Traditional Chinese medicine constitution identification method based on multi-mode bioelectric signal perception

Abstract

The invention provides a traditional Chinese medicine physique identification method based on multi-modal bioelectricity signal perception, which comprises the following steps: synchronously acquiring electrocardio, myoelectricity, electroencephalogram and skin electricity reaction signals of a subject, combining individual physiological parameters and historical physique labels, extracting time domain, frequency domain and nonlinear characteristics after multi-scale denoising and normalization, and fusing and modeling; a group causal map is constructed by adopting causal discovery algorithms such as PC and FCI, node weight dynamic adjustment is realized by utilizing individual parameters, an individualized dynamic causal map is generated, and causal reasoning path output and adaptive structure optimization are supported. And automation and accuracy of traditional Chinese medicine constitution identification are improved.

Description

Technical Field

[0001] This invention relates to the fields of biomedical signal processing and traditional Chinese medicine constitution identification, and in particular to a method for traditional Chinese medicine constitution identification based on multimodal bioelectric signal sensing. Background Technology

[0002] In the current fields of biomedical signal processing and TCM constitution identification, with the integration of multimodal physiological data acquisition, artificial intelligence and TCM constitution theory, the means of constitution classification are gradually developing from traditional questionnaires and subjective assessments to intelligent identification based on objective data. Traditional Chinese medicine (TCM) constitution identification systems largely rely on static machine learning models and physiological signal analysis. The common approach involves collecting single or limited types of bioelectrical signals (e.g., electrocardiogram, electroencephalogram, electrodermal conductance) from subjects, extracting features in the time and frequency domains, and then using a fixed-structure classification model (e.g., SVM, RF, KNN, or traditional neural networks) to determine constitution type. To improve model performance, some solutions introduce multimodal signal fusion and feature engineering; however, the models themselves are mostly static: the discrimination rules and parameters for each constitution type are fixed during training, and all samples are processed along a uniform path during inference, making it difficult to adjust the inference process for individual physiological differences or complex constitution states. Furthermore, recent attempts to introduce causal reasoning into constitution identification have largely employed fixed-structure causal graphs or Bayesian networks based on group samples. Once built, the inference path and method remain largely unchanged except for minor parameter adjustments, lacking the ability to dynamically adjust for individual differences. Regarding model interpretability, most rely on global feature importance ranking, static causal path display, or attention weight visualization, providing only limited, hard-coded explanations. Currently, existing methods for identifying constitution through causal reasoning have the following prominent technical defects and shortcomings: (1) Causal reasoning mechanisms are usually static structures. The network structure, causal path, and weights between nodes of most causal reasoning models are fixed during the training or modeling stage. During the reasoning process, the same path and explanation template are used for different physical samples, and the reasoning logic cannot be flexibly adjusted according to individual differences (such as age, medical history, physiological characteristics, etc.). This rigid structure leads to poor performance of the model for samples with blurred physical type boundaries, mixed physical types, or complex states; (2) Insufficient individualized dynamic adaptability: Existing models are unable to dynamically optimize the causal graph structure based on an individual's historical physical condition and real-time physiological parameters, and lack the ability to adaptively reason about high-level semantics such as specific subject-specific signal patterns and physical condition change trajectories. The causal paths learned by the group samples may not be applicable to individuals with special physiological and physical characteristics, thus affecting classification robustness and generalization ability; (3) The reasoning process has limited transparency. Although some causal models can output reasoning paths or feature importance, the explanation is usually based on globally unified nodes / paths, ignoring the individual's current physiological state and dynamic weight changes. The transparency and clinical reference value are low, making it difficult to meet the needs for traceable and rational explanation of complex cases or multidimensional health management. (4) Static models are not adaptable online when faced with changes in data distribution and long-term physical state evolution. They do not support periodic updates of structure or dynamic learning of causal relationships between newly emerging physical characteristics, which affects the stability of the model's life cycle. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method for identifying traditional Chinese medicine constitution based on multimodal bioelectric signal sensing.

[0004] The technical solution of this invention is implemented as follows: A method for identifying traditional Chinese medicine constitution based on multimodal bioelectrical signal sensing, comprising: S1: Collect multimodal bioelectrical signals from the subjects, including electrocardiogram signals, electromyogram signals, electroencephalogram signals, and skin conductance response signals, and simultaneously record individual physiological parameters and historical body type labels; S2: The collected multimodal bioelectric signals are preprocessed by wavelet transform denoising and Z-score normalization to eliminate environmental interference and individual baseline differences. S3: Extract time-domain, frequency-domain, and nonlinear features from the preprocessed signal, generate multi-dimensional feature vectors, and construct a feature fusion matrix by combining individual physiological parameters; S4: Construct an initial causal graph based on population sample data, use PC algorithm and causal discovery model to identify causal relationships between feature nodes, and define the causal path between body type and feature factors; S5: Input individual physiological parameters and historical physical condition information as dynamic adjustment factors into the causal graph, adjust node weights and path connection structure, and generate an individualized dynamic causal graph. S6: Input the current fused feature matrix into the dynamic causal graph, activate the matching causal subgraph path, perform causal inference calculation, and output the constitution classification result and the corresponding causal inference path explanation; S7: Determine whether the current constitution classification result meets the confidence threshold. If it does, output the final identification result and causal path explanation. Otherwise, trigger the graph structure fine-tuning mechanism to optimize the inference path. S8: Generates a visual report based on the identification results and causal path explanation, and provides identification basis and body conditioning suggestions through an interactive interface; S9: Regularly collect new sample data and update the population causal graph, and use an online learning mechanism to optimize the generalization ability and adaptability of the dynamic causal reasoning model.

[0005] This application provides a method for identifying traditional Chinese medicine constitution based on multimodal bioelectrical signal sensing, which has the following beneficial effects: This invention innovatively introduces an individualized dynamic causal graph and a self-adjusting edge weight mechanism. By combining real-time updated physiological parameters and historical physical information, it dynamically adjusts the causal inference path structure to achieve adaptive changes in the causal graph. Compared with traditional static causal inference models or fixed attention mechanisms, the dynamic causal structure can optimize the inference logic in real time for different physical types and individual characteristics. It effectively solves the problems of insufficient adaptability and rigid inference paths of existing causal inference models when facing complex physical states or group specificities, significantly improving the model's generalization ability and accuracy in tasks involving the identification of diverse individuals and physical types. This invention integrates multimodal bioelectrical data such as electrocardiogram (ECG), electromyography (EMG), electroencephalography (EEG), and electrodermal conductance (EDC) signals, and deeply fuses physiological parameters and historical constitution labels to propose a high-dimensional feature fusion matrix that combines physiological and cognitive levels. Based on this matrix, a dynamic causal graph can accurately extract individual differences and map them to the core causal path of constitution identification, effectively improving the identification model's ability to recognize physiological abnormalities, mixed constitutions, and special groups. Compared with existing single-modal or single-label inference methods, the accuracy of constitution classification is significantly improved, with significant advantages in scenarios where the boundaries of constitution types are ambiguous. The dynamic path activation and Top-K candidate selection algorithm, along with do-calculus causal contribution quantification, of this invention enable the output constitution classification to possess key path backtracking and traceable explanation capabilities. The system not only provides classification results but also simultaneously outputs causal path explanation text, presenting key feature nodes and their mechanisms of action in a fine-grained manner. This interpretability significantly enhances the trust of users and clinical experts in the system's decision-making basis, solving the problem that existing black-box models and static attention mechanisms struggle to provide clear medical explanations. This invention utilizes graph attention mechanisms and algorithms such as dynamic pruning and path enhancement to quickly filter high-value paths from massive causal paths. The average inference time is significantly reduced compared to traditional full-path computation, while energy consumption and computational resource usage are also significantly lowered. Furthermore, by quantifying the importance of nodes / paths, the model computation focuses on key information, improving its applicability to real-time processing of complex data in practical health management systems and supporting the real-time inference needs of edge computing and mobile healthcare scenarios. Attached Figure Description

[0006] Figure 1 This is a flowchart of a traditional Chinese medicine constitution identification method based on multimodal bioelectric signal sensing according to the present invention. Figure 2This is a sub-flowchart of a traditional Chinese medicine constitution identification method based on multimodal bioelectric signal sensing according to the present invention. Figure 3 This is another sub-flowchart of a traditional Chinese medicine constitution identification method based on multimodal bioelectric signal sensing according to the present invention. Detailed Implementation

[0007] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0008] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0009] When used herein, the singular forms of “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising / including” or “having,” etc., specify the presence of the stated features, wholes, steps, operations, components, parts, or combinations thereof, but do not preclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof. Meanwhile, the term “and / or” as used in this specification includes any and all combinations of the associated listed items.

