Traditional Chinese medicine physique dynamic change identification system

Abstract

The invention provides a traditional Chinese medicine physique dynamic change identification system. The system comprises a data acquisition module, a feature extraction module, a deep learning analysis module, a physique analysis and prediction module and a diagnosis report generation module. The data acquisition module is used for acquiring tongue image video streams, pulse condition waveforms and body surface infrared thermal imaging data and recording environment temperature and humidity, user behavior logs and medication information. The feature extraction module is used for extracting various feature data. And inputting the multi-dimensional data set into a deep learning analysis module, and calculating the transformation relation probability of the traditional Chinese medicine constitution types. The physique analysis and prediction module judges the current physique type based on the probability data and predicts the physique evolution trend in combination with historical data, and finally the diagnosis report generation module outputs a diagnosis report containing physique evolution analysis and conditioning suggestions. According to the method, multi-dimensional physiological, environment and behavior data can be integrated, dynamic identification of traditional Chinese medicine physique is realized, the accuracy of physique analysis is improved, and a personalized health management scheme is provided.

Description

Technical Field

[0001] This application relates to the field of medical diagnosis technologies, and particularly to a system and method for identifying the dynamic changes of traditional Chinese medicine constitutions. Background Art

[0002] In the prior art, the identification of traditional Chinese medicine constitutions mainly relies on the clinical experience of physicians, and evaluates the constitution types of individuals through methods such as inspection, inquiry, and palpation. In recent years, with the development of medical informatization, constitution identification methods based on image processing, signal analysis, and artificial intelligence technologies have been gradually applied. Common methods include tongue image analysis, pulse signal processing, and constitution classification models combined with questionnaires.

[0003] However, the existing traditional Chinese medicine constitution identification methods generally have the following problems: First, most methods only focus on single or limited physiological characteristics, lacking comprehensive analysis of multi-dimensional physiological, environmental, and behavioral factors, resulting in limited accuracy of constitution assessment; Second, the existing methods are mostly based on static data or short-term data, failing to fully consider the constitution evolution process and being difficult to effectively predict the dynamic change trend of individual constitutions; Third, although some studies use machine learning methods for constitution classification, there are still great limitations in modeling the conversion relationship between constitution types and personalized health management, and it is difficult to provide targeted conditioning suggestions.

[0004] In view of this, there is an urgent need for a system for identifying the changes of traditional Chinese medicine constitutions that can integrate multi-dimensional data and combine machine learning for dynamic analysis, so as to improve the accuracy of constitution assessment and achieve personalized constitution evolution prediction and health conditioning plans. Summary of the Invention

[0005] This application provides a system for identifying the dynamic changes of traditional Chinese medicine constitutions to improve the accuracy of constitution analysis and provide a personalized health management plan.

[0006] This application provides a system for identifying the dynamic changes of traditional Chinese medicine constitutions, including:

[0007] A data acquisition module, configured to collect the tongue image video stream data, pulse waveform data, and body surface infrared thermal imaging data of a user, and record the environmental temperature and humidity data, user behavior log data, and medication information data;

[0008] A feature extraction module, connected to the data acquisition module, is used to perform image segmentation on the tongue image video stream data, extract the tongue body and tongue coating regions, and obtain tongue image feature data including tongue color, tongue shape, and tongue coating features; perform wavelet transform on the pulse waveform data to extract pulse time-frequency data including waveform amplitude, frequency, and phase features; process the surface infrared thermal imaging data to obtain surface temperature distribution data, and jointly construct a multi-dimensional dataset with the tongue image feature data, pulse time-frequency data, and surface temperature distribution data according to the environmental temperature and humidity data, user behavior log data, and medication information data;

[0009] A deep learning analysis module, connected to the feature extraction module, is used to input the multi-dimensional dataset into a pre-trained deep learning model to obtain probability data representing the conversion relationship between different traditional Chinese medicine constitution types;

[0010] A constitution analysis and prediction module, connected to the deep learning analysis module, is used to determine the user's current traditional Chinese medicine constitution type based on the probability data, and predict the future traditional Chinese medicine constitution change trend of the user in combination with historical acquisition data, and generate a diagnostic report including constitution evolution analysis and conditioning suggestions.

[0011] The technical solution proposed in this application has the following beneficial technical effects:

[0012] (1) By collecting tongue image, pulse, and surface infrared thermal imaging data, and combining environmental temperature and humidity, user behavior, and medication information, a multi-dimensional dataset containing physiological, environmental, and behavioral factors is constructed. Compared with traditional single-index identification methods, the comprehensiveness and accuracy of constitution analysis are improved. (2) Image segmentation is used to extract tongue image features, wavelet transform is used to extract pulse time-frequency features, and surface temperature data is analyzed. Combined with deep learning technology, key features are automatically extracted, avoiding the subjectivity of manual identification and improving the automation level and efficiency of constitution feature extraction. (3) Combining historical data and the probability model of constitution conversion relationship, dynamic evolution prediction of the user's constitution is realized, which can provide personalized constitution change trend analysis, avoid the limitations of static constitution identification methods, and provide more targeted health management suggestions for users. (4) By inputting multi-dimensional data into a pre-trained deep learning model, calculating the conversion relationship between different traditional Chinese medicine constitution types, and combining historical data for trend prediction, the stability of constitution classification can be effectively improved, and scientific and reasonable conditioning suggestions can be provided for users, improving the intelligent level of health management. Description of the Drawings

[0013] Figure 1 is a schematic diagram of a system for identifying the dynamic changes of traditional Chinese medicine constitution provided by the first embodiment of this application. Detailed Embodiments

[0014] In the following description, numerous specific details are set forth to provide a thorough understanding of the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the connotation of the present application. Therefore, the present application is not limited by the specific implementations disclosed below.

[0015] The first embodiment of the present application provides a system for identifying the dynamic changes of traditional Chinese medicine constitutions. Please refer to Figure 1 , which is a schematic diagram of the first embodiment of the present application. The following will be combined with Figure 1 to elaborate in detail on a system for identifying the dynamic changes of traditional Chinese medicine constitutions provided by the first embodiment of the present application.

[0016] The system for identifying the dynamic changes of traditional Chinese medicine constitutions includes a data acquisition module 101, a feature extraction module 102, a deep learning analysis module 103, and a constitution analysis and prediction module 104.

[0017] The data acquisition module 101 is used to collect the tongue image video stream data, pulse waveform data, and body surface infrared thermal imaging data of the user, and record the environmental temperature and humidity data, user behavior log data, and medication information data.

[0018] The data acquisition module 101 is used to collect the multi-dimensional physiological, environmental, and behavioral data of the user, and perform structured storage on it for subsequent analysis and processing. This module includes multiple sub-units, corresponding to different data acquisition tasks respectively, to ensure that the acquired data is comprehensive, accurate, and has temporal continuity.

[0019] First of all, the data acquisition module 101 includes a tongue image video stream data acquisition unit. This unit uses a high-resolution camera or a dedicated tongue image acquisition device to dynamically capture the user's tongue image and record the state of the tongue body under different lighting conditions. This unit can combine polarized light technology and multi-spectral imaging technology to enhance the color and texture details of the tongue image and improve the identification ability of features such as the thickness of the tongue coating, the shape of the tongue body, and the wetness of the tongue surface. To ensure data consistency, the system can guide the user to perform the acquisition in a standardized lighting environment and use a deep learning-assisted framework for automatic calibration of the tongue body area to avoid the influence of factors such as shooting angle and light change on the data quality.

[0020] Secondly, the data acquisition module 101 also includes a pulse waveform data acquisition unit, which usually uses a high-precision pressure sensor, a photoplethysmogram (PPG) sensor, or an ultrasonic Doppler sensor to collect pulse signals from the radial artery or other parts. This unit can obtain continuous waveform data of the pulse at a certain sampling rate and store it in real time to ensure that the pulse data can be used for time-frequency feature analysis. This unit can adopt a multi-channel synchronous acquisition method to ensure the relative timing relationship of the pulse at different measurement points, thereby improving the accuracy and integrity of the data. In addition, the system can combine environmental factors, such as the user's body position and pressure conditions during measurement, to perform data correction to reduce measurement errors.

[0021] In addition, the data acquisition module 101 includes a body surface infrared thermal imaging data acquisition unit, which uses a high-sensitivity infrared thermal imaging sensor or a thermopile sensor to measure the temperature distribution on the user's body surface. This unit can capture the body surface heat distribution pattern and generate an infrared thermal image to reflect the temperature differences in different parts. The system can, based on the infrared thermal imaging data, combine a medical temperature correction algorithm to compensate for the influence of the external environmental temperature on the body surface temperature and calculate the temperature characteristics of areas such as the tongue, limbs, and torso.

[0022] To provide more comprehensive constitution identification information, the data acquisition module 101 also includes an environmental temperature and humidity recording unit, a user behavior log recording unit, and a medication information recording unit. The environmental temperature and humidity recording unit uses temperature and humidity sensors to monitor the temperature and humidity of the user's environment in real time to ensure that external environmental influences are considered in constitution analysis. The user behavior log recording unit is used to collect the user's living habits, exercise conditions, and sleep status, and the data can be obtained through smart wearable devices, mobile phone applications, or self-filled questionnaires. The medication information recording unit is used to record the user's medication history, including information such as drug types, dosages, and taking times, and can be stored through an electronic health record interface or user input for subsequent analysis of the relationship between constitution evolution and medication.

