Physiological function dynamic early warning method and device and medical equipment

Abstract

This application discloses a method, device, and medical equipment for dynamic early warning of physiological functions, relating to the field of artificial intelligence technology. The method includes: acquiring tongue images of a target object at multiple consecutive time points within a preset time period and determining the corresponding tongue image sequence; decoupling the tongue image sequence by features to obtain a first type of feature sequence and a second type of feature sequence, wherein the first type of feature sequence has a higher frequency of change in the time dimension than the second type of feature sequence; calculating the first evolution rate of the first type of feature sequence and the second evolution rate of the second type of feature sequence within the observation window, and calculating the difference metric between the first and second evolution rates; and generating corresponding physiological function early warning information based on the difference metric. The method provided in this application aims to improve the accuracy and timeliness of physiological function early warning by distinguishing high-frequency and low-frequency change features through dynamic feature decoupling.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, device, and medical equipment for dynamic early warning of physiological functions. Background Technology

[0002] With the rapid development of artificial intelligence and computer vision technologies, intelligent tongue diagnosis systems based on traditional Chinese medicine theory are increasingly widely used in auxiliary diagnosis and daily health management. Existing systems can output warning signals by recognizing and analyzing users' tongue images.

[0003] However, in practical applications, this type of technology has limitations and is difficult to achieve accurate early health warnings, ultimately restricting the timeliness and accuracy of health warnings. Summary of the Invention

[0004] The main purpose of this application is to propose a method, device, and medical equipment for dynamic early warning of physiological functions, which aims to improve the accuracy and timeliness of early warning of physiological functions by decoupling high-frequency and low-frequency change features through dynamic features.

[0005] To achieve the above objectives, the dynamic early warning method for physiological functions proposed in this application includes: Acquire tongue images of the target object at multiple consecutive time points within a preset time period, and determine the corresponding tongue image sequence; The tongue image sequence is decoupled to obtain a first type of feature sequence and a second type of feature sequence. The first type of feature sequence has a higher frequency of change in the time dimension than the second type of feature sequence. Calculate the first evolution rate of the first type of feature sequence within the observation window and the second evolution rate of the second type of feature sequence within the observation window, and calculate the difference metric between the first evolution rate and the second evolution rate. Based on the difference measurement values, corresponding physiological function early warning information is generated.

[0006] In some embodiments, acquiring tongue images of the target object at multiple consecutive time points within a preset time period and determining the corresponding tongue image sequence includes: Acquire original tongue images of a target object at multiple consecutive time points within a preset time period, wherein the original tongue images contain a preset reference object; Based on the preset reference objects in each of the original tongue images, a dynamic color correction matrix corresponding to each of the original tongue images is determined. Based on the dynamic color correction matrix, the corresponding original tongue image is subjected to color correction processing to obtain a calibrated tongue image; The clarity score of the calibrated tongue image is calculated, and the calibrated tongue image is filtered based on the clarity score. The filtered calibrated tongue image images are combined according to the time dimension to determine the tongue image sequence.

[0007] In some embodiments, the feature decoupling of the tongue image sequence to obtain a first type of feature sequence and a second type of feature sequence, wherein the first type of feature sequence has a higher frequency of change in the time dimension than the second type of feature sequence, includes: By using a pre-defined convolutional neural network, each tongue image in the tongue image sequence is mapped to a first feature space and a second feature space that are orthogonal to each other. The first feature space represents high-frequency evolution, and the second feature space represents low-frequency evolution. Each of the tongue images is obtained by acquiring a first feature vector in the first feature space and a second feature vector in the second feature space. The first feature vector contains image-based feature indicators, including at least one of tongue redness index, tongue gloss, and tongue coating moisture content. The second feature vector contains image-based feature indicators, including at least one of crack topology index, tongue thickness index, and spot solidification degree. The first feature vectors corresponding to each of the tongue images in the first feature space are combined according to the time dimension to obtain the first type of feature sequence; The second feature vectors corresponding to each of the tongue images in the second feature space are combined according to the time dimension to obtain the second type of feature sequence.

[0008] In some embodiments, the image-based feature indicators included in the first feature vector include: tongue redness index, tongue surface gloss, and tongue coating moisture content; the image-based feature indicators included in the second feature vector include: crack topology index, tongue thickness index, and spot solidification degree; obtaining the first feature vector corresponding to each of the tongue images in the first feature space, and the second feature vector corresponding to each of the tongue images in the second feature space, includes: Each of the tongue images is converted to the CIELAB color space, and the tongue redness index is calculated based on the weighted average of the a* channel in the CIELAB color space. Each of the tongue images is converted to the CIELAB color space, and the gloss of the tongue surface is calculated based on the local variance of the L* channel in the CIELAB color space. The water content of the tongue moss is calculated based on the reflectance distribution characteristics of the highlight areas in each of the tongue images. The total length and number of branches of the cracks in each of the tongue images are calculated based on the skeleton extraction algorithm to obtain the crack topology index. The ratio of the segmented area of ​​the tongue body to the area of ​​the circumscribed rectangle of the tongue body in each of the tongue images is calculated to obtain the fatness index; Based on feature point matching, the ecchymosis that is fixed across time points in each of the tongue images is identified, and the degree of ecchymosis is obtained.

[0009] In some embodiments, calculating the first evolution rate of the first type of feature sequence within the observation window and the second evolution rate of the second type of feature sequence within the observation window, and calculating the difference metric between the first evolution rate and the second evolution rate, includes: The first type of feature sequence and the second type of feature sequence are input into a preset time series prediction model. The temporal evolution patterns of the first type of feature sequence and the second type of feature sequence are extracted through two independent attention channels in the time series prediction model. The first derivative of the first type of feature sequence within the prediction window is calculated as the first evolution rate, and the first derivative of the second type of feature sequence is calculated as the second evolution rate. The first evolution rate and the second evolution rate are normalized respectively, and the absolute value of the difference between the normalized first evolution rate and the normalized second evolution rate is calculated to obtain the difference metric value.

[0010] In some embodiments, the input to the time series prediction model further includes static attribute data of the target object and known future variable data; The static attribute data includes at least one of physical condition, age, and gender, and the known future variable data includes at least one of solar term data and physiological cycle data. The time series prediction model is used to combine the static attribute data and the known future variable data to assist in predicting the first evolution rate and the second evolution rate.

[0011] In some embodiments, generating corresponding physiological function warning information based on the difference metric includes: Based on the relationship between the difference metric and the preset threshold, and combined with the changing trends of the first evolution rate and the second evolution rate, the target warning level corresponding to the target object is determined. The target warning level includes one of the following: a first-level warning that characterizes functional disturbance, a second-level warning that characterizes organic loss, and a third-level warning that characterizes critical signal. Generate physiological function warning information that includes the target warning level.

[0012] In some embodiments, determining the target warning level corresponding to the target object based on the relationship between the difference metric and a preset threshold, combined with the changing trends of the first evolution rate and the second evolution rate, includes: When the difference metric value is greater than the preset threshold, and the first evolution rate shows an upward trend while the second evolution rate approaches zero, the target warning level is determined to be the first level warning. When the difference metric value is greater than the preset threshold and the second evolution rate shows a continuous unidirectional change, the target warning level is determined to be the second level warning. When the second evolution rate reaches a preset extreme value and the rate of increase of the first evolution rate is greater than a preset value, the target warning level is determined to be the third level warning.

[0013] This application also provides a dynamic early warning device for physiological functions, which is configured to implement the dynamic early warning method for physiological functions described above.

[0014] This application also provides a medical device, which includes the dynamic early warning device for physiological functions described above.

