Wound healing intervention method and system based on multi-modal data

Abstract

The invention provides a wound healing intervention method and system based on multi-modal data, and the method comprises the steps: calculating the change trend of each feature along with time through a time sequence analysis algorithm if the time dimension of a standardized feature vector set is complete, and generating a dynamic change curve; according to the dynamic change curve, a preset threshold judgment algorithm is adopted, the curve slope and the fluctuation amplitude are analyzed, and a key turning point of the healing process is determined; aiming at the intervention opportunity, generating a structured instruction set containing a time point and an intervention type, and outputting the structured instruction set to a clinical decision support system to complete precise intervention deployment; obtaining new multi-modal data through a real-time data updating module, repeating the steps, and updating the unified analysis model and an intervention opportunity prediction result; and according to the updated prediction result, optimizing the intervention instruction set by adopting a dynamic adjustment algorithm, and generating a final intervention scheme.

Description

Technical Field

[0001] This invention relates to the field of medical diagnostic technology, and in particular to a wound healing intervention method and system based on multimodal data. Background Technology

[0002] Wound healing intervention research is a key area at the intersection of medicine and technology, aiming to optimize the wound recovery process, reduce complications, and improve patients' quality of life through scientific methods. This field is important because it directly impacts clinical treatment outcomes, especially in the management of chronic wounds and complex trauma, where precise intervention can significantly shorten healing time and reduce healthcare costs. With the rise of multimodal data analysis technologies, researchers have attempted to guide interventions by integrating multiple data sources, but current methods still face significant limitations.

[0003] Existing solutions typically rely on single data types, such as wound images or single biomarkers, making it difficult to comprehensively capture the dynamic changes in the healing process. This limitation leads to inaccurate judgment of intervention timing, often resulting in premature or delayed intervention, which affects healing outcomes. Furthermore, traditional methods lack systematic analysis of multi-source data fusion, restricting the accurate prediction of complex wound conditions.

[0004] The core challenge lies first in effectively integrating multimodal data, including wound image sequences, inflammatory marker concentrations, and tissue regeneration indicators. These data come from heterogeneous sources and have complex temporal dimensions, making it difficult to form a unified analytical framework to accurately describe the natural process of wound healing. The difficulty in data integration further leads to inaccurate predictions of intervention timing, and the system cannot accurately identify the critical points requiring intervention. Deviations in intervention timing can interfere with natural healing or delay treatment, making it difficult to achieve the precision medicine goal of on-demand intervention. Summary of the Invention

[0005] This invention provides a wound healing intervention method based on multimodal data, mainly comprising: The multimodal data acquisition module obtains raw data from wound image sequences, inflammatory marker concentrations, and tissue regeneration indicators to generate a multidimensional dataset containing timestamps. Feature extraction algorithms are used to perform edge detection and texture analysis on wound images, and normalization processing is performed on inflammatory markers and tissue regeneration indicators to obtain a standardized feature vector set; If the time dimension of the standardized feature vector set is complete, then the time series analysis algorithm is used to calculate the changing trend of each feature over time and generate a dynamic change curve. Based on the dynamic change curve, a preset threshold judgment algorithm is used to analyze the slope and fluctuation amplitude of the curve to determine the key turning point in the healing process. By using a deep learning network, inputting dynamic change curves and key inflection points, a multimodal fusion model is trained to generate a unified analysis model and output the healing process prediction results. If the confidence level of the prediction result is higher than the preset threshold, a time series prediction algorithm is used to analyze the healing process output by the unified analysis model and determine the best time for intervention. For the timing of intervention, a structured instruction set containing time points and intervention types is generated and output to the clinical decision support system to complete the precise deployment of intervention; By using the real-time data update module to acquire new multimodal data, repeat the above steps to update the unified analysis model and the prediction results of intervention timing; Based on the updated prediction results, a dynamic adjustment algorithm is used to optimize the intervention instruction set and generate the final intervention plan.

[0006] This invention provides a wound healing intervention system based on multimodal data, mainly comprising: The multimodal data acquisition module is used to acquire raw data from wound image sequences, inflammatory marker concentrations, and tissue regeneration indicators, and generate a multidimensional dataset containing timestamps. The feature extraction module is used to perform edge detection and texture analysis on wound images using feature extraction algorithms, and to perform normalization processing on inflammatory markers and tissue regeneration indicators to obtain a standardized feature vector set; The time series analysis module is used to calculate the changing trend of each feature over time and generate dynamic change curves if the time dimension of the standardized feature vector set is complete. The threshold judgment module is used to analyze the slope and fluctuation amplitude of the curve based on the dynamic change curve and a preset threshold judgment algorithm to determine the key turning point of the healing process. The deep learning module is used to train a multimodal fusion model by taking a dynamic change curve and key inflection points as input through a deep learning network, generating a unified analysis model, and outputting the healing process prediction results. The time series prediction module is used to analyze the healing process output by the unified analysis model and determine the best intervention time if the confidence level of the prediction result is higher than a preset threshold. The instruction generation module is used to generate a structured instruction set containing time points and intervention types for intervention timing, and output it to the clinical decision support system to complete the precise intervention deployment; The real-time data update module is used to acquire new multimodal data, repeat the above steps, and update the unified analysis model and intervention timing prediction results. The dynamic adjustment module is used to optimize the intervention instruction set and generate the final intervention plan based on the updated prediction results using a dynamic adjustment algorithm.

[0007] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention acquires wound image sequences, inflammatory marker concentrations, and tissue regeneration indicators through multimodal data acquisition. It uses a feature extraction algorithm to generate a standardized feature vector set, performs time series analysis to obtain dynamic change curves, and identifies key inflection points in the healing process. A multimodal fusion model is trained using a deep learning network to output healing process prediction results. Based on these predictions, the optimal intervention timing is determined, and a structured intervention instruction set is generated. Through real-time data updates and dynamic adjustments, this invention can continuously optimize the prediction model and intervention plan, achieving precise monitoring and personalized intervention of the wound healing process, effectively improving treatment outcomes and the efficiency of medical resource utilization. Attached Figure Description

[0008] Figure 1 This is a flowchart of a wound healing intervention method based on multimodal data according to the present invention.

[0009] Figure 2 This is a schematic diagram of a wound healing intervention method and system based on multimodal data according to the present invention.

[0010] Figure 3 This is another schematic diagram of a wound healing intervention method and system based on multimodal data according to the present invention.

[0011] Figure 4 This is a schematic diagram of the structure of a wound healing intervention method and system based on multimodal data according to the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0013] like Figure 1-4 This embodiment of a wound healing intervention method and system based on multimodal data may specifically include: S101. Through the multimodal data acquisition module, raw data are obtained from wound image sequences, inflammatory marker concentrations, and tissue regeneration indicators to generate a multidimensional dataset containing timestamps.

