A medical beauty postoperative repair effect evaluation method based on big data

Abstract

The application discloses a medical beauty postoperative repair effect evaluation method based on big data, and belongs to the technical field of medical evaluation, and specifically comprises the following steps: extracting a current feature vector containing skin texture, color parameter and physiological index after time sequence alignment and denoising and cleaning; inputting the current feature vector into a pre-trained feature mapping model to convert it into a standardized recovery representation code that eliminates individual physiological differences and has a unified dimension; fitting a multi-dimensional dynamic benchmark recovery curve set of the object based on the operation type label and the preoperative benchmark feature of the target object; calculating the feature deviation degree of the standardized recovery representation code and the multi-dimensional dynamic benchmark recovery curve set at the current time sequence node, combining the weight to generate a comprehensive evaluation index of the quantitative repair state; identifying abnormal recovery feature dimensions according to the comprehensive evaluation index, automatically associating the repair scheme knowledge base to match the corresponding intervention measures and generating a repair effect evaluation report.

Description

Technical Field

[0001] This invention relates to the field of medical assessment technology, specifically to a method for evaluating the post-medical aesthetic repair effect based on big data. Background Technology

[0002] With the booming development of the medical aesthetics industry and the increasing sophistication of public aesthetic awareness, post-operative care, as a crucial link in ensuring the effectiveness of cosmetic surgery and reducing the risk of complications, is playing an increasingly important role in the entire medical aesthetics service process. In current clinical practice, post-operative management systems typically cover the entire process from wound healing monitoring and physiological indicator tracking to final result acceptance. To effectively track the recovery progress of patients, modern medical institutions have generally established electronic medical record (EMR)-based file management systems. These systems utilize high-definition digital imaging equipment to regularly collect post-operative images and combine this with basic physiological monitoring methods to record key vital signs during the recovery period. These measures aim to provide patients with post-operative care guidance and phased recovery feedback through continuous information recording, thereby ensuring the overall quality and safety of medical aesthetics services to a certain extent and providing basic data retention for the prevention of medical disputes.

[0003] However, while existing postoperative management models meet basic information recording needs to a certain extent, they still face significant challenges in terms of accurate assessment and intelligent decision-making. The current problem is that existing medical aesthetic postoperative repair assessments mainly rely on visual observation or discrete qualitative feedback from medical staff, lacking objective, standardized, multi-dimensional data analysis models. This assessment method, highly dependent on human experience, is limited by individual differences and cognitive biases among the assessors, leading to significant dispersion and fluctuation in assessment conclusions for the same postoperative condition across different time dimensions or among assessors, making it difficult to establish unified quality control standards. Simultaneously, conventional systems have not yet deeply integrated big data mining and pattern recognition technologies, failing to extract features and transform value from massive amounts of historical postoperative data. This results in static and fixed assessment logic, unable to adaptively iterate and dynamically optimize based on individual recovery patterns and data accumulation, severely restricting the accuracy of assessment results and the ability to generate personalized repair plans. Summary of the Invention

[0004] The purpose of this invention is to provide a method for evaluating the repair effect after cosmetic surgery based on big data, thereby solving the problems in the background technology:

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A big data-based method for evaluating the effectiveness of post-medical aesthetic repair includes the following steps:

[0007] S1: Collect postoperative multimodal data of the target object, and extract the current feature vector containing skin texture, color parameters and physiological indicators after time alignment and noise removal;

[0008] S2: Input the current feature vector into the pre-trained feature mapping model and transform it into a standardized recovery representation code that eliminates individual basic physiological differences and has a unified dimension;

[0009] S3: Based on the surgical type label and preoperative baseline characteristics of the target object, retrieve similar case groups with satisfactory recovery results from the historical database, and fit and generate a set of multidimensional dynamic baseline recovery curves for the object.

[0010] S4: Calculate the feature deviation between the standardized recovery characterization code and the multidimensional dynamic benchmark recovery curve set at the current time node, and generate a comprehensive evaluation index for the quantitative repair status by combining the weights;

[0011] S5: Identify abnormal recovery feature dimensions based on the comprehensive evaluation index, automatically associate with the repair solution knowledge base to match corresponding intervention measures and generate a repair effect evaluation report.

[0012] As a further aspect of the present invention: In step S1, the process of collecting postoperative multimodal data of the target object and extracting the current feature vector containing skin texture, color parameters, and physiological indicators after temporal alignment and noise reduction is as follows:

[0013] High-definition image acquisition equipment was used to acquire postoperative skin texture and color related image data of the target subject, and postoperative physiological index data was collected through physiological sensing equipment. All data were collected synchronously with postoperative timestamps.

[0014] The collected multimodal data were time-aligned based on timestamps, and a smoothing filter method was used to remove noise from image data and physiological index data while retaining valid data information.

[0015] Skin texture features and color parameters are extracted from temporally aligned and denoised image data, and physiological indicators are extracted from processed physiological data. These are then integrated to construct the current feature vector.

[0016] As a further aspect of the present invention: in step S2, the process of inputting the current feature vector into the pre-trained feature mapping model and transforming it into a standardized recovery representation encoding that eliminates individual basic physiological differences and has a unified dimension is as follows:

[0017] Retrieve the preoperative baseline feature vector of the target object, perform element-by-element subtraction between the current feature vector and the preoperative baseline feature vector, and generate a difference feature vector reflecting the postoperative changes.

[0018] The differential feature vectors are loaded into a pre-trained feature mapping model, and the differential feature vectors are mapped to a high-dimensional feature space through a fully connected layer in the model to obtain latent feature vectors.

[0019] A layer normalization function is used to process the latent feature vector, transforming the numerical distribution of each dimension of the vector into a standard normal distribution form, and outputting a standardized recovery representation code.