[0015] In summary, the physiological function dynamic early warning method, device, and medical equipment provided in this application acquire tongue image sequences at multiple time points, decouple the features into high-frequency and low-frequency sequences, calculate the difference in evolution rate, and generate early warning information. It can distinguish high-frequency and low-frequency change features through dynamic feature decoupling, thereby improving the accuracy and timeliness of physiological function early warning. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0017] Figure 1 A flowchart of the first embodiment of the dynamic early warning method for physiological functions provided in this application; Figure 2 A flowchart of the second embodiment of the dynamic early warning method for physiological functions provided in this application; Figure 3 A flowchart of the third embodiment of the dynamic early warning method for physiological functions provided in this application; Figure 4 A flowchart of the fourth embodiment of the dynamic early warning method for physiological functions provided in this application; Figure 5 A flowchart of the fifth embodiment of the dynamic early warning method for physiological functions provided in this application; Figure 6 A flowchart of the sixth embodiment of the dynamic early warning method for physiological functions provided in this application; Figure 7 A flowchart of the seventh embodiment of the dynamic early warning method for physiological functions provided in this application.

[0018] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0020] Furthermore, if the embodiments of this application involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the use of "and / or" or "and / or" throughout the text includes three parallel solutions. For example, "A and / or B" includes solution A, solution B, or a solution that simultaneously satisfies A and B. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0021] It should be noted that step designations such as S100 and S200 are used in this document for the purpose of more clearly and concisely describing the corresponding content, and do not constitute a substantial limitation on the order. In specific implementation, those skilled in the art may execute S200 first and then S100, etc., but these should all be within the protection scope of this application.

[0022] With the rapid development of artificial intelligence and computer vision technologies, intelligent tongue diagnosis systems based on traditional Chinese medicine theory are increasingly widely used in auxiliary diagnosis and daily health management. Existing systems can output health warnings by recognizing and analyzing users' tongue images. However, in practical applications, this technology has significant limitations, making it difficult to achieve accurate early health warnings, ultimately restricting the timeliness and accuracy of these warnings. Specifically, the shortcomings of existing technologies are mainly reflected in the following two aspects: On the one hand, existing tongue image analysis generally uses a static cross-sectional method, resulting in low accuracy of early warnings. Most existing systems only classify single images acquired at isolated time points, ignoring the fact that human physiological function is a dynamic equilibrium process. A single sampling is highly susceptible to interference from short-term, non-pathological external factors (such as localized changes in the tongue image caused by drinking hot water or specific dietary habits). Existing systems struggle to effectively distinguish between such instantaneous normal physiological fluctuations and long-term pathological trends, easily leading to misjudgments, high false alarm rates, and reduced accuracy of health warnings.

[0023] On the other hand, existing systems generally employ a globally uniform feature evaluation model, resulting in poor timeliness of early warnings. In actual physiological function evolution, the evolutionary cycles of different characteristics vary objectively (for example, color features reflecting the state of Qi and blood often change rapidly, while morphological features reflecting organic states require long-term accumulation before deformation occurs). Existing technologies can uniformly stitch together and model all extracted image features without differentiation. Under this one-size-fits-all evaluation mechanism, early, weak local abnormal signals are easily diluted and masked by other long-term stable physiological features. The system often needs to wait until multiple features of the tongue appearance deteriorate significantly simultaneously before crossing the warning threshold to trigger an alarm, thus completely missing the optimal window for early intervention and causing a significant delay in health warnings.

[0024] In response, this application proposes a dynamic early warning method for physiological functions, which aims to improve the accuracy and timeliness of early warning of physiological functions by decoupling high-frequency and low-frequency change features through dynamic features.

[0025] In one embodiment of this application, as Figure 1 As shown, the dynamic early warning method for physiological functions includes steps S100 to S400.

[0026] Step S100: Obtain tongue images of the target object at multiple consecutive time points within a preset time period, and determine the corresponding tongue image sequence.

[0027] Step S200: Decouple the tongue image sequence by features to obtain a first type of feature sequence and a second type of feature sequence. The first type of feature sequence has a higher frequency of change in the time dimension than the second type of feature sequence.

[0028] Step S300: Calculate the first evolution rate of the first type of feature sequence within the observation window and the second evolution rate of the second type of feature sequence within the observation window, and calculate the difference metric between the first evolution rate and the second evolution rate.

[0029] Step S400: Generate corresponding physiological function warning information based on the difference measurement value.

[0030] For ease of understanding, the following explains some key terms in this embodiment: A tongue image sequence refers to a collection of tongue images acquired at multiple consecutive time points within a preset time period and arranged chronologically. This tongue image sequence is used to capture the dynamic process of changes in the physiological functions of the target object over time.

[0031] Feature decoupling refers to the process of separating multiple physiological features contained in a tongue image based on their differences in frequency of change over time, either mathematically or physically. Through feature decoupling, features with higher frequency of change can be distinguished from those with lower frequency of change without interference, allowing for independent analysis of their evolutionary patterns.

[0032] The first type of feature sequence refers to the set of features extracted from tongue image sequences through feature decoupling, which have a relatively high frequency of change in the time dimension. The first type of feature sequence can reflect the functional and short-term physiological state changes of the target object, such as the circulation of qi and blood, and the metabolism of body fluids.

[0033] The second type of feature sequence refers to the set of features extracted from tongue image sequences through feature decoupling, which have a relatively low frequency of change in the time dimension. The second type of feature sequence can reflect organic and long-term physiological structural changes of the target object, such as tongue morphology and crack distribution.

[0034] The first evolution rate refers to the speed at which a first-class characteristic sequence changes over time within a specific observation window. It quantifies the dynamic trend and intensity of high-frequency physiological characteristics. The second evolution rate refers to the speed at which a second-class characteristic sequence changes over time within a specific observation window. It quantifies the dynamic trend and intensity of low-frequency physiological characteristics. The difference measure is a quantitative index of the difference between the first and second evolution rates. This difference measure is used to characterize the degree of asynchrony between functional and organic changes in the target object and is a key basis for determining physiological dysfunction.

[0035] Physiological function early warning information refers to the prompts generated and output based on difference measurement values, indicating the current physiological state and potential risks of a target subject. These prompts may include warning levels, recommendations, etc., to guide early intervention.

[0036] In this embodiment, tongue images of the target object are acquired at multiple consecutive time points within a preset time period, and the corresponding tongue image sequence is determined. Various methods can be used to acquire these images. For example, handheld or fixed image acquisition devices can be used to photograph or record the target object's tongue at preset time intervals (e.g., once a day, once an hour, or more frequently), thereby obtaining a series of raw images. These raw images can undergo simple preprocessing such as cropping and scaling, and then be arranged according to the acquisition time sequence to form a tongue image sequence. In another implementation, a telemedicine platform can be used, where users upload tongue images taken at different time points, which are then timestamped and serialized upon receipt.

[0037] In this embodiment, the purpose of feature decoupling is to separate features reflecting different rates of physiological change in tongue images. One implementation method is to extract multi-dimensional feature indicators from each tongue image manually or through a preset image processing algorithm. For example, features considered to change rapidly, such as tongue color, thickness of tongue coating, and tongue surface moisture, can be extracted and classified as first-class features; while features considered to change slowly, such as tongue size, tongue shape, cracks, and ecchymosis, can be extracted and classified as second-class features. Then, these similar features extracted at different time points are combined to form their respective feature sequences.