[0023] All the acquisition data of the data acquisition module 101 will be integrated into a standardized database to ensure unified data format and with timestamps for subsequent data analysis modules to perform time series modeling and feature extraction. This module also supports remote data acquisition functions, allowing users to upload data in different environments and combining cloud storage technology to achieve long-term data accumulation, providing basic data support for the analysis of constitution evolution trends.

[0024] Furthermore, the data acquisition module includes a temperature-controlled light source unit, which is used to dynamically adjust the color temperature and brightness of the light source according to the ambient temperature when collecting tongue image video stream data, so as to adapt to the tongue image reflection spectral characteristics of different individuals; wherein, the temperature-controlled light source unit includes multiple LED light sources with different color temperatures and a temperature sensor. After the temperature sensor detects the ambient temperature, it selects an appropriate combination of light sources with different color temperatures according to a preset light compensation model, and adjusts the output power of the light source by means of pulse width modulation to reduce the influence of ambient light on the extraction of tongue image color features and improve the stability and consistency of tongue image data.

[0025] The temperature-controlled light source unit is used to dynamically adjust the color temperature and brightness of the light source according to the ambient temperature when collecting tongue image video stream data, so as to adapt to the tongue image reflection spectral characteristics of different individuals. The design of this unit takes into account the influence of lighting conditions on color feature extraction during tongue image acquisition, ensures stable tongue image data can be obtained in different environments, and thus improves the accuracy of subsequent feature extraction and constitution identification.

[0026] This unit consists of multiple LED light sources with different color temperatures and a temperature sensor. The LED light sources include low-color-temperature, medium-color-temperature and high-color-temperature light sources, which are used to provide a warm, standard and cold lighting environment respectively. The temperature sensor is used to detect the temperature of the tongue image acquisition environment in real time and transmit the temperature data to the control module. The control module calculates the optimal color temperature combination according to the preset light compensation model and the currently detected ambient temperature to match the natural reflection characteristics of the tongue surface. This light compensation model is established through experimental data and physiological spectral analysis, taking into account the tongue body colors of different individuals, the light adaptation ability, and the spectral shift of the light source under different temperature conditions.

[0027] After calculating the target color temperature, the control module selects the corresponding combination of LED light sources and adjusts the output power of the light source by means of pulse width modulation (PWM) to precisely control the light intensity. The PWM control method allows the color temperature stability of the light source to be maintained under different brightness conditions, avoiding the occurrence of color deviation during the brightness adjustment process. Specifically, when the ambient temperature is low, the system will preferentially increase the power output of the low-color-temperature light source and appropriately reduce the brightness of the high-color-temperature light source to simulate a more natural warm-tone lighting environment; when the ambient temperature is high, the system increases the proportion of the high-color-temperature light source, making the overall lighting tend to be cold-toned to offset the red enhancement effect that may be caused by warm light sources. In addition, the control module can also dynamically optimize the color temperature adjustment strategy according to historical data to adapt to the trend of ambient light change during long-term use, further improving the stability and consistency of tongue image acquisition.

[0028] Through the above method, the temperature-controlled light source unit can reduce the influence of ambient light on the extraction of tongue image color features, so that the tongue image data obtained by the same user under different acquisition environments can maintain high consistency. This light source adjustment method not only improves the adaptability of the system in various environments, but also reduces the deviation of tongue image features caused by changes in lighting conditions, thereby improving the reliability and accuracy of the traditional Chinese medicine constitution dynamic change identification system.

[0029] Furthermore, the data acquisition module includes an adaptive signal adjustment unit, which is used to dynamically adjust the sampling frequency, signal gain, and filtering parameters according to the pulse characteristics of the user individual when collecting pulse waveform data. Among them, the adaptive signal adjustment unit includes a pulse waveform analysis module and a signal control module. The pulse waveform analysis module calculates the pulse intensity, signal-to-noise ratio, and pulse waveform morphology based on the preliminarily collected pulse data, and inputs the calculation results into the signal control module. The signal control module dynamically adjusts the sampling rate, gain parameter, and filtering threshold of the sensor according to the preset adaptive adjustment algorithm to optimize the signal quality and ensure accurate and reliable acquisition of pulse data under different individuals and different physiological states.

[0030] The adaptive signal adjustment unit is used to dynamically adjust the sampling frequency, signal gain, and filtering parameters according to the pulse characteristics of the user individual when collecting pulse waveform data, so as to adapt to the pulse characteristics of different individuals, optimize the signal quality, and improve the accuracy and stability of pulse data. Since there are significant differences in the pulse waveforms among individuals, the traditional signal acquisition method with fixed parameters is difficult to take into account the physiological characteristics of different users. However, this unit realizes personalized optimization processing of pulse signals through an adaptive adjustment strategy.

[0031] This unit includes a pulse waveform analysis module and a signal control module. The pulse waveform analysis module quickly analyzes the received original signal at the initial stage of pulse data acquisition, and calculates key indicators such as pulse intensity, signal-to-noise ratio, and pulse waveform morphology. Pulse intensity is an important parameter to measure the stability of signal amplitude. If the intensity is too low, it may indicate insufficient sampling gain or poor sensor fit; the signal-to-noise ratio is used to measure the proportion of effective components in the signal. If the noise level is high, it may be affected by movement interference, power supply noise, or unstable sensor contact; the pulse waveform morphology reflects the fluctuation characteristics within the pulse cycle, such as the main wave peak, secondary wave peak, and pulse cycle length. These information can be used to judge the accuracy and integrity of signal acquisition.

[0032] The results calculated by the analysis module are input into the signal control module, which dynamically optimizes the signal acquisition parameters according to a preset adaptive adjustment algorithm. First, the adjustment of the sampling rate is based on the pulse waveform cycle length and the main peak frequency distribution. When the pulse is fast, the system increases the sampling frequency to ensure the integrity of waveform details; when the pulse is slow, the sampling frequency is appropriately reduced to reduce data redundancy and optimize storage and calculation efficiency. Second, the adjustment of the gain parameter is based on the amplitude distribution of the pulse signal. If a low signal amplitude is detected, the gain is appropriately increased to enhance signal distinguishability; if the amplitude is too high, the gain is reduced to prevent distortion caused by signal saturation. The adjustment of the filtering threshold is optimized by combining the noise level and the pulse morphology characteristics. When strong high-frequency noise is detected, the system automatically adjusts the cut-off frequency of the filter to remove non-physiological signal interference; when the overall signal quality is good, the filtering intensity is reduced to retain the detailed information of the pulse condition to the greatest extent.

[0033] The entire adaptive signal adjustment process runs dynamically during the real-time acquisition process to ensure that the signal parameters are always in the optimal state. Through this method, whether it is an individual with a weak pulse or a situation where the signal fluctuates greatly due to movement or environmental interference, the system can achieve accurate and stable acquisition of pulse condition data, thus providing high-quality data support for subsequent constitution identification.

[0034] Furthermore, the data acquisition module includes a dynamic temperature baseline adjustment unit. The dynamic temperature baseline adjustment unit is used to construct a personalized reference curve based on the user's historical temperature data when collecting body surface infrared thermography data, and dynamically adjust the temperature data during subsequent measurements to adapt to the individual body temperature change pattern. Among them, the dynamic temperature baseline adjustment unit includes a user temperature database, a trend analysis module, and a temperature compensation module. The user temperature database stores the body surface temperature data of the user at multiple time nodes. The trend analysis module performs time series modeling on the historical data, calculates the temperature change law of the user individual, and generates a personalized reference curve. The temperature compensation module adjusts the current temperature data according to the reference curve during each measurement to reduce the impact of individual differences on constitution identification and improve the stability and accuracy of long-term body temperature monitoring.

[0035] The dynamic temperature baseline adjustment unit is used to construct a personalized baseline curve based on the historical body temperature data of the user individual when collecting surface infrared thermal imaging data, and dynamically adjust the temperature data during subsequent measurements to make it more conform to the body temperature change pattern of the user himself. Since the body temperature of an individual is affected by various factors such as physiological rhythm, environmental temperature, and activity status, traditional surface temperature measurement methods often rely on a fixed normal temperature range for comparative analysis, ignoring the baseline differences between individuals. However, this unit can establish a baseline curve that conforms to individual characteristics through the analysis of the user's long-term body temperature data, and adaptively adjust the temperature data during the measurement process, thereby reducing the impact of individual differences on physical constitution identification and improving the stability and accuracy of long-term body temperature monitoring.

[0036] This unit includes a user temperature database, a trend analysis module, and a temperature compensation module. The user temperature database is used to store the surface temperature data of the user at multiple time nodes, covering the surface temperature information under resting state, active state, and different environmental temperature conditions, so as to construct a complete individual body temperature record. This database supports continuous data update. The new data measured each time will be stored and participate in subsequent analysis to continuously optimize the accuracy of the baseline curve.

[0037] The trend analysis module performs time series modeling on the user's historical temperature data, calculates the individual temperature change law, and extracts the temperature characteristics under different states based on statistical analysis methods. For example, this module can analyze the body temperature change pattern of the user in the circadian rhythm, identify the periodic fluctuation trend of the temperature, and calculate the temperature offset under different environmental conditions. At the same time, the system uses a sliding window to calculate the temperature change trend of the user within a specific time range to ensure that the baseline curve can reflect the long-term change of the body temperature and has the ability to adapt to short-term fluctuations. Based on these analysis results, the system generates a personalized body temperature baseline curve, which not only reflects the static temperature characteristics of the user, but also can dynamically adjust the adaptability to short-term environmental changes, making the calibration of subsequent measurement data more accurate.