[0020] As a further aspect of the present invention: the specific method for processing the latent feature vector using a layer normalization function to transform the numerical distribution of each dimension of the vector into a standard normal distribution form, and outputting the standardized recovery representation code, is as follows:

[0021] Calculate the arithmetic mean of all dimensions in the latent feature vector, and calculate the variance of the latent feature vector based on the sum of squared differences between the arithmetic mean and each dimension value;

[0022] The centered value is obtained by subtracting the arithmetic mean from the value of each dimension of the latent feature vector. The standard deviation is obtained by adding a smoothing term to the variance to prevent the denominator from being zero and then taking the square root. The normalized value is obtained by calculating the ratio of the centered value to the standard deviation.

[0023] Arrange all normalized values ​​in sequence to construct a standardized recovery characterization code. This code has the statistical properties of zero mean and unit variance, and outputs the standardized recovery characterization code.

[0024] As a further aspect of the present invention: In step S3, the process of retrieving similar case groups with satisfactory recovery results from a historical database based on the surgical type label and preoperative baseline characteristics of the target object, and fitting and generating a set of multidimensional dynamic baseline recovery curves for the object, is as follows:

[0025] Extract the surgical procedure category and preoperative baseline feature vector of the target object, and select a set of high-quality cases with consistent surgical procedure categories and historical final repair scores higher than the preset threshold from the historical database.

[0026] Calculate the Euclidean distance between the preoperative baseline feature vector of the target object and the preoperative features of each case in the high-quality case set, select the preset number of cases with the closest distance and their associated time-series standardized coding data, and construct a similar case group;

[0027] For each feature dimension of the time-series standardized coded data in similar case groups, a time-feature numerical coordinate system is constructed to map the data of all cases in the corresponding dimension to the coordinate axis.

[0028] The nonlinear least squares method is used to perform regression fitting operations on discrete data points on the coordinate axes, and the coefficients of the fitting function are calculated.

[0029] Based on the coefficients of the fitted function, function curves of each dimension that change continuously over time are generated, and the function curves of all dimensions are combined to form a set of multidimensional dynamic benchmark recovery curves.

[0030] As a further aspect of the present invention: the specific method for calculating the coefficients of the fitting function by performing regression fitting operations on discrete data points on the coordinate axes using the nonlinear least squares method is as follows:

[0031] Extract the postoperative recovery value and corresponding postoperative time node of each case in the similar case group. With the postoperative time node as the independent variable on the horizontal axis and the postoperative recovery value as the dependent variable on the vertical axis, merge the data of all cases to construct a discrete point set matrix under a unified coordinate system.

[0032] A nonlinear exponential function containing undetermined parameters is selected as the preset regression model structure. The numerical values ​​of the discrete point set matrix are substituted into the regression model structure to construct a residual sum of squares function that characterizes the deviation between the observed values ​​in the discrete point set matrix and the theoretical values ​​calculated by the regression model structure.

[0033] The Levenberg-Marquardt iterative algorithm is used to perform a minimization operation on the residual sum of squares function. The values ​​of the undetermined parameters are updated iteratively until the residual sum of squares function meets the convergence condition, and the coefficients of the finally determined fitting function are output.

[0034] As a further aspect of the present invention: in step S4, the process of calculating the feature deviation between the standardized recovery characterization code and the multidimensional dynamic benchmark recovery curve set at the current time-series node, and combining the weights to generate a comprehensive evaluation index for the quantified repair state, is as follows:

[0035] Based on the current timeline, extract the corresponding theoretical benchmark values ​​from each dimension curve of the multidimensional dynamic benchmark recovery curve set, and construct the theoretical benchmark vector;

[0036] Calculate the absolute value of the difference between the standardized recovery representation code and the theoretical baseline vector in the corresponding dimension, and generate the feature deviation vector;

[0037] The feature weight vectors based on the clinical importance of each dimension are retrieved, and a weighted summation operation is performed on the feature deviation vector and the feature weight vector to calculate the weighted deviation value that represents the overall degree of difference.

[0038] Substitute the weighted deviation value into the preset negative exponential transformation function model, calculate the mapping score of the value within the standard scoring range, and output a comprehensive evaluation index that quantifies the repair status.

[0039] As a further aspect of the present invention: in step S5, the process of identifying abnormal recovery feature dimensions based on the comprehensive evaluation index, automatically associating with the repair solution knowledge base to match corresponding intervention measures, and generating a repair effect evaluation report is as follows:

[0040] The comprehensive evaluation index is compared with the preset normal threshold. If it is determined that the standard is not met, the feature deviation vector is traversed and the dimension whose value exceeds the local warning line is locked as the anomaly recovery feature dimension.

[0041] Input the abnormal recovery feature dimension into the repair plan knowledge base, retrieve the clinical intervention strategy for that dimension through predefined mapping rules, and obtain the corresponding intervention measures;

[0042] The comprehensive evaluation index, abnormal recovery feature dimensions, and intervention measures are filled into a preset document template. A structured text description is constructed using a natural language generation algorithm to generate a repair effect evaluation report.

[0043] The present invention has the following beneficial effects:

[0044] This invention achieves objectivity and standardization in the assessment of post-medical aesthetic repair, effectively avoiding individual differences and cognitive biases inherent in manual assessment. By collecting multimodal data on skin texture, color parameters, and physiological indicators, and extracting comprehensive feature vectors through temporal alignment and noise reduction, this data is then transformed into standardized codes that eliminate individual differences using a pre-trained model, constructing a unified assessment benchmark. The resulting comprehensive assessment index quantifies the multidimensional recovery status into precisely comparable numerical values, overcoming the ambiguity of traditional visual observation and qualitative feedback. This ensures that assessment conclusions for the same post-operative condition from different time dimensions and assessment subjects tend to be consistent, providing core technical support for establishing unified quality control standards for medical aesthetic repair.