[0014] The data acquisition module obtains multimodal data from wound image sequences, inflammatory marker concentrations, and tissue regeneration indicators to obtain a raw dataset with timestamps. Image processing techniques were used to preprocess the wound image sequence, including denoising and segmentation, to obtain clear wound feature images; If the boundary clarity of the wound feature image is lower than a preset threshold, then an edge enhancement algorithm is applied to obtain an optimized wound feature dataset. By using time series analysis, trend features of inflammatory marker concentrations and tissue regeneration indicators over time were extracted to obtain a dynamic feature set. Based on the dynamically changing feature set, a clustering algorithm is applied to classify the multimodal data, determine the wound healing stage, and obtain the healing stage label; If the healing stage label is inconsistent with the preset standard, a comprehensive feature dataset is obtained by integrating the wound feature dataset and the dynamic change feature dataset through data integration processing. A regression analysis model was used to predict the future trend of wound healing on a comprehensive feature dataset, and the prediction results were obtained.

[0015] In one possible implementation, the data acquisition module obtains multimodal data from wound image sequences, inflammatory marker concentrations, and tissue regeneration indicators.

[0016] For example, wound images are taken daily using a high-resolution camera to record changes in wound size and color; simultaneously, the concentration of inflammatory markers such as C-reactive protein is measured using blood testing equipment, in mg / L; tissue regeneration indicators are obtained through tissue sample analysis to determine collagen deposition. The original dataset with timestamps can be recorded as: 2025-05-01, wound area 10cm². 2 C-reactive protein 5 mg / L, collagen deposition 20%.

[0017] Specifically, image processing techniques preprocess the wound image sequence. Denoising can be achieved using Gaussian filtering to reduce image noise; segmentation employs thresholding to separate the wound from the surrounding skin, resulting in a clear wound feature image. If the wound boundary clarity is below a preset threshold, such as a clarity score below 0.8, edge enhancement algorithms, such as Canny edge detection, can be applied to optimize the image boundaries and generate an optimized wound feature dataset. This processing significantly improves the visualization accuracy of wound edges, aiding in subsequent analysis.

[0018] For example, time series analysis can extract dynamic trend features for the concentration of inflammatory markers and tissue regeneration indicators. Assuming that C-reactive protein concentrations are recorded as 5, 4, 3, 2, 1, 1, and 1 mg / L over seven consecutive days, it indicates that inflammation is gradually subsiding; and that collagen deposition increases from 20% to 50%, indicating accelerated tissue regeneration. These trend features constitute a dynamic change feature set that reflects the wound healing process.

[0019] In one embodiment, based on a dynamically changing feature set, the K-means clustering algorithm is applied to classify multimodal data and determine the stage of wound healing.

[0020] For example, clustering results divide the data into the inflammatory phase, proliferative phase, and maturation phase. If the healing stage label is inconsistent with the preset standard, such as the data indicating the proliferative phase but the standard expecting the maturation phase, then data integration processing is performed to fuse the wound feature dataset and the dynamic change feature set.

[0021] For example, normalizing wound area, boundary clarity, and inflammatory marker concentration can be combined into a comprehensive feature dataset to improve data consistency.

[0022] Preferably, a regression analysis model, such as linear regression, is used to predict the comprehensive feature dataset. Assuming the comprehensive feature dataset includes wound area, C-reactive protein concentration, and collagen deposition, the model predicts that the wound area will decrease from 5 cm in the next 7 days. 2 Reduced to 2cm 2 This indicates a good healing trend. The prediction results provide a basis for clinical decision-making and optimize treatment plans.

[0023] It should be noted that the advantage of the above method lies in the comprehensive utilization of multimodal data, which significantly improves the accuracy of the healing stage assessment and the reliability of the prediction.

[0024] For example, image processing ensures accurate feature extraction, time series analysis captures dynamic changes, and clustering and regression analysis automate stage classification and trend prediction. These technologies collectively support refined management of wound healing.

[0025] S102. Using a feature extraction algorithm, edge detection and texture analysis are performed on the wound image, and normalization processing is performed on inflammatory markers and tissue regeneration indicators to obtain a standardized feature vector set.

[0026] An edge detection algorithm is used to process the wound image, extract boundary features, and combine them with texture analysis to generate a wound surface feature set; Normalization was applied to the concentrations of inflammatory markers and tissue regeneration indicators to generate a standardized feature vector set, thus obtaining the initial feature set. Based on the initial feature set, the principal component analysis algorithm is used to reduce the dimensionality of the wound surface feature set and the standardized feature vector set, and the main feature components are extracted to obtain the dimensionality-reduced feature set. If the variance of the feature components in the dimensionality-reduced feature set is lower than a preset threshold, then a feature selection algorithm is applied to filter highly relevant features to obtain an optimized feature set. Based on the optimized feature set, time series analysis was used to extract trends from the standardized feature vectors of inflammatory markers and tissue regeneration, resulting in a dynamic trend feature set. If the time series fluctuation of the dynamic trend feature set exceeds a preset threshold, the dynamic trend feature set is smoothed by the sliding window method to obtain a smoothed trend feature set. Based on the smooth trend feature set, a clustering algorithm is used to classify the wound surface features and dynamic trend features, determine the wound healing stage, and obtain the healing stage label; Based on the healing stage labels, a regression analysis model is used to predict the smooth trend feature set, determine the wound healing trend, and obtain the healing trend prediction results.

[0027] For example, in wound image processing, edge detection algorithms are used to extract wound boundary features. Edge detection delineates the contour of the wound by identifying regions in the image where grayscale changes are significant.

[0028] For example, using the Canny edge detection algorithm, the wound image is first denoised using Gaussian blur, then the gradient intensity is calculated to filter out boundary pixels. Assuming a wound image has a resolution of 512×512 pixels, after processing, boundary pixels account for approximately 5% of the total pixels, forming a boundary feature set. These features describe the shape and size of the wound, providing a basis for subsequent analysis.

[0029] In one possible implementation, texture analysis is used to generate a feature set of the wound surface. Texture analysis extracts features such as surface roughness or uniformity by statistically analyzing the gray-level co-occurrence matrix of image pixels.

[0030] For example, the gray-level co-occurrence matrix is ​​calculated on a wound image, and contrast and entropy values ​​are extracted to generate a feature vector containing 10 texture parameters. These parameters reflect the smoothness or irregularity of the wound surface, which helps to distinguish the healing stage.

[0031] Specifically, the normalization of inflammatory marker concentrations and tissue regeneration indicators involves mapping the raw data to a range of 0 to 1.

[0032] For example, inflammatory marker concentrations ranging from 50 to 200 ng / mL are normalized using min-max normalization to generate standardized feature vectors. This process eliminates dimensional differences, facilitating subsequent analysis. Assuming a patient's inflammatory marker concentration is 150 ng / mL, the normalized value is 0.67. Tissue regeneration indicators are processed similarly to form a standardized feature vector set.

[0033] Preferably, principal component analysis is used for dimensionality reduction. Principal component analysis projects high-dimensional features into a low-dimensional space through linear transformation, preserving the main information.

[0034] For example, principal component analysis can be performed on a 20-dimensional wound surface feature set and a standardized feature vector set. The first five principal components are extracted, and 90% of the variance is retained to obtain a dimensionality-reduced feature set. This method reduces computational complexity while preserving key information.

[0035] It should be noted that if the variance of the dimensionality-reduced feature set is below a threshold, such as below 85%, a feature selection algorithm is used. Feature selection filters highly correlated features by calculating the correlation between features and the healing stage.

[0036] For example, mutual information is used to select features with a correlation coefficient higher than 0.7 with the healing stage, forming an optimized feature set. This selection ensures that the analysis focuses on the most representative features. In one implementation, time series analysis is used to extract dynamic trend features.