[0038] In this embodiment, for the first type of feature sequence, the first evolution rate can be obtained within a preset observation window (e.g., data from the last three days or one week) by calculating the difference in feature values ​​at adjacent time points, the average rate of change, or the slope of the trend line. Similarly, for the second type of feature sequence, a similar method can be used to calculate the second evolution rate within the same or different observation windows. To capture deeper temporal dependencies, a time series prediction model incorporating a dual-stream time-varying encoder can be introduced. This model extracts the patterns of the two types of sequences through two independent attention channels and calculates the first derivative of the sequence within the prediction window as the evolution rate. Simultaneously, this time series prediction model can also integrate prior static or external variables such as the target object's age, physical condition, and solar terms to correct the prediction trajectory. After obtaining the two types of evolution rates, various mathematical methods can be used to calculate the difference between them. For example, the absolute difference or relative difference between the two rates can be directly calculated, or statistical methods (such as correlation coefficients or distance metrics) can be used to assess the degree of deviation between them.

[0039] In this embodiment, after obtaining the difference metric value, it can be compared with a preset threshold to determine whether there is any physiological dysfunction. For example, when the difference metric value exceeds a certain threshold, the system can trigger an alert. The alert information can be a simple prompt, such as "abnormal fluctuations in physiological function," or a more specific suggestion, such as "it is recommended to pay attention to recent lifestyle habits." The generation of the alert information can be based on a simple rule engine, which directly maps different alert levels or alert content according to the magnitude of the difference metric value. It is understandable that the core technical logic of using the difference metric value calculated between the first evolution rate and the second evolution rate, rather than comparing them separately with preset absolute thresholds, lies in accurately quantifying the degree of asynchronous evolution (disconnection) of physiological function. The human physiological system is in a dynamic equilibrium. If a fixed absolute threshold is used, it is easy to cause frequent false alarms or missed alarms due to differences in objective benchmarks such as individual age and basic physical condition. On the contrary, the difference metric value essentially introduces an adaptive relative evaluation mechanism: using the slow second evolution rate, which represents organic evolution, as the user's own dynamic anchor point to measure the degree of abnormal deviation of the first evolution rate, which represents functional disturbance. Because early-stage diseases often follow the objective law of functional disorder preceding physical change, the evolutionary trajectories of the two become significantly disconnected, leading to a sharp increase in the difference measurement value. Therefore, this design not only effectively eliminates the static threshold failure problem caused by individual physical baseline differences, but also achieves a technological leap from traditional absolute value exceeding alarm to relative imbalance trend early warning, enabling it to keenly capture very early and weak signals of the body moving from homeostasis to disorder.

[0040] In this application, high frequency and low frequency are used to represent two distinctly different physiological characteristics. The fundamental physical logic behind this is the objective temporal asymmetry between the evolution of human physiological functions and their external visual representation. Specifically, high-frequency changes correspond to the physiological functional state and microcirculatory metabolism of the human body (corresponding to the concepts of "qi and blood" and "body fluids" in Traditional Chinese Medicine). These physiological activities are extremely sensitive to internal and external stimuli. Subtle fluctuations in microvascular perfusion or local fluid metabolism can be rapidly reflected in changes in tongue color (e.g., redness) and luster within a very short time (e.g., several hours to several days). Conversely, low-frequency changes correspond to the organic structure and tissue morphology of the human body (corresponding to the concepts of "form" and "substance" in Traditional Chinese Medicine). Their evolution usually involves the proliferation, atrophy, or long-term accumulation of pathological damage to tissue cells. Such substantial changes in physical appearance (e.g., changes in tongue size, the formation of deep cracks, or the deposition of solidified ecchymosis) often require a long period (e.g., several months or even several years) to become visually apparent. Therefore, this application uses the evolution frequency on the time axis as the starting point for feature decoupling, which can perfectly map the objective physiological evolution law of Qi and blood first (fast variable) and form and substance changing later (slow variable), thereby realizing the precise physical separation of short-term functional disturbances and long-term organic damage at the algorithm level.

[0041] In summary, the physiological function dynamic early warning method, device, and medical equipment provided in this application acquire tongue image sequences at multiple time points, decouple the features into high-frequency and low-frequency sequences, calculate the difference in evolution rate, and generate early warning information. It can distinguish high-frequency and low-frequency change features through dynamic feature decoupling, thereby improving the accuracy and timeliness of physiological function early warning.

[0042] In one embodiment, such as Figure 2 As shown, step S100, which involves acquiring tongue images of the target object at multiple consecutive time points within a preset time period and determining the corresponding tongue image sequence, may include steps S110 to S140.

[0043] Step S110: Obtain original tongue images of the target object at multiple consecutive time points within a preset time period, wherein the original tongue images contain preset reference objects.

[0044] The process involves acquiring raw tongue images of a target object at multiple consecutive time points within a preset time period. These raw tongue images include a preset reference object, aiming to obtain raw visual data of the target object's tongue at different time points and simultaneously introduce a benchmark for calibration and quality control. The preset reference object can be a standard color chart, such as the X-RiteColorChecker Passport, placed within the camera's field of view during image acquisition; alternatively, it can be a dedicated calibration target with known spectral reflectance characteristics, integrated into the imaging device or placed near the tongue. Image acquisition can be achieved through an automated image acquisition system, such as a dedicated tongue imaging device, at fixed time intervals (e.g., daily or twice daily); or it can be achieved through a user-operated smartphone camera, with a guiding application to ensure consistent frame composition and lighting conditions during capture and to prompt the user about the inclusion of preset color calibration blocks. This continuous acquisition across the preset time period provides the necessary temporal depth for subsequent construction of a dynamic time series.

[0045] Step S120: Based on the preset reference objects in each of the original tongue images, determine the dynamic color correction matrix corresponding to each of the original tongue images.

[0046] Based on preset reference objects in each of the original tongue images, a dynamic color correction matrix is ​​determined for each original tongue image. The purpose is to compensate for color deviations caused by ambient lighting, camera settings, and other environmental factors, thereby ensuring color consistency across the entire time series. The dynamic color correction matrix can be determined using a multinomial regression model (e.g., a 3x3 or 3x9 matrix) to map the RGB values ​​of the color blocks identified in the original image to their known standard CIELAB or sRGB values, and dynamically calculate the coefficients of this matrix for each image. Alternatively, a neural network-based color correction model can be used. This model is trained on a large number of images with known color targets captured under different lighting conditions. The network takes the original image and the identified reference objects as input and directly outputs correction parameters or the corrected image.

[0047] Step S130: Perform color correction processing on the corresponding original tongue image based on the dynamic color correction matrix to obtain a calibrated tongue image.

[0048] The original tongue image is color-corrected based on the dynamic color correction matrix to obtain a calibrated tongue image. This calibrated image applies the calculated correction parameters to the entire image, thereby generating a color-accurate representation of the tongue. The color correction process can involve applying a defined 3x3 or 3x9 transformation matrix to the RGB pixel values ​​of the original tongue image, typically involving matrix multiplication and possible gamma correction. Alternatively, a lookup table (LUT) generated by the dynamic color correction matrix can be used, with each pixel's RGB value as an index to retrieve its corrected RGB value, achieving faster processing speeds.

[0049] Step S140: Calculate the clarity score of the calibration tongue image, and filter the calibration tongue image based on the clarity score. Combine the filtered calibration tongue images according to the time dimension to determine the tongue image sequence.

[0050] The goal of calculating sharpness scores for calibrated tongue images and then filtering them based on these scores is to identify and remove blurry or out-of-focus images, preventing these low-quality images from introducing noise and inaccuracies in subsequent feature extraction and analysis. Sharpness scores can be calculated using frequency domain analysis methods, such as calculating the energy of high-frequency components (e.g., Laplacian variance, FFT amplitude spectrum analysis); higher high-frequency energy indicates a sharper image. Another approach is to use spatial domain methods, such as the Tenengrad gradient operator or the variance of the modified Laplacian operator, which assess sharpness by measuring the sharpness of image edges. The filtering can be performed by setting a fixed sharpness score threshold; images below this threshold are discarded. Alternatively, an adaptive threshold based on the distribution of sharpness scores within image batches or a rolling window can be used to achieve more robust filtering under different conditions.

