A wound healing degree evaluation method and system based on multi-data analysis

Abstract

The present invention discloses a wound healing degree assessment method and system based on multi-data analysis, which relates to the field of multimodal data analysis technology. The method comprises the following steps: collecting a wound-related multimodal dataset, constructing a cross-modal extraction model and inputting the multimodal dataset into the model to obtain a preliminary feature set; processing the preliminary feature set using dynamic time warping technology to obtain a cross-modal feature set; constructing a causal relationship graph based on the cross-modal feature set to identify intervention paths, and correcting causal effects and eliminating false correlations through a backdoor adjustment method to obtain a causal purified cross-modal feature sequence; inputting the causal purified cross-modal feature sequence into an established wound healing prediction model, and outputting a wound healing prediction result based on the time series characteristics; and generating a real-time wound visualization report and providing feedback based on the wound healing prediction result. The present invention achieves a breakthrough from a single time point to continuous assessment through the wound healing prediction model.

Description

Technical Field

[0001] The present invention relates to the technical field of multimodal data analysis, and in particular to a wound healing degree evaluation method and system based on multi-data analysis. Background Art

[0002] Wound healing assessment, a key research area in medicine, has made significant progress in recent years with advances in multimodal data acquisition and analysis. Traditionally, wound healing assessment relies primarily on clinical observation and analysis of single data sets, such as evaluating changes in wound appearance through professional observation by a physician or quantifying healing progress by measuring wound area. With advancements in imaging technology, high-resolution RGB images combined with infrared thermal imaging have become widely used for wound boundary extraction and temperature distribution analysis.

[0003] While significant progress has been made in mainstream wound healing assessment, many shortcomings remain. First, mainstream wound healing assessments rely heavily on correlation analysis to assess relationships between variables, ignoring spurious correlations caused by confounding factors, which can lead to misjudgments of factors influencing healing. Furthermore, they lack the ability to purify causal relationships and dynamically predict multimodal features, making it difficult to reveal intervention pathways and providing continuous healing score curves and real-time feedback. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a wound healing degree evaluation method based on multi-data analysis to solve the problems of ignoring confounding factors and lacking causal purification and dynamic prediction capabilities of multimodal features.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a method for evaluating wound healing degree based on multi-data analysis, which comprises:

[0008] Collect wound-related multimodal datasets, build a cross-modal extraction model and input the multimodal dataset to obtain a preliminary feature set;

[0009] The preliminary feature set is processed using dynamic time warping technology to obtain a cross-modal feature set;

[0010] Based on the cross-modal feature set, a causal relationship graph is constructed to identify intervention paths. The causal effect is corrected through the backdoor adjustment method to eliminate false correlations and obtain a causal purified cross-modal feature sequence.

[0011] The causal purification cross-modal feature sequence is input into the established wound healing prediction model, and the wound healing prediction results are output based on the time series characteristics;

[0012] Generate real-time wound visualization reports and provide feedback based on wound healing prediction results.

[0013] As a preferred embodiment of the wound healing degree evaluation method based on multi-data analysis described in the present invention, the wound-related multimodal dataset includes wound image features, microenvironment time series, biomarker sparse sequence and confusion factor encoding.

[0014] As a preferred embodiment of the wound healing degree evaluation method based on multi-data analysis of the present invention, a cross-modal extraction model is constructed and a multi-modal data set is input to obtain a preliminary feature set, which specifically includes the following steps:

[0015] Convolutional neural networks and fully connected neural networks are used as the basic networks, and the construction of the cross-modal extraction model is completed after training;

[0016] Input the multimodal dataset into the constructed cross-modal extraction model and output a preliminary feature set.

[0017] As a preferred embodiment of the wound healing degree evaluation method based on multi-data analysis of the present invention, wherein: obtaining the cross-modal feature set refers to generating an alignment path using Euclidean distance and dynamic programming formula;

[0018] Gaussian process regression was used to interpolate the null values ​​of biomarker sparse sequences, and the microenvironment time series and confounding factor encoding were integrated into a cross-modal feature set.

[0019] As a preferred solution of the wound healing degree evaluation method based on multi-data analysis described in the present invention, wherein: a causal relationship diagram is constructed based on a cross-modal feature set to identify intervention paths, and the causal effect is corrected by a backdoor adjustment method to eliminate false correlations, thereby obtaining a causal purified cross-modal feature sequence, which specifically includes the following steps:

[0020] Use the Pearson correlation coefficient to calculate the correlation of cross-modal feature sets and construct a causal relationship diagram;

[0021] The causal effect is corrected by a backdoor adjustment formula to remove the spurious correlation of the confounding factor vector, and the sparse sequence of biomarkers is adjusted using linear regression to generate a causal purified cross-modal feature sequence.

[0022] As a preferred solution of the wound healing degree evaluation method based on multi-data analysis described in the present invention, the established wound healing prediction model is based on a long short-term memory network and is trained through historical data sets.

[0023] As a preferred embodiment of the wound healing degree evaluation method based on multi-data analysis of the present invention, the method of outputting the wound healing prediction result according to the time series characteristics refers to inputting the causal purification cross-modal feature sequence into the wound healing prediction model, using the time series characteristics to capture the temporal dependency, and outputting a continuous healing score curve;

[0024] The causal targets are analyzed in combination with the causal effect correction results, and the impact ratio is calculated based on the contribution of the confounding factor vector to generate a confounding factor impact report.