[0045] This invention leverages big data mining to achieve dynamic adaptive assessment, significantly improving assessment accuracy and personalized intervention capabilities. By retrieving high-quality historical similar cases and fitting a multidimensional dynamic benchmark curve, the assessment logic breaks free from static and rigid limitations, adaptively matching the recovery patterns of different surgical types and individual preoperative baselines. Combining feature deviation weighted calculation and abnormal dimension identification, it can accurately locate specific stages of recovery abnormalities. Then, through a repair plan knowledge base, it automatically matches intervention measures, achieving an upgrade from "generalized assessment" to "personalized precision guidance." This not only solves the pain point of traditional assessments being unable to adapt to individual recovery differences, but also provides early warning of potential recovery risks and offers targeted solutions, greatly improving the controllability of postoperative repair outcomes. Furthermore, the continuous accumulation of historical data enables iterative optimization of the assessment logic. Attached Figure Description

[0046] The invention will now be further described with reference to the accompanying drawings.

[0047] Figure 1 This is a flowchart illustrating a method for evaluating the post-medical aesthetic repair effect based on big data, according to the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Please see Figure 1 As shown, this invention is a method for evaluating the repair effect after cosmetic surgery based on big data, including the following steps:

[0050] S1: Collect postoperative multimodal data of the target object, and extract the current feature vector containing skin texture, color parameters and physiological indicators after time alignment and noise removal;

[0051] S2: Input the current feature vector into the pre-trained feature mapping model and transform it into a standardized recovery representation code that eliminates individual basic physiological differences and has a unified dimension;

[0052] S3: Based on the surgical type label and preoperative baseline characteristics of the target object, retrieve similar case groups with satisfactory recovery results from the historical database, and fit and generate a set of multidimensional dynamic baseline recovery curves for the object.

[0053] S4: Calculate the feature deviation between the standardized recovery characterization code and the multidimensional dynamic benchmark recovery curve set at the current time node, and generate a comprehensive evaluation index for the quantitative repair status by combining the weights;

[0054] S5: Identify abnormal recovery feature dimensions based on the comprehensive evaluation index, automatically associate with the repair solution knowledge base to match corresponding intervention measures and generate a repair effect evaluation report.

[0055] In one embodiment of the present invention, the process of collecting postoperative multimodal data of the target object in step S1, and extracting the current feature vector containing skin texture, color parameters and physiological indicators after time alignment and noise reduction is as follows:

[0056] During the procedure, a high-resolution industrial-grade CMOS image acquisition device, coupled with a shadowless ring light source, is used to capture multi-angle images of the surgical site. The shadowless light source is employed to effectively eliminate shadow interference caused by uneven ambient light, ensuring that subtle textures and color changes on the skin surface are accurately recorded. The image acquisition device continuously captures multiple original images containing red, green, and blue channel information to cover the skin's reflectivity at different wavelengths. Simultaneously, a patch-type skin resistance sensor and an infrared temperature probe are used, closely attached to the healthy skin surrounding the surgical area. These sensors can detect subtle changes in current and thermal radiation intensity on the skin surface in real time. Because changes in skin moisture directly affect conductivity, and inflammation leads to increased local temperature, these physical quantities can indirectly reflect the physiological state of subcutaneous tissue. During acquisition, all hardware devices are connected to the same high-precision clock source. Whenever the image sensor captures a frame or records a set of values, it is instantly marked with the current date (year, month, day, hour, minute, second, millisecond). This synchronous acquisition method based on a unified switching clock source ensures a strict temporal correspondence between visual and physiological data, providing a precise spatiotemporal reference for subsequent data fusion.

[0057] The collected raw data is first processed according to millisecond-level time stamps. Since the acquisition frequency of image data is usually lower than that of physiological sensor data, the acquisition time of the image data needs to be used as the main reference point. For multiple physiological data points between two image acquisition times, a linear interpolation method is used to calculate the values ​​that precisely correspond to the image times. This interpolation process utilizes the changing trend between two adjacent known data points to estimate the values ​​at intermediate times, thereby filling the time gap caused by sampling rate differences and achieving strict alignment of multimodal data. After alignment, noise reduction processing is performed on the interference signals in the data. For image data, Gaussian smoothing filtering is applied to the image data. The prime matrix is ​​convolved; this process reassigns the center pixel by calculating the weighted average of each pixel and its surrounding neighboring pixels; because noise usually manifests as high-frequency random abrupt changes while texture details have a certain continuity, this weighted average can effectively smooth out isolated noise points while preserving the main contour edges; for physiological index data, a moving average filtering method is used; a fixed-length time window is set and the arithmetic mean of all sampled values ​​is calculated within the window as the effective value at the center time; as the window moves continuously on the time axis, instantaneous spike interference caused by device jitter or poor contact can be filtered out, thereby restoring the true trend of physiological index changes.

[0058] After data preprocessing, key pathological features are analyzed from the images. First, the color images are converted to grayscale, and texture features are extracted based on the gray-level co-occurrence matrix (GLCM) principle. This principle reflects the roughness and directionality of the texture by statistically analyzing the probability of grayscale value combinations of pixel pairs with specific distances and orientations in the image. For example, the contrast value in the matrix can quantify the depth of grooves on the skin surface, i.e., the unevenness of scars. Simultaneously, the image is converted from the red-green-blue color space to the lightness-chrominance color space. This space conversion is necessary because the original color space is greatly affected by light intensity, while the lightness-chrominance space can separate the lightness component from the color component. The red-green color space is then extracted. The chromaticity component values ​​can accurately reflect the degree of redness and swelling of the skin, i.e., the distribution of hemoglobin, without being affected by the intensity of ambient light. At the same time, statistical characteristics are calculated from the processed physiological data sequence. The mean and variance of temperature and resistance data within a time window are calculated respectively. The mean represents the current physiological baseline level, while the variance reflects the stability of the physiological state. Finally, the extracted texture contrast values, chromaticity component values, and the mean and variance of physiological indicators are concatenated in a predetermined order. This concatenation method maps information from different physical dimensions into the same multidimensional mathematical space, thereby constructing a feature vector that can comprehensively characterize the current postoperative recovery state of the target object.