[0037] For example, analyzing 7 consecutive days of inflammatory marker concentration data, calculating the daily rate of change, and generating a trend feature vector. If the data shows a gradual decrease in concentration, it reflects a reduction in inflammation. This trend feature helps determine the healing process.

[0038] Understandably, if the fluctuation range of the time series is too large, for example, the rate of change exceeds 20%, then a sliding window method is used for smoothing.

[0039] For example, a 3-day window is used to calculate the average value, generating a smoothed trend feature set. This process reduces noise interference and makes the trend clearer.

[0040] For example, clustering algorithms are used to classify healing stages. K-means clustering is used to divide the smooth trend feature set and wound surface feature set into three classes, corresponding to the early, middle and late healing stages.

[0041] For example, the feature vector of a wound sample is classified as mid-stage healing, generating a corresponding label. This classification intuitively reflects the healing progress.

[0042] In one embodiment, a regression analysis model is used to predict healing trends.

[0043] For example, a linear regression model can be used to predict the concentration of inflammatory markers over the next 7 days based on a smoothed trend feature set. It is assumed that the model predicts the concentration will decrease to the normal range, indicating a good healing trend. This prediction provides a reference for clinical decision-making.

[0044] S103. If the time dimension of the standardized feature vector set is complete, then the time series analysis algorithm is used to calculate the changing trend of each feature over time and generate a dynamic change curve.

[0045] By processing the standardized feature vector set through the time series decomposition algorithm, the changing trend of each feature over time is extracted to obtain a set of dynamic change curves; If the trend component of the dynamically changing curve set is significant, the sliding window method is used to smooth the curve set to obtain a smooth curve set. Based on the set of smooth curves, calculate the periodic components of each feature, extract the periodic features, and obtain the set of periodic features; If the amplitude of the periodic feature set exceeds a preset threshold, the periodic feature set is analyzed in the frequency domain by Fourier transform to obtain the frequency domain feature set. Based on the frequency domain feature set, a clustering algorithm is used to group the features, determine the correlation between the features, and obtain the feature group set. By grouping features into sets, statistical indicators of each set of features are extracted to generate a feature statistics set. Based on the feature statistics set, a regression analysis model is used to predict the trend of feature changes over time, resulting in a trend prediction set.

[0046] For example, when processing a standardized feature vector set using a time series decomposition algorithm, it can be decomposed into three parts: trend, seasonality, and residual, which are used to extract the trend of feature changes over time.

[0047] For example, in wound healing monitoring, the concentration data of inflammatory markers change over time. The STL decomposition algorithm can be used to extract long-term trends, reflecting the overall direction of inflammation reduction. Suppose a patient's wound inflammatory marker concentration is measured weekly, with a data sequence of 10, 8, 6, 4, 3. After decomposition, the trend component shows a smooth decrease, indicating that the inflammation is gradually weakening. This method can clearly separate long-term changes from short-term fluctuations, facilitating subsequent analysis.

[0048] In one possible implementation, if the trend component of the dynamically changing curve set is significant, a sliding window method can be used for smoothing.

[0049] Specifically, for the aforementioned inflammatory marker data, a window size of 3 weeks was set, the average value within the window was calculated, and a set of smooth curves was generated.

[0050] For example, the original data 10, 8, 6 may become 9, 7.3, 6.7 after being smoothed using a sliding window, reducing the interference of short-term fluctuations and highlighting the trend. This smoothing process helps improve the stability of subsequent analysis.

[0051] It should be noted that the calculation of periodic components to extract periodic features can be achieved through autocorrelation analysis.

[0052] For example, during wound healing, tissue regeneration indicators may fluctuate regularly due to the patient's physiological cycle. Assuming tissue regeneration indicators are measured weekly, and the data sequence shows a peak every four weeks, autocorrelation analysis can confirm the four-week cycle length, generating a cyclical feature set. This cyclical feature reflects the regularity of the healing process and helps identify abnormal fluctuations.

[0053] Specifically, if the amplitude of the periodic feature set exceeds a preset threshold, frequency domain analysis can be performed using Fourier transform.

[0054] For example, if the periodic characteristic amplitude of a tissue regeneration index is 2, exceeding the threshold of 1.5, it is converted to the frequency domain, and the main frequency components are extracted, such as 0.25Hz corresponding to a 4-cycle period. This frequency domain feature set can reveal the intensity and frequency of periodic changes, facilitating in-depth analysis of the relationships between features.

[0055] In one embodiment, a clustering algorithm is used to group the frequency domain feature set and determine the feature correlation.

[0056] For example, based on K-means clustering, the frequency domain features of inflammatory markers and tissue regeneration indicators were divided into two groups. It was found that inflammatory markers had a lower frequency, while tissue regeneration indicators had a higher frequency, indicating that their dynamic change patterns differed. Feature grouping sets can clarify the correlation between features and provide a basis for subsequent modeling.

[0057] Preferably, statistical indicators of the feature group set are extracted to generate a feature statistical set, which can calculate the mean, variance, etc.

[0058] For example, the frequency domain characteristics of the inflammatory marker group had a mean of 0.2 Hz and a variance of 0.01, while the tissue regeneration index group had a mean of 0.3 Hz and a variance of 0.02. These statistical indicators quantify the distribution characteristics of the features and help to capture differences between groups.

[0059] For example, regression analysis models can be used to predict the trend of feature changes based on a feature statistical set. Assuming a linear regression model is used, inputting statistical indicators of an inflammatory marker group, predicting the trend over the next four weeks yields a predicted set of gradually decreasing concentrations. This prediction result can provide data support for assessing the wound healing stage and improve the predictability of monitoring.

[0060] Understandably, each step of the above method revolves around the dynamic changes in wound healing characteristics, progressing layer by layer through decomposition, smoothing, periodic extraction, frequency domain analysis, clustering, statistics, and prediction to form a complete analytical chain. The output of each step provides reliable input for subsequent steps, ensuring the logical consistency and practicality of the analysis results.

[0061] S104. Based on the dynamic change curve, a preset threshold judgment algorithm is used to analyze the slope and fluctuation amplitude of the curve to determine the key turning point of the healing process.

[0062] Using time series data and the sliding window method, the local slope and fluctuation amplitude of the dynamic change curve are calculated to obtain the slope sequence and amplitude sequence. If the absolute value of the slope in a continuous time window in the slope sequence exceeds a preset threshold, the slope mutation point is determined by difference calculation to obtain the mutation point set. Based on the set of mutation points, a threshold judgment algorithm is used to analyze the fluctuation amplitude corresponding to the mutation point. If the fluctuation amplitude exceeds the amplitude threshold, it is determined as a candidate turning point, and a set of candidate turning points is obtained. By using the candidate inflection point set, a density clustering algorithm is employed to merge candidate inflection points that are temporally adjacent, thus obtaining an optimized inflection point set. Based on the optimized inflection point set, time series segments corresponding to the inflection points are extracted from the dynamic change curve, and the trend characteristics of the segments are calculated to obtain the trend feature set. By using trend feature sets and employing time series segmentation algorithms, the healing process is divided into stages, resulting in a process stage set. Based on the process stage set, calculate the mean slope and amplitude distribution of each stage to obtain the stage feature set.