[0025] In a second aspect, the present invention provides a wound healing degree evaluation system based on multi-data analysis, comprising:

[0026] The feature module collects wound-related multimodal datasets, builds a cross-modal extraction model, and inputs the multimodal dataset to obtain a preliminary feature set;

[0027] The adjustment module processes the preliminary feature set using dynamic time warping technology to obtain a cross-modal feature set;

[0028] The causal module constructs a causal relationship graph based on the cross-modal feature set to identify the intervention path, and corrects the causal effect through the backdoor adjustment method to eliminate false correlations and obtain a causal purified cross-modal feature sequence;

[0029] The prediction module inputs the causal purification cross-modal feature sequence into the established wound healing prediction model and outputs the wound healing prediction results based on the time series characteristics;

[0030] The visualization module generates real-time wound visualization reports and provides feedback based on the wound healing prediction results.

[0031] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the wound healing degree evaluation method based on multi-data analysis as described in the first aspect of the present invention is implemented.

[0032] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the wound healing degree evaluation method based on multi-data analysis as described in the first aspect of the present invention.

[0033] The beneficial effects of the present invention are as follows: by constructing a causal graph, identifying intervention paths, and revealing the causal chain between confounding factors and healing, the scientific nature and explanatory power of causal inference are improved; the causal effect is corrected by the backdoor adjustment formula to reveal the true effect of the variable, providing a quantitative basis for data purification, and achieving the effect of enhancing the accuracy and credibility of the evaluation results; and generating a causal purification cross-modal feature sequence through linear regression processing, eliminating the interference of confounding factors, and providing feature input without false correlation; in addition, by processing the causal purification cross-modal feature sequence through a wound healing prediction model established based on LSTM, the dynamic changes are captured by using the characteristics of the time series, which can improve the accuracy and dynamics of healing process monitoring and achieve a breakthrough from a single time point to continuous evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0035] Figure 1 Flowchart of the wound healing degree evaluation method based on multi-data analysis in Example 1.

[0036] Figure 2 This is an architectural diagram of the wound healing degree evaluation system based on multi-data analysis in Example 1.

[0037] Figure 3 Schematic diagram of cross-modal feature processing in Example 1.

[0038] Figure 4 This is the causal relationship correction and prediction model diagram in Example 1. DETAILED DESCRIPTION

[0039] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0040] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0041] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0042] Example 1, reference Figures 1 to 4 This embodiment provides a wound healing degree evaluation method based on multi-data analysis, comprising the following steps:

[0043] S1. Collect wound-related multimodal datasets, build a cross-modal extraction model and input the multimodal dataset to obtain a preliminary feature set.

[0044] The specific steps include:

[0045] S1.1. First, a multimodal wound dataset was collected. The specific process was as follows: RGB images of the wound were captured daily using a high-definition camera, and infrared thermal images of the wound were captured daily using an infrared lens. The acquisition time was set at 8:00 AM each day to ensure consistent lighting and environmental conditions. The HD camera and infrared lens were fixed to the same device using a clamp, and the lens focus distance was adjusted to the same distance to ensure that the RGB images and infrared thermal images covered the same wound area. After capture, the RGB images and infrared thermal images were saved as image data. The RGB images were converted to grayscale using a standard brightness method. Gaussian blur was then applied to the grayscale images to remove noise. The Canny edge detection algorithm was used with a low threshold of 50 and a high threshold of 150 (the pixel gradient range is 0-255; the specific threshold extraction range can be customized). The wound boundary was extracted. A contour detection algorithm was used to extract the preliminary wound area and obtain the wound boundary coordinates, for example, [(100, 200), (101, 201), ..., (150, 220)], representing the polygonal vertices of the wound edge. Then, the final wound area is calculated using Green's formula based on the wound boundary coordinates, and the pixel unit area is output. After the pixel unit area is obtained, it needs to be converted to a physical unit area. The conversion requires a proportional coefficient, which is determined based on the shooting distance and resolution. Assuming the shooting distance is 10 cm, the RGB image resolution is 1920 × 1080 pixels, and the field of view width is about 19.2 cm, the width of each pixel is calculated first. The expression is:

[0046] ;

[0047] Convert pixel width to physical unit area, the expression is:

[0048] ;

[0049] in, Indicates the physical unit area (unit: cm 2 ), Represents the unit area of ​​pixels.

[0050] For example, if the pixel unit area 5000 pixels 2 ,but =5000×(0.01) 2 =5000×0.0001=0.5cm 2 .

[0051] The preprocessed RGB image outputs the wound boundary coordinates and the physical unit area of ​​the wound (e.g., 0.5 cm²). For infrared thermal imaging, linear normalization is performed to output a normalized temperature distribution (ranging from 0 to 1, e.g., 0.75 corresponds to 30°C). This completes wound image acquisition and preprocessing, outputting wound image features including wound boundary coordinates, wound area, and normalized temperature distribution.

[0052] A smart bandage records wound microenvironmental data from midnight to 11:00 PM daily, recording temperature (°C) and humidity (%RH) every hour. The bandage connects to a smartphone via Bluetooth, automatically transmitting hourly data to a cloud server to ensure real-time data. After collection, a sliding average filter is applied to the temperature and humidity data to remove noise and output smoothed values ​​(e.g., the temperature at a given hour is smoothed from 32.1°C to 31.8°C). Twenty-four smoothed data points are generated daily to form a microenvironmental time series. The output microenvironmental time series includes smoothed temperature series (e.g., [31.8, 32.0, 31.9, ...]°C) and humidity series (e.g., [65.2, 64.8, 65.0, ...]%RH).

[0053] Biomarker data are collected weekly using a portable hematology analyzer (e.g., the i-STAT handheld analyzer). Data collection includes CRP concentration (in mg / L), IL-6 concentration (in pg / mL), and white blood cell count (in 10^9 / L). Each collection is performed via fingerstick blood draw, with the results transferred to a smartphone via a USB port and then uploaded to a cloud server. After collection, biomarker data are normalized and output as normalized values ​​(e.g., CRP is normalized from 5 mg / L to 0.05). This generates a sparse sequence of biomarkers, including normalized CRP concentrations, IL-6 concentrations, and white blood cell counts (e.g., [0.05, 0.01, 0.65]).