[0059] In one embodiment of the present invention, step S2, which involves inputting the current feature vector into a pre-trained feature mapping model and transforming it into a standardized recovery representation encoding that eliminates individual basic physiological differences and has a unified dimension, is as follows:

[0060] The process begins by retrieving the baseline feature vector of the target individual from the electronic database, based on the individual's unique identifier. This baseline feature vector includes the individual's original skin tone, skin texture roughness, and physiological parameters related to basal metabolism. The computational unit aligns the current feature vector, which is collected and constructed in real time, with this historical baseline feature vector and performs element-wise subtraction. Specifically, it subtracts the corresponding dimension value from the baseline vector for each dimension of the current vector. For example, subtracting the pre-operative recorded inherent redness value from the currently collected skin redness value yields a difference. The principle behind this step is to eliminate inherent physiological differences between individuals. Because different individuals have different skin tones and constitutions, directly using absolute values ​​for evaluation could lead to misjudging naturally redder skin as a post-operative inflammatory reaction. The difference feature vector generated through this difference calculation only retains the incremental changes brought about by surgical trauma and the recovery process. This ensures that the data purely reflects the dynamic evolution of post-operative tissues without being interfered with by the individual's innate genetic characteristics or long-term physiological habits, thus providing a purified data foundation for objective evaluation.

[0061] After obtaining the differential feature vectors, they are input into a pre-trained deep neural network structure for deep feature extraction. When the data flows through the fully connected layers inside the model, the value of each input node is weighted and summed with the weight parameters connected to the next layer node. These weight parameters are continuously adjusted and optimized through backpropagation during training with a large amount of historical medical data. They represent the importance and combination relationship of different input features in judging the recovery status. Through this weighted summation and the superposition of bias terms, the differential features that were originally in a low-dimensional physical space are projected into a high-dimensional abstract feature space to form potential feature vectors. The reason for this high-dimensional mapping is that a single physical indicator often cannot directly represent the complex pathological repair logic. For example, a simple increase in temperature or a simple change in color may have limited meaning, but when the two are combined in a certain proportion, they may correspond to specific recovery states such as subcutaneous effusion or infection. The role of the fully connected layer is to capture these implicit correlations that are difficult to observe with the naked eye and exist between multiple indicators through nonlinear combination, thereby mining the potential semantic information that can deeply represent the microscopic state of tissue repair.

[0062] A layer normalization function is used to process the latent feature vectors, transforming the numerical distribution of each dimension of the vector into a standard normal distribution, and outputting a standardized recovery representation code. The specific method is as follows:

[0063] In the stage of statistical feature extraction for latent feature vectors, the first step is to determine the overall data distribution center of the vector. The computing unit iterates through and reads the floating-point values ​​stored in each dimension of the latent feature vector. These values ​​represent the abstract feature strengths that are mixed with skin texture, color, and physiological indicators after high-dimensional mapping. The values ​​of all dimensions in the vector are summed by accumulation to obtain a sum, and then the sum is divided by the total number of dimensions of the vector to obtain the arithmetic mean. This arithmetic mean physically represents the average activation level of all feature dimensions of the target object at the current moment, i.e., the centroid of the feature distribution. Next, in order to quantify the dispersion of the feature data relative to this centroid, the variance needs to be calculated. The computing unit iterates through each value in the vector again and calculates the difference between it and the arithmetic mean. In order to eliminate the influence of positive and negative signs and amplify the weight of larger deviations, these differences need to be squared. The squares of all the calculated differences are summed and the result is divided by the total number of dimensions to obtain the variance. This variance value objectively reflects the degree of fluctuation of the data within the feature vector, i.e., it indicates whether the restored features of the object tend to be stable and consistent or have extreme abnormal fluctuation points.

[0064] After obtaining the statistical baseline, the raw data undergoes numerical transformation to achieve centering and scaling. First, centering is performed by subtracting the arithmetic mean calculated in the previous step from the raw value of each dimension. This process shifts the distribution center of the entire feature data to the zero point of the coordinate axis, eliminating the overall offset caused by excessively high or low individual baseline values. Next, the standardized denominator is constructed. To avoid mathematical errors such as division by zero that could cause program crashes, a very small positive number is added to the calculated variance as a smoothing term. This smoothing term is typically... The values ​​are so small that they do not substantially affect the accuracy of the data but ensure numerical stability. Then, the square root of the variance after adding the smoothing term is taken to obtain the standard deviation. The standard deviation represents the average span of the data distribution. Finally, a normalization division operation is performed, that is, each value after centering is divided by the calculated standard deviation. The essence of this step is to scale the data of all dimensions proportionally. Whether the original data is tens of degrees Celsius of skin temperature or a few tenths of a millimeter of texture depth, after this division operation, they are all converted into relative values ​​in terms of standard deviation, thereby eliminating the order of magnitude difference between different physical dimensions.

[0065] Finally, the numerical values ​​processed by the above mathematical transformations are rearranged according to their index order in the original vector. This process strictly preserves the topological positional relationship between each feature dimension, ensuring that the semantic structure of the features is not destroyed. A new standardized recovery representation code is constructed through this sequential arrangement. Statistically, this code is characterized by the mean of all dimension values ​​being strictly equal to zero and the overall variance being equal to one. This specific statistical distribution means that the data has been mapped to a standard normal distribution space. In this space, the magnitude of the value no longer represents an absolute physical quantity but rather the degree of deviation of the feature from its own average level. For example, a value of zero indicates that the feature is at the average level, while a positive value indicates that it is above the average level. Outputting this code with uniform statistical properties provides a fair measurement basis for subsequent similarity retrieval. Because the masking effect of large numerical features on small numerical features is eliminated when calculating Euclidean distance, it ensures that changes in each dimension can be measured equally, thereby improving the accuracy of matching similar cases in big data.