[0010] Please see Figures 1-3 As shown, a method for identifying traditional Chinese medicine constitution based on multimodal bioelectrical signal sensing includes: S1: Collect multimodal bioelectrical signals from the subjects, including electrocardiogram signals, electromyogram signals, electroencephalogram signals, and skin conductance response signals, and simultaneously record individual physiological parameters and historical body type labels; S2: The collected multimodal bioelectric signals are preprocessed by wavelet transform denoising and Z-score normalization to eliminate environmental interference and individual baseline differences. S3: Extract time-domain, frequency-domain, and nonlinear features from the preprocessed signal, generate multi-dimensional feature vectors, and construct a feature fusion matrix by combining individual physiological parameters; S4: Construct an initial causal graph based on population sample data, use PC algorithm and causal discovery model to identify causal relationships between feature nodes, and define the causal path between body type and feature factors; S5: Input individual physiological parameters and historical physical condition information as dynamic adjustment factors into the causal graph, adjust node weights and path connection structure, and generate an individualized dynamic causal graph. S6: Input the current fused feature matrix into the dynamic causal graph, activate the matching causal subgraph path, perform causal inference calculation, and output the constitution classification result and the corresponding causal inference path explanation; S7: Determine whether the current constitution classification result meets the confidence threshold. If it does, output the final identification result and causal path explanation. Otherwise, trigger the graph structure fine-tuning mechanism to optimize the inference path. S8: Generates a visual report based on the identification results and causal path explanation, and provides identification basis and body conditioning suggestions through an interactive interface; S9: Regularly collect new sample data and update the population causal graph, and use an online learning mechanism to optimize the generalization ability and adaptability of the dynamic causal reasoning model.

[0011] Step S1: Collect multimodal bioelectrical signals from the subject, including electrocardiogram (ECG), electromyography (EMG), electroencephalography (EEG), and skin conductance signals, and simultaneously record individual physiological parameters and historical body type labels. Specifically, this includes: S1.1: Multimodal bioelectric signals of the subject are simultaneously acquired through a multi-channel biosignal acquisition device. The signals include electrocardiogram signals, electromyogram signals, electroencephalogram signals and skin conductance response signals, in order to obtain raw physiological data reflecting the autonomic nervous system and physical state. When performing sub-step S1.1, the subject is placed in a data acquisition environment with controllable interference signal levels, ensuring that environmental electromagnetic interference and mechanical vibration noise are below a preset threshold as the initial conditions for data acquisition. A multi-channel biosignal acquisition device (parameter settings: sampling rate 1000Hz, resolution 24bit, channel input impedance greater than 10MΩ) is used to simultaneously access the electrocardiogram signal channel, electromyography signal channel, electroencephalogram signal channel and skin conductance response signal channel to achieve synchronous acquisition of raw physiological data of the autonomic nervous system and body state; Furthermore, through a multi-channel synchronous clock driving mechanism (clock accuracy ≤ (seconds), hardware binding of sampling trigger signals for each channel is used to ensure consistency of all modal signals on the time axis, thereby achieving high-precision alignment of cross-modal signal data; Furthermore, a pre-amplifier analog filter is used to perform front-end processing on the input signals of each channel, and a bandpass filter (cutoff frequency) is applied to the ECG channel. Hz to Hz), electromyography channel loaded with bandpass filter (cutoff frequency) Hz to Hz), EEG channel loaded with bandpass filter (cutoff frequency) Hz to Hz), skin conduction response channel loaded with a low-pass filter (cutoff frequency) (Hz) to suppress DC drift and high-frequency noise components; Furthermore, the filtered analog signal is converted into a digital signal stream through an analog-to-digital conversion module, and then buffered and packetized under the real-time control of the FPGA to generate data packets with precise timestamps; Through the above synchronous acquisition and signal preprocessing, the multimodal physiological raw data of the previous step is transformed into a high-time-precision, low-noise, traceable multi-channel bioelectric signal raw dataset, which provides a signal foundation for subsequent feature extraction and dynamic causal inference. For example, in one subject's data collection, the multi-channel biosignal acquisition device was set to BIOPACMP160. The ECG channel used disposable Ag / AgCl electrodes with a sampling rate of 1000Hz. The EEG channel used 32-lead Contex electrode caps with a synchronized sampling rate of 1000Hz. The EMG channel used surface electrodes attached to the quadriceps and biceps muscles. The dermal electromyography (EDG) channel electrodes were fixed to the palm and the tip of the index finger. Before the data collection began, environmental electromagnetic interference was reduced to below [a certain level] in a shielded room. μVRMS. During the acquisition process, the FPGA clock source accuracy is... Seconds, ensuring cross-channel synchronization error is less than Microseconds. The signal input is loaded with preset filtering parameters at the acquisition end; the bandpass filtering range of the electromyographic signal is set to... Hz- Hz was used to suppress motion artifacts. After 30 seconds of resting state acquisition, the final raw data packet size was 120MB, and the measured signal-to-noise ratio of the ECG channel reached [value missing]. dB or higher, brainwave channels reach Above dB, the signal-to-noise ratios of electromyography and electrodermal channels are respectively... dB and dB, to meet the input requirements for data quality in subsequent steps S1.2 and beyond; S1.2: Perform R-wave detection and RR interval calculation on the acquired ECG signal, use the Pan-Tompkins algorithm to identify QRS complexes, and extract the heart rate variability (HRV) index to quantify the autonomic nervous system regulation state; The Pan-Tompkins algorithm was used to analyze the ECG digital signal sequence acquired by S1.1 (parameters: bandpass filter 0.5Hz-40Hz, differential filter order 5, sliding window length). (ms), to achieve preliminary candidate interval detection of QRS groups; Furthermore, a feature enhancement method based on first-order differential and square operations (sliding integral window parameter) is employed. (ms), to achieve R-wave main peak enhancement and obtain candidate R-wave peak sequences; Furthermore, an adaptive threshold decision algorithm is adopted (the initial threshold is the average amplitude of the candidate peaks). This allows for accurate localization of the R wave in electrocardiogram (ECG) signals and the generation of R wave timestamp vectors sorted by time series. ; Furthermore, based on the time difference between adjacent R-wave points, the RR interval vector is calculated. This enables the acquisition of cardiac cycle sequences; Furthermore, based on the RR interval vector, the average heart rate is calculated using time-domain analysis methods. Standard deviation Root mean square of the difference between adjacent periods Equal-time domain HRV index; Furthermore, the Welch power spectrum estimation method (parameter: overlap rate) is employed. Using a Hamming window with 256 points, calculate the power spectral density during the RR interval and integrate to obtain the low-frequency power. (0.04-0.15Hz) and high-frequency power (0.15-0.4Hz), and then calculate the LF / HF index to quantify the balance between the sympathetic and parasympathetic nervous systems; The above algorithm is used to process the raw ECG channel signal obtained in the previous step into a multidimensional heart rate variability (HRV) index set, thereby achieving a quantitative characterization of the autonomic nervous system's regulatory state. For example, in a single subject trial, the sampling rate for acquiring electrocardiogram (ECG) signals was set to [value missing]. The Pan-Tompkins filter bank parameters are bandpass 0.5Hz-40Hz, differential filtering with five-point differential, and a sliding integral window of 150ms. After candidate QRS detection, the adaptive threshold is initially set to the mean of the candidate peak values. 1840 R waves were detected (10-minute recording). The RR interval was calculated based on the R wave timestamp differences to obtain the average RR interval. Second, =65ms =58ms. The frequency domain indices were calculated using the Welch method: LF = 820ms², HF = 640ms², and LF / HF = 1.28, indicating that the subject was in a state of basic balance between sympathetic and parasympathetic activity. This set of indices was input into the subsequent causal mapping modeling module as an important feature describing autonomic nervous system function. S1.3: Time-domain envelope extraction and spectrum analysis of electromyographic signals are performed. Hilbert transform is used to obtain the signal envelope amplitude, and the median frequency MF and average power frequency MPF are calculated to assess muscle activation and fatigue status. S1.4: Perform frequency band energy analysis on EEG signals, extract the power spectral density of the δ, θ, α, β, and γ frequency bands based on Fast Fourier Transform (FFT), and calculate the energy proportion of each frequency band to reflect the activity characteristics of the cerebral cortex. S1.5: Differentiate and detect peak values ​​of the electrodermal response (SCR) signal to extract amplitude and frequency characteristics of the SCR, thereby quantifying the levels of emotion and stress response. S1.6: Simultaneously collect individual physiological parameters of the subjects, including age, gender, body mass index (BMI), blood pressure and underlying disease information, in order to construct personalized modeling input features; S1.7: Record the subject's historical constitution type label, which includes balanced constitution, qi deficiency constitution, yang deficiency constitution, yin deficiency constitution, phlegm-dampness constitution, damp-heat constitution, blood stasis constitution, qi stagnation constitution and special constitution, as a supervisory signal for training the causal reasoning path.