[0038] The temperature compensation module adjusts the current temperature data according to the personalized baseline curve during each measurement to make it match the user's historical temperature pattern, so as to reduce the measurement error. For example, when it is detected that there is a deviation between the currently measured temperature and the baseline curve, the system will calculate the average value of this deviation and make corrections in combination with factors such as environmental temperature and measurement time to ensure the accuracy of the temperature data. For short-term temperature fluctuations, this module can identify whether they belong to normal physiological changes to avoid misjudgment of abnormal body temperature caused by accidental factors. At the same time, this module can also adjust the compensation parameters by using the continuous measurement data of the user through a real-time feedback mechanism to ensure that the system can adapt to the natural change trend of the user's body temperature during long-term monitoring, thereby improving the reliability of physical constitution identification.

[0039] Through the above method, the dynamic temperature baseline adjustment unit can provide a more accurate measurement benchmark in individual body temperature monitoring, enabling the body temperature data to more accurately reflect the user's true physiological state, avoiding misjudgment caused by individual differences or environmental impacts, and improving the adaptability and accuracy of the physical constitution analysis system.

[0040] The feature extraction module 102, connected to the data acquisition module, is used to perform image segmentation processing on the tongue image video stream data, extract the tongue body and tongue coating regions, and obtain tongue image feature data including tongue color, tongue shape, and tongue coating features; perform wavelet transform on the pulse waveform data to extract pulse time-frequency data including waveform amplitude, frequency, and phase features; process the body surface infrared thermal imaging data to obtain body surface temperature distribution data, and jointly construct a multi-dimensional data set with the tongue image feature data, pulse time-frequency data, and body surface temperature distribution data according to the environmental temperature and humidity data, user behavior log data, and medication information data.

[0041] The feature extraction module 102 is used to analyze and extract features from the tongue image video stream data, pulse waveform data, and body surface infrared thermal imaging data acquired by the data acquisition module 101 to construct a multi-dimensional data set that can characterize the user's physical constitution state. This module includes multiple sub-units, and each sub-unit executes specific processing methods for different data types respectively to ensure that the acquired data has high feature resolution and stability for subsequent analysis.

[0042] In terms of tongue image data processing, this module first performs frame segmentation on the input tongue image video stream data to extract the key frames, so as to reduce redundant information and improve the calculation efficiency. Subsequently, image preprocessing is performed on the key frames, including color normalization, brightness correction, and denoising, to reduce the impact of lighting conditions on tongue image feature extraction. The preprocessed image is input into the image segmentation model, which can adopt the U-Net structure of deep learning or traditional image processing methods based on edge detection to accurately segment the tongue body and tongue coating regions. The tongue body region further extracts features such as tongue color and tongue shape. Among them, tongue color analysis is based on color histogram statistics, and tongue shape analysis uses contour detection algorithms to calculate parameters such as the length, width, and crack condition of the tongue body. The tongue coating region extracts tongue coating thickness, tongue coating coverage rate, and tongue coating color features. The tongue coating color can be classified through HSV color space conversion, and the tongue coating thickness can be evaluated using texture analysis methods. All the extracted tongue image feature data are subjected to feature vectorization processing for use by subsequent analysis modules.

[0043] In terms of pulse condition data processing, this module preprocesses the pulse waveform data, including baseline drift correction and denoising, to improve the signal stability. Subsequently, multi-scale analysis of the pulse signal is performed through wavelet transform to extract time-frequency domain features. Wavelet transform can use Daubechies wavelet or Morlet wavelet, enabling effective separation of high-frequency information (such as pulse oscillation changes) and low-frequency information (such as overall pulse trend) in the pulse waveform. Further calculate the amplitude, frequency, and phase characteristics of the pulse signal, where the amplitude reflects the strength of the pulse, the frequency represents the rate of pulse beating, and the phase characteristics can be used to evaluate the delay of pulse wave propagation. These characteristics can be used for pulse pattern analysis in constitution identification and are stored in the feature vector library.

[0044] In terms of the processing of body surface infrared thermal imaging data, this module first corrects the infrared images to compensate for the influence of environmental temperature changes during measurement on the data. The resolution of the infrared thermal map can be enhanced through bilinear interpolation or super-resolution algorithms to improve the spatial resolution of temperature data. Subsequently, temperature distribution analysis is performed on the thermal imaging data, and different regions' temperature characteristics are segmented by setting temperature thresholds, such as the temperature distribution of key constitution-related regions like the tongue, limbs, abdomen, etc. The system calculates indicators such as the average temperature, maximum temperature difference, and local temperature gradient of different regions to evaluate the user's body surface thermal state. These temperature characteristics, when combined with other physiological characteristics, can be used to analyze the cold or heat tendency of the constitution.

[0045] After the feature extraction module 102 completes the processing of tongue image, pulse condition, and thermal imaging data, it also needs to combine environmental temperature and humidity data, user behavior log data, and medication information data to construct a complete multi-dimensional data set. Environmental temperature and humidity data are used to evaluate the impact of the external climate on the constitution state, user behavior log data provides information on living habits such as sleep, exercise, and diet, and medication information data records the user's drug intervention situation. These data are all standardized and associated with physiological feature data using time series modeling methods to construct a constitution state representation that changes over time. Finally, the multi-dimensional data set output by this module can be used by the subsequent deep learning analysis module for constitution state determination and trend prediction to achieve more accurate traditional Chinese medicine constitution identification.

[0046] Furthermore, the feature extraction module includes a tongue image feature enhancement unit. The tongue image feature enhancement unit is used to, after performing image segmentation processing on the tongue image video stream data, adopt the local histogram equalization method based on adaptive contrast enhancement to optimize the color and texture features of the tongue body and tongue coating regions. At the same time, combined with a deep learning-based tongue image abnormality detection network, it automatically identifies features such as tongue cracks, tooth marks, and stasis points, and generates a high-resolution tongue image feature map to improve the accuracy and consistency of tongue image analysis.

[0047] The tongue image feature enhancement unit is used to further optimize the color and texture features of the tongue body and tongue coating regions after image segmentation processing of the tongue image stream data, so as to improve the accuracy and consistency of tongue image analysis. In tongue image acquisition, due to the influence of different lighting conditions, camera imaging characteristics, and individual differences, there may be large deviations in the contrast, color saturation, and local details of tongue images, making it difficult for traditional tongue image analysis methods to accurately extract key features. This unit uses the method of local histogram equalization with adaptive contrast enhancement, combined with a deep learning-based tongue image abnormality detection network, to achieve high-precision enhancement and automatic recognition of tongue images.

[0048] In the image optimization process, first, local histogram equalization processing is performed on the segmented tongue body and tongue coating regions to enhance the texture details of low-contrast regions and make the tongue image features clearer. The traditional global histogram equalization method may cause over-enhancement in some regions when enhancing the contrast, affecting the stability of actual features. This unit uses local histogram equalization technology to adaptively adjust the degree of contrast enhancement in different regions, thus avoiding information loss caused by overall equalization. In addition, this method combines an adaptive contrast enhancement strategy to dynamically adjust the enhancement amplitude according to the brightness distribution of different regions of the tongue body, so that darker regions can obtain appropriate brightness improvement, while brighter regions will not be over-enhanced, ensuring the color and contrast of the tongue image are balanced and stable.

[0049] In the process of tongue image abnormality detection, this unit uses a deep learning-based tongue image analysis model to automatically identify features such as tongue cracks, tooth marks, and petechiae. The deep learning model can accurately learn the morphology and distribution rules of different abnormal features by training a large number of tongue image datasets, and automatically detect the corresponding abnormal regions in the newly input tongue images. The detection of tongue cracks is based on edge enhancement and morphological analysis to accurately segment the fine cracks on the tongue surface; the detection of tooth marks uses tongue edge curvature analysis to identify the irregular indentation features on the tongue edge; the detection of petechiae combines color distribution and local texture analysis to identify the abnormal dark red or purple regions on the tongue surface.

[0050] In the stage of generating tongue image feature images, this unit reconstructs the optimized tongue body and tongue coating regions with high resolution to ensure that sufficient detailed information can be retained in subsequent analysis processes. Through high-precision enhancement processing of tongue images, this unit not only improves the clarity of tongue image features but also reduces data deviations caused by changes in acquisition conditions, making the tongue image data of the same user consistent at different times or in different environments. In this way, this unit can significantly improve the reliability of tongue image analysis, making the application of the constitution identification system in traditional Chinese medicine diagnosis more accurate and stable.

[0051] Furthermore, the feature extraction module includes a pulse dynamic feature modeling unit. The pulse dynamic feature modeling unit is used to perform time series modeling on the pulse waveform data after wavelet transform using a multi-scale recurrent neural network to extract short-term fluctuation features, long-term trend features, and pulse periodic change patterns, and analyze the spectral components of the pulse signal in combination with Fourier transform to establish a personalized pulse feature curve and achieve accurate pulse classification and constitution correlation analysis.