[0054] Confounding factor data were collected through an electronic questionnaire. Items included age (in years), medication history (e.g., antibiotic use, recorded as yes or no), and diabetes duration (in years). After collection, the confounding factor data were discretely coded: age was categorized as 0-30, 31-60, and 61+ (e.g., 45 years old was coded as 31-60); medication history was coded as 0 (no) or 1 (yes); and diabetes duration was categorized as 0-5, 6-10, and 11+ (e.g., 8 years was coded as 6-10). This generated a set of discrete values. The confounding factor codes were then output, including age category, medication history status, and diabetes duration (e.g., [31-60, 1, 6-10]).

[0055] Wound image features (boundary coordinates, wound area, temperature distribution), microenvironment time series (temperature, humidity), and biomarker sparse series (CRP, IL-6, white blood cell count) confounding factor encoding were integrated into a preprocessed multimodal dataset.

[0056] It is further explained that the integration of wound image features, microenvironment time series, biomarker sparse series and confounding factor encoding provides comprehensive information on visual, environmental, physiological and individual differences, enhancing the depth of healing evaluation.

[0057] S1.2. Build a cross-modal extraction model. First, we build the first part of the cross-modal extraction model to process image data (RGB images and infrared thermal images). We chose a convolutional neural network (ResNet-18, with 128 output dimensions) as the basis for processing image data because convolutional neural networks are both efficient and accurate in extracting image features.

[0058] The training process involves downloading the pre-trained weights for ResNet-18 (trained on the ImageNet dataset). This model consists of 18 convolutional layers, including four residual blocks. Each block consists of two convolutional layers with a kernel size of 3×3, and the final layer is a fully connected layer, outputting a 128-dimensional feature vector. RGB images (3 channels) and infrared thermal images (1 channel) are used as input. The input layer is adjusted to accommodate the two data types: the RGB image input remains 3-channel (224×224×3), while the infrared thermal image input is modified to a single channel (224×224×1). The kernel of the first convolutional layer is replaced (from 3 channels to 1 channel), and the weights are initialized to the average of the pre-trained weights. The fusion strategy is to fuse the feature vectors of the RGB and infrared thermal images through a weighted average to generate a unified image feature vector, thus completing the convolutional neural network. For example, the weighted average might be 0.6 for the RGB image and 0.4 for the infrared thermal image, with the RGB image receiving a higher weight due to its structural information. (The specific weight distribution can be customized based on your needs and scenario.)

[0059] Preferably, image feature extraction can extract multimodal features of wound images while maintaining computational efficiency, provide high-quality visual information input for subsequent cross-modal alignment, and enhance the ability to characterize the wound healing status.

[0060] After completing the first part of the cross-modal extraction model, we continued to build the second part of the cross-modal extraction model, which processes the biomarker and microenvironment data. We chose a fully connected neural network (two hidden layers, 64 neurons per layer) as the basis because it is suitable for processing numerical data and has a simple and efficient structure. We defined the input layer as 5-dimensional (3-dimensional biomarker + 2-dimensional microenvironment average), with 64 neurons in the first hidden layer, 64 neurons in the second hidden layer, and a 128-dimensional output layer. This completed the construction of the fully connected neural network. Based on the completed convolutional neural network and fully connected neural network components, we unified the output dimension to 128 dimensions to complete the integration.

[0061] Optimally, fully connected neural networks can mine deep patterns in biomarker and microenvironmental data, such as the potential correlation between inflammation severity and temperature changes. This provides numerical feature inputs consistent with the dimensions of the image feature vectors for subsequent cross-modal alignment, thereby enhancing the comprehensive analysis capabilities of multimodal data. Furthermore, by defining a unified feature output format for integration, computational barriers caused by inconsistent dimensions can be avoided, laying the foundation for a cross-modal self-supervised contrastive learning framework. This approach also places visual information (wound image features) and numerical information (biomarker and microenvironmental features) on an equal footing, promoting semantic alignment of the two modal data and thus improving the cross-modal extraction model's ability to comprehensively represent wound healing-related features.

[0062] Now, let's prepare training data. First, load a historical patient dataset from a cloud database. For example, this dataset contains multimodal data for 1,000 patients: RGB images, infrared thermal images, sparse sequences of biomarkers (CRP concentration, IL-6 concentration, white blood cell count), and microenvironment time series (temperature, humidity). Each patient's multimodal data is organized by time point. For example, Patient A has an RGB image and CRP concentration from Day 1. Pair the image data (RGB image and infrared thermal image) with the biomarker and microenvironment data for the same patient as positive sample pairs. For example, Patient A's image feature vector and biomarker sparse sequence constitute a positive sample pair. Randomly sample biomarker sparse sequences from the data of other patients as negative samples. For example, sample 10 biomarker sparse sequences from Patients B, C, and so on to form a negative sample set (batch size 10). This data preparation allows the cross-modal extraction model to learn to distinguish between data from the same patient (positive samples) and data from different patients (negative samples), providing a foundation for subsequent comparative learning.

[0063] Initial feature vectors were generated using the pre-built convolutional neural network and fully connected neural network components using positive and negative sample pairs. Specifically, the RGB image and infrared thermal image were fed into the convolutional neural network and fused using weighted averaging to form an image feature vector. The biomarker sparse sequence and the daily average of the microenvironment time series (e.g., [31.9, 65.0]) were concatenated into a 5-dimensional vector and fed into the fully connected neural network to generate the biomarker sparse sequence. Data from 1,000 patients were processed batch by batch, with 10 samples per batch, to generate image feature vectors and biomarker sparse sequences. The output initial feature vector set (1,000 pairs of image feature vectors and biomarker sparse sequences, along with their corresponding negative samples) was obtained. For each batch of 10 patient samples (positive and negative sample pairs), the dot product of the positive and negative sample pairs was calculated. The difference in similarity between the positive and negative samples was then quantified using a comparative loss function. Stochastic gradient descent was defined as the optimizer to optimize the parameters of the convolutional neural network and the fully connected neural network. A training threshold is set according to the task requirements (it can be set to 0.01-0.05, and the specific threshold can be adjusted according to your own needs). The average loss value is calculated every 100 iterations. When the average loss value is less than the training threshold for two consecutive times, it means that the cross-modal extraction model is completed.

