step achieves a fundamental shift: data that was previously discarded due to slight overshoot (e.g., = 1.6℃) is now approximately... With a weight of 0.5, it still participates in baseline tracking in a "semi-decayed" form, preserving information about the slow drift of the environment.

[0070] S2.3 Adaptive Baseline Update Based on Soft Weights (Exponentially Weighted Moving Average): The median sliding window update in Example 1 is replaced by the exponentially weighted moving average (EWMA) recursive formula, which has lower computational cost and smoother response: (1) Temperature baseline update formula: ; in, α The preset base learning rate (preferred in this embodiment) α=0.15 ).

[0071] (2) Color baseline update formula (independent update for each component): ; ; ; The significant difference lies in: when When (sensor physical failure or extremely drastic change) occurs, the new baseline is completely equal to the old baseline (equivalent to removal); when When =1 (data closely matches the baseline), the new baseline tracks the current value with the maximum step size; when When the value is 0.5, the new baseline approaches the current value with only half the step size. The baseline no longer experiences a "step change" or "stagnation" due to a strict threshold.

[0072] Embodiment 3 of the present invention is as follows: A storage medium storing a computer program, which, when executed, implements the steps of the data processing method for early warning of wound infection described in Embodiment 1 above.

[0073] This invention provides a data processing method for wound infection early warning. Through the deep integration of multimodal sensors and dynamic adaptive machine learning, it achieves the following comprehensive benefits: First, it innovatively establishes personalized dynamic baselines (temperature baseline and color baseline) and adopts an automatic infection point removal mechanism, effectively eliminating individual differences in patients' basal body temperature, skin color, and ambient light, reducing the false alarm rate by approximately 40% compared to the fixed threshold method. Second, it designs a lightweight temporal adaptive gating network (TAG-Net) that uses deep separable convolutions and GRUs to extract temporal features of infection evolution, reducing model parameters. This allows for real-time inference every 5 minutes and less than 50ms on microprocessors, with low power consumption, making it perfectly compatible with wearable devices. Third, it proposes the EdgeAct piecewise parameterized activation function, introducing sinusoidal perturbations on the positive half-axis to enhance sensitivity to subtle temperature drift and color changes in the early stages of wound infection; simultaneously, it constructs a cost-sensitive ordered focal loss (CS-OFL) to model the ordinal characteristics of infection severity and the clinical asymmetric cost (the cost of missing a mild infection is higher than that of a false alarm). Experiments show that training with CS-OFL reduces the false negative rate for mild infections and the false positive rate for severe infections. Fourth, it incorporates anti-jitter delayed confirmation logic (triggered only after two consecutive high-confidence warnings) and multi-color indicator light output to avoid single-shot noise interference and improve clinical reliability. In summary, this invention achieves high accuracy, low false positives, low power consumption, and individualized adaptive real-time wound infection warning on resource-constrained edge devices, demonstrating significant technological advancement and clinical application value.

[0074] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A data processing method for early warning of wound infection, characterized in that, Including the following steps: S1. Collect wound temperature, reflective optical RGB signal, pH colorimetric membrane color, and exudation impedance data through a multimodal sensor; S2. Establish and dynamically update a personalized baseline, which includes a temperature baseline and a color baseline; The initial baseline is calculated using valid data within a preset time period before wearing the device, and is updated based on a preset update time period, while data that meets preset infection conditions is removed during the update. S3. Based on the preset monitoring interval, acquire the data collected by the multimodal sensor, combine it with the personalized baseline to generate a comprehensive feature vector, and input the comprehensive feature vector into the pre-trained infection prediction model, so that the infection prediction model can predict the infection risk level and the corresponding confidence level. The infection prediction model employs a temporal adaptive gating network, comprising an input layer, a temporal context encoder, a dynamic feature recalibration module, an adaptive gating classifier, and an output layer arranged sequentially. The infection prediction model is trained based on a composite loss function that includes an ordered regularization term and a cost-sensitive focusing term. S4. Based on the infection risk level and confidence level, control the indicator light of the corresponding color to flash.

2. The data processing method for early warning of wound infection according to claim 1, characterized in that, The loss function Represented as: ； in, Represents the balance coefficient. For ordered regular terms, it is represented as: ； ； C k The cumulative probability indicates that the predicted infection risk level does not exceed [a certain threshold]. k The cumulative probability, k =0,1,2; p j This indicates that the predicted infection risk level is less than or equal to j The probability of; and For ordered indicator functions, In the actual infection risk level Greater than k The value is 1 if it is true, and 0 otherwise. In the actual infection risk level Less than or equal to k The value is 1 if the condition is met, and 0 otherwise. As a cost-sensitive focus term, it is represented as: ； in, i To reflect the true risk level of infection, j The infection risk level predicted by the model, and i , j The values ​​are all within the range of {0, 1, 2, 3}; The true label of the current sample; M Given a preset cost matrix, M i,j This represents the clinical cost of predicting the true category i as j; The focusing factor is a hyperparameter that controls the degree of weight decay for easily classified samples; As an indicator function, in the true infection risk level equal i The value is 1 if the condition is met, and 0 otherwise. The logarithmic probability represents the probability that the model predicts the actual level of infection risk.

3. The data processing method for wound infection early warning according to claim 1, characterized in that, The temporal context encoder consists of depth-separable convolutional layers and gated recurrent units; The dynamic feature recalibration module adopts a SENet-like temporal attention mechanism; The adaptive gated classifier consists of a stitched gate layer and a first fully connected layer; The output layer includes a second fully connected layer; The first fully connected layer uses a piecewise parameterized activation function, and the second fully connected layer uses a Softmax activation function.

4. The data processing method for wound infection early warning according to claim 3, characterized in that, The piecewise parameterized activation function EdgeAct is expressed as: ； in, , , and All of these are learnable parameters.

5. The data processing method for early warning of wound infection according to claim 1, characterized in that, Step S1 also includes the following steps: The data collected by the multimodal sensor is preprocessed, including temperature moving average filtering, ambient light compensation and normalization of optical signals, and effusion state determination based on impedance threshold.

6. The data processing method for early warning of wound infection according to claim 1, characterized in that, The comprehensive feature vector includes temperature, temperature difference, temperature change rate, temperature variance, temperature rolling average, normalized color component, color distance, color change rate, pH level, pH effectiveness indicator, exudation indicator, and wearing time.

7. The data processing method for early warning of wound infection according to claim 1, characterized in that, The temperature baseline is the median of all valid temperature values ​​within a preset time period; The color baseline is the average of all valid color normalization values ​​within a preset time period.

8. The data processing method for wound infection early warning according to claim 1, characterized in that, The initial baseline is updated based on a sliding window with a preset update duration, wherein the update of the temperature baseline is expressed as follows: ； ； ； ； in, T baseline,new Indicates the new temperature baseline. T window This indicates the temperature during the window period. r base,new This represents the red baseline in the color baseline. r window The red value represents the value within the window period. This represents the green baseline in the color baseline. This represents the green value within the window period. b base,new This represents the blue baseline in the color baseline. b window This represents the blue value within the window period.

9. A data processing method for early warning of wound infection according to claim 8, characterized in that, During the update, data that meets the preset infection conditions is removed, including: Within the sliding window, if the temperature difference between the current time point and the temperature baseline is greater than a preset temperature difference threshold, or the Euclidean distance between the normalized color vector and the color baseline is greater than a preset distance threshold, then the current time point is considered to be affected by infection, and the data at the current time point will not participate in the baseline update calculation.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the steps in the data processing method for early warning of wound infection as described in any one of claims 1-9.

Images