[0051] The selected calibrated tongue images are combined according to the time dimension to determine a tongue image sequence. The aim is to construct a high-quality, chronologically ordered tongue image sequence to accurately reflect the dynamic evolution of the target subject's physiological state over time. One approach is to store the selected images and their capture timestamps in a database and retrieve them chronologically. Alternatively, a data structure (such as a list or array of image objects) can be created where each element contains image data and its corresponding timestamp, ensuring the sequence is strictly ordered chronologically.

[0052] Through the above technical solution, this application can effectively solve the problems of color distortion and image blurring in original tongue images caused by changes in ambient lighting or differences in shooting angle. It provides reliable, consistent, and high-quality basic data for subsequent feature decoupling and dynamic evolution analysis, improving the accuracy and reliability of dynamic early warning of physiological functions.

[0053] In one embodiment, such as Figure 3 As shown, step S200, decoupling the features of the tongue image sequence to obtain a first type of feature sequence and a second type of feature sequence, wherein the first type of feature sequence has a higher change frequency in the time dimension than the second type of feature sequence, may include steps S210 to S240.

[0054] Step S210: Using a preset convolutional neural network, each tongue image in the tongue image sequence is mapped to a first feature space and a second feature space that are orthogonal to each other. The first feature space represents high-frequency evolution, and the second feature space represents low-frequency evolution.

[0055] This convolutional neural network (CNN) can be a deep learning model containing multiple convolutional layers, pooling layers, and fully connected layers. Its training objective is to learn to decompose an input image into two orthogonal feature representations. Orthogonality can be achieved by introducing an orthogonality constraint term into the loss function, for example, by minimizing the absolute value of the inner product of the vectors between the two feature spaces or by applying a Gram-Schmidt orthogonalization process. Alternatively, the CNN can employ a two-branch structure, with each branch responsible for learning a feature space. These branches maintain correlation by sharing a lower-level feature extraction layer, while at the top level, a specific loss function (such as mutual information minimization or orthogonal projection loss) forces the feature vectors output by the two branches to be orthogonal. This mechanism ensures the mathematical independence of high-frequency and low-frequency features, avoiding signal averaging problems caused by feature mixing.

[0056] Step S220: Obtain the first feature vector corresponding to each of the tongue images in the first feature space, and the second feature vector corresponding to each of the tongue images in the second feature space.

[0057] In the output layer of a convolutional neural network, two independent output branches can be designed, outputting feature vectors corresponding to the first and second feature spaces, respectively. These vectors can be fixed-dimensional numerical arrays used to quantify specific attributes of the image. Alternatively, a shared latent representation can be extracted in the encoder part of the convolutional neural network, and then mapped to the first and second feature spaces respectively through two independent decoder or projection layers to obtain the corresponding feature vectors. This is a concrete manifestation of feature decoupling, decomposing the original image information into two independent feature representations with different physical meanings.

[0058] The first feature vector includes image-based computational features such as at least one of the following: tongue redness index, tongue gloss, and tongue coating moisture content. These indicators are key quantitative parameters reflecting high-frequency changes in the tongue appearance (such as qi and blood dynamics and body fluid metabolism), and are typically sensitive to short-term physiological changes. In addition to tongue redness index, tongue gloss, and tongue coating moisture content, the first feature vector can also include other indicators reflecting rapid physiological changes, such as tongue color saturation, tongue moisture, and the rate of change in tongue coating thickness. These indicators can be directly extracted from the tongue image using image processing algorithms, such as color space conversion, texture analysis, and highlight detection, to ensure that the first feature vector accurately captures the rapidly changing functional signals represented by the concept of "qi and blood flowing first" in traditional Chinese medicine theory.

[0059] Meanwhile, the image-based feature indicators included in the second feature vector include at least one of the following: crack topology index, fatness / thinness index, and spot solidification degree. These indicators are key quantitative parameters reflecting the low-frequency evolution of the tongue image (such as tissue structure and morphological changes), and are usually sensitive to long-term pathological changes. In addition to the crack topology index, fatness / thinness index, and spot solidification degree, the second feature vector may also include other indicators of slow morphological changes, such as tongue morphological symmetry, sublingual vein morphology, and the degree of atrophy of tongue papillae. These indicators can be extracted from the tongue image using advanced computer vision techniques such as image segmentation, edge detection, skeleton extraction, and feature point matching, ensuring that the second feature vector can accurately capture the slowly changing organic signals represented by "form and substance change" in traditional Chinese medicine theory.

[0060] Step S230: Combine the first feature vectors corresponding to each of the tongue images in the first feature space according to the time dimension to obtain the first type of feature sequence.

[0061] The first type of feature sequence is a continuous arrangement of high-frequency evolution features on the time axis, used to analyze their dynamic trends. The first feature vectors corresponding to tongue images acquired at different time points can be simply concatenated according to their acquisition time order to form a multidimensional time series. Alternatively, when combining the sequences, methods such as sliding windows or exponentially weighted averaging can be used to smooth or weight the time series to highlight recent changes or reduce the influence of noise, providing a basic data structure for subsequent independent analysis of high-frequency evolution rates.

[0062] Step S240: Combine the second feature vectors corresponding to each of the tongue images in the second feature space according to the time dimension to obtain the second type of feature sequence.

[0063] The second type of feature sequence is a continuous arrangement of low-frequency evolution features on the time axis, used to analyze their dynamic change trends. Similar to the first type of feature sequence, the second feature vectors corresponding to tongue images acquired at different time points are simply concatenated according to their acquisition time order to form another multidimensional time series. When combining these sequences, considering the slow changes in low-frequency features, longer sampling intervals or stronger smoothing processes can be used to better capture their long-term trends, providing a basic data structure for subsequent independent analysis of low-frequency evolution rates.

[0064] Through the above technical solution, this application effectively solves the problem of early warning failure caused by feature coupling in the prior art. By mapping tongue images to mutually orthogonal first and second feature spaces through a pre-defined convolutional neural network, high-frequency evolution features (such as tongue redness index, tongue surface gloss, and tongue coating moisture content) are forcibly decoupled from low-frequency evolution features (such as crack topology index, tongue thickness index, and spot solidification degree). This orthogonal mapping mechanism ensures that high-frequency signals reflecting functional perturbations are not averaged or masked by low-frequency signals reflecting organic damage, thus enabling the capture of weak, asynchronous evolution signals even in the early stages of disease, when functional changes are significant but organic changes are not yet obvious. The independent construction of the first and second type feature sequences provides a clear data foundation for subsequent calculation of their evolution rates, enabling this method to identify potential physiological dysfunctions earlier and more accurately, thereby achieving the goals of "prevention before disease" and early intervention.

[0065] In one embodiment, such as Figure 4 As shown, step S220, obtaining the first feature vector corresponding to each of the tongue images in the first feature space and the second feature vector corresponding to each of the tongue images in the second feature space, may include steps S221 to S226.

[0066] Step S221: Convert each of the tongue images to the CIELAB color space, and calculate the tongue redness index based on the weighted average of the a* channel in the CIELAB color space.