[0064] The preprocessed multimodal dataset was fed into the constructed cross-modal extraction model to extract feature vectors. Specifically, the RGB image in the wound image feature was normalized to 0-1 (for example, a pixel value of 150 became 0.588). The normalized temperature distribution of the infrared thermal image was resized to 224×224×1. This was then fed into a convolutional neural network to generate two 128-dimensional feature vectors (one generated from the RGB image and the other from the infrared thermal image). These vectors were then weighted and fused into an image feature vector (for example, [0.13, -0.47, 0.79, ..., 1.22]). The biomarker sparse sequence (for example, [0.05, 0.01, 0.65]) and the daily average of the microenvironment time series (for example, [31.9, 65.0]) were concatenated into a 5-dimensional vector and fed into a fully connected neural network to generate a biomarker sparse sequence (for example, [0.23, -0.67, 0.89, ..., 0.46]).

[0065] The contrastive learning framework in the constructed cross-modal extraction model is used to align image feature vectors and biomarker sparse sequences. Specifically, the image feature vector and biomarker sparse sequence from the same patient are used as positive sample pairs, while the biomarker sparse sequences from 10 other patients are randomly selected as negative samples. A contrastive loss function is calculated between the image feature vector and the biomarker sparse sequence. The cross-modal extraction model is fine-tuned using stochastic gradient descent. After the multimodal dataset is updated (if there is no data update, the previous value is used for padding), for example, five iterations are performed. The latent space alignment between the image feature vector and the biomarker sparse sequence is adjusted to obtain a preliminary feature set, including the adjusted image feature vector (e.g., [0.14, -0.46, 0.80, ..., 1.21]) and the biomarker sparse sequence (e.g., [0.24, -0.66, 0.90, ..., 0.46]). A preliminary relationship is established through contrastive learning.

[0066] It is further explained that the combination of convolutional neural network (ResNet-18) and fully connected neural network can efficiently extract features of image and numerical data.

[0067] S2. The preliminary feature set is processed by using dynamic time warping technology to obtain a cross-modal feature set.

[0068] The specific steps include:

[0069] S2.1. Use dynamic time warping (DTW) to temporally align the wound image feature vector (daily data) and the biomarker sparse sequence (weekly data, including null values, for example, [valued, None, None, None, None, None, None, valued]). Specifically, first calculate the Euclidean distance between the wound image feature vector and the biomarker sparse sequence to generate a distance matrix (7×2, corresponding only to days 1 and 7 of the biomarker sparse sequence; days 2 to 6 do not contain biomarker feature sequences and are marked as pending). Create a 7×7 cumulative distance matrix to record the cumulative minimum distance from the starting point to each point. Initially set all elements in the cumulative distance matrix to infinity to indicate uncalculated path costs, and set the starting point value.

[0070] Apply the dynamic programming method to fill the cumulative distance matrix row by row and column by column starting from the starting value. The specific process is: use the dynamic programming recursive formula to calculate the value of each grid, the expression is:

[0071] ;

[0072] in, represents the cumulative distance matrix, Represents the row index, which represents the first row of the wound image feature vector sequence. sky, Represents the column index, indicating the first sky, Represents the first in the wound image feature vector sequence element, that is, the Day's wound image feature vector, Indicates the first element, that is, the Days of biomarker sparse sequences, represents the Euclidean distance between the wound image feature vector sequence and the biomarker sparse sequence, Indicates choosing the smallest cumulative distance from the three options, indicating choosing the alignment path with the lowest cost. Represents the cumulative distance matrix Middle position The minimum cumulative distance value, Represents the cumulative distance matrix Middle position The minimum cumulative distance value, Represents the cumulative distance matrix Middle position The minimum cumulative distance value.

[0073] If the value of the biomarker sparse sequence is null, it is set to infinity (invalid alignment is skipped), the end value of the cumulative distance matrix is ​​checked, and the filled cumulative distance matrix is ​​saved. Then, the time alignment path is determined by backtracking based on the filled cumulative distance matrix, specifically: Check The source, if from , then the wound image feature sequence and aligned biomarker sparse sequences If from , then the wound image feature sequence Aligning biomarker sparse sequences ( ) (to be the difference); if from , then the biomarker sparse sequence There is no corresponding wound image feature sequence.

[0074] Example alignment path: (But biomarker sparse sequences , skipped), actually from ; , trace back to The alignment path is arrive arrive arrive arrive arrive .

[0075] The result is: Day 1 wound image feature sequence Aligning biomarker sparse sequences ; Wound image feature sequence from day 2 to day 6 Aligning biomarker sparse sequences (To be interpolated); Seventh day wound image feature sequence Aligning biomarker sparse sequences The whole process outputs a preliminary aligned wound image feature vector sequence and a biomarker sparse sequence.

[0076] Advantageously, dynamic time warping (DWT) is used to temporally align the wound image feature vector sequence and the biomarker sparse sequence, effectively resolving the time axis misalignment caused by inconsistent data acquisition frequencies. This ensures the temporal correspondence between the two modalities. This alignment improves the accuracy of subsequent analysis and avoids feature matching errors caused by temporal misalignment.