[0063] For example, in time series data analysis, the sliding window method is often used to capture local features of dynamically changing curves. Suppose we are analyzing patient healing progress data monitored by a medical device, with the time series representing the daily percentage of wound healing. The sliding window is set to 7 days, and the local slope of the curve within the window is calculated to reflect the healing speed.

[0064] For example, within a certain window, the healing rate increases from 60% to 67%, with a slope of approximately 1% per day. Simultaneously, the fluctuation range is calculated using the standard deviation of the data within the window; assuming a standard deviation of 2% for a given window, it indicates a relatively stable healing process. This method effectively quantifies local trends and fluctuations, providing a foundation for subsequent analysis.

[0065] In one possible implementation, slope sequence analysis focuses on mutation point detection. If the absolute value of the slope exceeds 0.5% / day for three consecutive windows, it indicates a significant change in healing rate. Mutation points are identified by calculating the rate of change of the slope sequence through differentiation.

[0066] For example, if the slope suddenly changes from 0.2% / day to 0.8% / day, the difference exceeds the threshold and is marked as a mutation point. This method can accurately locate key changes in the healing process, facilitating further analysis.

[0067] Specifically, candidate inflection points are screened based on volatility. If the volatility of the 7-day window corresponding to a mutation point exceeds 3%, it is confirmed as a candidate inflection point.

[0068] For example, if the degree of healing fluctuates from 55% to 62% within a certain mutation point window, a range of 7%, the condition is met. This screening ensures that the inflection point is associated with significant changes, improving the reliability of the analysis.

[0069] Preferably, density clustering is used to optimize the inflection point set. Assuming that the time interval between multiple candidate inflection points is less than 3 days, they are merged into a single inflection point through density clustering.

[0070] For example, if two inflection points occur on day 10 and day 12 respectively, they can be merged into an optimized inflection point on day 11. This method reduces redundancy and improves the representativeness of inflection points.

[0071] In one embodiment, the trend feature set is extracted from the time series segment corresponding to the optimized inflection point. Assuming an inflection point occurs on day 20, data from the seven days before and after the inflection point are extracted, and the average slope and fluctuation range of the segment are calculated.

[0072] For example, a segment slope of 0.6% per day with a fluctuation range of 5% reflects the local healing trend. This feature set provides a quantitative basis for subsequent stage division.

[0073] Understandably, time series segmentation algorithms are used to divide the healing process into stages. Based on a trend feature set, slope and fluctuation thresholds are set to divide the sequence into stages such as accelerated healing and stable healing.

[0074] For example, segments with a slope greater than 0.4% / day are classified as the accelerated healing phase. This method clearly delineates the process, facilitating targeted analysis.

[0075] For example, the phase feature set is generated by calculating the mean slope and amplitude distribution for each phase. Assuming the accelerated healing phase has a mean slope of 0.7% / day and a standard deviation of 2.5% for the amplitude distribution, it reflects the phase characteristics. This feature set can be used to compare healing performance at different phases, providing a reference for medical decision-making.

[0076] It should be noted that each step of the above method is closely linked, forming a complete analysis chain from local feature extraction to stage division. The quantitative indicators of each step support each other, ensuring the logical consistency and practicality of the analysis results.

[0077] S105. Through a deep learning network, input dynamic change curves and key inflection points, train a multimodal fusion model, generate a unified analysis model, and output the healing process prediction results.

[0078] Time series analysis is used to obtain time series data from dynamic change curves. The sliding window method is then used to calculate local features and obtain feature sequences. Based on the feature sequence, key inflection points are extracted. If the feature values ​​of a continuous window in the feature sequence exceed a preset threshold, the inflection point is determined by difference calculation, and the inflection point set is obtained. By integrating feature sequences and inflection point sets through multimodal fusion, and using a deep learning network for feature mapping, a fused feature set is obtained. Based on the fused feature set, a deep learning network is used to train the model, and the parameters are iteratively optimized to obtain a healing process prediction model. By using a healing process prediction model and inputting a dynamic change curve, prediction results are generated, resulting in a healing process prediction sequence. Based on the predicted sequence of healing process, a trend feature extraction method is used to calculate the trend features of the predicted sequence and obtain a trend feature set. By using the trend feature set and a time series segmentation algorithm, the healing process is divided into stages, resulting in a process stage set.

[0079] For example, time series analysis is used in healing process studies to process data from dynamic change curves. Taking fracture healing as an example, the dynamic change curve may reflect changes in bone mineral density over time. Time series analysis creates a continuous time series by collecting daily bone mineral density data. The sliding window method can be used to calculate local features, such as the mean or rate of change of bone mineral density within the window. Assuming a window size of 7 days, calculating the weekly average bone mineral density yields a feature sequence reflecting local trends.

[0080] In one possible implementation, when extracting key inflection points, the feature values ​​of a continuous window are determined based on the feature sequence to see if they exceed a threshold.

[0081] For example, if a bone mineral density (BMD) change rate threshold is set to 5%, and the change rate exceeds 5% for seven consecutive days, then a turning point can be located using differential calculation. Differential calculation compares the change rates of adjacent windows to identify abrupt changes and form a set of turning points. For instance, days 30 and 60 might be nodes where BMD changes significantly.

[0082] Specifically, multimodal fusion integrates feature sequences and inflection point sets, and introduces a deep learning network for feature mapping. The feature sequences provide information on continuous changes, while the inflection point set marks key events. During fusion, the rate of change in bone density and the inflection point timestamps can be input into a convolutional neural network to generate a fused feature set, capturing time dependencies.

[0083] For example, the network can identify the association between the turning point on day 30 and rapid increases in bone density.

[0084] Preferably, when training the healing process prediction model using a deep learning network, the parameters are optimized by fusing feature sets. Assuming a Long Short-Term Memory (LSTM) network is used, and 60 days of historical bone density data are input, the model is trained to predict the healing trend for the next 30 days. After iterative optimization, the model can output a healing process prediction sequence, such as predicting that bone density will recover to 80% on day 90.

[0085] In one embodiment, trend feature extraction calculates trend features from the predicted sequence. For the predicted sequence, the slope of bone mineral density growth every 14 days can be calculated to form a trend feature set.

[0086] For example, the slope from day 61 to 75 shows a steady increase, indicating that healing has entered a consolidation phase.

[0087] Understandably, time series segmentation algorithms divide the healing process into stages based on a set of trend features. Assuming that healing is divided into three stages—initial, intermediate, and consolidation—based on slope changes, the initial stage might be days 1 to 30 with a lower slope; the intermediate stage might be days 31 to 60 with a higher slope; and the consolidation stage might be days 61 to 90 with a slope that tends to stabilize, thus obtaining a set of process stages.

[0088] For example, the process phase set can be further used to guide rehabilitation plans. The initial phase can include low-intensity activities, the intermediate phase can increase weight-bearing training, and the consolidation phase can optimize nutritional intake. This phased strategy is data-driven, ensuring the rehabilitation program is precise and efficient.

[0089] S106. If the confidence level of the prediction result is higher than the preset threshold, the time series prediction algorithm is used to analyze the healing process output by the unified analysis model and determine the best intervention time.