[0067] The tongue redness index measures the degree of redness of the tongue, reflecting its blood supply and thermal characteristics. The CIELAB color space is a device-independent color space where the a* channel represents the color range from green to red. By calculating the weighted average of the a* channel, the redness of the tongue can be objectively quantified, effectively avoiding the problem of traditional RGB color spaces being greatly affected by lighting conditions. For example, the a* values ​​of all pixels within the tongue region can be weighted and averaged, with weights determined based on the distance of the pixel from the center of the tongue or the pixel's brightness value, to highlight the color characteristics of the core tongue region. Furthermore, color histogram analysis of the tongue region can be performed to extract the peak value or specific percentile value of the a* channel as the tongue redness index.

[0068] Step S222: Convert each of the tongue images to the CIELAB color space, and calculate the gloss of the tongue surface based on the local variance of the L* channel in the CIELAB color space.

[0069] The glossiness of the tongue reflects its moisture level and reflective properties. In the CIELAB color space, the L* channel represents brightness. High tongue glossiness is usually characterized by dramatic brightness variations in an image, i.e., a large local variance. By calculating the local variance of the L* channel within the tongue region, the distribution characteristics of reflected light on the tongue surface can be effectively captured, thereby quantifying the glossiness. For example, a sliding window approach can be used to calculate the variance of L* values ​​within the window, and these variance values ​​can be averaged or statistically analyzed to obtain the overall tongue glossiness. Another approach is to evaluate glossiness by analyzing the shape, size, and intensity distribution of highlight areas on the tongue surface; the more concentrated and brighter the highlight areas, the higher the glossiness.

[0070] Step S223: Calculate the water content of the moss based on the reflectance distribution characteristics of the highlight areas in each of the tongue images.

[0071] Highlights on the tongue coating are formed by the reflection of light on its surface, and their reflectivity distribution is closely related to the water content of the tongue coating. For example, a moist tongue coating usually forms large, blurred-edge highlight areas, while a dry tongue coating may form smaller, sharp highlight spots. By analyzing the area, intensity, shape, and distribution of these highlight areas on the tongue surface, the water content of the coating can be inferred. Specifically, highlight detection can be performed on the image, followed by morphological analysis of the detected highlight areas to extract their geometric and brightness features, thereby establishing a mapping relationship with water content.

[0072] Step S224: Calculate the total length and number of branches of the cracks in each of the tongue images based on the skeleton extraction algorithm to obtain the crack topology index.

[0073] Crack topology indices are used to quantify the severity and complexity of tongue cracks. Skeleton extraction algorithms refine cracks in a tongue image into single-pixel-width skeleton lines, facilitating the measurement of crack geometry. By analyzing these skeleton lines, the total length of all cracks and the number of branch points between cracks can be calculated. The total length reflects the breadth of the crack, while the number of branches reflects its complexity. For example, morphological thinning algorithms or distance transform skeletonization algorithms can be used to extract the crack skeleton, then the skeleton pixels are traversed, and the number of neighboring pixels is counted to identify branch points and endpoints, thereby calculating the total length and the number of branches.

[0074] Step S225: Calculate the ratio of the segmented area of ​​the tongue body to the area of ​​the circumscribed rectangle of the tongue body in each of the tongue images to obtain the fatness index.

[0075] The tongue plumpness index is an indicator of tongue morphology, reflecting the fullness of the tongue. First, the tongue image needs to be segmented to obtain its precise contour and area. Then, the area of ​​the minimum bounding rectangle enclosing the tongue contour is calculated. The ratio of the segmented tongue area to the bounding rectangle area reflects the fullness of the tongue in a standardized way; the larger the ratio, the fuller (plump) the tongue, and vice versa. For example, a deep learning model can be used for tongue segmentation to obtain a tongue mask, and then the number of pixels in the mask can be calculated as the segmented tongue area, along with the minimum bounding rectangle area of ​​the mask.

[0076] Step S226: Identify the ecchymosis that exists across time points in each of the tongue images based on feature point matching, and obtain the degree of ecchymosis fixation.

[0077] Spot fixation is used to assess the stability and persistence of spots (such as ecchymosis) on the tongue, which is important for judging chronic pathological changes. By matching feature points between tongue images acquired at different time points, the position and morphological changes of specific spots can be tracked. If a spot can be stably identified and matched in images at multiple consecutive time points, and its position, size, color, and other features do not change significantly, its fixation is considered high. For example, feature point detection and matching algorithms such as SIFT, SURF, or ORB can be used to find and match the same set of feature points in tongue images at different time points, and then the matching results can be used to determine whether spots in a specific area persist and are stable.

[0078] Through the above technical solution, this application can provide an accurate and objective method for calculating feature indicators for the first type of feature sequence and the second type of feature sequence.

[0079] In one embodiment, such as Figure 5 As shown, step S300, which calculates the first evolution rate of the first type of feature sequence within the observation window and the second evolution rate of the second type of feature sequence within the observation window, and calculates the difference metric between the first evolution rate and the second evolution rate, may include steps S310 to S330.

[0080] Step S310: Input the first type of feature sequence and the second type of feature sequence into a preset time series prediction model, and extract the temporal evolution patterns of the first type of feature sequence and the second type of feature sequence through two independent attention channels in the time series prediction model.

[0081] Time series forecasting models are used to analyze and predict time-series data. They capture temporal characteristics such as trends, seasonality, periodicity, and irregular fluctuations to predict future data points. In this scheme, the time series forecasting model is used to process first-class and second-class feature sequences to identify and predict their respective dynamic evolution patterns, providing a more stable and accurate basis for subsequent rate calculations and reducing the interference of short-term noise. Time series forecasting models can employ deep learning-based models, such as Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRUs), or Transformer-based models. These models excel at handling long-sequence data and can automatically learn complex temporal dependencies. Alternatively, time series forecasting models can also employ statistical models, such as Autoregressive Integral Moving Average (ARIMA), Exponential Smoothing (ETS), or Prophet models. These models build predictive models through statistical analysis of historical data.

[0082] Attention mechanisms are techniques in neural networks that simulate human attention focus allocation. They allow the model to assign different weights to different parts of the input sequence, thus highlighting important information. Independent attention channels mean that independent attention mechanisms are designed or configured for the first and second types of feature sequences, ensuring they do not interfere with each other when extracting their respective temporal evolution patterns. This ensures that the high-frequency changing first-type feature sequences and the low-frequency changing second-type feature sequences can be independently focused on and analyzed by the model, avoiding mutual interference or masking during feature extraction, thereby more accurately capturing their unique evolution patterns. Independent attention channels can be implemented within the time series prediction model by setting separate attention layers or attention heads for the first and second types of feature sequences. For example, in the Transformer model, each feature sequence can be assigned an independent self-attention module, which considers only the internal dependencies of the sequence when calculating attention weights. Alternatively, independent attention channels can also employ a two-branch network structure, with each branch containing an independent attention mechanism to process the first and second types of feature sequences respectively, and the outputs of the two branches are then merged into subsequent layers of the model for integration.

[0083] Step S320: Calculate the first derivative of the first type of feature sequence within the prediction window as the first evolution rate, and calculate the first derivative of the second type of feature sequence as the second evolution rate.

[0084] The first derivative mathematically represents the rate of change of a function. In time series analysis, it quantifies the speed and direction of change of data points over time. A prediction window refers to the range within which a model predicts a future period based on historical data. By calculating the first derivative within the prediction window, the trend and speed of change of the characteristic sequence over the future can be quantitatively represented, i.e., the rate of evolution. This allows the system to detect changes in physiological functions in advance, rather than making delayed judgments based solely on changes that have already occurred. The first derivative can be calculated using numerical differentiation methods, such as the central difference method or the forward / backward difference method, calculating the ratio of the difference between characteristic values ​​at adjacent time points within the prediction window to the time interval; alternatively, the first derivative can also be calculated using curve fitting methods, performing curve fitting (such as polynomial fitting or spline fitting) on ​​the characteristic sequence data within the prediction window, and then differentiating the fitted function to obtain a smoother and more accurate instantaneous rate of change.