[0077] S2.2. For biomarker sparse sequences, use the Gaussian process regression method to interpolate the null values ​​of the biomarker sparse sequence to generate daily continuous values. Example: Known biomarker sparse sequence , biomarker sparse sequence , interpolation on the third day to obtain the sparse sequence of biomarkers Days 2-6 were interpolated to generate a continuous sparse sequence of biomarkers (7 days, 128 dimensions).

[0078] S2.3. Based on the continuous biomarker sparse sequence and wound image feature vector sequence, the microenvironment time series (temperature, humidity) is integrated using the mean calculation method to generate an aligned cross-modal feature set. Specifically, the daily mean of the microenvironment time series is calculated to generate a 7-day 2D microenvironment time series. Since the confounding factor encoding is static, it is copied to the 7-day time axis to form a 7-day × 3-dimensional microenvironment time series. The wound image feature vector sequence, biomarker sparse sequence, microenvironment time series, and confounding factor vector are integrated by splicing them along the time axis. The time error verification method is used to confirm that the length is all 7 days. Once the error is correct, the aligned cross-modal feature set is prepared.

[0079] Optimally, the generated cross-modal feature set integrates wound image feature vector sequences (reflecting wound appearance and temperature distribution), biomarker sparse sequences (reflecting inflammatory and immune status), microenvironment time series (reflecting the healing environment), and confounding factor vectors, providing a multi-dimensional representation of the wound state. This comprehensiveness facilitates more accurate assessment of wound healing progress. For example, the aligned cross-modal feature set can be used to analyze the relationship between wound area changes and inflammatory indicators and microenvironmental conditions, thereby supporting clinical decision-making.

[0080] S3. Based on the cross-modal feature set, a causal relationship graph is constructed to identify the intervention path, and the causal effect is corrected through the backdoor adjustment method to eliminate false correlations and obtain a causal purified cross-modal feature sequence.

[0081] The specific steps include:

[0082] S3.1. First, the confounding factor vector, biomarker sparse sequence, and wound image feature vector sequence were extracted from the aligned cross-modal feature set. The microenvironment time series was not used because the microenvironment time series reflects the external conditions of wound healing (such as temperature 31.9°C and humidity 65.0%) and directly affects the healing rate (for example, high temperature may slow healing), rather than indirectly interfering with the biomarker sparse sequence through confounding factor data.

[0083] Using a historical data set (e.g., wound healing records of 1,000 patients), the Pearson correlation coefficient is used to calculate the correlation between variables. The expression is:

[0084] ;

[0085] in, represents the Pearson correlation coefficient, which indicates the strength and direction of the linear correlation between two variables. Represents a value of the confounding factor vector, such as age, example The age code of each patient is 1 (31-60 years old), then =1, Represents the sample mean. For example, if the mean age code of 1000 patients is 1.2, then =1.2, Represents a dimension in a sparse sequence of biomarkers, such as CRP concentration. For example, The CRP concentration of a patient is 0.24, then =0.24, Represents the sample mean of a sparse sequence of biomarkers. For example, the mean CRP concentration of 1000 patients is 0.20, then =0.20.

[0086] For example, the Pearson correlation coefficient between age and CRP concentration is 0.6, indicating a strong positive correlation. The correlation coefficient between medication history and IL-6 concentration is 0.45, indicating a moderate correlation. This calculation is performed for each variable pair (e.g., age and CRP concentration, medication history and IL-6 concentration, and CRP concentration and healing rate).

[0087] Based on correlation results and domain knowledge, the causal direction between variables is determined. Example 1: The correlation coefficient between age and CRP concentration is 0.6. Domain knowledge suggests that increasing age may lead to an increased inflammatory response (increased CRP concentration), so the path is determined to be "age points to CRP concentration." The correlation coefficient between CRP concentration and healing rate (area change) is -0.5. Increased inflammation may slow healing, so the path is expanded to "age points to CRP concentration points to healing rate." Example 2: The correlation coefficient between medication history and IL-6 concentration is 0.45. Antibiotic use may suppress inflammation (decreased IL-6 concentration), so the path is determined to be "medication history points to IL-6 concentration." The correlation coefficient between IL-6 concentration and healing rate is -0.4. Reduced inflammation may accelerate healing, so the path is determined to be "medication history points to IL-6 concentration points to healing rate."

[0088] Based on the correlation results and causal directions, a causal graph is constructed. The nodes in the causal graph are: confounding factor data (age, medication history), the dimensions of the biomarker sparse sequence (CRP concentration, IL-6 concentration), and the healing rate of the wound image feature vector sequence. The edges in the causal graph are: key intervention pathways, such as "age to CRP concentration to healing rate" and "medication history to IL-6 concentration to healing rate." The causal graph is output to clearly identify the potential interference pathways of the confounding factor vector on the biomarker sparse sequence and the wound image feature vector sequence.

[0089] Preferably, the Pearson correlation coefficient is used to calculate the correlation between the confounding factor vector, the sparse sequence of biomarkers, and the sequence of wound image feature vectors (healing rate) (for example, the correlation coefficient between age and CRP concentration is 0.6). Combined with domain knowledge (for example, increasing age leads to increased inflammation), a causal graph is constructed. This identifies the potential interference pathways of the confounding factor vector on the sparse sequence of biomarkers and the sequence of wound image feature vectors (healing rate), avoiding the misjudgment of causal relationships based solely on correlation. For example, before constructing the causal graph, one might mistakenly believe that CRP concentration directly causes decreased healing rate (correlation coefficient -0.5), while the causal graph reveals age as an intermediate factor. This provides a structured foundation for subsequent causal effect correction and ensures a well-defined intervention pathway (for example, "medication history to IL-6 concentration to healing rate").

