Intelligent monitoring system for traditional Chinese medicine constitution identification

Abstract

This invention relates to the field of TCM constitution monitoring technology and discloses an intelligent monitoring system for TCM constitution identification. The system's constitution assessment model construction module acquires constitution characteristic inputs, constitution classification outputs, and constitution deviation interference quantities based on historical TCM diagnosis and treatment data, and constructs a constitution state equation. The dynamic feature extraction module performs sliding segmentation of the constitution deviation interference quantities by setting a time window, generating interference quantity time segments, and then extracts the periodic features and probability distribution features of each interference quantity. The constitution state prediction module predicts the interference quantity based on the current diagnosis and treatment time point, combined with periodic feature matching, and matches its occurrence probability according to the probability distribution features, finally outputting the predicted interference quantity and its corresponding occurrence probability. This system, by integrating dynamic interference factor analysis and probability prediction, solves the problems of strong subjectivity and insufficient dynamic monitoring in traditional constitution identification, realizing intelligent and dynamic monitoring of TCM constitution.

Description

Technical Field

[0001] This invention relates to the field of traditional Chinese medicine constitution monitoring technology, specifically a traditional Chinese medicine constitution identification and intelligent monitoring system. Background Technology

[0002] Traditional Chinese medicine (TCM) constitution identification is a crucial foundation for personalized health management within the TCM theoretical system. Its core lies in determining an individual's constitution type through a comprehensive analysis of physiological characteristics, lifestyle habits, and pathological manifestations, providing a basis for disease prevention and health management. Traditional TCM constitution identification primarily relies on the physician's experience in the four diagnostic methods of observation, auscultation, inquiry, and palpation, determining the constitution type through subjective judgment. This method is significantly influenced by the physician's individual experience, knowledge base, and subjective perception; different physicians may arrive at different identification results for the same subject, leading to insufficient consistency and objectivity in constitution assessment.

[0003] With the increasing demand for health management, traditional constitution identification methods are gradually showing their limitations in large-scale applications. Human constitution is not fixed but rather undergoes dynamic change, continuously influenced by various factors such as diet, sleep patterns, emotional fluctuations, and environmental changes. These factors can be considered as variables that disrupt constitution. Traditional methods are mostly static assessments, reflecting only the constitution at a specific point in time. They struggle to capture the dynamic trends of constitution over time and cannot identify potential disruptive factors that could lead to imbalances in constitution.

[0004] Some digital TCM constitution identification tools attempt to assist in constitution assessment through standardized questionnaires or basic physiological indicator tests. However, these tools often focus only on fixed constitution characteristic indicators and lack systematic analysis of interfering factors. They fail to effectively distinguish between temporary constitution fluctuations and persistent constitution shifts, and cannot identify the periodic patterns of interfering factors, resulting in significant deficiencies in predicting constitution changes. For example, some interfering factors, such as seasonal changes and work-rest cycles, have significant time periodicity, which existing tools struggle to capture, making it difficult for constitution assessment results to reflect the true dynamic changes in constitution. Furthermore, existing models often ignore the probability distribution characteristics of interfering quantities during construction, failing to quantify the probability of different interfering factors affecting constitution status, leading to a lack of specificity and foresight in health recommendations based on assessment results. These problems make traditional and existing technologies significantly inadequate in meeting the needs of precise and dynamic constitution monitoring, urgently requiring an intelligent monitoring system capable of integrating dynamic interfering factor analysis. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent monitoring system for identifying constitution in Traditional Chinese Medicine, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a Traditional Chinese Medicine constitution identification and intelligent monitoring system, the system comprising:

[0007] The constitution assessment model construction module is used to obtain the constitution characteristic input, constitution classification output, and constitution deviation interference based on historical TCM diagnosis and treatment data, and to construct the constitution state equation based on the constitution characteristic input, constitution classification output, and constitution deviation interference.

[0008] The dynamic feature extraction module is used to set a time window, perform sliding segmentation processing on the body mass shift interference amount based on the time window, generate interference amount time segments, and analyze the interference amount time segments to extract the periodic features and probability distribution features of each body mass shift interference amount.

[0009] The physical condition prediction module is used to obtain the current diagnosis and treatment time point, predict the interference amount based on the periodic features, match the occurrence probability of the predicted interference amount based on the probability distribution features, and output the predicted interference amount and the corresponding occurrence probability.

[0010] Preferably, the historical TCM diagnosis and treatment data includes tongue appearance characteristics, pulse appearance characteristics, symptoms during consultation, constitution classification labels, environmental temperature and humidity data, and emotional state scores;

[0011] The method for quantifying the interference of physical constitution deviation includes: calculating the difference between the pulse rate fluctuation value in the pulse characteristics and the reference pulse rate value as the first deviation, and setting the ratio of the first deviation to the reference pulse rate value as the pulse stability index.

[0012] The output of the constitution classification is represented as a probability distribution vector of various TCM constitution types.

[0013] Preferably, the input of physical characteristics includes tongue appearance characteristics, pulse appearance characteristics, and symptoms observed during medical consultation.

[0014] Preferably, the method for constructing the physical state equation includes:

[0015] The input quantity of physical characteristics, the output quantity of physical classification, and the quantity of physical offset interference are converted into frequency domain feature vectors through time-frequency conversion technology.

[0016] Ignoring the interference of body constitution offset, a basic body constitution response function is established based on the transformed body constitution feature input and body constitution classification output.

[0017] After introducing the physical deviation interference, the basic physical response function is extended into a physical state equation that includes the influence coefficient of the interference.

[0018] Preferably, the execution process of the physical state equation includes:

[0019] Receive the current physical characteristics input and the predicted interference;

[0020] The preset interference coefficients are invoked, and the predicted interference amount is weighted and corrected according to the interference coefficients; the interference coefficients are environmental temperature and humidity, emotional state, and pulse stability.