[0012] Step S2: Preprocessing the collected multimodal bioelectrical signals involves wavelet transform denoising and Z-score normalization to eliminate environmental interference and individual baseline differences. Specifically, this includes: S2.1: Channel alignment and time synchronization processing are performed on the acquired raw electrocardiogram, electromyography, electroencephalogram and skin conductance signals to eliminate time offset between multimodal signals and obtain synchronized bioelectric signal sequences. The raw ECG, EMG, EEG, and skin conductance digital signals acquired in step S1 are analyzed using a full-channel timestamp parsing algorithm (parameter: timestamp resolution). (microseconds) to decode and uniformly format the timestamp information of the acquired data packets in order to establish a unified time reference system for cross-modal signals; Furthermore, a coarse-grained alignment method based on the cross-correlation function (delayed search range ±) is used. ms, step size (ms), calculate the estimated time delay between each channel and the reference channel; Furthermore, a fine-grained synchronization algorithm based on Dynamic Time Warping (DTW) is adopted (parameter: window constraint width). (Frame), and perform nonlinear time axis registration on the coarsely aligned multimodal signal to generate a corrected time index mapping matrix to correct minor mismatches caused by acquisition hardware delay or signal shape differences in different channels; Furthermore, using the cubic spline interpolation method (parameter: interpolation step size) (seconds), resample all channel signals to a unified sampling rate matrix to ensure that multi-channel signals have synchronized data values ​​at every sampling point; Furthermore, an initial amplitude normalization process is performed (method: linear amplitude stretched to the [-1,1] interval) to eliminate the differences in output dimensions between different types of sensors and avoid the subsequent synchronization accuracy being affected by proportional imbalance. Through the above channel alignment and timing synchronization processing, the original bioelectric signal is transformed into a synchronized signal sequence under a unified time axis, thus achieving a rigorous input guarantee for cross-modal feature extraction and wavelet denoising. For example, in a simultaneous acquisition including 12-lead ECG, 8-channel EMG, 32-channel EEG, and 2-channel TESA, the signal sampling rate is 1. Hz, timestamp resolution is Microseconds. When using coarse alignment with cross-correlation, the maximum correlation delay of the electromyography-electrocardiography channel is... ms, EEG-ECG channel delay is ms, the skin conductance-electrocardiogram channel delay is After DTW fine-tuning, the cross-channel residual time error is less than ms. μs. After uniform sampling using cubic spline interpolation, all channels have corresponding values ​​at each sampling point every millisecond, and the amplitudes of each channel are distributed in the range [-1, 1] after amplitude normalization. Validation results show that this dataset achieves a high temporal correlation coefficient in the subsequent wavelet decomposition synchronization performance metrics. The above demonstrates that the synchronization process meets the technical requirements for high-precision multimodal signal fusion. S2.2: Based on the wavelet transform algorithm, each signal channel in the synchronized bioelectric signal sequence is decomposed into multiple scales, and wavelet coefficients at each scale are extracted to identify and separate high-frequency noise components in the signal; S2.3: Perform threshold quantization on the extracted wavelet coefficients and use a soft threshold function to compress the high-frequency coefficients to suppress non-stationary noise interference and obtain the denoised reconstructed signal; S2.4: Perform Z-score normalization on the denoised reconstructed signal, calculate the standardized value based on the historical mean and standard deviation of each channel signal to eliminate the influence of individual baseline differences on the signal amplitude, and obtain the standardized signal matrix; S2.5: Perform quality assessment on the standardized signal matrix, calculate the signal-to-noise ratio and signal integrity index, and mark abnormal signal channels. If any channel has substandard quality, trigger a resampling or signal compensation mechanism to ensure that the preprocessed signal meets the input requirements of the subsequent feature extraction module.