[0085] Step S330: Normalize the first evolution rate and the second evolution rate respectively, and calculate the absolute value of the difference between the normalized first evolution rate and the normalized second evolution rate to obtain the difference metric value.

[0086] The first and second evolution rates are normalized respectively. Normalization is a data preprocessing technique aimed at transforming data of different dimensions or ranges to a uniform scale, making them comparable. Since the first and second type of feature sequences may have different numerical ranges and physical meanings, their original evolution rate values ​​may not be directly comparable. Normalization eliminates this dimensional difference, allowing for a fair comparison of the degree of change between the two rates when subsequently calculating the difference metric. The normalization can employ Min-Max normalization, linearly scaling the data to a fixed interval (e.g., [0,1] or [-1,1]); or it can employ Z-score normalization (standardization), converting the data into a distribution with a mean of 0 and a standard deviation of 1.

[0087] In one embodiment, the input to the time series prediction model further includes static attribute data of the target object and known future variable data; wherein, the static attribute data includes at least one of physical condition, age, and gender, and the known future variable data includes at least one of solar term data and physiological cycle data; the time series prediction model is used to combine the static attribute data and the known future variable data to assist in predicting the first evolution rate and the second evolution rate.

[0088] This technical feature aims to expand the input dimensions of time series prediction models, enabling them to rely not only on the tongue image feature sequences themselves but also to fully consider the inherent characteristics of the target object and the potential influence of external environmental or physiological events. For example, static attribute data and known future variable data can be used as additional feature vectors, concatenated or fused with the first and second type feature sequences at the input layer of the time series prediction model; alternatively, a multimodal fusion module can be designed to process the time series data and these auxiliary data separately, and then the processed information can be fused in the middle layer of the model to capture their interactions in a more sophisticated way.

[0089] Static attribute data refers to relatively stable individual characteristics of a target object that are not easily changed rapidly over time. These characteristics reflect an individual's basic physiological state and susceptibility to disease. For example, physical constitution data can be obtained through questionnaires, traditional Chinese medicine constitution identification scales, or genetic testing, and then encoded as numerical values ​​or one-hot vectors input models. Age data can be directly input as numerical values ​​or segmented encoding; gender data can be encoded as a binary variable. In addition to the above examples, static attribute data can also include other physiological indicators that are not easily changed rapidly, such as height, weight, BMI, and family history of genetic diseases.

[0090] Known future variable data refers to external environmental or physiological event data whose future trends or periodic patterns are known at the prediction time. For example, solar term data can be automatically generated based on dates and encoded as periodic features (e.g., through sine / cosine transformation); physiological cycle data (e.g., women's menstrual cycles) can be obtained through user records, monitoring by smart wearable devices, or prediction algorithms, and represented as the number of days or cycle phases. In addition to the above examples, known future variable data can also include predictable external variables such as weather data (e.g., temperature, humidity), environmental pollution indices, and seasonal epidemic data.

[0091] The time series forecasting model is used to combine the static attribute data and the known future variable data to assist in predicting the first evolution rate and the second evolution rate.

[0092] The above technical solutions effectively address the problem of neglecting individual differences and future events when predicting evolution rates, thus significantly improving the accuracy and adaptability of predictions. Specifically, static attribute data such as constitution, age, and gender enable the model to customize the analysis of evolution rates for different target groups, enhancing the model's individual adaptability. This is because differences in constitution directly affect the speed at which physiological functions respond to internal and external disturbances, age is associated with the declining trend of the body's metabolism and recovery capabilities, and gender involves specific health risks caused by differences in hormone levels. Simultaneously, known future variable data, such as seasonal data and physiological cycle data, allow the model to proactively adjust the calculation of evolution rates within the prediction window, reducing prediction errors caused by external events or periodic factors. For example, changes in seasonal changes may cause fluctuations in environmental temperature and humidity, affecting overall physiological state, and physiological cycles involve regular hormonal changes. By simultaneously integrating inherent individual characteristics and foreseeable future events, time series prediction models can more comprehensively capture the dynamic evolution patterns of the first and second types of feature sequences, providing a more reliable basis for the generation of subsequent difference measurements. This optimizes the accuracy and timeliness of physiological function early warnings, making the warning information more personalized and forward-looking.

[0093] In one embodiment, such as Figure 6 As shown, step S400, generating corresponding physiological function warning information based on the difference measurement value, may include steps S410 and S420.

[0094] Step S410: Based on the relationship between the difference measurement value and the preset threshold, and combined with the changing trends of the first evolution rate and the second evolution rate, determine the target warning level corresponding to the target object.

[0095] The target warning level includes one of the following: Level 1 warning that characterizes functional disturbances, Level 2 warning that characterizes organic damage, and Level 3 warning that characterizes critical signals.

[0096] Step S420: Generate physiological function warning information containing the target warning level.

[0097] A preset threshold is a critical criterion used to determine whether such a difference reaches a level requiring an alert. This preset threshold can be set based on extensive clinical data and expert experience. For example, statistical analysis of tongue image data from healthy individuals and those with different disease states can determine the boundary values ​​for different physiological states. Furthermore, this preset threshold can also be dynamically optimized using machine learning algorithms. For instance, using support vector machines (SVM) or decision tree models, trained based on historical alert data and actual health outcomes, the threshold can be adaptively adjusted to improve the accuracy and sensitivity of the alert.

[0098] The first evolution rate reflects the rapid changes in high-frequency characteristics (such as tongue redness index, tongue gloss, and tongue coating moisture content) over time, while the second evolution rate reflects the relatively slow changes in low-frequency characteristics (such as crack topology index, tongue thickness index, and spot solidification degree) over time. Combining the trends of both allows for a more comprehensive understanding of the dynamic evolution pattern of the target's physiological functions. The trend can be determined in various ways. For example, time series analysis methods, such as moving averages, exponential smoothing, or linear regression, can be used to analyze the evolution rate data at continuous time points to identify its upward, downward, or stabilizing trends.

[0099] The target warning levels are a quantitative classification of the degree of physiological dysfunction in the target individual, aiming to provide intervention recommendations at different levels. The first, second, and third warning levels correspond to physiological abnormalities of different natures and severity. The first warning level represents functional disturbances, typically referring to mild imbalances or short-term fluctuations in physiological function, which may be recoverable through lifestyle adjustments, such as digestive disorders caused by short-term stress or improper diet. The second warning level represents organic damage, typically referring to persistent damage or degeneration at the physiological structure or tissue level, which may require medical intervention, such as chronic inflammation or early organ dysfunction. The third warning level represents a critical signal, typically referring to a severe pathological state requiring urgent medical intervention, such as an acute illness attack or a precursor to organ failure. These warning levels can correspond to specific clinical guidelines or interventions, thus providing clear action guidance for users or healthcare professionals.

[0100] Generating physiological function warning information containing the target warning level refers to outputting the determined warning level in a user-understandable manner so that the user or medical staff can take appropriate action. This information can be generated and presented in various forms. For example, structured text information can be generated and sent to the user via mobile application, SMS, or email. Through the above technical solution, this application overcomes the limitations of relying solely on a single difference metric for early warning. By comprehensively considering the relationship between the difference metric and a preset threshold, and combining the dynamic trends of the first evolution rate (high-frequency characteristic changes) and the second evolution rate (low-frequency characteristic changes), this application achieves refined hierarchical early warning for physiological dysfunction. This multi-dimensional, hierarchical early warning strategy significantly improves the accuracy and practicality of dynamic early warning of physiological function, contributing to the goals of disease prevention and early intervention.