[0021] The corrected interference quantity and the physical characteristics input quantity are input together into the basic physical response function, and the target physical classification probability distribution vector is output.

[0022] Preferably, the dynamic feature extraction module performs the following steps:

[0023] Set a fixed-length time window and sliding step size;

[0024] Traverse the time series data of body offset interference according to the sliding step size, and extract continuous time segments as interference time segments.

[0025] Spectral analysis was performed on the time segments of each interference quantity, and the period corresponding to the dominant frequency component was extracted as a periodic feature.

[0026] The frequency of a specific interference value in each interference time segment is statistically analyzed, and its ratio to the total number of segments is calculated as a probability distribution characteristic.

[0027] Preferably, the physical condition prediction module performs the following steps:

[0028] Calculate the time interval between the current diagnosis / treatment time point and the end point of the previous interference time segment;

[0029] The time interval is matched with the periodic characteristics of the interference amount of each body mass deviation by difference;

[0030] The interference quantity corresponding to the periodic feature with the smallest absolute difference is selected as the predicted interference quantity;

[0031] The probability of occurrence of the predicted interference quantity is retrieved from the probability distribution feature library.

[0032] Preferably, the system further includes:

[0033] The constitution mechanism model module is used to establish a constitution state migration equation based on the theory of constitution transformation in traditional Chinese medicine and output a constitution mechanism assessment value.

[0034] The real-time feature analysis module is used to process real-time collected tongue, pulse, and environmental data through a deep temporal network to generate real-time physical condition assessment values.

[0035] The fusion decision module is used to dynamically allocate weight coefficients based on the error between the physical fitness mechanism assessment value and the real-time physical fitness assessment value, and generate a fusion physical fitness assessment result.

[0036] Preferably, the system further includes:

[0037] The intervention parameter optimization module is used to construct an action space that includes parameters of traditional Chinese medicine compatibility and acupuncture plan parameters by using the output of constitution classification as the decision state quantity.

[0038] The probability distribution of intervention parameters output by the strategy generation network is used to select the optimal combination of intervention parameters based on the dual-channel output of the value assessment network.

[0039] Preferably, the intervention parameter optimization module further performs:

[0040] Receive the predicted interference amount and occurrence probability output by the physical condition prediction module;

[0041] Monitor the actual physical condition classification output after implementing the optimal combination of intervention parameters;

[0042] Calculate the intervention bias between the actual body constitution classification output and the expected body constitution classification output;

[0043] When the intervention deviation exceeds the preset tolerance threshold, the dose weight coefficient of the Chinese medicine compatibility parameters and the stimulation intensity adjustment coefficient of the acupuncture scheme parameters are dynamically adjusted based on the probability of the predicted interference.

[0044] Compared with the prior art, the beneficial effects of the present invention are:

[0045] This intelligent monitoring system for TCM constitution identification achieves multi-dimensional improvements in the scientific rigor and practicality of TCM constitution identification through the collaborative operation of multiple modules. The constitution assessment model construction module breaks through the limitations of traditional constitution identification, which relies solely on fixed characteristic indicators. It incorporates constitution characteristic inputs, constitution classification outputs, and constitution deviation interferences into a unified analytical framework, constructing a constitution state equation. This process fully considers the complexity of constitution formation and change, moving beyond isolated views of constitution characteristic indicators to revealing the intrinsic relationships between various factors through equations. This shifts constitution assessment from empirical judgment to a systematic analysis based on data correlation, making the classification of constitution types more closely reflect the true state of individual physiological characteristics.

[0046] The dynamic feature extraction module uses a sliding segmentation method within a time window to process the disturbance amount of physical fitness deviation, generating time segments of the disturbance amount and analyzing its periodicity and probability distribution characteristics. This effectively solves the problem of insufficient capture of dynamic disturbance factors by traditional methods. The sliding segmentation method of the time window can adapt to the characteristics of physical fitness disturbance factors changing over time. By analyzing different time segments, periodic disturbance factors such as seasonal changes and work-rest patterns can be accurately identified, clarifying their impact on physical fitness. At the same time, the extraction of probability distribution characteristics can quantify the probability of different disturbance factors occurring, making the originally vague impact of disturbance factors perceptible and analyzable, providing richer reference data for subsequent prediction of physical fitness.

[0047] The physical condition prediction module outputs the predicted interference amount and probability of occurrence based on periodic and probability distribution characteristics, further enhancing the foresight and targeting of physical condition monitoring. By matching periodic characteristics, the system can predict potential physical condition interference factors that may occur at specific time stages, avoiding the neglect of potential risks in traditional static assessments. Matching the probability of occurrence distinguishes the degree of influence of different interference factors, helping users prioritize high-probability interference factors and take preventative measures in advance. This prediction method combines the regularity and uncertainty of physical condition changes, making the prediction results closer to the individual's actual situation. It provides dynamic guidance for personalized health management, allowing physical condition conditioning suggestions to be adjusted in advance based on predicted interference factors, better adapting to the dynamic changes in individual physical condition. Attached Figure Description

[0048] Figure 1 This is a schematic diagram illustrating the working principle of the intelligent monitoring system for TCM constitution identification described in this invention.

[0049] Figure 2 A flowchart for data definition and quantification in a Traditional Chinese Medicine constitution identification intelligent monitoring system;

[0050] Figure 3 A flowchart for executing the physical state equation;

[0051] Figure 4 A flowchart illustrating the operation of the dynamic feature extraction module. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Please see Figure 1 This invention provides a TCM constitution identification and intelligent monitoring system, which includes three core components: a constitution assessment model construction module, a dynamic feature extraction module, and a constitution status prediction module.