[0013] Step S3: Based on the preprocessed signal, extract time-domain, frequency-domain, and nonlinear features to generate a multi-dimensional feature vector, and construct a feature fusion matrix by combining individual physiological parameters. For example... Figure 2 As shown, it specifically includes: S3.1: Temporal features were extracted from the preprocessed electrocardiogram (ECG), electromyography (EMG), electroencephalogram (EEG), and electrodermal response (TEG) signals. Mean, variance, skewness, and kurtosis statistical methods were used to obtain the temporal feature vectors of each signal. S3.2: Perform Fast Fourier Transform (FFT) processing on the electrocardiogram (ECG), electromyogram (EMG), electroencephalogram (EEG), and electrodermal response (EDS) signals to extract frequency domain features, including power spectral density, main band energy ratio, and spectral centroid, and generate frequency domain feature vectors; For the preprocessed signal after extracting time-domain features in step S3.1, the Fast Fourier Transform (FFT) algorithm is used (parameter: sampling rate). Hz, FFT points (The window function type is Hanning window) to realize the time-frequency domain conversion of multimodal bioelectric signals and obtain the complex spectrum data corresponding to each channel; Furthermore, the power spectral density (PSD) estimation method (parameter: Welch segment length) is used. Points, overlap rate %), calculate the energy distribution of each signal channel at each frequency component, and obtain the power spectral density function. ; Furthermore, by setting the frequency band of interest (ECG is...) Hz, EMG is Hz, EEG is Hz, GSR is (Hz) Perform integration on the power spectral density curve to calculate the energy value of each main frequency band. And calculate its relative to the total energy. The ratio is used to reflect the contribution of a specific frequency band to the overall energy; Furthermore, using the spectral centroid calculation method, the power spectrum is used as a weight to perform a weighted average of the frequencies, thus obtaining the spectral centroid. It is used to characterize the central location of energy distribution and reflect the dominant frequency characteristics of the signal; By sequentially combining indicators such as power spectral density, main band energy ratio, and spectral centroid of each channel, a frequency domain feature vector is generated, which serves as the input for subsequent nonlinear feature extraction and feature fusion. For example, at a sampling rate of In the processing of 12-lead ECG signals at Hz, 4 seconds of data are extracted from each lead (total). Using a 1024-point FFT and Hanning window framing, the Welch method employs a segment length of 256 points with 50% overlap in power spectral density estimation. The dominant band energy ratio is calculated for the 0.05–150 Hz range, yielding the R value for lead II. The centroid of the spectrum is Hz; In the analysis of 32-channel EEG signals (0.5~45Hz), the energy proportion R of the α band The centroid of the spectrum is Hz. Verification results show that this frequency domain feature can accurately reflect the energy concentration characteristics and dominant activity frequency bands of different modal signals, providing effective feature input for causal path mapping of physical constitution identification; S3.3: Based on the time series data of the multimodal bioelectric signals, calculate the sample entropy, approximate entropy, and Lyapunov exponent to extract nonlinear dynamic features and form a nonlinear feature vector; Based on the frequency domain feature vector obtained in step S3.2 and the original time-series signal data, the sample entropy calculation method is adopted (parameter: embedding dimension). Similarity tolerance ,in (As the standard deviation of the signal), enabling quantitative assessment of the complexity level of bioelectrical signals in different modalities; Furthermore, the sample entropy is calculated using the following formula. : in For all lengths The number of matching templates, For length is The number of matching templates; Furthermore, an approximate entropy calculation method is employed (parameter: embedding dimension). (The tolerance threshold is consistent with the sample entropy) to characterize the regularity of signal patterns, and the approximate entropy is calculated using the following formula. : in To match probability statistics functions, For similarity tolerance; Furthermore, the Wolf algorithm is used to calculate the Lyapunov exponent (parameter: phase space reconstruction delay time). Point, Embedded Dimension This allows for a quantitative description of the chaotic characteristics of a signal and estimation of the maximum Lyapunov exponent. ; Furthermore, the three nonlinear features—sample entropy, approximate entropy, and Lyapunov exponent—are normalized (method: Min-Max normalization to interval). ), and concatenate them in order of feature type to form a multimodal nonlinear dynamics feature vector; The above methods transform the signal data processed in the frequency and time domains into quantitative indicators with complex, nonlinear and chaotic characteristics, thereby achieving a comprehensive characterization of bioelectric signals at the nonlinear dynamics level. For example, at a sampling rate of In a 32-channel EEG analysis at Hz, 10 seconds of data were taken from each channel. (Point), embedding dimension set to The tolerance is set to the signal standard deviation. The average sample entropy was calculated to be The average approximate entropy is The phase space (delay time) is reconstructed using the Wolf algorithm. Point, Embedded Dimension And by tracking the divergence rate of adjacent trajectories, the mean of the maximum Lyapunov exponent is obtained. / second. The above three types of nonlinear features, after Min-Max normalization, are concatenated to form a length of... The nonlinear feature vector significantly improved the model's sensitivity to and classification and explanation capabilities for abnormal brain function activity patterns in the subsequent causal reasoning model of physical constitution classification. S3.4: The time-domain feature vector, frequency-domain feature vector, and nonlinear feature vector are concatenated to generate a unified multidimensional feature vector to characterize the comprehensive physiological response characteristics of the subject in the current state; The time-domain feature vector, frequency-domain feature vector, and nonlinear feature vector obtained through steps S3.1, S3.2, and S3.3 are then processed using a feature vector normalization localization algorithm (parameter: normalization interval). This enables scale matching of all feature components under a unified dimension. Furthermore, the feature type identifier encoding method (parameter: time-domain feature label value) is used. The frequency domain feature label value is The nonlinear feature label value is The three sets of feature vectors are labeled to retain the feature source information, which facilitates subsequent causal path backtracking analysis. Furthermore, a vector concatenation operator is used to cascade the three types of features. Let the time-domain feature vector be... The frequency domain eigenvector is The nonlinear eigenvector is The following connection formula generates a multidimensional feature vector. : in, This indicates a vector concatenation operation, with the output dimension being the sum of the lengths of the three types of features; Furthermore, to address the distribution structure of the concatenated high-dimensional vector, a covariance matrix sparsity constraint method is employed (parameters: L1 norm regularization coefficient is...). ), optimize the correlation structure between different feature components, and obtain the sparsified coefficient matrix. ; Furthermore, the principal component retention ratio threshold is set (parameter: cumulative contribution rate ≥ The optimized feature matrix is ​​subjected to principal component transformation to generate a multidimensional comprehensive feature vector that retains the main information under low redundancy conditions. ; By using this feature splicing and redundancy reduction method, the representations of different time-frequency scales and dynamic complexity are unified into a multi-dimensional feature space, realizing a feature representation that comprehensively reflects the subject's current physiological response pattern; For example, in a 12-lead ECG (time-domain feature dimension) 8-channel electromyography (frequency domain characteristic dimension) ), 32-channel EEG and 2-channel skin conductance response (nonlinear characteristic dimension) In the data collection, the first step is to perform Min-Max normalization on each group of features. Set the range and add a type label. Then press The combined generation length is The initial multidimensional feature vectors are obtained. A covariance matrix is ​​constructed from the concatenated result, and L1 regularization sparsity is applied. Reduce the proportion of redundant related terms to Below %. When the cumulative contribution rate reaches Under the condition of % selection Each principal component generates a simplified yet informative multidimensional comprehensive feature vector. This vector significantly improves classification stability and computational efficiency in subsequent causal modeling and individualized dynamic map adjustment; S3.5: The multidimensional feature vector is fused with individual physiological parameters, including age, gender, basal metabolic rate and historical body type information, to construct a feature fusion matrix, so as to enhance the model's adaptability to individual differences and the interpretability of the identification results.

[0014] Step S4: Construct an initial causal graph based on population sample data, use a PC algorithm and a causal discovery model to identify causal relationships between feature nodes, and define the causal path between constitution type and feature factors. For example... Figure 3 As shown, it specifically includes: S4.1: Standardize the multidimensional feature vectors of the group samples and use the Z-score normalization method to eliminate the dimensional differences between different modal signals in order to obtain a feature dataset under a uniform scale, which serves as the input to the causal discovery model; The feature fusion matrix after the S3.5 step is standardized to eliminate the differences in the units and numerical ranges of different modal features, ensuring that the causal discovery model receives data input of a uniform scale. The Z-score normalization method (parameters: based on the mean and standard deviation of all population samples) is used to achieve centering and variance normalization of each feature column in the feature fusion matrix; Furthermore, the standardized value of each feature component is calculated using the following formula. : in, For the first The first sample The original values ​​of each feature, For the first The mean of the feature in the population sample. Its standard deviation; Furthermore, to avoid extreme values ​​causing bias in the calculation of the mean and standard deviation, the eigenvalues ​​are truncated and corrected using the three-standard-deviation principle. Values ​​outside the range are replaced with truncated boundary values ​​to stabilize the standardization effect. Furthermore, to address the high dimensionality of multimodal features, a block-based Z-score strategy is used to normalize the feature fusion matrix according to the signal source category (such as ECG, EMG, EEG, ESC, and physiological parameters) to preserve the relative proportions of features within the modality. By performing numerical precision quantization on the standardized feature matrix (parameter: retain 6 decimal places), the consistency and repeatability of the input data for subsequent PC algorithms are ensured. Through the Z-score standardization and block processing steps described above, the feature fusion matrix is ​​transformed into a standardized feature dataset with a unified scale and outlier correction, thereby achieving the expected technical effect of eliminating dimensional differences and improving the stability of causal relationship inference.