[0066] In one embodiment of the present invention, the process of retrieving similar case groups with satisfactory recovery results from a historical database based on the surgical type label and preoperative baseline characteristics of the target object, and fitting and generating a set of multidimensional dynamic baseline recovery curves for the object, is as follows:

[0067] In the initial stage of establishing the benchmark model, the computing unit first reads the surgical procedure category label and the preoperative benchmark feature vector collected and stored before the surgery from the target subject's electronic medical record. The surgical procedure category label clarifies the specific cosmetic procedure the subject underwent, such as double eyelid surgery or rhinoplasty, while the preoperative benchmark feature vector digitally records the subject's original innate physiological conditions such as skin elastic modulus, subcutaneous fat thickness, and basal metabolic rate. The computing unit uses this information as a retrieval index to access a historical database. This database stores a massive amount of past case data, but not all data is of reference value. To ensure the accuracy of the generated benchmark curve... Representing an ideal recovery state rather than a mediocre or failed repair process, the calculation unit executes a strict screening logic. First, it filters out all cases whose surgical procedures are inconsistent with the current target, retaining only data records of similar surgeries. Next, it reads the historical final repair scores of the remaining cases. These scores are objective scores given by professional physicians based on clinical outcomes after the recovery period of past cases. The calculation unit compares these scores with a preset high-score threshold, selecting only those high-quality cases with scores above the threshold. The principle behind this screening step is to eliminate noisy data with many complications or poor recovery results, thereby defining a reference sample pool composed of successful cases.

[0068] After identifying the high-quality case set, it is necessary to further find the precise reference group with the smallest individual differences from the target object. The computing unit retrieves the preoperative baseline feature vector of the target object as the anchor point and retrieves the preoperative feature data of each case in the high-quality case set one by one. Although all cases belong to the same type of surgery and have recovered well, the differences in physical condition among individuals can have a huge impact on the recovery speed, so it is necessary to perform distance measurement in the feature space. The computing unit uses the Euclidean distance algorithm to calculate the geometric distance between the target vector and the historical case vectors in multidimensional space. This algorithm calculates the average difference between the two vectors in each corresponding dimension. The similarity between the two cases is quantified by taking the square root of the sum of the square roots. The smaller the distance value, the closer the skin condition and physiological function of the two subjects were before the operation. The calculation unit sorts all the high-quality cases in ascending order according to the distance value and selects a preset number of cases at the top of the list. These selected cases constitute a group of similar cases. Then, the calculation unit retrieves all time-series standardized coding data of these similar cases collected at different time points after the operation through the database index. This process not only obtains static identity information, but more importantly, it obtains the dynamic recovery trajectory data of these subjects with similar physical conditions every day after the operation.

[0069] After acquiring the massive dataset of similar case groups, it needs to be transformed from a chaotic, discrete state into an ordered analytical view. Since standardized recovery characterization typically contains dozens or even hundreds of feature dimensions, each representing a different physiological meaning such as redness or healing degree, all data cannot be mixed together. The computation unit initiates a traversal process to establish independent processing channels for each feature dimension in the time-series standardized coded data. For any specific dimension, the computation unit constructs a two-dimensional time-feature value coordinate system. The horizontal axis of this coordinate system is defined as postoperative recovery time, usually in hours or days, while the vertical axis is defined as the feature value under that dimension. The computation unit maps the values ​​of that dimension recorded at different time points for all cases in the similar case group as discrete data points and projects them onto this coordinate system. For example, if there are fifty similar cases and each case recorded data at ten time points, then five hundred discrete points will appear in the coordinate system. The principle of this mapping process is to deconstruct and visualize abstract high-dimensional vector data. It presents the recovery trends of all similar patients under that dimension in the form of a point cloud on the coordinate axis, thus preparing the data for finding common patterns of change.

[0070] The nonlinear least squares method is used to perform regression fitting operations on discrete data points on the coordinate axes, and the coefficients of the fitting function are calculated; the specific content is as follows:

[0071] In the data preparation phase of regression fitting, the computational unit first traverses each independent case file in the previously constructed similar case group; from each case file, it precisely extracts the recorded postoperative recovery value and the precise postoperative time point corresponding to that value; these two data constitute a set of causal observations, where the passage of time is the independent variable driving the recovery process, and the recovery value is the dependent variable changing with time; therefore, the computational unit defines the postoperative time point as the horizontal axis coordinate of a Cartesian coordinate system and the postoperative recovery value as the vertical axis coordinate; although these data come from different individuals, they have all undergone standardization and belong to the same type of surgery, therefore... These can be viewed as sample observations under the same population; the computing unit merges and summarizes hundreds or thousands of pairs of time and numerical data from all cases; this merging operation mathematically constructs a huge discrete point set matrix, each row of which stores a specific coordinate point; the physical principle of this process is to piece together fragmented recovery fragments scattered across different individuals into a complete panoramic view; by mapping all data points to the same unified coordinate system, a dense point cloud is formed. This point cloud can intuitively reflect the general changing trend of this type of surgical group during the recovery period, providing sufficient sample support for subsequent mathematical modeling.

[0072] After data integration, the computational unit needs to select a suitable mathematical model to describe the direction of the point cloud bands. Based on the natural law that tissue repair in biology usually follows a rapid-then-slow pattern, the computational unit selected a nonlinear exponential function containing multiple undetermined parameters as the preset regression model structure. The exponential function was chosen because it can well fit the nonlinear characteristics of rapid inflammation subsidence in the early stage of wound healing and slow tissue remodeling in the later stage. The computational unit substitutes the time coordinate of each point in the discrete point set matrix into this preset exponential function model and calculates the corresponding theoretical prediction value using the current initial parameters. Then, the computational unit subtracts this theoretical prediction value from the actual observed values ​​recorded in the matrix to obtain the deviation value. In order to eliminate the mutual cancellation of positive and negative deviations and amplify the weight of larger errors, the computational unit squares each deviation value. Then, the squared deviation values ​​of all data points are summed to construct a residual sum of squares function. The value of this function intuitively represents the overall deviation between the current model curve and the actual data point cloud. The purpose of constructing this function is to transform the complex curve fitting problem into a pure mathematical extremum optimization problem, that is, to find a set of parameters that minimizes the value of this residual sum of squares function.