[0054] The constitution assessment model construction module first extracts tongue appearance features, pulse features, and symptoms from historical TCM diagnosis and treatment data as input parameters for constitution characteristics, collects constitution classification labels as output parameters, and quantifies environmental temperature and humidity data and emotional state scores to form constitution offset interference quantities. Based on this data, the constitution state equation is constructed using frequency domain feature transformation technology to convert time-domain features into frequency-domain feature vectors to establish a basic constitution response function, which is then expanded into a complete state equation by introducing interference quantity influence coefficients. The dynamic feature extraction module uses a sliding time window technique to segment the time-series interference quantity data, extracting periodic features and probability distribution features through spectral analysis and probabilistic statistical methods. The constitution state prediction module generates predicted interference quantities and their probability of occurrence based on the matching results between the current diagnosis and treatment time point and the periodic features, and finally outputs the constitution assessment results by combining real-time collected physiological characteristic data.

[0055] Example 1: See Figure 2 In the tongue image feature acquisition stage, high-resolution digital imaging equipment was used to capture tongue images under standard lighting conditions. During image acquisition, subjects were required to naturally extend their tongues, avoiding excessive force or contraction to ensure a natural tongue shape. The shooting angle was kept vertical, and the distance was controlled between 20 and 30 centimeters to obtain clear tongue surface details. The acquired images underwent preprocessing, including color correction, background removal, and region segmentation, dividing the tongue into four regions: tip, middle, root, and sides. The color characteristics of each region were extracted using RGB and HSV color space conversion, yielding the chroma, saturation, and brightness distribution of the tongue body and coating. Texture feature analysis employed wavelet transform and local binary mode algorithms to quantify the thickness, moisture, and cracking degree of the tongue coating. The final generated tongue image feature matrix contains three categories of indicators: color distribution, texture features, and region morphology, totaling 32 quantified parameters.

[0056] Pulse feature acquisition utilized a multi-channel intelligent pulse diagnostic instrument. The sensor array was positioned at the radial artery's cun, guan, and chi positions. During acquisition, the subject remained seated with arms naturally flat, avoiding speaking or movement. The instrument recorded pulse wave signals at a sampling rate of 200 frames per second for 3 minutes to ensure data stability. The raw pulse wave signals underwent filtering and noise reduction to eliminate interference from respiratory movements and muscle tremors. Time-domain feature extraction included four dimensions: pulse rate, pulse amplitude, pulse momentum, and pulse shape. Pulse rate was calculated based on the average interval between adjacent pulse wave peaks; pulse amplitude reflected the intensity variation of the pulse pulsation; pulse momentum described the slope characteristics of the pulse's rise and fall; and pulse shape was identified through waveform decomposition to recognize typical pulse characteristics such as wiry, slippery, and hesitant. Frequency-domain analysis employed Fast Fourier Transform to extract the dominant frequency and harmonic components of the pulse wave signal, reflecting the periodicity of qi and blood circulation. The final generated pulse feature vector contained 18 quantified indicators, covering key features in both the time and frequency domains.

[0057] Symptom characteristics were collected through a structured electronic questionnaire designed based on the "Classification and Determination of Traditional Chinese Medicine Constitution." The questionnaire covered six major categories of symptoms: cold and heat, sweating, sleep, diet, bowel movements, and emotions. Each category included several specific questions. For example, the cold and heat category included sub-items such as feeling cold, feeling hot, and hand and foot temperature; the sleep category included sub-items such as difficulty falling asleep, frequent dreams, and early awakening. Each question had a five-level rating option, ranging from 0 to 4 points from "none" to "severe." The questionnaire data was converted into a 56-dimensional Boolean feature vector, where continuous scores were converted into binary variables through thresholding. For example, a sleep quality score of 3 or higher was marked as "positive for sleep disorder," otherwise it was marked as negative. This structured processing facilitated machine learning models in identifying the correlation between symptom combination patterns and constitution types.

[0058] The constitution classification output is represented by probability distributions of nine basic TCM constitution types: balanced, qi-deficient, yang-deficient, yin-deficient, phlegm-dampness, damp-heat, blood stasis, qi stagnation, and special constitution. The probability distribution vectors are generated based on the output layer activation values ​​of a multi-classification model, which are normalized to a probability distribution summing to 1 using the Softmax function. This representation method reflects the tendency and transitional state of constitution types; for example, an individual may simultaneously possess characteristics of 60% phlegm-dampness and 30% damp-heat, reflecting the mixed nature and dynamic changes of constitution types.

[0059] Environmental temperature and humidity data were collected in real time via IoT sensors deployed in the subjects' usual living environments, such as bedrooms and offices. Data collection was conducted once per minute, recording three parameters: temperature, relative humidity, and dew point temperature. Data preprocessing included outlier removal and moving average smoothing to eliminate noise from instantaneous fluctuations. Analysis extracted the temperature and humidity trends over the two hours prior to treatment, calculating statistics such as mean, fluctuation amplitude, and rate of change. These parameters quantify the impact of the external environment on physical condition; for example, a sudden drop in temperature may exacerbate symptoms of Yang deficiency.

[0060] The emotional state scoring was assessed using a modified Traditional Chinese Medicine (TCM) emotional scale, which includes seven basic emotional states: joy, anger, worry, pensiveness, grief, fear, and fright. Each emotion has two dimensions: intensity score (0-10 points) and duration (hours). The assessment combined interviews and questionnaires to understand the subjects' recent (within one week) emotional changes and stressful events. Scoring data were standardized to eliminate baseline differences between individuals. The association analysis between emotional state and constitution type was based on historical data mining, such as the high correlation between Qi stagnation constitution and worry, or the potential link between damp-heat constitution and irritability.