[0015] For example, in a sample set containing 500 subjects, the multidimensional feature fusion matrix has a dimension of 500×208, with 48 columns for electrocardiogram (ECG) features, 64 columns for electromyography (EMG) features, and 96 columns for nonlinear features related to electroencephalography (EEG) and electrodermal conductance (EDC). The mean value of the ECG feature in column 12 was calculated. for Standard deviation for its original value Standardized results of sample calculation: When performing a three-standard-deviation truncation, if the standardized value of a certain EEG alpha wave energy ratio characteristic is... It is then truncated to the upper limit. To prevent extreme values ​​from interfering with causal relationship modeling, the modal block Z-score method was used to normalize the ECG, EMG, EEG, and SCEA features and physiological parameters respectively. The mean of all feature columns then approximates... The standard deviation is close to After numerical quantization and retention to 6 decimal places, a standardized feature data matrix of 500×208 was obtained. It was verified that it improved the robustness of edge weight estimation in the causal structure learning process of PC algorithm and reduced the spurious causal connection rate caused by heteroscedasticity of feature distribution. S4.2: Based on the normalized feature dataset, the PC algorithm is applied to perform causal structure learning, identify undirected edges and directional causal relationships between feature nodes, and generate a preliminary causal graph structure. Based on the uniformly scaled and standardized feature dataset obtained through step S4.1, the PC (Peter-Clark) algorithm (parameter: significance level) is used. The independence test method is the Pearson correlation coefficient test) to realize causal structure learning, so as to identify the possible conditional independence relationship between feature nodes; Furthermore, feature nodes are paired to form variable pairs using the conditional independence test function, and the conditional correlation coefficient is calculated at a fixed significance level. And its corresponding p value, to determine whether to remove the undirected connection edge between the variable pair; Furthermore, based on the conditional independence test rules, for variable pairs with p-values ​​greater than the significance level, an edge deletion operation is performed to obtain the undirected skeleton graph structure. The skeleton diagram only retains the connection relationships of feature nodes that are statistically correlated; Furthermore, based on the skeleton graph, the V-structure identification criterion is applied during the edge orientation stage of the PC algorithm, i.e., when a node... and All with nodes Connected, but and When there is no boundary between them, if Not here and In the conditionally independent set, then... → and → The direction is determined as a directed edge; Furthermore, through a rule propagation process (parameter: looping until no new directions are generated), directions are sequentially assigned to the remaining undirected edges using rules to prevent the formation of directed cycles, thereby generating a preliminary causal graph structure that satisfies the DAG (Directed Acyclic Graph) constraints. ; By using the structural constraints and directional induction of the PC algorithm, the statistically significant dependencies in the standardized feature dataset are mapped to the preliminary causal graph node connections and directional definitions, thus realizing the first construction of the causal topology between feature nodes; For example, in containing Number of subjects, feature dimension: In the standardized feature dataset, the PC algorithm performs a process on each feature pair. Calculate the Pearson conditional correlation coefficient if, given a set of condition variables... Its p-value is greater than Then determine and Conditionally independent, remove connected edges. For example, for ECG HRV feature nodes. Characteristic nodes of alpha wave energy ratio in EEG Given electromyographic (MF) characteristics The conditional correlation coefficient under the given conditions is p value Therefore, the undirected edge between the two is deleted. Finally, the skeleton graph is generated after V-structure determination and direction propagation. , containing approximately The total number of directed and undirected edges decreases The verification results show that the causal structure captures the stable dependencies between features in the population data, providing an accurate structural basis for subsequent identification of confounding factors and construction of label-guided causal graphs; S4.3: Introduce causal discovery models such as the FCI (Fast Causal Inference) algorithm to model intervention variables and identify potential confounding factors in the preliminary causal graph, so as to optimize the causal path direction and connectivity between feature nodes; Based on the preliminary cause-effect graph structure generated in step S4.2 The FCI (Fast Causal Inference) algorithm was used (parameter: significance level). The conditional independence test method is Fisher's Z test. An intervention variable modeling mechanism is introduced to identify and correct potential confounding factors. Through the latent variable hypothesis expansion phase, the FCI algorithm is used to... Perform a conditional independence test on each pair of related nodes, setting the size of the condition set to no more than [value missing]. In a given set of conditions Calculate Fisher's Z statistic ; Furthermore, based on The corresponding p-value is calculated. If the p-value is greater than the significance level and there is an unexplained independence relationship, a double-headed latent confounder edge is introduced between the corresponding node pairs to identify the possible influence path of the hidden variable. Furthermore, the Possible-D-SEP set enumeration method in the backward stage is adopted to generate a separation set that each node pair may depend on, and an iterative independence test is performed based on the set to correct the erroneous directed edge directions introduced by confounding or measurement errors. Furthermore, by using the edge constraint propagation rules (R1-R10 rule set) to redirect some edges in the case of introducing potential confounding labels, and introducing Partial Ancestral Graph (PAG) representation for edges with uncertain causal directions, the logical consistency of the final causal structure is guaranteed under the condition of the existence of latent variables. Through the latent confounding identification and intervention variable simulation processing using the FCI algorithm, The structure is optimized to include a PAG structure with potential hybrid identifiers and some directed edges, enabling precise correction of causal path direction and connection relationships; For example, in a population sample containing 500 subjects and 208 features, a significance level is set. The maximum size of the condition set is During the FCI latent variable hypothesis expansion phase, for ECG HRV characteristic nodes... Characteristic nodes of the energy ratio of EEG beta waves In the condition set {electromyography MF} Calculate under {age, gender} = Substitute into the formula to calculate = The corresponding p value is The condition for removing directed edges to introduce potential confounding labels was met. Subsequently, the Possible-D-SEP process identified unobserved confounding effects of this node, and a double-arrow symbol was assigned to the PAG to correct the edge direction. After a full FCI execution, 37 new potentially confounding labeled edges were added to the causal graph, and the directed edge adjustment ratio was 12%. This verified that the path direction error rate was reduced to less than 3% when including hidden variables, providing a more stable and interpretable structural foundation for the subsequent label-guided causal model construction. S4.4: Using the constitution type label as the top-level node of the causal graph, and combining the feature factor node and the causal path, a label-guided causal graph model is constructed to clarify the causal mapping relationship between constitution classification and bioelectric signal characteristics. S4.5: Based on the constructed causal graph model, extract the key causal paths and node influence weights to generate the structural parameters of the initial causal graph, providing a benchmark reference for subsequent individualized dynamic graph adjustments.