[0090] If the confidence level of the predicted sequence is higher than the preset threshold, the predicted sequence of the healing process is analyzed by time series prediction algorithm to obtain the trend change set. Based on the trend change set, feature extraction methods are used to calculate the dynamic changes of sequence features and obtain a dynamic feature set; If the feature values ​​in the dynamic feature set exceed the preset range, the key fluctuation points of the feature sequence are determined by difference calculation to obtain the fluctuation point set. Based on the set of fluctuation points, a clustering algorithm is used to divide the healing process into stages, resulting in a set of stage features. If the distribution of the stage feature set conforms to the preset pattern, then the correlation between the stage features and the intervention timing is analyzed by the sequence matching method to obtain the timing matching set; Based on the timing matching set, a sorting algorithm is used to determine the priority sequence of the best intervention timing, thus obtaining the intervention timing set; By using the intervention timing set, an intervention time series of the healing process is generated, resulting in an intervention sequence.

[0091] For example, in scenarios where the confidence level of the predicted sequence is higher than a preset threshold, the healing process can be analyzed using time series forecasting algorithms. The core of time series forecasting algorithms lies in using historical data to predict future trends, such as inferring the healing status over a future period based on autoregressive models and the periodicity and volatility of the healing process.

[0092] In one possible implementation, assuming the confidence threshold of the predicted sequence is 0.9, if the confidence of a certain sequence reaches 0.95, then exponential smoothing can be used, combined with the healing data of the past 7 days, to predict the trend change in the next 3 days, generating a trend change set that reflects the acceleration or deceleration of the healing speed.

[0093] Specifically, feature extraction of trend change sets can be achieved using the sliding window method. The sliding window analyzes local features of the sequence over a fixed time period, such as a 24-hour window, to extract key indicators in the healing process, such as the rate of change of the healing area.

[0094] Preferably, if the healing area within a certain window decreases by 10% daily, this can be recorded as a dynamic feature, generating a dynamic feature set. This method can capture subtle changes in the healing process, providing data support for subsequent analysis.

[0095] In one embodiment, if the feature values ​​in the dynamic feature set exceed a preset range, for example, the rate of change in the healing area exceeds 15% or is less than 5%, then critical fluctuation points are determined through differential calculation. Differential calculation identifies abnormal change points by comparing feature values ​​at adjacent time points.

[0096] For example, if the rate of change in the healing area suddenly drops from 10% to 3% on a given day, the difference is significant and can be marked as a fluctuation point, generating a set of fluctuation points. This method helps to identify abnormal events in the healing process and provides a basis for determining the timing of intervention.

[0097] For example, for clustering algorithms targeting fluctuating point sets, K-means clustering can be used to group the fluctuating points according to feature similarity. Assuming the fluctuating point set contains 30 points, they can be clustered into 3 groups, representing the accelerated healing, stable, and slowing-down stages, respectively, generating a stage feature set.

[0098] It should be noted that the advantage of clustering lies in summarizing complex fluctuations into interpretable stages, which facilitates subsequent analysis of the correlation between intervention timing and the overall situation.

[0099] Understandably, if the stage feature set conforms to a preset pattern, such as the acceleration stage accounting for more than 50%, the timing of intervention can be analyzed using sequence matching methods. Sequence matching identifies the optimal intervention point by comparing the similarity between stage features and historical intervention data.

[0100] For example, if historical data indicates that intervention is most effective during the middle of the acceleration phase, a timing-matching set can be generated. This approach ensures that the timing of intervention is highly correlated with the actual state of the healing process.

[0101] In one possible implementation, the timing matching set can be used to determine the intervention priority through a sorting algorithm.

[0102] For example, based on historical scores of intervention effectiveness, multiple potential intervention opportunities can be ranked, with priority given to those scoring above 80 points, to generate a set of intervention opportunities. This ranking method can optimize resource allocation and improve the targeting of interventions.

[0103] Preferably, the intervention timing set can generate an intervention time series of the healing process.

[0104] For example, if the intervention timing set includes three highest priority timings, on days 5, 7, and 10 respectively, an intervention sequence is generated to guide actual operations. This serialized output provides a clear timeline for healing process management, improving operational efficiency.

[0105] S107. Generate a structured instruction set containing time points and intervention types for the timing of intervention, and output it to the clinical decision support system to complete the deployment of precise intervention.

[0106] If the confidence level of the predicted sequence is higher than the preset threshold, time series analysis is used to generate a trend sequence of the healing process, thus obtaining a trend dataset. Based on the trend dataset, feature extraction methods are used to calculate the changes in dynamic features, generate feature sequences, and obtain a dynamic feature set. If the feature values ​​in the dynamic feature set exceed the preset range, the key fluctuation points of the feature sequence are calculated using the fluctuation analysis method to obtain the fluctuation point set. Based on the set of fluctuation points, a clustering method is used to divide the healing process into stages, generate a stage feature sequence, and obtain a stage feature set; If the distribution of the stage feature set conforms to the preset pattern, then the correlation between stage features and intervention timing is analyzed by sequence matching method to generate timing association set; Based on the timing association set, a sorting method is used to generate a sequence of instructions containing time points and intervention types, resulting in a structured instruction set; The structured instruction set is output to the clinical decision support system to complete the intervention deployment.

[0107] For example, in predicting the healing process, time series analysis methods can be used to generate trend sequences. Time series analysis captures the periodicity and trends of the healing process through historical data, such as changes in the cell proliferation rate during wound healing. Suppose wound healing data records the daily percentage reduction in wound area; analysis reveals an increasing trend in the rate of area reduction over seven consecutive days, generating a trend dataset reflecting the accelerated healing phase. This method ensures accurate trend capture, providing a reliable foundation for subsequent analysis.

[0108] Specifically, feature extraction methods extract dynamic features from trend datasets, such as healing rate and tissue regeneration intensity.

[0109] In one embodiment, the extracted features include daily wound edge smoothness and blood flow rate of change. Calculations showed that blood flow significantly increased on day 5, indicating that healing had entered an active phase. The dynamic feature set, by quantifying these changes, clearly presents the dynamic evolution of the healing process, laying the foundation for fluctuation analysis.

[0110] In one possible implementation, fluctuation analysis identifies key fluctuation points in the feature sequence. If the dynamic feature set shows a sudden 20% increase in blood flow change rate on day 6, exceeding a preset range of 15%, it is marked as a key fluctuation point. By analyzing abnormal changes in multiple features, the fluctuation point set locates turning points in the healing process, such as the transition from the inflammatory phase to the proliferative phase, providing a basis for stage division.

[0111] Preferably, the clustering method divides the healing process into stages based on the set of fluctuation points. Assuming the set of fluctuation points includes characteristic changes on days 3 and 6, the healing process is divided into the inflammatory phase, the proliferative phase, and the remodeling phase using a clustering algorithm, generating a stage feature set. The stage feature set describes the characteristic distribution of each stage; for example, blood flow is consistently high during the proliferative phase, providing data for subsequent timing matching.

[0112] For example, sequence matching methods analyze the correlation between stage characteristics and intervention timing to generate timing association sets.

[0113] In one embodiment, matching revealed that phototherapy intervention was suitable for the high blood flow phase of the proliferative stage, generating a timing association set to clarify the intervention type and time point. This method uses pattern matching to accurately associate stage characteristics with intervention needs.