[0061] Establishing a baseline pulse rate requires seven consecutive days of morning resting pulse rate data, collected within one hour of waking up, while maintaining a fasting and resting state. Daily pulse rate data are processed using a moving average to obtain a smoothed baseline curve. The difference between the real-time pulse rate and the baseline value reflects the immediate state of the autonomic nervous system, while the ratio provides a standardized stability assessment. When the pulse stability index exceeds a threshold, the system automatically marks it as a disturbance event, indicating possible emotional fluctuations, fatigue, or external factors. This dynamic monitoring mechanism can capture subtle changes in physical condition, providing a data foundation for subsequent disturbance prediction.

[0062] The entire data acquisition and processing workflow follows standardized operating procedures to ensure consistency across different time points and operators. The feature extraction algorithm has been optimized to balance computational efficiency and feature representation capabilities. The design of quantitative indicators takes into account both the theoretical connotations of Traditional Chinese Medicine and the needs of modern data analysis; for example, the "string pulse" feature is transformed into a combination of frequency domain energy distribution and time domain slope features. This multi-dimensional feature representation provides reliable data support for the comprehensive assessment and dynamic monitoring of physical condition.

[0063] Example 2: See Figure 3 The processing of tongue image features employs a two-dimensional discrete cosine transform (DCT) technique. The acquired tongue image is first divided into 8×8 pixel blocks, and a DCT is performed on each block. After the transform, 20 coefficients in the low-frequency region are retained; these coefficients contain key information about the tongue's color distribution and coating texture. High-frequency coefficients are discarded to reduce noise interference. Finally, the low-frequency coefficients of all blocks are recombinated according to their spatial location to form a frequency domain vector describing the overall tongue features. This vector captures the global color pattern and local texture variations of the tongue image; for example, the common characteristic of a red tongue with little coating in individuals with Yin deficiency is manifested by an abnormal increase in coefficients in specific frequency bands.

[0064] Frequency domain transformation of pulse characteristics is achieved through short-time Fourier transform. Continuously acquired pulse signals are segmented into 3-second segments, and each segment undergoes a Fourier transform after applying a Hanning window function. The transform result generates a time-frequency spectrum, from which energy values ​​of three key frequency bands are extracted: 0.1-0.5Hz reflects the baseline pulse rate, 0.5-2Hz corresponds to respiratory rhythm, and 2-10Hz characterizes cardiovascular regulatory activity. The average energy and peak frequency of each frequency band form a 12-dimensional feature vector. This processing can distinguish different types of pulse abnormalities, such as the characteristic energy accumulation of a wiry pulse in the 6-8Hz frequency band. Symptom features from medical history are directly input into the model as Boolean variables. The 56 symptom features are organized into a binary vector, where the presence or absence of each symptom is represented by 1 or 0. This vector is concatenated with the frequency-transformed tongue and pulse features to form a complete constitution feature input vector. This hybrid feature representation includes both quantitative physiological data and subjective symptom descriptions, such as the correlation pattern between cold intolerance symptom markers and frequency domain features of Yang deficiency constitution.

[0065] The basic constitution response function employs a three-layer fully connected neural network architecture. The number of nodes in the input layer corresponds to the total dimension of the feature vectors, receiving concatenated mixed features. The first hidden layer contains 128 nodes, using the ReLU activation function for non-linear transformation. The second hidden layer has 64 nodes, extracting key feature patterns through dimensionality reduction. The output layer uses 9 nodes corresponding to the nine constitution types, and the Softmax function transforms the node outputs into a probability distribution. The network is trained using a historical medical dataset, and the weight parameters are optimized through backpropagation to make the output probability distribution approximate the actual constitution classification labels.

[0066] The environmental temperature and humidity interference coefficient is modeled using a multivariate Gaussian distribution. The collected temperature and humidity data form a two-dimensional feature space, and the covariance matrix of different temperature and humidity combinations and constitution fluctuations in historical data is calculated. When real-time temperature and humidity data are input, the Mahalanobis distance between the data and the center of the ideal constitution environment is calculated. This distance value is normalized to an interference coefficient in the 0-1 range using a Sigmoid function. For example, in the assessment of phlegm-dampness constitution, a high humidity environment will cause the interference coefficient to approach 1, amplifying the impact of dampness on the constitution. The emotional state interference coefficient adopts a weighted scoring mechanism. Seven emotional states are assigned basic influence weights, and the sensitivity of each emotion to a specific constitution is determined through historical data analysis. Real-time emotional scores are multiplied by their corresponding weights, summed, and then mapped to the standard interference coefficient range through a linear transformation. For example, in the assessment of Qi stagnation constitution, the weights of "worry" and "thought" are set to twice that of other emotions, reflecting the pathological mechanism of worry and thought damaging the spleen and causing depression in traditional Chinese medicine theory. The pulse stability interference coefficient is directly quantified. The pulse stability index calculated in Example 1 has been adjusted from its original ratio to a coefficient value in the 0-1 range. This coefficient is directly related to the autonomic nervous system's regulatory state. When pulse rate fluctuations increase, the coefficient approaches 1, indicating a need for enhanced interference compensation.

[0067] During equation execution, the system first performs the same frequency domain transformation on the real-time acquired tongue and pulse data. The tongue image undergoes the same block segmentation and discrete cosine transform to extract low-frequency coefficients at the same locations. The pulse signal is subjected to short-time Fourier analysis using the same window length and frequency band division. The transformed frequency domain features are concatenated with the diagnostic symptom vector to form the current constitution feature input.