[0052] The pulse dynamic feature modeling unit is used to deeply analyze the pulse waveform data to extract pulse features on different time scales, so as to achieve accurate modeling of the pulse state and establish an association with constitution information. Since the pulse waveform is affected by individual physiological states, external environments, and measurement methods, its time series features have significant short-term volatility and long-term trend changes. Traditional analysis methods based on fixed parameters are difficult to accurately capture these complex dynamic features. This unit combines wavelet transform, multi-scale time series modeling, and spectral analysis techniques to perform multi-level analysis on pulse data to obtain high-precision and personalized pulse feature descriptions.

[0053] In the data processing stage, first perform wavelet transform on the pulse waveform data to decompose the signal into time-frequency components of different frequency bands, so as to extract high-frequency, low-frequency, and intermediate-frequency information respectively. The high-frequency component is used to capture the short-term fluctuation features of the pulse, including the subtle amplitude changes of the pulse, the relative amplitude ratio between the main wave and the secondary wave, etc.; the low-frequency component reflects the long-term trend of the pulse, including the change pattern of the basic waveform, the overall hemodynamic characteristics, etc.; the intermediate-frequency component is used to analyze the features of the pulse in different states, such as the relative time ratio between the systolic and diastolic periods, the acceleration curve of the pulse, etc. In this way, this unit can separate different information components of the pulse data for subsequent analysis.

[0054] In the time series modeling process, this unit uses a multi-scale recurrent neural network to perform structured modeling on the decomposed pulse signal. This modeling method can effectively capture the long-term and short-term dependencies in time series data, enabling the system to analyze the change trend of the pulse signal rather than just extracting the features of the static waveform. In terms of short-term fluctuation modeling, this unit analyzes the key points within each pulse cycle, including the relative positions of the main peak, secondary peak, and notch point, and extracts the time, amplitude, and phase information of these key points to establish a short-term feature vector. In terms of long-term trend modeling, this unit calculates the change trend between multiple pulse cycles and learns the individualized pulse pattern through historical pulse data to predict the future pulse evolution trend. This process ensures that pulse analysis can adapt to the physiological state changes of different individuals rather than relying on fixed rules for classification.

[0055] In the spectrum analysis stage, this unit combines Fourier transform to analyze the frequency components of the pulse signal in order to extract stable spectral features. Fourier transform can identify the periodic components in the pulse signal, such as the basal heart rate, harmonic distribution, low-frequency and high-frequency power ratio, etc. These parameters can be used to further classify different types of pulse states. In addition, this unit evaluates the overall stability of the pulse by calculating the energy ratio of different frequency bands, and combines the time-domain analysis results to generate a personalized pulse feature curve.

[0056] Through the above method, this unit can extract the key features of the pulse at different time scales and provide a comprehensive pulse description through time series modeling and spectrum analysis. This method not only improves the classification accuracy of pulse data, but also ensures the individual pulse modeling ability, enabling the constitution analysis system to achieve more accurate traditional Chinese medicine constitution identification and health assessment based on accurate pulse features.

[0057] Furthermore, the feature extraction module includes a refined pulse analysis unit and a dynamic constitution identification unit. The refined pulse analysis unit is used to analyze the overall pulse body and pulse potential during the analysis of pulse waveform data, and extract specific pulse features including stasis, blood stasis, water, dampness, phlegm, cold, heat, etc., including the amplitude difference of the pulse waveform, the ratio of the main wave to the secondary wave, the peak-to-peak distance, and the pulse propagation delay parameters. Combining non-linear signal processing methods, it calculates the weight distribution of different pulse features to improve the refined analysis ability of the constitution state.

[0058] The dynamic constitution identification unit conducts time series analysis on the user's constitution data based on the theory of five evolutive phases and six climatic factors or the twenty-four solar terms model, combines environmental temperature and humidity, seasonal changes, and climate trends, and dynamically adjusts the calculation parameters of the constitution identification model to adapt to the constitution change patterns under different climate conditions, and optimizes the conditioning plan to provide more accurate constitution assessment and personalized health management suggestions.

[0059] The refined pulse analysis unit is used to deeply analyze specific pulse characteristics related to traditional Chinese medicine theory during the analysis of pulse waveform data, in order to provide more accurate constitution analysis. Traditional pulse analysis usually only makes an overall assessment of the pulse body and pulse force, while ignoring the dynamic changes of specific pulse characteristics such as stasis, blood stasis, water, dampness, phlegm, cold, and heat. This unit extracts key waveform parameters through a high-precision pulse waveform measurement device, including data such as the amplitude difference of the pulse waveform, the ratio of the main wave to the secondary wave, the peak-to-peak interval, and the pulse propagation delay, so as to accurately describe different types of pulse characteristics. For example, when judging the state of blood stasis obstruction, the system calculates the peak ratio of the pulse wave and hemodynamic changes to identify whether there is blood stasis in the blood flow; when detecting dampness pathogen, it combines the changes in the low-frequency components of the waveform to analyze the state of body fluid metabolism. This unit further combines non-linear signal processing methods to denoise, enhance features, and perform pattern recognition on pulse data to ensure the stability and reliability of data analysis. In addition, by constructing a pulse characteristic weight model, the system can automatically calculate the influence degree of different pulse characteristics on constitution analysis to improve the accuracy and individual adaptation ability of the analysis.

[0060] The dynamic constitution identification unit conducts time series analysis on the user's constitution state based on the theory of five evolutive phases and six climatic factors or the twenty-four solar terms model, and combines environmental temperature and humidity, seasonal changes, and climate trends to optimize the calculation parameters for constitution identification. Since the theory of five evolutive phases and six climatic factors emphasizes the influence of natural climate on the human constitution, and the twenty-four solar terms reflect the physiological regulation mechanism of the human body's adaptation to different climate changes, this unit calculates the potential influence of meteorological factors such as temperature, humidity, air pressure, and wind force on the constitution by analyzing the user's historical data and combining real-time climate information. For example, when the cold air prevails in winter, the system can identify that users with yang deficiency constitution are more likely to be affected by cold pathogens and appropriately adjust the analysis parameters to increase the influence weight of cold pathogens on the constitution; in the hot and humid season, for users with damp-heat constitution, the system can dynamically adjust the characteristic weight of dampness-related pulses to improve the identification accuracy. In addition, this unit combines the user's living habits, schedule adjustments, and dietary preferences to construct an individualized constitution evolution trend model to achieve dynamic identification.

[0061] The system adopts a multi-level calculation method during the analysis process. First, it conducts trend analysis on the individual's historical data, then combines external environmental parameters for dynamic correction, and in the final decision-making stage, uses a comprehensive constitution matching model to accurately evaluate the user's constitution state. In order to improve the temporal consistency of the data, this unit uses a sliding window algorithm to smooth short-term and long-term data and automatically correct outliers to ensure the stability of constitution evolution analysis. The system can predict the change trend of the user's constitution state before and after each solar term and provide corresponding health regulation suggestions. For example, it reminds users to appropriately increase warm and tonic foods before the Beginning of Winter, and recommends a light diet before the Summer Solstice to avoid the accumulation of dampness and heat.

[0062] Through this design, the system not only improves the refinement of pulse condition analysis to make it more in line with traditional Chinese medicine diagnosis standards, but also introduces the models of the five evolutive phases and six climatic factors and solar terms, making the constitution identification more dynamic and personalized. This method enables the constitution analysis system to more accurately adapt to the changes in the human body state under different environments, thus providing users with a more scientific and personalized constitution assessment and regulation plan.

[0063] The deep learning analysis module 103, connected to the feature extraction module, is used to input the multi-dimensional data set into a pre-trained deep learning model to obtain probability data representing the conversion relationship between different traditional Chinese medicine constitution types.

[0064] The deep learning analysis module 103 is used to model and analyze the multi-dimensional data set generated by the feature extraction module 102 to identify the conversion relationship between different traditional Chinese medicine constitution types and provide quantitative probability data. This module mainly includes steps such as data preprocessing, model calculation, constitution conversion probability calculation, and output optimization to ensure that the input data can be accurately parsed and finally generate a prediction result that conforms to the evolution law of traditional Chinese medicine constitution.

[0065] During the data preprocessing process, this module first receives the multi-dimensional data set output by the feature extraction module 102, which includes tongue image features, pulse condition time-frequency features, body surface infrared temperature distribution, environmental temperature and humidity information, user behavior logs, and medication records. To improve the stability of model calculation, the system performs standardization processing on the input data. For example, continuous variables are normalized so that their value ranges are kept within a fixed range to prevent the influence of too large or too small numerical values on model convergence. At the same time, for missing data, the system can use interpolation algorithms or regression models based on historical data for completion to avoid affecting the analysis results due to incomplete data. In addition, since the input data includes time series information, the system uses the sliding window technique to construct time series features to ensure that the model can capture the trend of constitution change over time.

[0066] In the model calculation stage, the deep learning analysis module 103 inputs the preprocessed data into a pre-trained deep learning model, which can be based on a Recurrent Neural Network (RNN), Long Short-Term Memory Network (LSTM), or Transformer structure to fully explore the dynamic evolution relationships in time-series data. Through feature mapping of multiple hidden layers, the model converts the input high-dimensional data into low-dimensional physical state feature vectors and calculates the state transition relationships between different physical types using the trained weight parameters. During the calculation process, the system performs supervised learning on the model by combining the labeled training data to ensure that the probability data output by the model can reflect the actual possibility of physical transformation. For example, given tongue image, pulse condition, and body surface temperature features, the model can predict the probability of a user's physical type transitioning from qi deficiency to yin deficiency and determine the long-term evolution direction of the physical type based on the long-term data trend.