[0016] Step S5: Individual physiological parameters and historical physical condition information are input into the causal graph as dynamic adjustment factors to adjust node weights and path connection structures, generating an individualized dynamic causal graph. Specifically, this includes: S5.1: Perform feature encoding processing on individual physiological parameters and historical physical condition information to obtain an individualized regulatory factor vector; S5.2: Calculate the initial weight adjustment coefficients of each causal graph node based on the individualized adjustment factor vector, and use a nonlinear activation function to map the importance of the nodes to reflect the degree of influence of individual differences on the feature nodes; Based on the individualized adjustment factor vector obtained in step S5.1, a node weight initialization mapping method is adopted (parameter: total number of nodes). The initial weight matrix is ​​the node weight baseline matrix of the population causal graph, which realizes the mapping operation of the adjustment factor vector and the initial weight matrix in a one-to-one correspondence. Furthermore, the initial weight adjustment coefficient for each node is calculated using the weighted linear transformation formula. : in, To adjust the gain coefficient of the mapping factor, The first adjustment factor vector One portion, For bias terms; Furthermore, the calculated adjustment coefficients are input into the nonlinear activation function. In the middle (parameter: activation function type is Sigmoid or Tanh, selected according to the role of the node in the causal inference path), a nonlinear mapping of node importance is implemented: in, For the adjusted number Each node's weight; Furthermore, in the weight normalization stage, vector normalization operation (parameter: L2 norm) is used to normalize the adjusted weights of all nodes to eliminate the influence of scale differences in the adjustment factors of different individuals; Furthermore, the normalized node weight vector is written into the node attribute field of the dynamic causal graph to serve as the prior weight basis for the dynamic adjustment of path connections in S5.3. By using nonlinear mapping and normalization, the high-dimensional information of individual modulatory factors is transformed into standardized and discriminative node weight data, thereby achieving the quantification effect of node importance that reflects individual differences. For example, in a containing In the initial causal graph of feature nodes, the length of the individualized adjustment factor vector is... A certain node The adjustment factor component value ,Pick = , = Then calculate the adjustment factor: The adjusted weights are calculated using the Sigmoid activation function: After all node weights are calculated, L2 normalization is performed, assuming the normalization factor is 1. Then the final weight of the node is: The results were written into the dynamic causal graph to verify that the probability of path retention of this node decreased due to its low weight during subsequent dynamic pruning of path connections, thus reflecting its secondary role in individual causal reasoning and improving the model's explanatory accuracy in personalized constitution identification tasks. S5.3: Perform dynamic pruning and enhancement operations on the path connection structure in the initial causal graph, and execute the graph attention mechanism based on the individualized adjustment factor vector to generate causal subgraph connection relationships adapted to the current individual; Based on the adjusted node weight vector obtained in step S5.2, it is used as a prior importance parameter for dynamic optimization of path connectivity and input into the edge attribute table of the initial causal graph. A dynamic pruning algorithm is then employed (parameter: weight threshold). (Adaptive calculation based on the mean and standard deviation of node weight distribution) is used to remove low-contribution edges to reduce the interference of redundant connections on the inference path; Furthermore, by calculating the comprehensive retention score for each edge... Taking into account both the importance weights of the nodes at both ends of the edge and the historical salience of the edge in the group causal graph, the formula is as follows: in, and These are the adjusted weights of the nodes at both ends of the edge. The historical significance coefficient of this edge; Furthermore, if Below the threshold Perform edge pruning to remove irrelevant paths from the current individualized subgraph structure to reduce computational overhead. Furthermore, through path enhancement algorithms (parameter: enhancement ratio) Based on the quantile of the target node weight distribution, high-value edges are enhanced, i.e., their connection strength is increased and their priority weight in graph reasoning is improved. The formula for increasing connection strength is: in The original edge connection strength; Furthermore, a graph attention mechanism is implemented based on individualized adjustment factor vectors (parameter: the attention scoring function adopts the scaled dot-product form), assigning dynamic attention coefficients to the retained edges. The specific calculation formula is as follows: in, and These are the query vector and key vector, respectively, encoded based on node weights. For vector dimensions; Furthermore, the dynamic attention coefficient is written into the edge attribute weight field, and the priority of causal path selection is reordered according to the coefficient, finally generating a causal subgraph connection relationship that adapts to the current individual state; By combining dynamic pruning and enhanced graph attention mechanism, the connection structure of the group causal graph is adjusted into a sparse and efficient subgraph that reflects individual specificity, thus achieving the expected technical effect of reducing irrelevant reasoning paths and highlighting key causal flows. For example, in a containing Feature nodes, In the initial causal graph of the edge, the range of node weights obtained by mapping the individualized adjustment factor vector is: The weighted mean is The standard deviation is Adaptive threshold = When calculating the overall retention score, for example, if the weights of the nodes at both ends of an edge are respectively... and The historical significance coefficient is The overall retention score is then calculated as follows: If the value is higher than the threshold, it is retained and its connection strength is multiplied during the path enhancement phase. During the graph attention allocation phase, the inner product of the query vector and the key vector generated by the weight encoding of the node corresponding to this edge is... Dimension = The attention coefficient is then calculated as follows: After full graph calculation and sorting, the proportion of edges retained is reduced to %, average path length shortened %, Individualized reasoning computation time reduced This percentage verifies the effectiveness of this step in improving reasoning efficiency and maintaining key causal relationships; S5.4: Combining the adjusted node weights with the optimized path connection structure, construct an individualized dynamic causal graph model to form a causal reasoning framework with individual specificity; S5.5: Perform topological structure verification and causal consistency verification on the generated individualized dynamic causal graph to ensure that the reasoning path is logically consistent with the TCM constitution classification standard.

[0017] Step S6: Input the current fused feature matrix into the dynamic causal graph, activate the matching causal subgraph path, perform causal inference calculation, and output the constitution classification result and the corresponding causal inference path explanation. Specifically, this includes: S6.1: Input the feature fusion matrix into the individualized dynamic causal graph, and calculate the path activation degree of the current feature vector in the causal graph based on the causal weights and connections between nodes in the graph to obtain the initial causal path activation distribution; S6.2: The Top-K path selection algorithm is used to sort the initial causal path activation distribution and select the K causal paths with the highest activation as the candidate inference path set to reduce computational complexity and focus on key causal logic; For the initial causal path activation distribution data matrix calculated by step S6.1, the Top-K path filtering algorithm (parameter: K value is set based on the ranking quantile of the initial distribution activation) is used to sort all candidate paths by importance, thereby filtering the candidate path set; Based on the initial path activation distribution vector Calculate the activation ranking position of each path using the corresponding path index, and generate a sorted index matrix using a descending sorting algorithm. And mark the index numbers of the first K paths; Furthermore, through the index matrix Obtain the corresponding activation value vector Activation threshold is used (Parameters: adaptively set based on the mean and standard deviation of the activation distribution, formula) ,in The mean, Standard deviation, (As an adjustment coefficient) the first K paths are filtered again to remove redundant paths below the threshold to ensure the effectiveness of the candidate set; Furthermore, for the remaining highly activated paths, their normalized importance scores are calculated. This ensures the comparability and normalization of scores across different paths. Furthermore, the candidate inference path set structure (including: path sequence, node sequence, edge connection relationship, normalized score) is output to the S6.3 causal effect estimation module to achieve seamless connection of data stream; By combining Top-K screening with two-level filtering based on activation threshold and score normalization, the global initial activation distribution result calculated by S6.1 is transformed into a set of high-value candidate paths that focus on key causal logic, thereby achieving the expected technical effect of reducing the computational complexity of subsequent causal effect estimation and improving the focus of reasoning. For example, in dynamic causal graph inference computation, it is assumed that the initial total number of paths is... The initial activation distribution has a mean of [value]. The standard deviation is Let the adjustment coefficient be... = Then the activation threshold is calculated. = Sort the paths in descending order of activation and select the top K= Path, remove those with activation levels lower than The remaining path A high-value path. For activation level... The normalized score of path p is calculated as follows: assuming the total activation of the candidate set is... ,but = ≈ Verification showed that the selected candidate set covered over 95% of the causal explanation feature nodes, reducing subsequent inference time by approximately 64% compared to the full path, while improving causal classification accuracy by approximately 3.7%. S6.3: Perform causal effect estimation for each path in the candidate causal path set, and calculate the causal contribution of each path to the body constitution classification result based on do-calculus to quantify the influence weight of different paths on the final identification result; S6.4: A weighted causal path fusion strategy is adopted, which combines the causal contribution of each path to perform a weighted summation of the path outputs and generate a comprehensive causal reasoning result to improve the robustness and interpretability of the identification results. S6.5: Generate a constitution classification decision based on the comprehensive causal reasoning results, and trace the key feature nodes and causal relationships in the activation path in reverse, and output the corresponding causal reasoning path explanation to enhance the interpretability of the identification process.