[0068] In the interference prediction processing stage, the three types of predicted interference quantities—environmental, emotional, and pulse—obtained from the dynamic feature extraction module are multiplied by their corresponding interference coefficients. Environmental interference is multiplied by a temperature and humidity interference coefficient, emotional interference by an emotional state interference coefficient, and pulse interference by a pulse stability interference coefficient. The weighted and corrected interference quantities form a three-dimensional interference vector.

[0069] In the feature fusion stage, the 3D interference vector and the constitution feature input vector are concatenated again. The expanded feature vector is then input into the pre-trained basic constitution response function network. The hidden layers of the network learn the interaction between the interference quantity and constitution features through nonlinear transformations; for example, the thermal features of Yin deficiency constitution are enhanced under high temperature conditions. The 9-dimensional probability vector generated by the output layer is the final constitution assessment result, containing the probability of occurrence of nine constitution types.

[0070] The entire execution process adopts a modular design, allowing for independent optimization of stages such as frequency domain conversion, interference coefficient calculation, and network forward propagation. The system incorporates a feature dimension verification mechanism to ensure data format matching at each stage. Execution logs record intermediate results for each evaluation, including frequency domain coefficient values, the interference coefficient calculation process, and activation values ​​at each network layer, providing a traceable basis for model optimization. This design balances the holistic perspective of traditional Chinese medicine theory with the precision of modern data processing, enabling multi-factor dynamic assessment of physical constitution.

[0071] Example 3: See Figure 4 This study utilizes time-series data analysis techniques to process physical imbalance interference, establishing a quantitative model of periodic and probability distribution characteristics, and using a time-matching algorithm to predict interference levels. The dynamic feature extraction module employs a sliding window algorithm to process continuous time-series interference data. The window length is set to 24 hours, and the sliding step size is configured to 1 hour; this parameter selection comprehensively covers diurnal rhythm variations. Data segments within each window are standardized to eliminate dimensional differences. Hourly averages are calculated for temperature and humidity data, the highest daily value is used for emotional state scores, and a moving average is used for pulse stability indicators. The processed data segments are stored in a structured format, containing three basic fields: timestamp, interference type, and numerical value.

[0072] The spectral analysis process performs a Fast Fourier Transform (FFT) on the interference sequence within each 24-hour window. A Hanning window function is applied before the transform to reduce spectral leakage, and the sampling frequency is set to once per hour based on the data characteristics. The top three dominant frequency components are extracted from the transform results, and their corresponding period values ​​are used as the periodicity characteristics of that window. The dominant period of temperature and humidity interference typically exhibits a 24-hour diurnal rhythm and a 12-hour semi-diurnal cycle, while emotional state interference may show a psychological cycle of about 7 days, and pulse stability commonly shows a short cycle of 4-6 hours related to diet and sleep patterns. These periodic characteristics are encoded and stored as triples:

[0073]

[0074] Wherein: T win f represents the periodicity of the current time window. primary f secondary and f tertiaryThese correspond to the top three frequency components in the power spectral density ranking. The physical meaning of these frequency components reflects the fluctuation patterns of different interference quantities, for example... The hour represents the period during which the diurnal temperature range affects physical condition.

[0075] For all historical window data of each disturbance type, a non-parametric probability density function is calculated. Temperature and humidity disturbances are considered as a two-dimensional joint distribution, while emotional state and pulse stability are treated as one-dimensional distributions. A Gaussian kernel is selected as the kernel function, and the bandwidth parameter is automatically determined using the Silverman rule. The density estimation results are discretized into percentiles, generating a cumulative distribution function lookup table. The probability of occurrence of a specific disturbance value x is calculated using linear interpolation.

[0076] P(x)=F cdf (x upper )-F cdf (x lower )

[0077] Wherein: F cdf Let x represent the cumulative distribution function. upper and x lower These are two discrete quantiles adjacent to x. This processing can accurately reflect the probability of extreme disturbance events, such as the possibility of continuous high temperatures or sudden emotional fluctuations.

[0078] The time-matching algorithm of the physical condition prediction module adopts a multi-level retrieval strategy, with the current diagnosis and treatment time point t as the reference. current Compared to the end point t of the previous interference time segment last The time interval Δt = t current -t last The calculation is accurate to the minute. The system maintains a periodic feature index, storing typical periodic patterns categorized by interference type and season. The matching process first searches within data of the same type of interference and the same season, calculating Δt relative to the storage period T. stored The absolute difference |Δt-T stored | Select the period corresponding to the smallest difference as the basis for prediction.

[0079] For the selected periodic feature T matched Retrieve all historical data that satisfy |Δt-T matched The instances of |<∈ are used, where ∈ is the time tolerance threshold, which is set to 10% of the period length by default. The disturbance values ​​in these instances form a candidate set, and the probability distribution of the current predicted value is calculated using kernel density estimation. The final output is the disturbance value with the highest probability density. and their corresponding probabilities It also provides a list of alternative values, sorted in descending order of probability.

[0080] The raw data layer stores complete time-series records, the feature layer stores extracted periodic features and probability distribution parameters, and the index layer constructs a fast query tree based on time and interference types. The data update mechanism is set to automatically remove expired segments and add new segments as the sliding window advances, maintaining a dynamic update of the feature library. This design can adapt to the drift in periodic patterns caused by seasonal changes and changes in individual habits.

[0081] The frequency resolution during spectrum analysis is set to The minimum identifiable period is 24 hours. Significance testing of frequency components uses an adaptive threshold, requiring the power spectral density of the dominant frequency to be at least three times the noise floor. For discrete interference quantities such as emotional states, linear interpolation is performed before spectral analysis to ensure sequence continuity.