[0101] In one embodiment, such as Figure 7As shown, step S410, determining the target warning level corresponding to the target object based on the relationship between the difference measurement value and the preset threshold, combined with the changing trends of the first evolution rate and the second evolution rate, may include steps S411 to S413.

[0102] Step S411: When the difference measurement value is greater than the preset threshold, and the first evolution rate shows an upward trend while the second evolution rate approaches zero, the target warning level is determined to be the first level warning.

[0103] An upward trend in the first evolution rate indicates a sustained increase in the rate of change of high-frequency features during the observation period, suggesting intensified functional activity or disturbances. This can be determined by comparing the mean or median of the first evolution rate within the current observation window with the corresponding values ​​in the previous one or more observation windows, or by performing linear regression analysis on the first evolution rate sequence; if the slope is positive and significant, it is considered an upward trend. A second evolution rate approaching zero indicates that the rate of change of low-frequency features is extremely small or negligible, suggesting relative stability in structure or morphology. This can be determined by checking whether the absolute value of the second evolution rate is less than a preset minimum threshold, or by using statistical tests to determine whether the mean of the second evolution rate sequence is not significantly different from zero.

[0104] Step S412: When the difference measurement value is greater than the preset threshold and the second evolution rate shows a continuous unidirectional change, the target warning level is determined to be the second level warning.

[0105] The second evolution rate exhibits a continuous unidirectional change, indicating that the rate of change of low-frequency features continues to increase or decrease over a relatively long period, suggesting a gradual change in structure or morphology. This can be determined by trend analysis of the second evolution rate sequence, for example, by whether its mean continuously increases or decreases over multiple consecutive observation windows, or by using non-parametric tests (such as the Mann-Kendall trend test) to determine whether a significant monotonic trend exists.

[0106] Step S413: When the second evolution rate reaches a preset extreme value and the rate of increase of the first evolution rate is greater than a preset value, the target warning level is determined to be the third level warning.

[0107] The second evolution rate reaching a preset extreme value indicates that the rate of change of low-frequency features has reached a critical high or low point, suggesting significant and potentially serious structural changes. This can be triggered by setting an absolute upper or lower limit, which is triggered when the second evolution rate exceeds or falls below that value, or by setting a percentile as the extreme value based on historical data or clinical guidelines. An increase in the first evolution rate greater than a preset value indicates an abnormal increase, signifying a sudden and significant increase in the rate of change of high-frequency features, usually accompanied by severe structural changes. This preset value is not limited here and depends on the needs of the practical application. This can be determined by comparing the deviation of the current first evolution rate from the historical average or baseline value; if the deviation exceeds a preset statistical threshold, it is judged as an abnormal increase, or by setting a steepness threshold, which is triggered when the slope of the first evolution rate exceeds the threshold within a short period of time.

[0108] The first, second, and third warning levels represent physiological functional risks of varying severity and nature. The first warning level indicates functional disturbances and aims to alert to early, reversible physiological imbalances; the second warning level indicates organic damage and aims to indicate more serious conditions involving progressive structural changes; and the third warning level indicates a critical signal and aims to warn of acute, life-threatening conditions or severe progression of diseases.

[0109] Through the above technical solution, this application can provide refined physiological function early warning for the target object based on the relationship between the difference measurement value and the preset threshold, combined with the specific changing trends of the first and second evolution rates. When the difference measurement value is greater than the preset threshold, and the first evolution rate shows an upward trend while the second evolution rate approaches zero, the system can accurately identify early functional disturbances. At this time, high-frequency features change rapidly while low-frequency features are relatively stable, which conforms to the pathological evolution law of "qi and blood first, form and substance change later" in traditional Chinese medicine, avoiding the early functional abnormality signal being masked by stable morphological features, thus issuing a first-level warning. When the difference measurement value is greater than the preset threshold, and the second evolution rate shows a continuous unidirectional change, the system can capture the gradual process of organic damage. Even if the morphological changes have not yet reached a significant level, it can still issue a second-level warning in a timely manner, providing a time window for early intervention. When the second evolution rate reaches a preset extreme value, and the first evolution rate rises abnormally, the system can accurately judge the critical signal. At this time, both functional and organic changes have reached a critical state, requiring immediate attention, thus issuing a third-level warning. Overall, this application effectively solves the problems of vague warning information and inability to accurately distinguish different types of health risks in the prior art by establishing clear rules for determining the warning level. It significantly improves the accuracy, timeliness and pertinence of physiological function warnings, and helps to achieve more effective disease prevention and health management.

[0110] This application also provides a dynamic early warning device for physiological functions, which is configured to implement the dynamic early warning method for physiological functions described above.

[0111] It should be noted that the specific embodiments of the dynamic early warning method for physiological functions are described above. Since this dynamic early warning device for physiological functions adopts all the technical solutions of all the above embodiments, it has at least all the beneficial effects brought about by the technical solutions of the above embodiments.

[0112] The dynamic early warning device for physiological functions can be a standalone hardware module, such as an integrated circuit (ASIC), a field-programmable gate array (FPGA), or a microcontroller unit (MCU), specifically designed to process tongue image data and execute early warning algorithms. Alternatively, the device can be a software system running on a general-purpose computing platform (such as a server, personal computer, or embedded system), which implements the various steps of the method through programming. In some embodiments, the device can also be a combined hardware and software system, where some computationally intensive tasks are accelerated by dedicated hardware, while control logic and data management are handled by software.

[0113] The dynamic early warning device for physiological functions is configured to implement a dynamic early warning method for physiological functions, aiming to ensure that the device can completely, accurately, and efficiently execute all steps of the dynamic early warning method for physiological functions. This can be achieved in various ways, for example, by pre-installing or loading the software program, algorithm model, and parameters implementing the method into the device. Alternatively, by designing the hardware circuit structure of the device to directly support specific computational tasks in the method, such as image processing units, neural network inference accelerators, etc. Or, by firmware programming, directly burning the logic of the method into the device's memory so that it can be executed upon startup.

[0114] This application also provides a medical device, which includes the dynamic early warning device for physiological functions described above.

[0115] It should be noted that the specific embodiments of the dynamic early warning device for physiological functions are described above. Since this medical device adopts all the technical solutions of all the above embodiments, it has at least all the beneficial effects brought about by the technical solutions of the above embodiments.

[0116] Medical devices refer to various instruments, equipment, appliances, materials, or other items used for the treatment, monitoring, or relief of diseases or injuries, or for health management and physiological function support. Specifically, such medical devices can be a dedicated tongue image analyzer that integrates image acquisition, processing, and early warning functions; they can also be part of a multifunctional health monitoring platform, such as an intelligent physical examination machine, in which tongue image analysis is a module; they can also be portable handheld devices that allow users to conduct self-monitoring at home; or they can serve as an auxiliary diagnostic terminal for hospital information systems (HIS) or electronic medical record systems (EMR).

[0117] In one feasible implementation, it is assumed that the system continuously acquires the original tongue image of user A containing a standard color card for 30 consecutive days. After processing with a dynamic color correction matrix and removing images based on clarity scores, the system generates a tongue image sequence by strictly aligning the images according to timestamps. The sequence is then input into a preset convolutional neural network and mapped to two mutually orthogonal feature spaces. In the first feature space, the weighted mean of the a* channel of CIELAB (tongue redness index), the local variance of the L* channel (tongue gloss), and the moss moisture content calculated based on high reflectivity are extracted to form a first feature vector. In the second feature space, a skeleton extraction algorithm (crack topology index), the ratio of the segmented area to the circumscribed rectangle area (fatness index), and feature point matching (spot solidification degree) are used to form a second feature vector. The two types of vectors are then spliced ​​together along the time axis to generate the first and second type of feature sequences, respectively. Next, combining the user's static attributes and external solar term variables, the two types of sequences are input into the dual-stream time-varying encoder model. The first derivatives of the two types of sequences within the prediction window are calculated through independent attention channels to obtain the first evolution rate and the second evolution rate. After normalization, the absolute value of the difference is taken as the difference metric. Finally, based on the difference metric and the rate trend, a hierarchical output is triggered: when the difference metric is greater than a preset threshold and the first evolution rate increases while the second evolution rate approaches zero, a first-level warning is output; when the difference metric is greater than a preset threshold and the second evolution rate shows a continuous unidirectional change, a second-level warning is output; and when the second evolution rate reaches a preset extreme value and the first evolution rate increases abnormally, a third-level warning is output.