[0018] Step S7: Determine whether the current constitution classification result meets the confidence threshold. If it does, output the final identification result and causal path explanation; otherwise, trigger the graph structure fine-tuning mechanism to optimize the inference path. Specifically, this includes: S7.1: Calculate the confidence level of the physical constitution classification results output by the current dynamic causal graph, and generate a confidence index based on the classification probability distribution and causal path weights to quantify the credibility of the current identification results; S7.2: Compare the generated confidence index with the preset confidence threshold to determine whether the current identification result meets the output standard. If it does, proceed to the result output process; otherwise, trigger the map structure fine-tuning mechanism. S7.3: When the confidence level does not reach the threshold, based on the residual analysis of the current fused feature matrix and the causal inference path, identify nodes and edge structures with high uncertainty in the causal path and generate a list of graph fine-tuning suggestions; Based on the condition that the confidence level in step S7 does not reach the threshold, the comprehensive causal inference result output from S6.5 and the original fused feature matrix are used as input data to establish a residual signal calculation model to quantify the degree of deviation between the predicted value and the expected label. A residual vector generation method (parameter: residual calculation formula based on the standard Euclidean distance definition) is used to calculate the absolute error value of each physical constitution classification output, forming a node-level residual distribution matrix. Furthermore, a path residual aggregation algorithm is used (parameter: aggregation function type is weighted average, weights are determined by the importance scores of each node in the path). (Assignment), calculate the mean residual of each causal reasoning path, and obtain the path-level uncertainty index. The polymerization process satisfies the following formula: in Let be the absolute value of the residual at the i-th node in the path. Score the importance of this node; Furthermore, a node uncertainty scoring method is employed (parameter: the scoring function is the product of the absolute value of the node residual and its connectivity in the global causal graph) to rank the nodes based on their sensitivity in uncertainty analysis and output a node priority list. This scoring is defined as: in Let n be the absolute value of the residual at node n. Let n be the degree centrality of node n; Furthermore, by using a structural uncertainty screening algorithm (parameter: the edge weight change threshold is adaptively set based on the historical mean ± 2σ), the weight fluctuation of edges in the path is detected, edge structures with high instability are marked, and cross-correlation analysis is performed with the list of high uncertainty nodes to screen the candidate set of subgraph structures that urgently need optimization; Through the multi-level chain diagnostic process of residual calculation, path aggregation, node scoring and edge volatility detection, the results are transformed into a graph fine-tuning suggestion list containing high-priority nodes, edges and their causal path numbers, to achieve the expected technical effect of targeted structural optimization. For example, in a dynamic causal graph inference, the dimension of the fused feature matrix is... The output constitution classification is "damp-heat constitution" with a confidence score of 0.72, which is below the threshold of 0.85. Using Euclidean distance residual calculation, the absolute difference between the predicted value (0.65) and the label value (1.0) of a certain node n is 0.35, and its importance score is 0.08. Therefore, the contribution of this node in the path residual aggregation calculation is... After aggregation, the path p of a certain path = The value exceeds the uncertainty threshold of 0.1 and is therefore marked as an abnormal path. In the node uncertainty score, the degree centrality of this node is 5, therefore... = The node is prioritized into the top 5. In edge volatility detection, the standard deviation of the edge weight between node n and its neighbor m is 0.043, which is higher than the set threshold of 0.04, and is therefore marked as an unstable edge. The final fine-tuning suggestion list contains three items: the node, its neighboring edges, and their path numbers, which are used as input to the S7.4 local causality discovery optimization module. After executing this fine-tuning suggestion list, the classification confidence score for the next inference increases to 0.88, significantly enhancing the model's adaptive capability. S7.4: Based on the graph fine-tuning suggestion list, the local causal discovery algorithm is used to reconstruct the selected subgraph structure in the dynamic causal graph, adjust the causal strength and connection relationship between nodes, and generate an optimized causal inference path; S7.5: The reconstructed dynamic causal graph is re-input into the causal reasoning process, and the reasoning calculation is executed again to obtain the updated constitution classification results and causal path explanation, thereby improving the accuracy of the identification results and the model's adaptability.

[0019] Step S8: Generate a visual report based on the identification results and causal path explanation, and provide identification evidence and constitution conditioning suggestions through an interactive interface. Specifically, this includes: S8.1: Perform structured analysis on the constitution classification results and causal path explanation information output by the causal reasoning model to generate a renderable visual data model; S8.2: Based on a structured visualization data model, a dynamic causal path graph interface is built using the D3.js graphics library to display key feature nodes and their causal flow relationships; S8.3: Semantically map the results of the physical constitution classification with the key influencing factors in the causal path to generate a natural language version of the identification basis explanation text; Using the structured visualization data model and dynamic causal path graph interface generated by S8.1 and S8.2 as input objects, a semantic mapping algorithm (parameter: domain knowledge base is a pre-built TCM constitution characteristics-representation mapping dictionary) is adopted to realize the semantic binding between constitution classification results and key influencing factors in causal paths; Furthermore, by using the feature label parsing method (parameter: label parsing rules are based on multimodal signal feature coding structure), the physiological meaning of each key factor and its correlation strength with body type are extracted to obtain a factor-meaning comparison table; Furthermore, a natural language template generation algorithm is used (parameter: the template syntax rules include a set of TCM constitution determination terms and causal relationship description phrases) to embed the constitution classification labels and entries in the factor-meaning lookup table into the template in causal path order to generate basic identification basis sentence segments; Furthermore, a weight-driven language ranking and merging algorithm is used (parameters: ranking weight is the causal contribution, merging rule is the merging threshold of adjacent low contribution factors). The basic identification criteria sentences and paragraphs are prioritized and information is merged to ensure that the explanatory text is concise and highlights the key points. Furthermore, a grammatical correctness and fluency optimization module (parameter: syntactic adjustment strategy based on BiLSTM language model) is used to automatically correct and optimize the sorted sentence segments, generating natural language explanation text that conforms to the logic of medical expression; Through the above semantic mapping, template filling, importance ranking and syntactic optimization processing methods, the physical constitution classification labels and causal path data of the previous step are transformed into identification basis explanatory text in natural language form, so as to achieve the expected technical effect of human-computer readability and clinical interpretability of the output results. For example, in a damp-heat constitution identification task, the constitution classification result is damp-heat constitution. The top three key factors with the highest activation in the causal path are high β wave power ratio (contribution). Increased frequency of skin conductance response (contribution) ) and the increase in the low-frequency / high-frequency HRV ratio (contribution) In this task, the domain knowledge base maps a high β-wave power ratio to "overexcitation of the cerebral cortex," an increased skin conductance frequency to "enhanced sympathetic activity," and an increased HRV low-frequency / high-frequency ratio to "autonomic nervous system balance with sympathetic inhibition." The natural language template is defined as "Factor A (medical meaning X) was detected, indicating characteristic Y, which matches constitution characteristic Z." After template filling, the initial sentence segments are formed: ① A high β-wave power ratio (overexcitation of the cerebral cortex) indicates an active nervous system, consistent with the characteristics of a damp-heat constitution; ② An increased skin conductance frequency (enhanced sympathetic activity) suggests the body is in a state of high stress response; ③ An increased HRV low-frequency / high-frequency ratio (autonomic nervous system balance with sympathetic inhibition) indicates that cardiovascular regulation is biased towards sympathetic dominance. After sorting by causal contribution weight, the sentence segment order is adjusted to ①②③, and sentences with contributions lower than [a certain value] are merged. The factor description was performed. After correction by the BiLSTM language model, the final generated identification basis description text was: "Comprehensive analysis shows that the subject's cerebral cortex activity level was significantly increased, the sympathetic nervous system was in a state of sustained high activation, and the cardiovascular autonomic regulation function showed sympathetic dominance. The overall characteristics are consistent with the damp-heat constitution." This example verifies that chain semantic mapping and natural language generation strategies can significantly improve the professionalism and readability of the output content while maintaining the accuracy of causal explanation. S8.4: Based on the body constitution type and the TCM conditioning guidelines knowledge base, match the corresponding lifestyle suggestions and intervention plans to generate a personalized body constitution conditioning suggestion report; S8.5: Integrates visualized causal path maps and treatment suggestion reports through web or mobile interactive interfaces, providing users with interactive parameter adjustment and path tracing functions.

[0020] Step S9: Periodically collect new sample data and update the population causal graph, and use an online learning mechanism to optimize the generalization ability and adaptability of the dynamic causal inference model. Specifically, this includes: S9.1: Preprocess the newly acquired multimodal bioelectric signal samples by performing wavelet transform denoising and Z-score normalization to eliminate environmental interference and individual baseline differences, and obtain standardized multimodal signal data. S9.2: Extract time-domain, frequency-domain, and nonlinear features from standardized multimodal signal data, generate multidimensional feature vectors, and construct a feature fusion matrix by combining individual physiological parameters and historical physical labels, which serves as the input dataset for causal graph updates; S9.3: The PC algorithm and causal discovery model are used to identify causal relationships in the newly added feature fusion matrix, extract the causal paths between feature nodes in the newly added samples, and generate local causal subgraphs as the basis for expanding the population causal graph. S9.4: Merge the local causal subgraph with the existing group causal graph, perform graph matching and conflict resolution to maintain the consistency and logical integrity of causal paths, and obtain the updated group causal graph; S9.5: Based on the updated population causal graph, an online learning mechanism is implemented to adjust the graph node weights and path connection probabilities, optimize the generalization ability and inference path adaptability of the dynamic causal inference model, and improve the model's recognition accuracy and stability on new samples.