[0082] When the standard deviation of the periodic characteristics for three consecutive windows is less than a threshold, a stable period is entered, and the probability distribution is updated using a moving average. When a periodic abrupt change is detected, a fluctuating period is entered, a new probability distribution model is established, and the transition state is marked. This mechanism can promptly capture the impact of changes in lifestyle or sudden environmental changes. A comprehensive confidence score is defined. Where α is the weighting coefficient, which is 0.5 by default. When S conf When the value is below 0.7, the system triggers a manual review flag, indicating that a comprehensive judgment needs to be made in conjunction with other information.

[0083] The entire implementation process emphasizes the dynamic and probabilistic nature of time-series characteristics. It utilizes a sliding window mechanism for continuous feature updates, captures cyclical patterns through spectral analysis, and quantifies uncertainty through probability distribution. The time-matching algorithm considers both cyclical patterns and real-time status, and the prediction results include probability assessments, providing multi-dimensional references for assessing physical condition. The data processing workflow is designed for incremental updates, adapting to the needs of data accumulation and pattern evolution in long-term monitoring.

[0084] Example 4: This example describes the implementation of a multi-model fusion decision-making mechanism in a TCM constitution identification intelligent monitoring system. By combining traditional TCM theoretical models with modern real-time data analysis, a dynamic weight adjustment fusion evaluation system is constructed. The constitution mechanism model module constructs a state transition network based on the transformation rules in *TCM Constitution Theory*. Nine basic constitution types serve as nodes, and the connections between nodes are categorized into three types: mutually generating relationships (e.g., the transformation from Qi deficiency to Yang deficiency), mutually restraining relationships (e.g., the constraint between damp-heat and Yang deficiency), and spontaneous transformations (e.g., the evolution from phlegm-dampness to blood stasis). The migration rule base contains 72 state transition paths, each with a set basic transition probability. The transition probability matrix is ​​initialized using an expert knowledge base and then updated using Bayesian data based on historical cases. For example, the initial transition probability from Qi deficiency to Yang deficiency is set to 0.25. When 28 out of 100 cases of Qi deficiency transformation develop into Yang deficiency, the probability is updated to 0.28. The state transition equation adopts a discrete-time Markov chain model. The current physical state vector is multiplied by the transition probability matrix to output the physical evolution trend in the next three months, forming a nine-dimensional mechanism evaluation vector.

[0085] The real-time feature analysis module employs a bidirectional long short-term memory (LSTM) network to process time-series data. The input layer receives three types of real-time data streams: tongue image feature sequences are collected every 6 hours, forming a 24-dimensional × 12-frame time-series matrix; pulse waveforms are sampled for 5 seconds every 30 minutes, and generated into an 18-dimensional × 48-frame feature cube through wavelet transform; environmental temperature and humidity data are recorded every minute and downsampled to the hourly average. The network structure includes a forward and backward layer with 64 memory units, and the memory unit state updates utilize a gating mechanism. The output layer, through a temporally distributed fully connected layer, maps the features of each time step to a 9-dimensional constitution assessment vector, ultimately using the output of the last time step as the real-time assessment value. This network can capture short-term fluctuations in physiological parameters, such as the impact of postprandial pulse changes on the assessment of damp-heat constitution.

[0086] The fusion decision module employs an adaptive weight allocation mechanism, with initial weights set at 0.4 for the mechanistic model and 0.6 for real-time analysis, reflecting an emphasis on dynamic data. The weight adjustment algorithm monitors the difference between the two assessment outputs, initiating an adjustment process when the Euclidean distance of three consecutive assessments exceeds a threshold D. The difference is calculated using a piecewise function: |mechanistic value - real-time value| < 0.1 is considered consistent; 0.1-0.3 is acceptable; and > 0.3 is considered significant discrepancy. The adjustment strategy includes three directions: if the difference between the real-time analysis module's assessment results and the manual review exceeds 0.2 for three consecutive assessments, its weight is reduced by 0.1; when sudden changes in environmental data cause abnormal output from the mechanistic model, its weight is reduced based on temperature and humidity deviation; if the difference continues to narrow, the initial weights are restored in 0.05 step increments. The fusion calculation uses a weighted summation formula: final physical fitness assessment value = W1 × mechanistic value + W2 × real-time value, where W1 + W2 = 1.

[0087] Table 1: Examples of physical condition transfer rules.

[0088] Initial constitution type Target body type Conversion type Description of conversion conditions Transition probability range Qi deficiency constitution Yang deficiency constitution Mutual generation Long-term fear of cold and cold limbs 0.20-0.30 Phlegm-dampness constitution damp-heat constitution Spontaneous conversion Overeating spicy and fatty foods 0.15-0.25 Yin deficiency constitution Blood stasis constitution Spontaneous conversion Persistent five-center heat 0.10-0.20 Qi stagnation constitution Blood stasis constitution Mutual restraint Depression lasting more than three months 0.25-0.35 peaceful temperament Qi deficiency constitution Spontaneous conversion Long-term overwork 0.05-0.15

[0089] The tongue image feature sequence was filtered using median filtering to eliminate transient interference, and the pulse waveform was analyzed using correlation coefficients to remove abnormal heartbeat cycles. Environmental data was analyzed using a sliding window standard deviation to detect abrupt changes. The network training employed an incremental learning strategy, fine-tuning the model parameters weekly with new data to maintain adaptability to individual constitution changes. In summer, the weight for the transformation to damp-heat constitution was increased by 0.1; in winter, the probability of transformation to yang deficiency constitution was increased by 0.15; and during the plum rain season, the tendency for transformation to phlegm-dampness constitution was strengthened by 0.2. Special transition probability matrices were used for special solar terms such as the winter solstice and summer solstice, reflecting the TCM theory of "correspondence between man and nature."