[0067] During the calculation of the physical transformation probability, the system calculates the transition probability matrix from the current physical state to other physical states based on the feature vectors output by the deep learning model. The construction of this matrix uses the Markov chain method, that is, the conditional probability of the future state is calculated using the joint distribution of the historical state and the current state to depict the dynamic evolution relationships between different traditional Chinese medicine physical types. The size of the transition probability matrix depends on the predefined number of physical classifications. For example, if the system differentiates nine traditional Chinese medicine physical types, the matrix dimension is 9×9, and each element represents the transition probability from one physical state to another. By continuously optimizing the model parameters, this matrix can be dynamically adjusted as the user data is updated to improve the prediction accuracy.

[0068] In the output optimization stage, the deep learning analysis module 103 combines historical data to adjust the transition probability to reduce the misjudgment rate and improve the generalization ability of the model. The system can adopt the Bayesian update method to incorporate the newly collected data into the calculation process, enabling the transition probability to be continuously updated as the individual characteristics of the user change. In addition, to improve the adaptability of the model to different populations, the system can adjust the model weights through an adaptive learning mechanism according to individual characteristics. For example, if a user's pulse condition feature is stable over a long period with a small change range, the system will reduce the weight of its pulse signal in the physical transformation prediction and assign higher weights to the tongue image and behavior data.

[0069] After the above processing, the deep learning analysis module 103 finally outputs the probability data of physical transformation, which can be used as the input of the physical analysis and prediction module 104 to further determine the user's current physical state and predict future physical changes in combination with the historical trend. Through the above steps, this module realizes the dynamic identification of physical types based on deep learning, provides a quantifiable and traceable method for physical evolution analysis, and provides reliable data support for subsequent personalized health conditioning.

[0070] Furthermore, the deep learning model includes a feature encoding sub-model, a time series modeling sub-model, and a physique conversion relationship prediction sub-model;

[0071] Among them, the feature encoding sub-model is used to receive the tongue image feature data, pulse condition time-frequency data, body surface temperature distribution data, environmental temperature and humidity data, user behavior log data, and medication information data output by the feature extraction module, and perform feature transformation on different types of data through a non-linear mapping function, converting them into a unified high-dimensional feature vector. At the same time, the correlation weight between features is calculated to remove redundant information and retain the key features affecting physique conversion;

[0072] The time series modeling sub-model is used to receive the output data of the feature encoding sub-model, construct time series samples based on a sliding time window, use a multi-scale dynamic change analysis method to quantify the change trend of physique features in different time periods, combine trend smoothing processing and outlier detection methods to optimize the temporal continuity of physique data, and calculate the short-term fluctuations and long-term evolution laws of physique features;

[0073] The physique conversion relationship prediction sub-model is used to receive the output data of the time series modeling sub-model, construct the conversion relationship between different traditional Chinese medicine physique types based on a multi-state probability distribution matrix, predict the conversion possibility between different physique states by calculating the state transition probability distribution, and combine individual adjustment factors to dynamically correct the prediction results to adapt to the individual physique evolution patterns of different users, and finally output the probability data representing the conversion relationship between different traditional Chinese medicine physique types.

[0074] The deep learning model works collaboratively through three sub-models: feature encoding, time series modeling, and physique conversion relationship prediction to accurately analyze and predict the dynamic changes in a user's physique. First, the feature encoding sub-model is used to process and integrate different types of input data, including tongue image features, pulse condition time-frequency features, body surface temperature distribution, environmental temperature and humidity, user behavior logs, and medication information. These data sources are extensive, and the data formats and distributions are diverse. Therefore, this sub-model uses a non-linear mapping function to perform feature transformation on all data, projecting it into a unified high-dimensional feature vector space. During the transformation process, this sub-model uses a method for calculating the correlation weight between features to remove redundant information and ensure that the key features most representative of physique conversion are retained. For example, if there is a high correlation between the change in the color of a user's tongue image and the body surface temperature distribution, the system will assign higher weights to these features, enabling them to play a greater role in subsequent physique modeling and prediction processes.

[0075] After the feature encoding is completed, the time series modeling sub-model receives the high-dimensional feature vectors and organizes them into time series samples to analyze the changing trends of physical characteristics in the time dimension. Since the evolution of physical constitution is usually not a linear change but a complex process affected by multiple factors, this sub-model uses a sliding time window to construct a dynamic time series, enabling the system to identify short-term and long-term change patterns. For example, in a short period, a user's pulse condition may show temporary deviations due to emotional fluctuations or environmental changes, while the long-term trend may show a gradual transformation of the physical constitution from qi deficiency to yang deficiency. To optimize the continuity of time series data, this sub-model combines a trend smoothing processing algorithm to denoise and detect anomalies in the physical constitution data, so as to eliminate the interference caused by sudden fluctuations. For example, if a user's pulse waveform is abnormal at a certain time, but this trend does not continue in subsequent time windows, the system can identify this data point as noise and correct it through interpolation or historical trend matching to ensure the smoothness of the physical constitution evolution trend.

[0076] After the time series modeling is completed, the physical constitution conversion relationship prediction sub-model is responsible for modeling and inferring the future evolution of the user's physical constitution. This sub-model constructs a multi-state probability distribution matrix to describe the conversion relationships between different traditional Chinese medicine physical constitution types, and calculates the transition probabilities of physical constitution states based on the user's historical physical constitution data. For example, for a certain user, within the past six months, the probability of the conversion of their physical constitution from yin deficiency to yang deficiency is relatively high, while the probability of conversion to qi deficiency is relatively low. Then the system will, based on this trend, predict the likelihood of the user's physical constitution continuing to develop towards yang deficiency in the next few months and adjust the prediction weights. To further optimize the prediction accuracy, this sub-model also introduces an individualized adjustment factor to dynamically correct the state transition probabilities to adapt to the individual physical constitution evolution patterns of different users. For example, for users with a relatively rapid increase in age, the system may increase the weight of the conversion probabilities of physical constitution types related to aging, while for users with significant changes in work and rest patterns, the system will consider the impact of lifestyle adjustments on physical constitution conversion. Finally, this sub-model outputs the conversion relationship probability data between different traditional Chinese medicine physical constitution types, providing a quantitative basis for subsequent physical constitution analysis and prediction.

[0077] Through this hierarchical modeling method, the entire deep learning model can effectively integrate multi-source data, combine the changing trends of time series, and accurately predict the future development direction of an individual's physical constitution based on probability calculations, enabling the physical constitution identification system to provide more targeted health management suggestions and adapt to the long-term physical constitution evolution patterns of different individuals.

[0078] Furthermore, the time series modeling sub-model constructs an individualized physical constitution evolution curve based on the historical data of individual physical constitution characteristics, and combines a multi-scale dynamic change analysis method to calculate the short-term fluctuations and long-term trends of physical constitution characteristics;

[0079] The time series modeling sub-model includes a trend smoothing processing unit, an individualized time weighting unit, and an outlier correction unit. Among them, the individualized time weighting unit is used to dynamically assign data weights for different time windows to reflect the different degrees of influence of an individual's physical fitness over time. The data weights are calculated according to the following formula 1:

[0080]

[0081] where w t represents the influence weight of time t on the current physical fitness state; t0 is the current time point; x t is the physical fitness characteristic value at time t; σ x is the standard deviation of the physical fitness characteristic data; ∈ is a regularization factor to avoid the denominator approaching zero; T is the length of the time series; α is the time decay coefficient, which determines the decay rate of the influence of historical data on the current physical fitness state; β is the physical fitness characteristic change sensitivity parameter, which determines the influence degree of the change amplitude of an individual's physical fitness characteristics in time weighting;

[0082] The trend smoothing processing unit receives the weights calculated by the individualized time weighting unit and calculates the weighted physical fitness characteristic value according to the following formula 2:

[0083]

[0084] where is the weighted physical fitness characteristic value at the current time point t; x i is the physical fitness characteristic value at the historical time point i; w i is the influence weight of time point i on the current physical fitness state, calculated according to formula 1; T is the length of the time series;

[0085] The outlier correction unit performs outlier detection on the weighted physical fitness characteristic value based on the local deviation detection method. By calculating the mean μ Δx and the standard deviation σ Δx of the historical physical fitness change rate, an outlier determination threshold θ is set. When the following formula 3 is satisfied, the data point is identified as an outlier and corrected using an interpolation method based on the historical trend:

[0086] |x t ―x t―1 |>μ Δx +θ·σ Δx (3)

[0087] where x t is the physical fitness characteristic value at time t; x t―1is the physical characteristic value at time t-1; θ is the abnormality detection sensitivity parameter, which controls the strictness of the judgment of abnormal points; if an abnormal point is detected, the nearest K normal data points are used for interpolation correction to ensure the stability and rationality of the physical change trend.

[0088] The time series modeling sub-model constructs an individualized physical evolution curve, performs multi-scale analysis on historical physical characteristic data to calculate short-term fluctuations and long-term trends, and combines dynamic time weighting and outlier correction methods to improve the stability and reliability of the data. Since individual physical fitness is affected by many factors, the changes in physical characteristic data at different time points may have trend changes, periodic fluctuations, and occasional anomalies. Therefore, this sub-model uses a series of mathematical modeling methods to reasonably process the data to ensure that the final physical condition assessment and prediction are more accurate.