[0021] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0022] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying traditional Chinese medicine constitution based on multimodal bioelectrical signal sensing, characterized in that, Includes the following steps: S1: Collect multimodal bioelectrical signals from the subjects, including electrocardiogram signals, electromyogram signals, electroencephalogram signals, and skin conductance response signals, and simultaneously record individual physiological parameters and historical body type labels; S2: Preprocess the collected multimodal bioelectric signals to generate preprocessed signals; S3: Based on the preprocessed signal, extract time-domain, frequency-domain, and nonlinear features, generate multi-dimensional feature vectors, and construct a feature fusion matrix by combining individual physiological parameters; S4: Construct an initial causal graph based on population sample data, use PC algorithm and causal discovery model to identify causal relationships between feature nodes, and define the causal path between body type and feature factors; S5: Input individual physiological parameters and historical physical condition information as dynamic adjustment factors into the initial causal graph, adjust node weights and path connection structure, and generate an individualized dynamic causal graph; S6: Input the current fused feature matrix into the individualized dynamic causal graph, activate the matching causal subgraph path, perform causal inference calculation, and output the constitution classification result and the corresponding causal inference path explanation; S7: Determine whether the current constitution classification result meets the confidence threshold. If it does, output the final identification result and causal path explanation. Otherwise, trigger the graph structure fine-tuning mechanism. S8: Generates a visual report based on the identification results and causal path explanation, and provides identification basis and body conditioning suggestions through an interactive interface.

2. The method for identifying and warning of hazardous sources at construction sites according to claim 1, characterized in that, Following step S8, the following is also included: S9: Regularly collect new sample data and update the population causal graph, and use an online learning mechanism to optimize the generalization ability and adaptability of the dynamic causal reasoning model.

3. The method for identifying traditional Chinese medicine constitution based on multimodal bioelectrical signal perception according to claim 1, characterized in that, Step S1 specifically includes: Multimodal bioelectrical signals of the subjects are acquired synchronously using a multi-channel biosignal acquisition device. These signals include electrocardiogram (ECG), electromyography (EMG), electroencephalography (EEG), and electrodermal response (TEG) signals, thereby obtaining raw physiological data reflecting the autonomic nervous system and physical state. R-wave detection and RR interval calculation were performed on the acquired electrocardiogram signals to identify QRS complexes and extract heart rate variability indices. Time-domain envelope extraction and spectrum analysis were performed on the electromyographic signal to obtain the signal envelope amplitude, and the median frequency and average power frequency were calculated. Perform frequency band energy analysis on EEG signals and calculate the energy proportion of each frequency band; Differential processing and peak detection were performed on the skin conductance response (SCR) signal to extract the amplitude and frequency characteristics of the SCR. Simultaneously collect individual physiological parameters of subjects to construct personalized modeling input features; Record the subjects' historical physical type labels as a supervisory signal for training causal reasoning paths.

4. The method for TCM constitution identification based on multimodal bioelectrical signal sensing according to claim 3, characterized in that, In step S1, the sampling rate of the multi-channel biosignal acquisition device is 1000Hz, the resolution is 24bit, the channel input impedance is greater than 10MΩ, and the synchronous acquisition of each signal channel is achieved by high-precision clock driving. The cutoff frequencies of the pre-amplifier analog filter include 0.5-40Hz for ECG channel bandpass, 20-450Hz for EMG channel bandpass, 0.1-100Hz for EEG channel bandpass, and 5Hz for ESK channel low-pass.

5. The method for identifying traditional Chinese medicine constitution based on multimodal bioelectrical signal sensing according to claim 1, characterized in that, Step S2 specifically includes: The collected raw electrocardiogram (ECG), electromyogram (EMG), electroencephalogram (EEG), and electrodermal response (EDS) signals were processed by channel alignment and temporal synchronization to obtain synchronized bioelectrical signal sequences. Each signal channel in the synchronized bioelectric signal sequence is decomposed into multiple scales, and wavelet coefficients at each scale are extracted. Threshold quantization is performed on the extracted wavelet coefficients, and high-frequency coefficients are compressed to obtain the denoised reconstructed signal. Normalization processing is performed on the denoised reconstructed signal, and the standardized value is calculated based on the historical mean and standard deviation of each channel signal to obtain the standardized signal matrix; The standardized signal matrix is ​​subjected to quality assessment, the signal-to-noise ratio and signal integrity index are calculated, and abnormal signal channels are marked. If a channel has substandard quality, a resampling or signal compensation mechanism is triggered.

6. The method for TCM constitution identification based on multimodal bioelectrical signal sensing according to claim 5, characterized in that, In step S2, the channel alignment in signal preprocessing adopts timestamp parsing and cross-channel cross-correlation coarse alignment combined with dynamic time warping fine adjustment, and the amplitude is normalized to the [-1,1] interval.

7. The method for identifying traditional Chinese medicine constitution based on multimodal bioelectrical signal sensing according to claim 1, characterized in that, Step S3 specifically includes: Temporal feature extraction was performed on the preprocessed electrocardiogram (ECG), electromyography (EMG), electroencephalogram (EEG), and electrodermal response (EDS) signals to obtain the temporal feature vectors of each signal. Fast Fourier Transform (FFT) processing is performed on the electrocardiogram (ECG), electromyogram (EMG), electroencephalogram (EEG), and electrodermal response (EDS) signals to extract frequency domain features and generate frequency domain feature vectors. Based on the time series data of the multimodal bioelectric signals, the sample entropy, approximate entropy and Lyapunov exponent are calculated, and nonlinear dynamic features are extracted to form a nonlinear feature vector. The time-domain feature vector, the frequency-domain feature vector, and the nonlinear feature vector are concatenated to generate a multidimensional feature vector; The multidimensional feature vectors are fused with individual physiological parameters to construct a feature fusion matrix.

8. The method for identifying traditional Chinese medicine constitution based on multimodal bioelectrical signal sensing according to claim 1, characterized in that, Step S4 specifically includes: The multidimensional feature vectors of the group samples are standardized to obtain a normalized feature dataset. Based on the normalized feature dataset, causal structure learning is performed to identify undirected edges and directional causal relationships between each feature node, generating a preliminary causal graph. A causal discovery model is introduced to model intervention variables and identify potential confounding factors in the preliminary causal diagram; By using the body type label as the top-level node of the causal graph and combining it with feature factor nodes and causal paths, a label-guided causal graph model is constructed. Based on the causal graph model, key causal paths and node influence weights are extracted to generate the structural parameters of the initial causal graph.

9. A method for identifying traditional Chinese medicine constitution based on multimodal bioelectrical signal sensing according to claim 8, characterized in that, In step S4, the causal graph construction process includes Z-score normalization of the feature fusion matrix, using the PC algorithm to complete the skeleton graph of statistical correlation relationships, and then combining the FCI algorithm to identify potential confounding factors and adjust the direction of some edges. The constitution type label is set as the top-level output node of the causal graph.

10. The method for TCM constitution identification based on multimodal bioelectrical signal sensing according to claim 1, characterized in that, Step S5 specifically includes: Individual physiological parameters and historical physical information are processed by feature encoding to obtain individualized regulatory factor vectors; The initial weight adjustment coefficients of each causal graph node are calculated based on the individualized adjustment factor vector, and the node importance is mapped using a nonlinear activation function. Dynamic pruning and enhancement operations are performed on the path connection structure in the initial causal graph, and a graph attention mechanism is executed based on the individualized adjustment factor vector to generate causal subgraph connection relationships; By combining the adjusted node weights with the optimized path connection structure, an individualized dynamic causal graph model is constructed. The generated individualized dynamic causal graph is subjected to topological structure verification and causal consistency verification.

Images