[0090] S3.2. Extract the target variable, dependent variable, and confounding factor vector from the aligned cross-modal feature set. The target variable, healing rate, is represented by the area change in the wound image feature vector sequence, for example, from 0.5 cm² on day 1 to 0.4 cm² on day 7, defined as a positive value (healing rate > 0) or a negative value (worsening). The dependent variable is a specific dimension in the sparse sequence of biomarkers, such as CRP concentration (e.g., 0.05) or IL-6 concentration (e.g., -0.65).

[0091] Use a historical data set (such as wound healing records of 1,000 patients) to estimate the prior probability. Specifically, calculate the distribution of the confounding factor vector statistically. For example: in the age segment, 0-30 accounts for 0.3, 31-60 accounts for 0.4, and 61+ accounts for 0.3; in the medication history, 0 accounts for 0.6 and 1 accounts for 0.4.

[0092] Next, the conditional probability of healing speed is calculated for a given biomarker sparse sequence and a confounding factor vector, and a conditional probability table is defined. Specifically, the observed values ​​of the variables biomarker sparse sequence, confounding factor vector, and healing speed are extracted from the historical data set. This observation is obtained by querying the database and extracting the original data by patient ID and time point, and then preprocessing and integrating them to construct the conditional probability table. Create a table to list the combinations of biomarker sparse sequences and confounding factor vectors, and record the distribution of the corresponding healing speeds. For example, the biomarker sparse sequence has 3 values ​​(low, medium, and high), the confounding factor vector has 3 values ​​(0-30, 31-60, 61+), and the healing speed has 2 values ​​(positive and negative). The table size is 3 times 3 times 2. Use the Bayesian method to estimate the conditional probability by frequency counting, and its expression is:

[0093] ;

[0094] in, Indicates the sparse sequence of a given biomarker (e.g. CRP concentration) and confounding factor vector (e.g. age), the speed of healing probability.

[0095] Summarize the probabilities of all combinations and form a table, also known as a conditional probability table.

[0096] The causal effect is calculated using the backdoor adjustment formula for each patient data, which is expressed as:

[0097] ;

[0098] in, Represents a sparse sequence of dependent variable biomarkers In case of intervention, represents the intervention estimate, and the target variable is the healing speed The probability of occurrence, represents the estimated prior probability.

[0099] Example, determining sparse sequences of biomarkers and healing speed Specific values, for example, =CRP concentration=0.05, = healing speed > 0. For all confounding factor vectors Sum the values, assuming =Age, has 3 segments: , ; , ; , .calculate . So after adjustment =0.05 pairs The causal effect for β > 0 is 0.67.

[0100] Compare the causal effects before and after adjustment to eliminate the influence of confounding factor data. Specifically, without adjustment: Direct calculation ,For example, It is believed that CRP concentration is negatively correlated with healing speed (the influence coefficient is assumed to be -0.5). After adjustment: , indicating that the true effect of CRP concentration is neutral (the effect coefficient is corrected to -0.1). Analysis: The negative correlation is mainly caused by the confounding factor vector (such as age 61+), rather than the CRP concentration itself. Output the corrected causal effect value, The effect coefficient of CRP concentration on healing rate was corrected from -0.5 to -0.1 after standardization and conversion through regression analysis.

[0101] Ideally, a backdoor adjustment formula removes interference from confounding factor vectors (e.g., age 61+), preventing spurious correlations. For example, a CRP concentration of 0.05 was negatively correlated with healing rate (coefficient -0.5) without adjustment, but became neutral (coefficient -0.1), indicating that the true impact of CRP concentration was masked by confounding factors. This provided a quantitative causal effect value (e.g., from -0.5 to -0.1), providing a precise basis for subsequent data adjustments and clinical decision-making.

[0102] S3.3. Correct the results by causal effects and adjust the data in the aligned cross-modal feature set to generate a causal purified cross-modal feature sequence. Specifically, use the linear regression method with the confounding factor vector as a covariate to calculate the effect of the confounding factor vector on each dimension of the biomarker sparse sequence. The expression is:

[0103] ;

[0104] in, Represents the sparse sequence of biomarkers Tiandi The value of the dimension, for example, CRP concentration is 0.245, represents the age code in the confusion factor vector, such as 1 (31-60), Medication history in the confounding factor vector, for example, 1, represents the duration of diabetes in the confounding factor vector, such as 1 (6-10), represents the intercept, represents the regression coefficient of age coding, represents the effect of age coding, represents the regression coefficient of medication history, indicating the influence of medication history, represents the regression coefficient of diabetes duration, represents the effect of diabetes duration, Represents the residual.

[0105] For example, for the CRP concentration dimension, assume the regression results:

[0106] ;

[0107] The confounding factor data is [31-60, 1, 6-10], that is, =1, =1, = 1. The contribution calculation process is:

[0108] .

[0109] For each dimension of the biomarker sparse sequence, the contribution of the confounding factor data is subtracted.

[0110] Example: The original value of CRP concentration on day 3 is 0.245, the confounding factor vector contribution is 0.045, and the adjusted CRP is:

[0111] .

[0112] The adjusted biomarker sparse sequence, wound image feature vector sequence, and microenvironment time series are concatenated in dimensional order to form a causal-purified cross-modal feature sequence. For example, a 7-day × 128-dimensional biomarker sparse sequence might be [[0.20, -0.655, ...], ...], removing interference from age and medication history. A 7-day × 128-dimensional wound image feature vector sequence might be [[0.14, -0.46, ...], ...]. A 7-day × 2-dimensional microenvironment time series might be [[31.9, 65.0], ...]].

[0113] Optimally, causal graphs identify key intervention pathways, enhancing the accuracy of causal inference; backdoor adjustments eliminate spurious correlations, revealing the true effect of sparse biomarker sequences on healing speed; and linear regression generates causally purified cross-modal feature sequences to improve data quality. Together, these approaches ensure the reliability of causal relationships extracted from cross-modal feature sets, laying a solid foundation for subsequent analysis and application.