[0118] It should be noted that the dynamic early warning method for physiological functions and the early warning information generated provided in this application are only intended to provide users with objective data references and management assistance for their daily health status, and do not directly lead to disease diagnosis conclusions, nor do they constitute a disease diagnosis method for medical purposes.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and not to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and they should all be covered within the scope of the specification of this application. In particular, as long as there is no technical conflict, the various technical features mentioned in the various embodiments can be combined in any way. This application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the specification.

Claims

1. A method for dynamic early warning of physiological functions, characterized in that, The dynamic early warning method for physiological functions includes: Acquire tongue images of the target object at multiple consecutive time points within a preset time period, and determine the corresponding tongue image sequence; The tongue image sequence is decoupled to obtain a first type of feature sequence and a second type of feature sequence. The first type of feature sequence has a higher frequency of change in the time dimension than the second type of feature sequence. Calculate the first evolution rate of the first type of feature sequence within the observation window and the second evolution rate of the second type of feature sequence within the observation window, and calculate the difference metric between the first evolution rate and the second evolution rate. Based on the difference measurement values, corresponding physiological function early warning information is generated.

2. The dynamic early warning method for physiological functions as described in claim 1, characterized in that, The step of acquiring tongue images of the target object at multiple consecutive time points within a preset time period and determining the corresponding tongue image sequence includes: Acquire original tongue images of a target object at multiple consecutive time points within a preset time period, wherein the original tongue images contain a preset reference object; Based on the preset reference objects in each of the original tongue images, a dynamic color correction matrix corresponding to each of the original tongue images is determined. Based on the dynamic color correction matrix, the corresponding original tongue image is subjected to color correction processing to obtain a calibrated tongue image; The clarity score of the calibrated tongue image is calculated, and the calibrated tongue image is filtered based on the clarity score. The filtered calibrated tongue image images are combined according to the time dimension to determine the tongue image sequence.

3. The dynamic early warning method for physiological functions as described in claim 1, characterized in that, The step of decoupling the tongue image sequence to obtain a first type of feature sequence and a second type of feature sequence, wherein the first type of feature sequence has a higher frequency of change in the time dimension than the second type of feature sequence, includes: By using a pre-defined convolutional neural network, each tongue image in the tongue image sequence is mapped to a first feature space and a second feature space that are orthogonal to each other. The first feature space represents high-frequency evolution, and the second feature space represents low-frequency evolution. Each of the tongue images is obtained in the first feature space and in the second feature space. The first feature vector contains at least one of the image-based feature indicators, namely, tongue redness index, tongue gloss, and tongue coating moisture content. The second feature vector contains at least one of the image-based feature indicators, namely, crack topology index, tongue thickness index, and spot solidification degree. The first feature vectors corresponding to each of the tongue images in the first feature space are combined according to the time dimension to obtain the first type of feature sequence; The second feature vectors corresponding to each of the tongue images in the second feature space are combined according to the time dimension to obtain the second type of feature sequence.

4. The dynamic early warning method for physiological functions as described in claim 3, characterized in that, The first feature vector includes image-based feature indicators such as tongue redness index, tongue gloss, and tongue coating moisture content; the second feature vector includes image-based feature indicators such as crack topology index, tongue thickness index, and spot solidification degree. The step of obtaining the first feature vector corresponding to each of the tongue images in the first feature space and the second feature vector corresponding to each of the tongue images in the second feature space includes: Each of the tongue images is converted to the CIELAB color space, and the tongue redness index is calculated based on the weighted average of the a* channel in the CIELAB color space. Each of the tongue images is converted to the CIELAB color space, and the gloss of the tongue surface is calculated based on the local variance of the L* channel in the CIELAB color space. The water content of the tongue moss is calculated based on the reflectance distribution characteristics of the highlight areas in each of the tongue images. The total length and number of branches of the cracks in each of the tongue images are calculated based on the skeleton extraction algorithm to obtain the crack topology index. The ratio of the segmented area of ​​the tongue body to the area of ​​the circumscribed rectangle of the tongue body in each of the tongue images is calculated to obtain the fatness index; Based on feature point matching, the ecchymosis that is fixed across time points in each of the tongue images is identified, and the degree of ecchymosis is obtained.

5. The dynamic early warning method for physiological functions as described in claim 1, characterized in that, The step of calculating the first evolution rate of the first type of feature sequence within the observation window and the second evolution rate of the second type of feature sequence within the observation window, and calculating the difference metric between the first evolution rate and the second evolution rate, includes: The first type of feature sequence and the second type of feature sequence are input into a preset time series prediction model. The temporal evolution patterns of the first type of feature sequence and the second type of feature sequence are extracted through two independent attention channels in the time series prediction model. The first derivative of the first type of feature sequence within the prediction window is calculated as the first evolution rate, and the first derivative of the second type of feature sequence is calculated as the second evolution rate. The first evolution rate and the second evolution rate are normalized respectively, and the absolute value of the difference between the normalized first evolution rate and the normalized second evolution rate is calculated to obtain the difference metric value.

6. The dynamic early warning method for physiological functions as described in claim 5, characterized in that, The input to the time series prediction model also includes static attribute data of the target object and known future variable data; The static attribute data includes at least one of physical condition, age, and gender, and the known future variable data includes at least one of solar term data and physiological cycle data. The time series prediction model is used to combine the static attribute data and the known future variable data to assist in predicting the first evolution rate and the second evolution rate.

7. The dynamic early warning method for physiological functions as described in claim 1, characterized in that, The generation of corresponding physiological function early warning information based on the difference measurement value includes: Based on the relationship between the difference metric and the preset threshold, and combined with the changing trends of the first evolution rate and the second evolution rate, the target warning level corresponding to the target object is determined. The target warning level includes one of the following: a first-level warning that characterizes functional disturbance, a second-level warning that characterizes organic loss, and a third-level warning that characterizes critical signal. Generate physiological function warning information that includes the target warning level.

8. The dynamic early warning method for physiological functions as described in claim 7, characterized in that, The step of determining the target warning level corresponding to the target object based on the relationship between the difference metric value and the preset threshold, combined with the changing trends of the first evolution rate and the second evolution rate, includes: When the difference metric value is greater than the preset threshold, and the first evolution rate shows an upward trend while the second evolution rate approaches zero, the target warning level is determined to be the first level warning. When the difference metric value is greater than the preset threshold and the second evolution rate shows a continuous unidirectional change, the target warning level is determined to be the second level warning. When the second evolution rate reaches a preset extreme value and the rate of increase of the first evolution rate is greater than a preset value, the target warning level is determined to be the third level warning.

9. A dynamic early warning device for physiological functions, characterized in that, The physiological function dynamic early warning device is configured to implement the physiological function dynamic early warning method as described in any one of claims 1 to 8.

10. A medical device, characterized in that, The medical device includes the dynamic early warning device for physiological functions as described in claim 9.

Images