[0090] The fusion output comprises a three-layer structure: the constitution classification corresponding to the maximum probability of the main constitution type; the minor constitution types with a probability > 0.3 for transitional constitutions; and an abnormal fluctuation warning triggered when the probability of the main type decreases by more than 0.2 within a week. The output interface simultaneously displays the raw output values ​​of the mechanistic model and the real-time analysis, as well as the fusion weighting coefficients used, for professionals to refer to and make decisions.

[0091] Throughout the implementation process, a quality monitoring log was established to record the input data, intermediate results, reasons for weight adjustments, and final output for each evaluation. The log data was used for periodic model validation; if the fusion results deviated from clinical diagnosis by more than 15% across 20 consecutive evaluations, a model refactoring process was triggered. This design maintains the system's continuous optimization capabilities, balancing the advantages of traditional Chinese medicine theory with modern data analysis.

[0092] Example 5: This example describes the implementation of an intervention program optimization and dynamic adjustment mechanism within a TCM constitution identification intelligent monitoring system. A reinforcement learning framework is constructed to transform constitution assessment results into personalized TCM intervention parameters, and a feedback-based continuous optimization process is established. The intervention parameter optimization module employs a dual-network architecture. The strategy generation network receives a 65-dimensional decision state input, composed of a constitution classification probability vector and key symptom features. The first hidden layer contains 256 nodes, employing random deactivation to prevent overfitting; the second hidden layer outputs the mean vectors of 18-dimensional TCM compatibility parameters and 12-dimensional acupuncture program parameters. The value assessment network has dual output channels, predicting the short-term effect score within 3 days after intervention implementation and the long-term stability index after 3 weeks. Network training uses a historical intervention case dataset, with each case containing initial constitution state, intervention parameters, and implementation effect data. During training, experience replay technology is used, prioritizing the sampling of cases with larger effect deviations for focused learning.

[0093] The selection of principal herbs employs a hierarchical classification method, first determining broad categories such as qi-tonifying herbs, heat-clearing herbs, and dampness-removing herbs, and then selecting specific herbs within each category. The compatibility of assistant herbs considers their meridian tropism and synergistic effects, using a compatibility strength coefficient to adjust the principal-assistant relationship. Dosage gradients are divided into three levels: light, medium, and heavy, corresponding to 50%, 100%, and 150% of the standard dose, respectively. Herb processing methods include options such as raw use, processed use, and stir-frying, represented by unique heat codes. The administration time is set at three time slots: morning, noon, and evening, with independent dosage proportions allocated to each time slot.

[0094] The adjustable parameters of the acupuncture treatment plan constitute a complete treatment system. Acupoint combinations are grouped and coded, dividing commonly used acupoints into main acupoint groups, accessory acupoint groups, and special acupoint groups, with a maximum of four acupoints selected in each group. Stimulation techniques include three basic categories: twisting and rotating tonification / sedation, lifting and thrusting tonification / sedation, and even tonification / sedation, each with its own intensity level. Needle retention time is divided into three levels: short (15-20 minutes), standard (30 minutes), and long (40-60 minutes). Needle depth is set as a relative value based on the characteristics of the acupoint; for example, the Guanyuan acupoint uses an adjustable range of 80%-120% of the standard depth. Treatment frequency supports three modes: once daily, once every other day, and twice weekly.

[0095] The integrated processing of predicted interference quantities employs a probability-weighted method. Three types of predicted interference quantities—environmental, emotional, and pulse-related—obtained from the constitution prediction module are multiplied by their respective probabilities and then summed to obtain a comprehensive interference quantity estimate. This value is mapped to the adjustment space of different intervention parameters: environmental interference mainly affects drug dosage and needle retention time; emotional interference focuses on acupoint selection and stimulation techniques; and pulse-related interference adjusts treatment frequency and acupuncture depth. When the interference quantity exceeds a threshold, a parameter compensation mechanism is triggered, such as automatically increasing the dosage of heat-clearing drugs by 10%-15% for patients with damp-heat constitution in hot weather. A multi-dimensional evaluation system is established for intervention effect monitoring. Short-term effect evaluation is conducted on the third day after intervention, collecting changes in tongue and pulse characteristics and the degree of symptom improvement. Long-term stability monitoring continues for three weeks, recording fluctuations in constitution type and the frequency of symptom recurrence. The deviation between the actual constitution classification output and the expected target is calculated using the vector angle method, comparing the directional consistency of the nine-dimensional probability distribution vector. When the deviation angle exceeds 25 degrees, it is considered a significant difference, requiring the initiation of the parameter adjustment process. The dynamic correction of the dose weighting coefficient employs fuzzy logic rules, classifying intervention deviations into three fuzzy sets: small, medium, and large. The predicted probability of interference occurrence is categorized into three levels: low, medium, and high. The correction rule base contains 27 combination rules, such as "medium deviation + high probability → increase dose by 30%" or "large deviation + low probability → change the principal herb." Coefficient adjustments consider the characteristics of the medicinal materials; tonifying drugs use linear increments, while toxic drugs have a maximum safety threshold. The adjustment of acupuncture stimulation intensity is linked to pulse stability; when pulse fluctuations exceed the baseline by 20%, the stimulation intensity is automatically reduced by one level.

[0096] Basic optimization involves fine-tuning parameters, adjusting dosage and stimulation intensity while keeping the principal herb and main acupoints unchanged. Intermediate optimization allows changing the auxiliary herbs and acupoints, but retains the core treatment approach. Advanced optimization is initiated after two consecutive ineffective fine-tuning attempts, regenerating a complete intervention plan. Each optimization retains historical parameter records to avoid cyclical adjustments between similar states. The system maintains a contraindication knowledge base to prevent the generation of plans that violate contraindications for traditional Chinese medicine compatibility or acupuncture.