[0114] Understandably, the sorting method generates a structured instruction set based on a timing-related set. Assuming the timing-related set includes both phototherapy and drug intervention, the sorting algorithm generates an instruction sequence based on priority, such as phototherapy on day 7 and drug intervention on day 9. The structured instruction set clearly lists the time points and intervention types, facilitating clinical implementation.

[0115] It should be noted that the structured instruction set is output to the clinical decision support system to complete the intervention deployment.

[0116] For example, the instruction set is transmitted to the system via API, automatically generating an intervention plan, such as scheduling phototherapy on day 7, and the system reminds medical staff to execute it. This approach ensures efficient and precise intervention deployment, improving the efficiency of healing management.

[0117] S108. Obtain new multimodal data through the real-time data update module, repeat the above steps, and update the unified analysis model and intervention timing prediction results.

[0118] The real-time data update module acquires newly collected physiological signals and image data from multimodal data sources to generate the original dataset. If the integrity of the original dataset meets the preset standards, then the data cleaning method is used to remove noise and standardize the data format to obtain a cleaned dataset. Based on the cleaned dataset, feature extraction methods are used to calculate the temporal features of physiological signals and the spatial features of image data to generate a feature dataset. If the dimensionality of the feature dataset meets the preset range, then the structure of the feature dataset is optimized by dimensionality reduction methods to obtain an optimized feature set; By optimizing the feature set and using regression analysis, the parameters of the unified analysis model are updated to obtain the updated model. If the prediction error of the updated model is lower than the preset threshold, the trend of the optimized feature set is calculated by time series analysis to obtain the trend sequence. Based on the trend sequence, a classification method is used to identify key time points in the trend sequence and generate a time point set; If the distribution of the time point set conforms to the preset pattern, then the correlation between the time points and the intervention type is analyzed by sequence matching method to obtain the timing association set; By using the timing association set and a sorting method, intervention types are arranged according to the priority of time points to generate an intervention instruction sequence, resulting in a structured instruction set; Based on the structured instruction set, the instruction sequence is output to the clinical decision support system through data transmission methods to complete the intervention deployment.

[0119] Specifically, the real-time data update module acquires physiological signals and image data from multimodal data sources to generate the raw dataset.

[0120] For example, in clinical monitoring scenarios, the module can acquire heart rate signals from an electrocardiogram (ECG) device and lung image data from a CT scanner. An ECG acquires heart rate data 1000 times per second, and the CT image resolution is 512x512 pixels.

[0121] It should be noted that the data source must be synchronized to avoid timestamp discrepancies affecting subsequent analysis. The completeness of the original dataset must meet preset standards, such as a missing rate of less than 5%.

[0122] Specifically, if the heart rate signal is missing less than 50 points and the image data has no obvious artifacts, it is processed by data cleaning methods.

[0123] In one possible implementation, the cleaning method includes median filtering of the heart rate signal to remove noise and grayscale normalization of the image data to ensure pixel values ​​are within the range of 0-255. Cleaning the dataset ensures the accuracy of subsequent feature extraction. The feature extraction method computes temporal and spatial features from the cleaned dataset.

[0124] For example, the mean and standard deviation of the RR interval are extracted from heart rate signals, which are 800 milliseconds and 50 milliseconds, respectively; the edge sharpness and texture entropy of lung nodules are extracted from image data. The feature dataset dimension must be within a preset range, such as less than 100 features in total.

[0125] Preferably, principal component analysis is used for dimensionality reduction, retaining 95% of the variance to generate an optimized feature set, thereby reducing computational complexity. Regression analysis is then used to update the parameters of the unified analysis model.

[0126] In one embodiment, the model is a linear regression model that takes an optimized feature set as input and outputs predicted lung function parameters, such as forced vital capacity. The prediction error of the updated model must be less than 5%.

[0127] Understandably, low error indicates a strong ability of the model to adapt to new data. Time series analysis methods calculate the trend of changes in the optimized feature set; for example, a decrease in the RR interval from 800 milliseconds to 750 milliseconds within 24 hours indicates an increased heart rate. Classification methods identify key time points in the trend sequence.

[0128] For example, a random forest classifier detects the onset of increased heart rate, such as at hour 6. The distribution of the time point set must conform to a pattern, such as time point intervals greater than 1 hour. Sequence matching methods analyze the association between time points and intervention types.

[0129] In one possible implementation, if the heart rate increases in the 6th hour, the matched intervention type is oxygen flow adjustment. The timing-related set is sorted by time point priority using a sorting method, such as prioritizing oxygen adjustment in the 6th hour before drug intervention in the 8th hour. The data transmission method outputs the structured instruction set to the clinical decision support system.

[0130] For example, the instruction set is transmitted in JSON format, containing the time point 6:00 and the intervention type "increase oxygen flow to 3L / min". Transmission must ensure low latency, such as less than 100 milliseconds. This method ensures that instructions arrive at the system in a timely manner, optimizing intervention deployment efficiency.

[0131] S109. Based on the updated prediction results, a dynamic adjustment algorithm is used to optimize the intervention instruction set and generate the final intervention plan.

[0132] By using data acquisition methods, updated prediction results are obtained from the predictive analysis module to generate a prediction dataset. If the completeness of the prediction dataset meets the preset criteria, key prediction features are extracted using data filtering methods to obtain the feature prediction set. Based on the feature prediction set, a dynamic adjustment algorithm is used to optimize the parameters of the intervention instructions and generate an optimized instruction set. By using sequence analysis methods, the time distribution characteristics of the optimized instruction set are analyzed to obtain the instruction time series; If the distribution of the instruction time series conforms to the preset pattern, the execution order of the intervention instructions is adjusted by the priority sorting method to generate a sorted instruction set; Based on the sorting instruction set, a format conversion method is used to generate a structured final intervention plan, resulting in a plan dataset. The protocol dataset is output to the clinical decision support system via data transmission methods to complete protocol deployment.

[0133] Specifically, the updated prediction results are obtained from the predictive analysis module through data acquisition methods to generate a prediction dataset.

[0134] In one possible implementation, the predictive analytics module generates a health risk prediction for the next 24 hours based on the patient's real-time physiological data, such as heart rate and blood pressure.

[0135] For example, a patient's prediction results show a 70% probability of abnormal heart rate and a 60% probability of abnormal blood pressure. These prediction results are integrated into a prediction dataset through a data interface, which includes fields such as timestamps and risk probabilities.

[0136] It should be noted that the data acquisition method must ensure the real-time nature and completeness of the data, typically by using an API interface to periodically retrieve data. If the completeness of the prediction dataset meets the preset standards, key prediction features are extracted using data filtering methods to obtain a feature prediction set.

[0137] Preferably, the integrity criteria include a data missing rate of less than 5% and a time span covering at least 12 hours.

[0138] In one embodiment, the data filtering method focuses on high-risk predictive features, such as records with a heart rate abnormality probability exceeding 50%.

[0139] For example, for a patient's dataset, the probability values ​​of abnormal heart rate and blood pressure are retained after filtering to form a feature prediction set containing approximately 100 high-risk records. This filtering ensures that subsequent analysis focuses on key information. Based on the feature prediction set, a dynamic adjustment algorithm is used to optimize the parameters of the intervention instructions, generating an optimized instruction set.