[0073] The Levenberg-Marquardt iterative algorithm is used to find the optimal solution for the constructed residual sum of squares function calculation unit. This is a numerical optimization algorithm specifically designed to solve nonlinear least squares problems. It cleverly combines the global stability of gradient descent and the local fast convergence of Gauss-Newton's method. During the operation, the algorithm first calculates the Jacobian matrix of the residual function with respect to the undetermined parameters to determine the direction and step size of parameter adjustment. In each iteration, the calculation unit updates the values ​​of the undetermined parameters based on the calculated step size and recalculates the new residual sum of squares. If the new residual sum of squares is less than the previous result, it indicates that the parameter adjustment direction is correct and the model curve... The line is converging towards the center of the data point cloud; the computation unit will continuously repeat this process of calculating gradients, updating parameters, and evaluating residuals; as the number of iterations increases, the model curve will more and more accurately fit the distribution trend of discrete data points; when the change in the sum of squared residuals is less than a very small preset threshold or the number of iterations reaches the upper limit, the determination function satisfies the convergence condition; at this point, it means that the algorithm has found the global optimum or a local optimum; the computation unit then stops iterating and locks the parameter values ​​at the current moment as the final determined coefficients of the fitting function; this set of coefficients gives the exponential function a specific shape so that it can represent the standard recovery trajectory of this type of similar population.

[0074] After obtaining the fitting function coefficients for each dimension, the computational unit can generate a complete benchmark model. Based on these determined coefficients, the computational unit can calculate the corresponding function values ​​at any consecutive moment on the time axis, thus drawing a smooth and continuous curve. This curve is no longer a discrete point but a mathematical trajectory that can cover the entire recovery cycle. It represents the standard recovery process that a group of people with similar physical conditions to the target object should have in this feature dimension under ideal conditions. The computational unit will repeat the above drawing process for each feature dimension to generate its corresponding function curve. Since human recovery is a complex process with the synergistic effect of multiple factors, a single-dimensional curve cannot summarize the whole picture. Therefore, the computational unit will logically combine all the generated function curves according to the index order of the feature dimensions. This combination is not a simple superposition but a construction of a multi-dimensional function set. Each curve in this set corresponds to a component in the standardized recovery representation encoding, thus forming a multi-dimensional dynamic benchmark recovery curve set. This set serves as a multi-dimensional ruler, providing a comprehensive and dynamically changing reference standard for evaluating the current complex recovery state of the target object.

[0075] In one embodiment of the present invention, step S4, which involves calculating the feature deviation between the standardized recovery characterization code and the multidimensional dynamic benchmark recovery curve set at the current time-series node, and combining this deviation with weights to generate a comprehensive evaluation index for the quantified repair status, is as follows:

[0076] In the initial stage of the evaluation calculation, the computing unit first reads the precise timestamp recorded during the current multimodal data acquisition; this timestamp represents the specific postoperative time point of the target object, such as the 72nd hour postoperatively; the computing unit uses this time value as an independent variable index to access the previously generated set of multidimensional dynamic baseline recovery curves; this set contains dozens or even hundreds of function curves for different physiological characteristic dimensions, each curve representing the standard trend of that specific dimension under ideal conditions; the computing unit starts parallel computing threads to substitute the current time value into the analytical expression of each function curve for calculation; through this... The function mapping operation can accurately calculate the theoretical value that the target object should achieve at each subdivided feature at this moment; for example, after substituting time into the redness and swelling curve, the standard redness and swelling reduction value is obtained; for the healing curve, the standard closure value is obtained. The computing unit arranges and combines these calculated theoretical values ​​according to the inherent order of the feature dimensions. The principle of this process is to discretize and sample the continuous mathematical model at a specific time section to construct a theoretical benchmark vector with a fixed length and dimension. This vector acts as a standard ruler to evaluate the current state and provides accurate reference coordinates for subsequent difference comparisons.

[0077] After constructing the theoretical baseline vector, the computational unit retrieves the standardized recovery characterization code generated in step S2 as the actual observation object. Since this code and the theoretical baseline vector follow the same feature dimension definition and arrangement order during construction, they are strictly aligned mathematically. The computational unit executes the dimension-wise difference calculation logic, that is, subtracting the theoretical value at the corresponding position in the theoretical baseline vector from the actual value of each dimension in the standardized recovery characterization code. Since the deviation of the recovery state may be positive or negative, for example, excessively high local temperature represents inflammation while excessively low temperature may represent necrosis, both of which are abnormal, the computational unit performs absolute value calculation on all differences in order to uniformly measure the magnitude of the deviation. This step eliminates the directional interference caused by the sign and only retains the magnitude of the value, thus truly reflecting the distance between the actual state and the ideal state. After this series of calculations, a feature deviation vector containing deviation information of all dimensions is generated. Each element in this vector quantifies the degree to which the corresponding physiological indicator deviates from the normal recovery trajectory, providing fine-grained basic data for comprehensive evaluation.

[0078] After obtaining the feature deviation vector, it needs to be transformed into a single value that reflects the overall severity. However, in clinical medicine, different physiological characteristics have completely different weights in judging the recovery effect. For example, slight skin color difference may only affect appearance, while deep subcutaneous fluid accumulation or abnormal high fever may indicate a serious risk of infection. Therefore, it is not possible to simply add up all deviation values. The calculation unit retrieves a preset feature weight vector from the knowledge base. Each value in this vector is an importance coefficient set by senior medical experts based on clinical experience. The calculation unit uses a weighted summation algorithm to perform a dot product operation between the feature deviation vector and the feature weight vector. Specifically, the deviation value of each dimension is multiplied by its corresponding weight coefficient, and then all the product results are summed together. Through this operation, the deviation of key indicators with significant clinical significance will be amplified by the weight, thus significantly increasing the final result value, while the deviation of secondary indicators will be reduced by the weight. The principle of this process is to simulate the professional judgment logic of doctors through mathematical weighting to calculate a weighted deviation value that represents the degree of overall difference. The larger this value is, the further the overall recovery of the target object is from the ideal state.