[0097] The entire implementation process emphasizes the personalization and dynamic adaptability of the intervention plan. The strategy generation network learns the optimal intervention strategy under different physical conditions, and the value assessment network predicts the short-term and long-term effects of the plan. The integration of predicted interference amounts makes the plan forward-looking, and effect monitoring and parameter adjustment form a closed-loop optimization. The correction of dosage and stimulation intensity considers both the current effect deviation and the possibility of interference, achieving precise treatment. The multi-level optimization mechanism balances the continuity and innovation of the treatment plan, flexibly adjusting detailed parameters while maintaining the core treatment ideas. The intervention cases accumulated during system operation continuously enrich the training data, driving the model to continuously improve itself.

[0098] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0099] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart monitoring system for identifying traditional Chinese medicine constitution, characterized in that, include: The constitution assessment model construction module is used to obtain the constitution characteristic input, constitution classification output, and constitution deviation interference based on historical TCM diagnosis and treatment data, and to construct the constitution state equation based on the constitution characteristic input, constitution classification output, and constitution deviation interference. The dynamic feature extraction module is used to set a time window, perform sliding segmentation processing on the body mass shift interference amount based on the time window, generate interference amount time segments, and analyze the interference amount time segments to extract the periodic features and probability distribution features of each body mass shift interference amount. The physical condition prediction module is used to obtain the current diagnosis and treatment time point, predict the interference amount based on the periodic features, match the occurrence probability of the predicted interference amount based on the probability distribution features, and output the predicted interference amount and the corresponding occurrence probability. The historical TCM diagnosis and treatment data includes tongue appearance characteristics, pulse appearance characteristics, symptoms during consultation, constitution classification labels, environmental temperature and humidity data, and emotional state scores. The method for quantifying the interference of physical constitution deviation includes: calculating the difference between the pulse rate fluctuation value in the pulse characteristics and the reference pulse rate value as the first deviation, and setting the ratio of the first deviation to the reference pulse rate value as the pulse stability index. The output of the constitution classification is represented as a probability distribution vector of various TCM constitution types; The input of physical characteristics includes tongue appearance characteristics, pulse appearance characteristics, and symptoms observed during medical consultation. The method for constructing the physical state equation includes: The input quantity of physical characteristics, the output quantity of physical classification, and the quantity of physical offset interference are converted into frequency domain feature vectors through time-frequency conversion technology. Ignoring the interference of body constitution offset, a basic body constitution response function is established based on the transformed body constitution feature input and body constitution classification output. After introducing the body weight deviation interference, the basic body weight response function is extended into a body weight state equation that includes the interference influence coefficient. The execution process of the physical state equation includes: Receive the current physical characteristics input and the predicted interference; The preset interference coefficients are invoked, and the predicted interference amount is weighted and corrected according to the interference coefficients; the interference coefficients are environmental temperature and humidity, emotional state, and pulse stability. The corrected interference quantity and the physical characteristics input quantity are input together into the basic physical response function, and the target physical classification probability distribution vector is output.

2. The intelligent monitoring system for TCM constitution identification as described in claim 1, characterized in that, The dynamic feature extraction module performs the following steps: Set a fixed-length time window and sliding step size; Traverse the time series data of body offset interference according to the sliding step size, and extract continuous time segments as interference time segments. Spectral analysis was performed on the time segments of each interference quantity, and the period corresponding to the dominant frequency component was extracted as a periodic feature. The frequency of a specific interference value in each interference time segment is statistically analyzed, and its ratio to the total number of segments is calculated as a probability distribution characteristic.

3. The intelligent monitoring system for TCM constitution identification as described in claim 1, characterized in that, The physical condition prediction module performs the following steps: Calculate the time interval between the current diagnosis / treatment time point and the end point of the previous interference time segment; The time interval is matched with the periodic characteristics of the interference amount of each body mass deviation by difference; The interference quantity corresponding to the periodic feature with the smallest absolute difference is selected as the predicted interference quantity; The probability of occurrence of the predicted interference quantity is retrieved from the probability distribution feature library.

4. The intelligent monitoring system for TCM constitution identification as described in claim 1, characterized in that, Also includes: The constitution mechanism model module is used to establish a constitution state transition equation based on the TCM constitution transformation theory. The constitution state transition equation uses nine basic TCM constitution types as nodes, constructs a transition probability matrix based on the mutual generation, mutual restraint and spontaneous transformation relationship of constitutions, processes the current constitution state vector through a discrete-time Markov chain model, and outputs a constitution mechanism evaluation value. The real-time feature analysis module is used to process real-time collected tongue, pulse, and environmental data through a deep temporal network to generate real-time physical condition assessment values. The fusion decision module is used to dynamically allocate weight coefficients based on the error between the physical fitness mechanism assessment value and the real-time physical fitness assessment value, and generate a fusion physical fitness assessment result.

5. The intelligent monitoring system for TCM constitution identification as described in claim 1, characterized in that, Also includes: The intervention parameter optimization module is used to construct an action space that includes parameters of traditional Chinese medicine compatibility and acupuncture plan parameters by using the output of constitution classification as the decision state quantity. The probability distribution of intervention parameters output by the strategy generation network is used to select the optimal combination of intervention parameters based on the dual-channel output of the value assessment network.

6. The intelligent monitoring system for TCM constitution identification as described in claim 5, characterized in that, The intervention parameter optimization module also performs: Receive the predicted interference amount and occurrence probability output by the physical condition prediction module; Monitor the actual physical condition classification output after implementing the optimal combination of intervention parameters; Calculate the intervention bias between the actual body constitution classification output and the expected body constitution classification output; When the intervention deviation exceeds the preset tolerance threshold, the dose weight coefficient of the Chinese medicine compatibility parameters and the stimulation intensity adjustment coefficient of the acupuncture scheme parameters are dynamically adjusted based on the probability of the predicted interference.

Images