[0140] Specifically, the dynamic adjustment algorithm adjusts the intensity and frequency of intervention instructions based on the risk level of the feature prediction set.

[0141] For example, for a patient with a 70% probability of heart rate abnormality, the algorithm might suggest ECG monitoring every 4 hours, while adjusting to every 6 hours when the probability is 50%. The optimized instruction set therefore includes parameters such as specific monitoring frequency and drug dosage, logically directly related to the predicted risk. Through sequence analysis, the temporal distribution characteristics of the optimized instruction set are analyzed to obtain the instruction time series.

[0142] In one embodiment, the sequence analysis method checks the uniformity of the distribution of the instruction set over 24 hours.

[0143] For example, an optimization instruction set contains 8 monitoring instructions. Analysis shows that their time intervals are 3-4 hours, which meets the requirement of uniform distribution. The instruction time series therefore records the execution time of each instruction, such as 08:00, 12:00, etc., for subsequent sorting. If the distribution of the instruction time series conforms to a preset pattern, the execution order of the intervention instructions is adjusted using a priority sorting method to generate a sorted instruction set.

[0144] Understandably, the default mode requires that the instruction interval not exceed 6 hours and that high-risk instructions be executed first.

[0145] For example, for a set of instructions that includes ECG monitoring and medication adjustments, the prioritization method places heart rate abnormality-related monitoring instructions first, ensuring execution at 08:00, while lower-priority blood pressure monitoring is scheduled for 12:00. The prioritized instruction set thus reflects a balance between time and risk. Based on the prioritized instruction set, a format conversion method is used to generate a structured final intervention plan, resulting in a plan dataset.

[0146] For example, the sorting instruction set is converted into an XML format to generate a structured scheme containing instruction type, execution time, and parameters, such as "ECG monitoring, 08:00, lasting 30 minutes". This format facilitates parsing and execution by clinical systems. The scheme dataset is then output to the clinical decision support system via data transmission methods to complete scheme deployment.

[0147] In one possible implementation, the data transmission method uses the encrypted HTTP protocol to push the solution dataset to the hospital's decision support system.

[0148] For example, after receiving a dataset containing eight instructions, a hospital system automatically assigns it to the corresponding medical staff's workstations to ensure timely intervention. This approach guarantees the real-time nature and security of the intervention plan.

[0149] This invention provides a wound healing intervention system based on multimodal data, mainly comprising: The multimodal data acquisition module is used to acquire raw data from wound image sequences, inflammatory marker concentrations, and tissue regeneration indicators, and generate a multidimensional dataset containing timestamps. The feature extraction module is used to perform edge detection and texture analysis on wound images using feature extraction algorithms, and to perform normalization processing on inflammatory markers and tissue regeneration indicators to obtain a standardized feature vector set; The time series analysis module is used to calculate the changing trend of each feature over time and generate dynamic change curves if the time dimension of the standardized feature vector set is complete. The threshold judgment module is used to analyze the slope and fluctuation amplitude of the curve based on the dynamic change curve and a preset threshold judgment algorithm to determine the key turning point of the healing process. The deep learning module is used to train a multimodal fusion model by taking a dynamic change curve and key inflection points as input through a deep learning network, generating a unified analysis model, and outputting the healing process prediction results. The time series prediction module is used to analyze the healing process output by the unified analysis model and determine the best intervention time if the confidence level of the prediction result is higher than a preset threshold. The instruction generation module is used to generate a structured instruction set containing time points and intervention types for intervention timing, and output it to the clinical decision support system to complete the precise intervention deployment; The real-time data update module is used to acquire new multimodal data, repeat the above steps, and update the unified analysis model and intervention timing prediction results. The dynamic adjustment module is used to optimize the intervention instruction set and generate the final intervention plan based on the updated prediction results using a dynamic adjustment algorithm.

[0150] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. The present invention has been described in detail with reference to preferred embodiments. Those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications and substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A wound healing intervention method based on multimodal data, characterized in that, The method includes: S101. Through the multimodal data acquisition module, raw data are obtained from wound image sequences, inflammatory marker concentrations, and tissue regeneration indicators to generate a multidimensional dataset containing timestamps; S102. Perform edge detection and texture analysis on the wound image, and perform normalization processing on inflammatory markers and tissue regeneration indicators to obtain a standardized feature vector set; S103. If the time dimension of the standardized feature vector set is complete, then the time series analysis algorithm is used to calculate the changing trend of each feature over time and generate a dynamic change curve. S104. Based on the dynamic change curve, analyze the slope and fluctuation amplitude of the curve to determine the key turning points in the healing process. S105. Input the dynamic change curve and key inflection points, train the multimodal fusion model, generate a unified analysis model, and output the healing process prediction results. S106. If the confidence level of the prediction result is higher than the preset threshold, analyze the healing process output by the unified analysis model to determine the best time for intervention. S107. Generate a structured instruction set containing time points and intervention types for the timing of intervention, and output it to the clinical decision support system to complete the deployment of precise intervention.

2. The wound healing intervention method based on multimodal data according to claim 1, characterized in that, The generation of the timestamp-containing multidimensional dataset includes: The data acquisition module obtains multimodal data from wound image sequences, inflammatory marker concentrations, and tissue regeneration indicators to obtain a raw dataset with timestamps. Image processing techniques were used to preprocess the wound image sequence, including denoising and segmentation, to obtain clear wound feature images; If the boundary clarity of the wound feature image is lower than a preset threshold, then an edge enhancement algorithm is applied to obtain an optimized wound feature dataset. By using time series analysis, trend features of inflammatory marker concentrations and tissue regeneration indicators over time were extracted to obtain a dynamic feature set. Based on the dynamically changing feature set, a clustering algorithm is applied to classify the multimodal data, determine the wound healing stage, and obtain the healing stage label; If the healing stage label is inconsistent with the preset standard, a comprehensive feature dataset is obtained by integrating the wound feature dataset and the dynamic change feature dataset through data integration processing. A regression analysis model was used to predict the future trend of wound healing on a comprehensive feature dataset, and the prediction results were obtained.

3. The wound healing intervention method based on multimodal data according to claim 1, characterized in that, The standardized feature vector set obtained includes: An edge detection algorithm is used to process the wound image, extract boundary features, and combine them with texture analysis to generate a wound surface feature set; Normalization was applied to the concentrations of inflammatory markers and tissue regeneration indicators to generate a standardized feature vector set, thus obtaining the initial feature set. Based on the initial feature set, the principal component analysis algorithm is used to reduce the dimensionality of the wound surface feature set and the standardized feature vector set, and the main feature components are extracted to obtain the dimensionality-reduced feature set. If the variance of the feature components in the dimensionality-reduced feature set is lower than a preset threshold, then a feature selection algorithm is applied to filter highly relevant features to obtain an optimized feature set. Based on the optimized feature set, time series analysis was used to extract trends from the standardized feature vectors of inflammatory markers and tissue regeneration, resulting in a dynamic trend feature set. If the time series fluctuation of the dynamic trend feature set exceeds a preset threshold, the dynamic trend feature set is smoothed by the sliding window method to obtain a smoothed trend feature set. Based on the smooth trend feature set, a clustering algorithm is used to classify the wound surface features and dynamic trend features, determine the wound healing stage, and obtain the healing stage label; Based on the healing stage labels, a regression analysis model is used to predict the smooth trend feature set, determine the wound healing trend, and obtain the healing trend prediction results.