[0079] Finally, to make this abstract mathematical deviation value more intuitive and readable, it needs to be transformed into a general scoring format. The calculation unit uses a preset negative exponential transformation function model to process this weighted deviation value. The negative exponential model is chosen because it conforms to the human cognitive pattern of scoring: the smaller the deviation, the higher the score, and the downward trend of the score will gradually slow down after the deviation increases to a certain extent. The calculation unit substitutes the weighted deviation value as an independent variable into the negative exponential function formula. This formula usually includes a decay coefficient to control the rate of score decline and a scaling factor to map the result to a standard scoring range of zero to one hundred. During the calculation process, if the weighted deviation value is close to zero, it means that the actual state is almost in line with the ideal state, and the function output will approach the full score; conversely, if the deviation value is large, the function output will rapidly decay and approach zero. Through this non-linear mapping transformation, the calculation unit finally outputs a comprehensive evaluation index that quantifies the repair status. This index, as a standardized percentage score, not only eliminates the data differences between different surgical procedures but also allows doctors and patients to intuitively grasp the current recovery level and the quality of the results at a glance.

[0080] In one embodiment of the present invention, step S5, wherein the process of identifying abnormal recovery feature dimensions based on the comprehensive evaluation index, automatically associating with the repair scheme knowledge base to match corresponding intervention measures, and generating a repair effect evaluation report, is as follows:

[0081] In the logic judgment and feature screening stage, the calculation unit first receives the quantitative comprehensive assessment index of the repair status output from the previous stage. This is a value between zero and one hundred, which intuitively reflects the patient's current overall recovery level. The calculation unit strictly compares this index with a pre-set normal threshold. This threshold is usually set by medical experts based on large sample statistical data, for example, set at eighty points. If the calculation unit determines that the current assessment index is lower than this threshold, it means that the recovery effect has not reached the ideal standard and there is a potential pathological risk or delay. At this time, in order to find out the specific reason for the low score, the calculation unit will start a traversal procedure for the feature deviation vector. This vector stores... The calculation unit reads each element in the vector one by one and compares the value of the element with the local warning line value corresponding to that dimension. The reason for setting the local warning line is that different physiological indicators have different sensitivities to abnormalities. For example, a slight increase in body temperature may be more dangerous than a slight change in skin color, so differentiated judgment criteria are required. When the calculation unit finds that the deviation value of a certain dimension exceeds its corresponding local warning line, it will immediately lock the dimension and mark it as an abnormal recovery feature dimension. The principle of this process is to trigger micro-diagnosis through macro indicators, so as to accurately locate the shortcomings that reduce the overall recovery effect.

[0082] After identifying the specific abnormal dimension, the computational unit inputs it as a retrieval index into a pre-built repair solution knowledge base. This knowledge base is a structured database storing massive amounts of clinical medical knowledge, including various pathological models and their corresponding standard treatments, ranging from common complications to rare rejection reactions. The computational unit performs matching operations according to predefined mapping rules. These mapping rules are essentially a series of rigorous logical judgment statements that establish causal connections from symptom characteristics to intervention methods. For example, when the input abnormal dimension label is subcutaneous hematoma and the deviation is moderate, the mapping rule will point to a strategy of applying heat combined with blood-activating and stasis-removing drugs; while when the deviation is severe, the rule may point to puncture and drainage surgery. The computational unit performs multi-condition cross-retrieval in the knowledge base by parsing the semantic labels and numerical strength of the abnormal dimension. This process is not just a simple keyword matching but a reasoning process based on medical logic, ensuring that the selected intervention measures are accurate and targeted. Through this automated retrieval mechanism, the computational unit can simulate the diagnostic thinking of senior doctors and quickly select the most suitable clinical intervention strategy for the current patient's condition from the vast medical knowledge graph, thus providing a scientific basis for personalized repair.

[0083] In the final report generation stage, the computing unit is responsible for integrating all the aforementioned scattered data into a human-readable professional document. First, the computing unit retrieves a preset document template; this template defines the standard format framework of the assessment report, including fixed sections such as patient basic information, comprehensive score, abnormality details analysis, and medical recommendations. The computing unit then fills the calculated comprehensive assessment index into the score section and automatically generates an overall evaluation statement based on the score range. Next, a natural language processing algorithm is used to process the abnormal recovery feature dimensions and intervention data; this algorithm transforms abstract feature codes and rigid mathematical values ​​into easily understandable language. The algorithm provides textual descriptions; for example, it transforms a redness deviation value of 0.8 into a descriptive statement indicating obvious redness and swelling at the surgical site. Simultaneously, the algorithm organizes the matched interventions into clear and logical guidance steps and fills them into the medical advice section. During the filling process, the algorithm also performs grammar checks and semantic polishing to ensure the generated text is logically coherent, professionally worded, and avoids the harshness of machine generation. The principle behind this process is to convert structured data streams into unstructured natural language text, thereby generating a detailed and focused assessment report of the repair effect, enabling doctors and patients to quickly understand the current recovery status and the next course of action.