[0089] In the time weighting process, the system adopts a dynamic time weight allocation strategy to reflect the influence of data in different time windows on the current physical condition. The weight calculation follows the above formula 1.

[0090] In formula 1, w t Indicates the weight of the influence of time t on the current physical condition. The larger the value, the stronger the influence of the data at that time point on the current physical condition. t0 is the current time point, that is, the time point when the system needs to calculate the physical condition. The influence of all historical data points is calculated relative to this time point.

[0091] x t is the physical characteristic value collected at time t, such as pulse, tongue, body temperature and other measurement data. x It is the standard deviation of physical characteristic data, which is used to measure the overall volatility of physical characteristic data in time series and can usually be calculated based on historical data.

[0092] ∈ is the regularization factor, which prevents the denominator from approaching zero during the calculation process, and is usually set to 10 ―6 To ensure computational stability.

[0093] The parameter α is the time decay coefficient, which controls the time decay rate of the data. This parameter determines the contribution of earlier data to the current state. The recommended value range is 0.01≤α≤0.10. When α is small, the impact of long-term data is greater, and when α is large, the system pays more attention to recent data.

[0094] β is the sensitivity parameter of physical characteristics change, which is used to adjust the impact of short-term changes on weights. This parameter determines the impact of short-term fluctuations in data on the current physical condition. The recommended value range is 0.1≤β≤1. A larger β is suitable for individuals with more drastic short-term fluctuations in physical characteristics, while a smaller β is suitable for individuals with long-term stable physical characteristics.

[0095] After the weight calculation is completed, the system further uses the weighted moving average method to calculate the physical characteristic value at the current time point, and the calculation formula is shown in Formula 2.

[0096] In Formula 2, is the weighted physical characteristic value at the current time point t, that is, the final physical characteristic value obtained after considering the influence of different historical data points on the current state. x i is the physical characteristic data collected at the historical time point i, and w i is the individualized time weight calculated at the time point i, which is calculated by Formula (1). T is the length of the time series, that is, the historical time window used by the system to calculate the current physical state. Generally, an appropriate length can be selected according to the characteristics of physical evolution. For example, for short-term physical fluctuation analysis, T can be set to 7 days, while for long-term trend analysis, T can be set to 30 days or longer.

[0097] After calculating the weighted physical characteristic value, the system further performs outlier detection on the data to eliminate abnormal data points that may be caused by measurement errors, environmental impacts, or short-term abnormal fluctuations. The outlier detection uses the local deviation detection method, calculates the mean and standard deviation of the change rate of historical data, and judges whether the data is abnormal based on the set outlier determination threshold. The calculation formula is shown in Formula 3.

[0098] In Formula 3, x t is the physical characteristic value at the current time point, and x t―1 is the physical characteristic value at the previous time point t−1. The difference between the two represents the physical change amplitude between the current time point and the previous time point. μ Δx is the mean of the historical physical characteristic change rate, which can be obtained by calculating the change rates of multiple time points. σ Δx is the standard deviation of the historical physical characteristic change rate, which is used to measure the stability of physical changes.

[0099] θ is the outlier detection sensitivity parameter, and the recommended value range is 2≤θ≤3. A larger θ indicates a more lenient judgment of outliers, while a smaller θ makes the system more sensitive to abnormal fluctuations.

[0100] When an outlier is detected, the system does not directly discard the data point, but uses an interpolation method based on the historical trend to correct it. Specifically, the system will select the nearest K normal data points and use linear interpolation or curve fitting methods to calculate the corrected value, making the physical change trend smoother and reducing misjudgments or sudden fluctuations caused by abnormal data.

[0101] Through the above method, the time series modeling sub-model can effectively integrate historical data, dynamically adjust time weights, smooth the physical change trend, and automatically detect and correct outliers, thus ensuring the accuracy and reliability of physical evolution analysis. This method not only improves the accuracy of physical analysis, but also enhances the adaptability of the system at different individuals and different time scales, enabling traditional Chinese medicine physical identification to more accurately reflect the individual's health status and change trend.

[0102] The following is the reference implementation code of the deep learning model:

[0103] import torch

[0104] import torch.nn as nn

[0105] import torch.optim as optim

[0106] import numpy as np

[0107] class FeatureEncodingModule(nn.Module):

[0108] """

[0109] Feature Encoding Sub-Model:

[0110] This module receives various physical-related data (tongue image features, pulse features, temperature, environment, etc.),

[0111] converts them into a unified high-dimensional feature vector through a non-linear mapping function, and calculates the correlation between features.

[0112]

[0113]

[0114] Time Series Modeling Sub-Model:

[0115] This module receives the data after feature encoding, constructs a time series based on a sliding time window, and adopts a multi-scale dynamic change analysis method to extract short-term fluctuations and long-term trends.

[0116]

[0117] Physical Transformation Relationship Prediction Sub-Model:

[0118] This module receives the results of time series modeling and predicts the physical transformation relationship based on a multi-state probability distribution matrix.

[0119]

[0120]

[0121] Overall model:

[0122] It includes three sub - models: feature encoding, time - series modeling, and constitution conversion prediction.

[0123]

[0124] To train this deep - learning model, first, a dataset containing multi - dimensional features such as tongue image features, pulse - condition time - frequency data, and body - surface temperature distribution needs to be prepared and normalized. Then, the data is divided into a training set and a validation set to ensure the generalization ability of the model. The cross - entropy loss function is used to measure the accuracy of constitution category prediction, and the Adam optimizer is used for gradient update. During training, data is input in batches. First, the high - dimensional representation is extracted through the feature - encoding sub - model, then it is sent to the time - series modeling sub - model for dynamic trend analysis, and finally, the constitution - conversion relationship prediction sub - model calculates the conversion probability of constitution categories. In each training epoch, the loss is calculated and backpropagated, and the weights are adjusted to optimize the prediction performance. At the same time, the accuracy and stability of the model are evaluated on the validation set to ensure that the training effect meets the expectations.

[0125] The constitution analysis and prediction module 104, connected to the deep - learning analysis module, is used to determine the user's current traditional Chinese medicine (TCM) constitution type based on the probability data, and combine historical acquisition data to predict the user's future TCM constitution change trend, generating a diagnosis report including constitution evolution analysis and conditioning suggestions.

[0126] The constitution analysis and prediction module 104 is used to receive the constitution conversion probability data generated by the deep - learning analysis module 103, determine the user's current TCM constitution type based on these data, and at the same time predict the user's future constitution change trend by combining historical data, so as to generate a diagnosis report including constitution evolution analysis and conditioning suggestions. This module needs to combine multiple data sources and adopt a multi - step calculation process to ensure the accuracy and interpretability of constitution determination and trend prediction.

[0127] After receiving the constitution conversion probability data output by the deep - learning analysis module 103, the system first makes a determination of the current constitution type. The constitution determination is based on the current state distribution of the conversion probability matrix. By calculating the maximum possibility of different constitution categories, the user's current main constitution type is determined. If the conversion probability of a certain constitution category is the largest and meets the threshold requirement, the system directly determines that category as the current constitution. If the conversion probabilities are relatively dispersed, the system will combine the user's most recent constitution determination result and use a weighted - average method to calculate the current constitution state to ensure the stability of constitution judgment.

[0128] To improve the accuracy of constitution determination, this module makes corrections by combining historical data. The historical data includes tongue images, pulse conditions, body surface temperature, environmental factors, and behavior data over a period of time. Through time series analysis methods, it evaluates the long-term change trend of the user's constitution, and uses the Bayesian update method to adjust the current constitution determination result. For example, if the constitution assessment results of a certain user have shown the qi deficiency constitution in the past few months, and the current data leans towards the yang deficiency constitution but the change range is small, the system may maintain the original constitution classification or provide transitional classification suggestions to avoid misjudgment caused by short-term data fluctuations.

[0129] After completing the determination of the current constitution type, this module further predicts the constitution change trend. The prediction process is based on the transition probability matrix provided by the deep learning analysis module 103, and combines the user's historical data to calculate the constitution evolution probability at different future time nodes through a Markov chain or a time series regression model. The prediction results can cover the constitution evolution trends in the short term (within a few weeks), medium term (within a few months), and long term (more than one year). If there are significant changes in the user's behavior data or medication information, the system will automatically adjust the weights of the prediction model to make it more in line with the user's individual situation. For example, if the system detects that the user has recently increased constitution conditioning measures, such as exercise, diet adjustment, or traditional Chinese medicine intervention, the system will appropriately increase the influence weight of the positive conditioning effect during trend prediction to reflect the possibility of constitution improvement.

[0130] After obtaining the constitution trend prediction results, this module generates a diagnostic report on constitution evolution analysis and conditioning suggestions. The content of the report includes a detailed description of the user's current constitution type, the constitution changes over a period of time in the past, the analysis of the possibility of future constitution changes, and corresponding conditioning suggestions. The conditioning suggestions can generate a personalized health management plan based on the traditional Chinese medicine constitution classification standard through rule matching or expert system methods. The plan may cover aspects such as diet, exercise, daily routine, emotional regulation, and traditional Chinese medicine conditioning. For the constitution change trends that need to be focused on, the system can provide targeted warning information. For example, if a user's constitution has been developing towards the yin deficiency direction for a long time, it may be necessary to strengthen yin-nourishing measures, or if a user's qi deficiency constitution shows a tendency to transform into a excess syndrome, it may be necessary to adjust the tonifying strategy.