[0114] S4. Input the causal purified cross-modal feature sequence into the established wound healing prediction model, and output the wound healing prediction result based on the time series characteristics.

[0115] The specific steps include:

[0116] S4.1. We first began to establish a wound healing prediction model, selecting a long short-term memory network (LSTM) as the basic architecture, configured with two layers, 128 units per layer, and the input being a causal purified cross-modal feature sequence.

[0117] Prepare a training set for the wound healing prediction model, including a causally cleansed cross-modal feature sequence generated from a historical dataset (e.g., 1,000 patients) and corresponding healing score labels, such as a 7-day continuous score of [10, 20, 30, 40, 50, 60, 70] (range 0-100). The goal of training is to adjust the parameters of the wound healing prediction model so that the predicted score of the wound healing prediction model is close to the actual healing score.

[0118] Optimally, the 7-day time series accurately reflects the dynamics of the healing process, e.g., 10% healing on day 1 and 45% healing on day 5, avoiding the limitations of single-point predictions. Furthermore, based on causal purification of cross-modal feature sequences, the predictions eliminate interference from confounding factor vectors (e.g., age and medication history), resulting in a closer approximation to the true healing trend.

[0119] The causal purification cross-modal feature sequence in the training set is input into the wound healing prediction model, and the mean square error between the predicted score and the true score of the wound healing prediction model is calculated. For example, the prediction is [10, 20, 30], the true score is [12, 22, 32], and the error is 4. The weights and biases of the two layers of 128 units are updated by gradient descent, and the iteration is repeated until the loss converges (for example, an error threshold is set, and when the mean square error is less than the error threshold, it indicates that the training is completed, or a maximum number of iterations is set, and when the maximum number of iterations is reached, it indicates that the training is completed).

[0120] The causal purified cross-modal feature sequence is input into the wound healing prediction model. The wound healing prediction model uses time series characteristics (7 days of data) to capture the temporal dependencies between the wound image feature vector sequence, the biomarker sparse sequence, and the microenvironment time series, and outputs a continuous healing score curve, such as 10 (10% healing) on ​​day 1, 45 (45% healing) on ​​day 5, and 55 (55% healing) on ​​day 7.

[0121] For example, if the input is [[0.14,-0.46,...,0.20,-0.655,...,31.9,65.0],...], the wound healing prediction model outputs a 7-day score, for example, [10,20,30,40,45,50,55].

[0122] S4.2. Based on the causal effect correction results (e.g., a CRP concentration-healing rate coefficient of -0.1), combined with the continuous healing score curve output by the wound healing prediction model, a causal target is determined. For example, if, in the microenvironment time series, the tissue oxygen partial pressure is >25 mmHg, and there is a nonlinear relationship between CRP concentration and the healing rate score (e.g., the score decreases more slowly with increasing CRP concentration), this is recorded as a causal target, also known as an intervention target. This is also recorded as a medical recommendation, e.g., when the tissue oxygen partial pressure is >25 mmHg, the effect of CRP concentration on healing rate is weakened.

[0123] Ideally, combining causal effect correction results (e.g., a CRP concentration influence coefficient of -0.1) with the wound healing prediction model's predictive score allows analysis of causal targets, revealing key influencing factors (e.g., the interaction between tissue oxygen tension and CRP concentration), going beyond simple prediction to provide causal insights. Furthermore, medical recommendations (e.g., adjusting oxygen tension rather than simply lowering CRP) can be generated, providing specific guidance for clinical intervention and enhancing practicality.

[0124] S4.3. Calculate the impact ratio based on the contribution of the confounding factor vector to the biomarker sparse sequence (e.g., the adjustment for CRP concentration is 0.045 (0.245 - 0.20)). The impact ratio is calculated as the ratio of the adjustment to the original value. For example, if the original CRP concentration is 0.245 and the adjustment is 0.20, the reduction is 0.045. This is calculated as 0.045 / 0.245 ≈ 0.1837, or approximately 18.37%. This is then summarized as a textual description, such as "the confounding factor vector interference is reduced by 18.37%," to report the confounding factor impact.

[0125] The continuous healing score curve, intervention target and confounding factor impact report are the complete wound healing prediction results.

[0126] S5. Generate a real-time wound visualization report and provide feedback based on the wound healing prediction results.

[0127] The specific steps include:

[0128] A visualization report is generated based on the wound healing prediction results and multimodal datasets. Specifically, the following steps are performed: Area change is calculated from the wound image feature vector sequence (e.g., 0.5 cm² on day 1, 0.4 cm² on day 7), highlighting the healing area in green. Area change is annotated, for example, with "-0.1 cm²." Seven-day temperature data (e.g., [31.9, 32.0, 31.8, ...]) is extracted from the microenvironment time series to generate a thermal image with a color scale ranging from 0-40°C. CRP concentrations (e.g., [5, 4.8, 4.5, 4, 3.5, 3.2, 3] mg / L) are extracted from the biomarker sparse sequence, and a daily trend graph is generated through linear interpolation. Temperature and humidity are extracted from the microenvironment time series (e.g., [[31.9, 65.0], [32.0, 64.8], ...]), and a 7-day change curve is plotted. Use a healing score curve (e.g., [10, 20, 30, 40, 45, 50, 60]) and annotate daily percentages, such as 10% on day 1 and 60% on day 7. Summarize the results of causal analysis, for example, "The miscalculation of CRP due to age has been corrected, and the actual healing rate has increased by 15%" (based on adjustments for the effects of confounding factors, such as 18% increased to the example value of 15%); "It is recommended to pay attention to oxygen partial pressure, the current value of 22 mmHg is below the threshold" (based on the intervention target).