[0084] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for evaluating the repair effect after cosmetic surgery based on big data, characterized in that, Includes the following steps: S1: Collect postoperative multimodal data of the target object, and extract the current feature vector containing skin texture, color parameters and physiological indicators after time alignment and noise removal; S2: Input the current feature vector into the pre-trained feature mapping model, and transform it into a standardized recovery representation code that eliminates individual basic physiological differences and has a uniform dimension; the specific process is as follows: Retrieve the preoperative baseline feature vector of the target object, perform element-by-element subtraction between the current feature vector and the preoperative baseline feature vector, and generate a difference feature vector reflecting the postoperative changes. The differential feature vectors are loaded into a pre-trained feature mapping model, and the differential feature vectors are mapped to a high-dimensional feature space through a fully connected layer in the model to obtain latent feature vectors. A layer normalization function is used to process the latent feature vector, which transforms the numerical distribution of each dimension of the vector into a standard normal distribution form and outputs a standardized recovery representation code. S3: Based on the surgical type label and preoperative baseline characteristics of the target subject, retrieve similar case groups with satisfactory recovery results from the historical database, and fit and generate a set of multidimensional dynamic baseline recovery curves for the subject; the specific process is as follows: Extract the surgical procedure category and preoperative baseline feature vector of the target object, and select a set of high-quality cases with consistent surgical procedure categories and historical final repair scores higher than the preset threshold from the historical database. Calculate the Euclidean distance between the preoperative baseline feature vector of the target object and the preoperative features of each case in the high-quality case set, select the preset number of cases with the closest distance and their associated time-series standardized coding data, and construct a similar case group; For each feature dimension of the time-series standardized coded data in similar case groups, a time-feature numerical coordinate system is constructed to map the data of all cases in the corresponding dimension to the coordinate axis. The nonlinear least squares method is used to perform regression fitting operations on discrete data points on the coordinate axes, and the coefficients of the fitting function are calculated. Based on the coefficients of the fitted function, function curves of each dimension that change continuously over time are generated, and function curves of all dimensions are combined to form a set of multidimensional dynamic benchmark recovery curves. S4: Calculate the feature deviation between the standardized recovery characterization code and the multidimensional dynamic benchmark recovery curve set at the current time node, and generate a comprehensive evaluation index for the quantitative repair status by combining the weights; S5: Identify abnormal recovery feature dimensions based on the comprehensive evaluation index, automatically associate with the repair solution knowledge base to match corresponding intervention measures and generate a repair effect evaluation report.

2. The method for evaluating the post-medical aesthetic repair effect based on big data according to claim 1, characterized in that, In step S1, the process of collecting postoperative multimodal data of the target object and extracting the current feature vector containing skin texture, color parameters, and physiological indicators after temporal alignment and denoising is as follows: High-definition image acquisition equipment was used to acquire postoperative skin texture and color related image data of the target subject, and postoperative physiological index data was collected through physiological sensing equipment. All data were collected synchronously with postoperative timestamps. The collected multimodal data were time-aligned based on timestamps, and a smoothing filter method was used to remove noise from image data and physiological index data while retaining valid data information. Skin texture features and color parameters are extracted from temporally aligned and denoised image data, and physiological indicators are extracted from processed physiological data. These are then integrated to construct the current feature vector.

3. The method for evaluating the post-medical aesthetic repair effect based on big data according to claim 1, characterized in that, The specific method for processing latent feature vectors using a layer normalization function to transform the numerical distribution of each dimension of the vector into a standard normal distribution form and outputting the standardized recovery representation code is as follows: Calculate the arithmetic mean of all dimensions in the latent feature vector, and calculate the variance of the latent feature vector based on the sum of squared differences between the arithmetic mean and each dimension value; The centered value is obtained by subtracting the arithmetic mean from the value of each dimension of the latent feature vector. The standard deviation is obtained by adding a smoothing term to the variance to prevent the denominator from being zero and then taking the square root. The normalized value is obtained by calculating the ratio of the centered value to the standard deviation. Arrange all normalized values ​​in sequence to construct a standardized recovery characterization code. This code has the statistical properties of zero mean and unit variance, and outputs the standardized recovery characterization code.

4. The method for evaluating the post-medical aesthetic repair effect based on big data according to claim 1, characterized in that, The specific method for calculating the coefficients of the fitting function by performing regression fitting operations on discrete data points on the coordinate axes using the nonlinear least squares method is as follows: Extract the postoperative recovery value and corresponding postoperative time node of each case in the similar case group. With the postoperative time node as the independent variable on the horizontal axis and the postoperative recovery value as the dependent variable on the vertical axis, merge the data of all cases to construct a discrete point set matrix under a unified coordinate system. A nonlinear exponential function containing undetermined parameters is selected as the preset regression model structure. The numerical values ​​of the discrete point set matrix are substituted into the regression model structure to construct a residual sum of squares function that characterizes the deviation between the observed values ​​in the discrete point set matrix and the theoretical values ​​calculated by the regression model structure. The Levenberg-Marquardt iterative algorithm is used to perform a minimization operation on the residual sum of squares function. The values ​​of the undetermined parameters are updated iteratively until the residual sum of squares function meets the convergence condition, and the coefficients of the finally determined fitting function are output.

5. The method for evaluating the post-medical aesthetic repair effect based on big data according to claim 1, characterized in that, In step S4, the process of calculating the feature deviation between the standardized recovery characterization code and the multidimensional dynamic benchmark recovery curve set at the current time node, and combining the weights to generate a comprehensive evaluation index for the quantified repair status, is as follows: Based on the current timeline, extract the corresponding theoretical benchmark values ​​from each dimension curve of the multidimensional dynamic benchmark recovery curve set, and construct the theoretical benchmark vector; Calculate the absolute value of the difference between the standardized recovery representation code and the theoretical baseline vector in the corresponding dimension, and generate the feature deviation vector; The feature weight vectors based on the clinical importance of each dimension are retrieved, and a weighted summation operation is performed on the feature deviation vector and the feature weight vector to calculate the weighted deviation value that represents the overall degree of difference. Substitute the weighted deviation value into the preset negative exponential transformation function model, calculate the mapping score of the value within the standard scoring range, and output a comprehensive evaluation index that quantifies the repair status.

6. The method for evaluating the post-medical aesthetic repair effect based on big data according to claim 1, characterized in that, In step S5, the process of identifying abnormal recovery feature dimensions based on the comprehensive evaluation index, automatically associating with the repair solution knowledge base to match corresponding intervention measures, and generating a repair effect evaluation report is as follows: The comprehensive evaluation index is compared with the preset normal threshold. If it is determined that the standard is not met, the feature deviation vector is traversed and the dimension whose value exceeds the local warning line is locked as the anomaly recovery feature dimension. Input the abnormal recovery feature dimension into the repair plan knowledge base, retrieve the clinical intervention strategy for that dimension through predefined mapping rules, and obtain the corresponding intervention measures; The comprehensive evaluation index, abnormal recovery feature dimensions, and intervention measures are filled into a preset document template. A structured text description is constructed using a natural language generation algorithm to generate a repair effect evaluation report.

Images