4. The wound healing intervention method based on multimodal data according to claim 1, characterized in that, The generated dynamic change curve includes: By processing the standardized feature vector set through the time series decomposition algorithm, the changing trend of each feature over time is extracted to obtain a set of dynamic change curves; If the trend component of the dynamically changing curve set is significant, the sliding window method is used to smooth the curve set to obtain a smooth curve set. Based on the set of smooth curves, calculate the periodic components of each feature, extract the periodic features, and obtain the set of periodic features; If the amplitude of the periodic feature set exceeds a preset threshold, the periodic feature set is analyzed in the frequency domain by Fourier transform to obtain the frequency domain feature set. Based on the frequency domain feature set, a clustering algorithm is used to group the features, determine the correlation between the features, and obtain the feature group set. By grouping features into sets, statistical indicators of each set of features are extracted to generate a feature statistics set. Based on the feature statistics set, a regression analysis model is used to predict the trend of feature changes over time, resulting in a trend prediction set.

5. The wound healing intervention method based on multimodal data according to claim 1, characterized in that, The key turning points in determining the healing process include: Using time series data and the sliding window method, the local slope and fluctuation amplitude of the dynamic change curve are calculated to obtain the slope sequence and amplitude sequence. If the absolute value of the slope in a continuous time window in the slope sequence exceeds a preset threshold, the slope mutation point is determined by difference calculation to obtain the mutation point set. Based on the set of mutation points, a threshold judgment algorithm is used to analyze the fluctuation amplitude corresponding to the mutation point. If the fluctuation amplitude exceeds the amplitude threshold, it is determined as a candidate turning point, and a set of candidate turning points is obtained. By using the candidate inflection point set, a density clustering algorithm is employed to merge candidate inflection points that are temporally adjacent, thus obtaining an optimized inflection point set. Based on the optimized inflection point set, time series segments corresponding to the inflection points are extracted from the dynamic change curve, and the trend characteristics of the segments are calculated to obtain the trend feature set. By using trend feature sets and employing time series segmentation algorithms, the healing process is divided into stages, resulting in a process stage set. Based on the process stage set, calculate the mean slope and amplitude distribution of each stage to obtain the stage feature set.

6. The wound healing intervention method based on multimodal data according to claim 1, characterized in that, The output healing process prediction results include: Time series analysis is used to obtain time series data from dynamic change curves. The sliding window method is then used to calculate local features and obtain feature sequences. Based on the feature sequence, key inflection points are extracted. If the feature values ​​of a continuous window in the feature sequence exceed a preset threshold, the inflection point is determined by difference calculation, and the inflection point set is obtained. By integrating feature sequences and inflection point sets through multimodal fusion, and using a deep learning network for feature mapping, a fused feature set is obtained. Based on the fused feature set, a deep learning network is used to train the model, and the parameters are iteratively optimized to obtain a healing process prediction model. By using a healing process prediction model, inputting a dynamic change curve, generating prediction results, and obtaining a healing process prediction sequence; Based on the predicted sequence of healing process, a trend feature extraction method is used to calculate the trend features of the predicted sequence and obtain a trend feature set. By using the trend feature set and a time series segmentation algorithm, the healing process is divided into stages, resulting in a process stage set.

7. The wound healing intervention method based on multimodal data according to claim 1, characterized in that, Determining the optimal timing for intervention includes: If the confidence level of the predicted sequence is higher than the preset threshold, the predicted sequence of the healing process is analyzed by time series prediction algorithm to obtain the trend change set. Based on the trend change set, feature extraction methods are used to calculate the dynamic changes of sequence features and obtain a dynamic feature set; If the feature values ​​in the dynamic feature set exceed the preset range, the key fluctuation points of the feature sequence are determined by difference calculation to obtain the fluctuation point set. Based on the set of fluctuation points, a clustering algorithm is used to divide the healing process into stages, resulting in a set of stage features. If the distribution of the stage feature set conforms to the preset pattern, then the correlation between the stage features and the intervention timing is analyzed by the sequence matching method to obtain the timing matching set; Based on the timing matching set, a sorting algorithm is used to determine the priority sequence of the best intervention timing, thus obtaining the intervention timing set; By using the intervention timing set, an intervention time series of the healing process is generated, resulting in an intervention sequence.

8. The wound healing intervention method based on multimodal data according to claim 1, characterized in that, The needle's precise intervention deployment includes: If the confidence level of the predicted sequence is higher than the preset threshold, time series analysis is used to generate a trend sequence of the healing process, thus obtaining a trend dataset. Based on the trend dataset, feature extraction methods are used to calculate the changes in dynamic features, generate feature sequences, and obtain a dynamic feature set. If the feature values ​​in the dynamic feature set exceed the preset range, the key fluctuation points of the feature sequence are calculated using the fluctuation analysis method to obtain the fluctuation point set. Based on the set of fluctuation points, a clustering method is used to divide the healing process into stages, generate a stage feature sequence, and obtain a stage feature set; If the distribution of the stage feature set conforms to the preset pattern, then the correlation between stage features and intervention timing is analyzed by sequence matching method to generate timing association set; Based on the timing association set, a sorting method is used to generate a sequence of instructions containing time points and intervention types, resulting in a structured instruction set; The structured instruction set is output to the clinical decision support system to complete the intervention deployment.

9. A wound healing intervention method based on multimodal data according to claim 1, characterized in that, Also includes: By using the real-time data update module to acquire new multimodal data, repeat the above steps S101~S107 to update the unified analysis model and intervention timing prediction results; Based on the updated prediction results, a dynamic adjustment algorithm is used to optimize the intervention instruction set and generate the final intervention plan.

10. A wound healing intervention system based on multimodal data, characterized in that, The system includes: The multimodal data acquisition module is used to acquire raw data from wound image sequences, inflammatory marker concentrations, and tissue regeneration indicators, and generate a multidimensional dataset containing timestamps. The feature extraction module is used to perform edge detection and texture analysis on wound images using feature extraction algorithms, and to perform normalization processing on inflammatory markers and tissue regeneration indicators to obtain a standardized feature vector set; The time series analysis module is used to calculate the changing trend of each feature over time and generate dynamic change curves if the time dimension of the standardized feature vector set is complete. The threshold judgment module is used to analyze the slope and fluctuation amplitude of the curve based on the dynamic change curve and a preset threshold judgment algorithm to determine the key turning point of the healing process. The deep learning module is used to train a multimodal fusion model by taking a dynamic change curve and key inflection points as input through a deep learning network, generating a unified analysis model, and outputting the healing process prediction results. The time series prediction module is used to analyze the healing process output by the unified analysis model and determine the best intervention time if the confidence level of the prediction result is higher than a preset threshold. The instruction generation module is used to generate a structured instruction set containing time points and intervention types for intervention timing, and output it to the clinical decision support system to complete the precise intervention deployment; The real-time data update module is used to acquire new multimodal data, repeat the above steps, and update the unified analysis model and intervention timing prediction results. The dynamic adjustment module is used to optimize the intervention instruction set and generate the final intervention plan based on the updated prediction results using a dynamic adjustment algorithm.

Images