[0131] The diagnostic report can be presented in a visual way. For example, the constitution change trend can be shown through a line chart, the comprehensive score of the current constitution characteristics can be shown through a radar chart, or personalized health guidance can be provided through text analysis. Users can access the report through mobile devices or online health platforms and make corresponding adjustments according to the suggestions. In addition, the system can combine the user's feedback data to dynamically adjust the parameters of the constitution prediction model to achieve personalized learning, thereby improving the accuracy and reliability of future predictions.

[0132] The overall design of the constitution analysis and prediction module 104 ensures the accuracy of constitution determination, the rationality of prediction of the trend of constitution change, and the personalization of conditioning suggestions, enabling the system to effectively assist users in long-term constitution management and enhancing the scientific nature and pertinence of health intervention.

[0133] Furthermore, the constitution analysis and prediction module includes an individual constitution trend modeling unit. The individual constitution trend modeling unit constructs an individualized constitution evolution curve based on the user's long-term constitution data, and predicts the possible future constitution change trend of the user by calculating the conversion probability between constitution types. Among them, the individual constitution trend modeling unit combines environmental changes, lifestyle adjustments, and drug intervention information to dynamically adjust the constitution trend prediction parameters, so as to improve the adaptability to individual constitution changes and ensure the accuracy and reliability of the prediction results.

[0134] The individual constitution trend modeling unit is used to construct a personalized constitution evolution curve based on the user's long-term constitution data, and predict the possible future constitution change trend by calculating the conversion probability between constitution types. Since an individual's constitution is affected by multiple factors, including age growth, seasonal changes, lifestyle adjustments, and drug interventions, relying solely on data at a single time point cannot accurately reflect the development trend of the constitution. Therefore, this unit uses time series modeling methods and combines long-term constitution data to generate a constitution change trajectory that conforms to individual characteristics, so as to improve the accuracy and reliability of the prediction.

[0135] During the constitution trend modeling process, this unit first extracts key constitution indicators from the user's historical data, including pulse condition, tongue condition, body surface temperature distribution, and other physiological characteristics, and constructs a constitution evolution curve in the time dimension. The constitution evolution curve is based on time series modeling methods, and by calculating the change trend of the constitution state at different time points, it identifies the long-term development direction of the constitution. For example, if a user's constitution has gradually changed from the qi-deficiency type to the yang-deficiency type in the past six months, this unit can analyze the driving force of this change based on historical data and predict whether this trend will continue or reverse in the next few months.

[0136] To further enhance the adaptability of prediction, this unit not only relies on physiological data but also comprehensively considers the impacts of environmental factors, lifestyle habits, and drug interventions on physical constitution changes. In terms of environmental factor analysis, this unit combines the climate data of the user's location to calculate the impact of seasonal factors on physical constitution changes. For example, in the cold season, users with yang deficiency constitution may be more likely to experience a drop in body temperature, while a damp-heat environment may exacerbate the symptoms of phlegm-dampness constitution. In terms of lifestyle habit adjustment, this unit analyzes factors such as the user's diet, exercise, and daily routine, evaluates the contribution of these factors to the physical constitution trend, and dynamically adjusts the prediction parameters. For example, users who persist in exercise for a long time may gradually improve their qi deficiency state, while irregular daily routines may lead to aggravated yin deficiency. In terms of drug intervention, this unit uses the user's medication records to calculate the intervention effects of different drugs and optimizes the prediction of physical constitution changes in combination with individual differences. For example, using warming and tonifying drugs may help improve qi deficiency constitution, but for users with damp-heat constitution, it may exacerbate certain symptoms. Therefore, a differential parameter adjustment strategy needs to be adopted for different users during the prediction process.

[0137] When calculating the transition probabilities between physical constitution types, this unit adopts the state transition matrix method to model the transition probabilities between different physical constitution categories and dynamically adjusts the transition weights in combination with the individual's long-term data. For example, for a certain user, their past data shows that the transition probability from yang deficiency to qi deficiency is relatively high, while the transition probability from phlegm-dampness to damp-heat is relatively low. Then the system will assign a higher weight to the former during prediction to ensure that the prediction conforms to the actual physical constitution evolution law of the user. In addition, this unit can update the transition probabilities in real time. As the user's physical constitution data continues to accumulate, the system can continuously optimize the physical constitution trend prediction model to make it more in line with the individual's actual situation.

[0138] Through the above methods, this unit can comprehensively consider various influencing factors to accurately predict the future change trend of an individual's physical constitution and provide more targeted physical constitution conditioning suggestions for users. At the same time, since the system can dynamically adjust the prediction parameters and correct them according to the user's latest data, it can ensure the accuracy and reliability of the physical constitution analysis results and make them applicable to the application scenarios of long-term health management and traditional Chinese medicine physical constitution identification.

[0139] Furthermore, the physical constitution analysis and prediction module includes an intelligent conditioning plan generation unit. The intelligent conditioning plan generation unit matches appropriate conditioning measures from the physical constitution feature database based on the user's current physical constitution state and the predicted physical constitution evolution trend, and generates personalized health management suggestions in combination with the traditional Chinese medicine knowledge base. Among them, the intelligent conditioning plan generation unit is used to dynamically adjust the conditioning plan based on the user's feedback information and provide early warning prompts when the user's physical constitution deviates from the normal trend to optimize the conditioning strategy and improve the pertinence of health intervention.

[0140] The intelligent conditioning plan generation unit is used to match appropriate conditioning measures from the physical characteristic database based on the user's current physical condition and the predicted physical evolution trend, and generate personalized health management suggestions in combination with the traditional Chinese medicine knowledge base. Since an individual's physical condition can change due to various factors such as time, environmental factors, living habits, and drug interventions, a single fixed conditioning plan often fails to meet the individual's needs in the long term. Therefore, this unit can automatically match and adjust the most suitable conditioning measures according to the user's dynamic physical information to improve the accuracy and effectiveness of health interventions.

[0141] During the process of matching the conditioning plan, this unit first obtains the user's current physical data, including tongue image characteristics, pulse characteristics, body surface temperature distribution, and other relevant physiological parameters. At the same time, in combination with the individual's historical physical change trend, it calculates the possible future physical evolution direction of the user. For example, if a user's current physical condition is biased towards qi deficiency type, and the physical trend prediction shows that the qi deficiency state may further worsen, the system will tend to recommend a conditioning plan for tonifying qi and provide corresponding suggestions in terms of diet, rest, exercise, etc. This matching process relies on the physical characteristic database, which stores standardized conditioning measures corresponding to different physical types, covering various methods such as diet therapy, exercise, rest adjustment, emotional regulation, and traditional Chinese medicine intervention.

[0142] When generating personalized health management suggestions, this unit further combines the traditional Chinese medicine knowledge base to ensure that the recommended conditioning measures conform to the theory of traditional Chinese medicine and can be optimized according to the user's individual characteristics. For example, for a user with qi deficiency constitution and weak spleen and stomach, the system may recommend a mild qi-tonifying diet, such as Chinese yam and Chinese dates. For a qi deficiency user with dampness pathogen at the same time, it may recommend conditioning methods for strengthening the spleen and removing dampness, such as coix seed and tangerine peel. In addition, in terms of exercise intervention, this unit will recommend a suitable exercise intensity according to the user's physical characteristics. For example, for users with yin deficiency constitution, it may recommend mild exercises such as Taijiquan and Baduanjin, while for users with yang deficiency constitution, it may recommend moderate aerobic exercises to strengthen yang qi.

[0143] To ensure the effectiveness and pertinence of the conditioning plan, this unit supports dynamic adjustment based on the information feedback by the user. When the user executes the conditioning plan for a period of time, the system will automatically collect feedback data, including changes in the user's physical condition, the implementation of the conditioning plan, and subjective feelings, etc. If it is detected that the user's physical condition fails to improve as expected or there are discomfort situations, the system will automatically adjust the plan. For example, if a user shows symptoms of getting angry after executing the recommended warming and tonifying conditioning plan, the system will re-evaluate the physical condition and appropriately reduce the degree of warming and tonifying, and increase heat-clearing or balancing conditioning measures to ensure that the plan better meets the individual needs.

[0144] When the user's physical condition deviates from the normal trend, this unit can also provide warning prompts to help the user adjust the health management strategy in a timely manner. For example, if the system detects that the changing trend of the user's physical condition deviates from the established goal, such as a user with qi and blood deficiency failing to improve qi deficiency after conditioning but showing a tendency of yin deficiency instead, the system will actively remind the user to adjust the conditioning method and provide corresponding warning suggestions. In addition, this unit can adaptively optimize the user's conditioning plan in combination with changes in environmental factors, such as seasonal transitions and abnormal climates, to ensure that the user can obtain a more accurate health management plan that conforms to the current physical condition.

[0145] Through the above method, this unit can provide intelligent support in the aspects of personalized conditioning plan generation, user feedback optimization, and dynamic adjustment strategy, ensuring that the conditioning plan can be continuously optimized as the user's physical condition changes, so as to achieve more accurate health management and improve the application value and reliability of the traditional Chinese medicine physical constitution identification system.

[0146] Although this application is disclosed above with preferred embodiments, it is not intended to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the protection scope of this application should be subject to the scope defined by the claims of this application.