[0129] Visual reports are presented through a web interface, displaying wound RGB images, infrared thermal images, biomarker trend charts, microenvironment data curves, and healing score curves in graphical form. Causal analysis results are listed in text form.

[0130] This embodiment also provides a wound healing degree evaluation system based on multi-data analysis, including:

[0131] The feature module collects wound-related multimodal datasets, builds a cross-modal extraction model, and inputs the multimodal dataset to obtain a preliminary feature set;

[0132] The adjustment module processes the preliminary feature set using dynamic time warping technology to obtain a cross-modal feature set;

[0133] The causal module constructs a causal relationship graph based on the cross-modal feature set to identify the intervention path, and corrects the causal effect through the backdoor adjustment method to eliminate false correlations and obtain a causal purified cross-modal feature sequence;

[0134] The prediction module inputs the causal purification cross-modal feature sequence into the established wound healing prediction model and outputs the wound healing prediction results based on the time series characteristics;

[0135] The visualization module generates real-time wound visualization reports and provides feedback based on the wound healing prediction results.

[0136] This embodiment also provides a computer device, which is suitable for the case of a wound healing degree evaluation method based on multi-data analysis, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the wound healing degree evaluation method based on multi-data analysis proposed in the above embodiment.

[0137] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0138] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the wound healing degree evaluation method based on multi-data analysis as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0139] In summary, the present invention improves the scientific nature and explanatory power of causal inference by: constructing a causal graph, identifying intervention paths, and revealing the causal chain between confounding factors and healing; correcting the causal effect through the backdoor adjustment formula to reveal the true effect of the variable, providing a quantitative basis for data purification, and enhancing the accuracy and credibility of the evaluation results; and generating a causal purification cross-modal feature sequence through linear regression processing, eliminating the interference of confounding factors, and providing feature input without false correlation; in addition, by processing the causal purification cross-modal feature sequence through a wound healing prediction model established based on LSTM, and using the characteristics of time series to capture dynamic changes, the accuracy and dynamics of healing process monitoring can be improved, achieving a breakthrough from a single time point to continuous evaluation.

[0140] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A wound healing degree evaluation method based on multi-data analysis, characterized by: include, collecting a wound-related multimodal dataset, constructing a cross-modal extraction model, and inputting the multimodal dataset to obtain a preliminary feature set, wherein the wound-related multimodal dataset includes wound image features, microenvironment time series, biomarker sparse series, and confounding factor encoding; The preliminary feature set is processed using dynamic time warping technology to obtain a cross-modal feature set; Based on the cross-modal feature set, a causal relationship graph is constructed to identify the intervention path, and the causal effect is corrected through the backdoor adjustment method to eliminate false correlations, thereby obtaining a causal purified cross-modal feature sequence. The specific steps include: Use the Pearson correlation coefficient to calculate the correlation of cross-modal feature sets and construct a causal relationship diagram; The causal effect is corrected through a backdoor adjustment formula to remove the spurious correlation of the confounding factor vector, and the sparse sequence of biomarkers is adjusted using linear regression to generate a causal purified cross-modal feature sequence; The causal purification cross-modal feature sequence is input into the established wound healing prediction model, and the wound healing prediction results are output based on the time series characteristics; Generate real-time wound visualization reports and provide feedback based on wound healing prediction results.

2. The wound healing degree evaluation method based on multi-data analysis according to claim 1, characterized in that: Construct a cross-modal extraction model and input the multimodal dataset to obtain a preliminary feature set. The specific steps include the following: Convolutional neural networks and fully connected neural networks are used as the basic networks, and the construction of the cross-modal extraction model is completed after training; Input the multimodal dataset into the constructed cross-modal extraction model and output a preliminary feature set.

3. The wound healing degree evaluation method based on multi-data analysis according to claim 2, characterized in that: Obtaining the cross-modal feature set refers to generating an alignment path using Euclidean distance and a dynamic programming formula; Gaussian process regression was used to interpolate the null values ​​of biomarker sparse sequences, and the microenvironment time series and confounding factor encoding were integrated into a cross-modal feature set.

4. The wound healing degree evaluation method based on multi-data analysis according to claim 3, characterized in that: The established wound healing prediction model is based on a long short-term memory network and is trained using historical data sets.

5. The wound healing degree evaluation method based on multi-data analysis according to claim 4, characterized in that: Outputting the wound healing prediction result according to the time series characteristics refers to inputting the causal purification cross-modal feature sequence into the wound healing prediction model, using the time series characteristics to capture the temporal dependency, and outputting a continuous healing score curve; The causal targets are analyzed in combination with the causal effect correction results, and the impact ratio is calculated based on the contribution of the confounding factor vector to generate a confounding factor impact report.

6. A wound healing degree evaluation system based on multi-data analysis, based on the wound healing degree evaluation method based on multi-data analysis according to any one of claims 1 to 5, characterized in that: include, The feature module collects wound-related multimodal datasets, builds a cross-modal extraction model, and inputs the multimodal dataset to obtain a preliminary feature set; The adjustment module processes the preliminary feature set using dynamic time warping technology to obtain a cross-modal feature set; The causal module constructs a causal relationship graph based on the cross-modal feature set to identify the intervention path, and corrects the causal effect through the backdoor adjustment method to eliminate false correlations and obtain a causal purified cross-modal feature sequence; The prediction module inputs the causal purification cross-modal feature sequence into the established wound healing prediction model and outputs the wound healing prediction results based on the time series characteristics; The visualization module generates real-time wound visualization reports and provides feedback based on the wound healing prediction results.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the wound healing degree evaluation method based on multi-data analysis according to any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the wound healing degree evaluation method based on multi-data analysis according to any one of claims 1 to 5 are implemented.

Images