Claims

1. A system for identifying dynamic changes in TCM constitution, characterized in that: include: The data collection module is used to collect the user's tongue video stream data, pulse waveform data and body surface infrared thermal imaging data, and record the environmental temperature and humidity data, user behavior log data and medication information data; A feature extraction module, connected to the data acquisition module, is used to perform image segmentation processing on the tongue image video stream data, extract the tongue body and tongue coating area, and obtain tongue image feature data including tongue color, tongue shape and tongue coating features; perform wavelet transform on the pulse waveform data, and extract pulse time-frequency data including waveform amplitude, frequency and phase features; process the body surface infrared thermal imaging data to obtain body surface temperature distribution data, and construct a multidimensional data set together with the tongue image feature data, pulse time-frequency data and body surface temperature distribution data based on the environmental temperature and humidity data, user behavior log data and medication information data; A deep learning analysis module, connected to the feature extraction module, for inputting the multidimensional data set into a pre-trained deep learning model to obtain probability data characterizing the conversion relationship between different TCM constitution types; The constitution analysis and prediction module is connected to the deep learning analysis module, and is used to determine the user's current TCM constitution type based on the probability data, and predict the user's future TCM constitution change trend in combination with the historical collected data, and generate a diagnostic report including constitution evolution analysis and conditioning suggestions.

2. The TCM constitution dynamic change identification system according to claim 1 is characterized in that: The data acquisition module includes a temperature-controlled light source unit, which is used to dynamically adjust the color temperature and brightness of the light source according to the ambient temperature when collecting tongue image video stream data, so as to adapt to the tongue image reflection spectrum characteristics of different individuals; wherein, the temperature-controlled light source unit includes a plurality of LED light sources with different color temperatures and a temperature sensor. After the temperature sensor detects the ambient temperature, it selects a light source combination with an appropriate color temperature according to a preset illumination compensation model, and adjusts the output power of the light source through pulse width modulation to reduce the influence of ambient light on the extraction of tongue image color features, thereby improving the stability and consistency of tongue image data.

3. The TCM constitution dynamic change identification system according to claim 1 is characterized in that: The data acquisition module includes an adaptive signal adjustment unit, which is used to dynamically adjust the sampling frequency, signal gain and filtering parameters according to the pulse characteristics of the individual user when collecting pulse waveform data, wherein the adaptive signal adjustment unit includes a pulse waveform analysis module and a signal control module, the pulse waveform analysis module calculates the pulse intensity, signal-to-noise ratio and pulse waveform morphology based on the preliminary collected pulse data, and inputs the calculation results into the signal control module, and the signal control module dynamically adjusts the sampling rate, gain parameter and filtering threshold of the sensor according to a preset adaptive adjustment algorithm to optimize the signal quality and ensure that the pulse data collection under different individuals and different physiological conditions is accurate and reliable.

4. The TCM constitution dynamic change identification system according to claim 1 is characterized in that: The data acquisition module includes a dynamic temperature baseline adjustment unit, which is used to build a personalized baseline curve based on the user's historical temperature data when collecting body surface infrared thermal imaging data, and dynamically adjust the temperature data in subsequent measurement processes to adapt to the individual body temperature change pattern. The dynamic temperature baseline adjustment unit includes a user temperature database, a trend analysis module and a temperature compensation module. The user temperature database stores the user's body surface temperature data at multiple time nodes. The trend analysis module performs time series modeling on the historical data, calculates the temperature change law of the individual user, and generates a personalized baseline curve. The temperature compensation module adjusts the current temperature data according to the baseline curve during each measurement to reduce the impact of individual differences on physical identification and improve the stability and accuracy of long-term body temperature monitoring.

5. The TCM constitution dynamic change identification system according to claim 1 is characterized in that: The feature extraction module includes a tongue feature enhancement unit, which is used to optimize the color and texture features of the tongue body and tongue coating area by using a local histogram equalization method based on adaptive contrast enhancement after image segmentation processing of the tongue image video stream data. At the same time, combined with a tongue abnormality detection network based on deep learning, the features of tongue fissures, tooth marks, and petechiae are automatically identified, and high-resolution tongue feature images are generated to improve the accuracy and consistency of tongue image analysis.

6. The TCM constitution dynamic change identification system according to claim 1, characterized in that: The feature extraction module includes a pulse dynamic feature modeling unit, which is used to use a multi-scale recursive neural network to perform time series modeling on the pulse waveform after wavelet transforming the pulse waveform data, so as to extract short-term fluctuation characteristics, long-term trend characteristics and pulse periodic change patterns, and combine Fourier transform to analyze the frequency spectrum components of the pulse signal to establish a personalized pulse characteristic curve, so as to achieve accurate pulse classification and constitution association analysis.

7. The TCM constitution dynamic change identification system according to claim 1 is characterized in that: The physical fitness analysis and prediction module includes an individual physical fitness trend modeling unit, which constructs an individualized physical fitness evolution curve based on the user's long-term physical fitness data, and predicts the user's possible physical fitness change trend in the future by calculating the conversion probability between physical fitness types. The individual physical fitness trend modeling unit dynamically adjusts the physical fitness trend prediction parameters in combination with environmental changes, lifestyle adjustments and drug intervention information to improve adaptability to individual physical fitness changes and ensure the accuracy and reliability of the prediction results.

8. The TCM constitution dynamic change identification system according to claim 1 is characterized in that: The physical constitution analysis and prediction module includes an intelligent conditioning program generation unit, which matches appropriate conditioning measures from the physical constitution feature database based on the user's current physical condition and predicted physical constitution evolution trend, and generates personalized health management suggestions in combination with the traditional Chinese medicine knowledge base. The intelligent conditioning program generation unit is used to dynamically adjust the conditioning plan based on the user's feedback information, and provide early warning prompts when the user's physical constitution deviates from the normal trend, so as to optimize the conditioning strategy and improve the targeted health intervention.

9. The TCM constitution dynamic change identification system according to claim 1, characterized in that: The deep learning model includes a feature encoding sub-model, a time series modeling sub-model and a body constitution conversion relationship prediction sub-model; The feature encoding sub-model is used to receive the tongue feature data, pulse time-frequency data, body surface temperature distribution data, environmental temperature and humidity data, user behavior log data and medication information data output by the feature extraction module, and perform feature transformation on different types of data through a nonlinear mapping function, converting them into a unified high-dimensional feature vector, and calculating the correlation weights between the features to remove redundant information and retain the key features that affect the physical fitness conversion; The time series modeling submodel is used to receive the output data of the feature encoding submodel, and construct a time series sample based on a sliding time window, quantify the changing trend of physical characteristics in different time periods by using a multi-scale dynamic change analysis method, optimize the temporal continuity of physical data by combining trend smoothing and outlier detection methods, and calculate the short-term fluctuation and long-term evolution law of physical characteristics; The constitution conversion relationship prediction submodel is used to receive the output data of the time series modeling submodel, and construct the conversion relationship between different TCM constitution types based on the multi-state probability distribution matrix, predict the conversion possibility between different constitution states by calculating the state transition probability distribution, and dynamically correct the prediction results in combination with the individualized adjustment factor to adapt to the individual constitution evolution patterns of different users, and finally output the probability data representing the conversion relationship between different TCM constitution types.

10. The TCM constitution dynamic change identification system according to claim 1, characterized in that: The time series modeling sub-model constructs a personalized physical evolution curve based on the historical data of individual physical characteristics, and calculates the short-term fluctuations and long-term trends of physical characteristics in combination with a multi-scale dynamic change analysis method; The time series modeling sub-model includes a trend smoothing processing unit, an individualized time weighting unit and an outlier correction unit, wherein the individualized time weighting unit is used to dynamically allocate data weights of different time windows to reflect the different degrees of influence of individual physical fitness changes over time, and the data weights are calculated according to the following formula 1: Among them, w t represents the influence weight of time t on the current physical condition; t0 is the current time point; x t is the physical characteristic value at time t; σ x is the standard deviation of the physical characteristic data; ∈ is the regularization factor to prevent the denominator from approaching zero; T is the length of the time series; α is the time decay coefficient, which determines the decay rate of the influence of historical data on the current physical state; β is the sensitivity parameter of physical characteristic change, which determines the influence of the change amplitude of individual physical characteristics in time weighting; The trend smoothing processing unit receives the weight calculated by the individualized time weighting unit, and calculates the weighted physical characteristic value according to the following formula 2: in, is the weighted physical characteristic value at the current time point t; x i is the physical characteristic value at historical time point i; w i is the influence weight of time point i on the current physical condition, calculated according to formula 1; T is the length of the time series; The outlier correction unit performs outlier detection on the weighted physical characteristic values ​​based on the local deviation detection method, and calculates the mean value μ of the historical physical characteristic change rate. Δx and standard deviation σ Δx , set the abnormal judgment threshold θ, when the following formula 3 is satisfied, the data point is identified as an abnormal value, and the interpolation method based on historical trends is used for correction: |x t ―x t―1 |>m Δx +θ·s Δx (3) Among them, x t is the physical characteristic value at time t; x t―1 is the physical characteristic value at time t-1; θ is the abnormality detection sensitivity parameter, which controls the strictness of the judgment of abnormal points; if an abnormal point is detected, the nearest K normal data points are used for interpolation correction to ensure the stability and rationality of the physical change trend.

Images