A Deep Learning-Based Syndrome Quantification Health Assessment Model and its Construction Method

Abstract

This invention discloses a deep learning-based syndrome quantification health assessment model, namely the Deep TCM-XGB fusion model and its construction method. This method encodes the Five Elements theory, the meridian-exterior relationship, and the holistic view of "three causes and one condition" in Traditional Chinese Medicine (TCM) into a structured prior knowledge matrix. It embeds this matrix into a Transformer architecture through an adaptive soft-gated attention mechanism and combines multimodal spatiotemporal feature fusion and Mixup data augmentation to achieve collaborative analysis of multi-source information from meridian electrical signals, spatiotemporal environment, and demographic characteristics. The model outputs an integrated diagnostic result of syndrome classification and severity quantification. By constructing a meridian association network and calculating the strength of associations between meridians based on attention weights, it presents meridian synergy, constraint, and exterior-exterior relationship patterns, providing interpretable evidence consistent with TCM theory for the model's diagnostic decisions. Experiments show that this invention possesses excellent generalization ability, robustness, and interpretability with limited clinical samples, providing an effective technical solution for promoting the development of TCM syndrome differentiation towards precision, standardization, and objectivity.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of TCM intelligent health assessment and artificial intelligence, specifically to a syndrome-based quantitative health assessment model and its construction method based on deep learning. Background Technology

[0002] Traditional Chinese medicine (TCM), as a unique system of systematic and empirical medicine, relies on the comprehensive collection of pathological information from patients through the four diagnostic methods of observation, auscultation and olfaction, inquiry, and palpation. In recent years, with the rapid development of biosensor technology and artificial intelligence, the objectification and digitization of TCM health assessment has become an irreversible trend. Among these methods, meridian detection, as an important branch of TCM objectification, assesses the state of the body's Qi and blood by measuring the bioelectrical resistance values ​​of acupoints, and has accumulated rich standardized clinical data. However, the essence of TCM theory lies in the holistic view of "harmony between man and nature" and "adapting to the three factors," emphasizing that the state of the body's meridians is not isolated but profoundly regulated by multidimensional spatiotemporal and individual factors. This regulatory mechanism is a complex nonlinear coupling system: in the temporal dimension, the Five Elements theory, with its core principles of mutual generation and restraint among the five elements (wood, fire, earth, metal, and water), combined with the theory of meridian flow corresponding to the time of day, systematically explains the mutual generation, restraint, and synergistic relationships among the twelve meridians of the human body based on the attributes of the Five Elements; in the spatial and individual dimensions, the dryness, dampness, cold, and heat of the regional environment, the Yin-Yang differences brought about by gender, and the waxing and waning of Qi and blood due to age all constitute key variables determining the state of the human body. Therefore, only by integrating microscopic meridian data with macroscopic Five Elements theory, demographic characteristics, and geographical environment can an individual's physical condition be truly assessed. This unique spatiotemporal coupling mechanism is a key feature that distinguishes Traditional Chinese Medicine (TCM) from modern medical health assessment systems and is also the foundation for achieving syndrome differentiation and treatment.

[0003] Current machine learning and deep learning methods applied to TCM-assisted assessment mainly face the following problems: First, the overfitting dilemma of "small-scale sample - high dimensionality". The acquisition cost of high-quality labeled medical data is extremely high, resulting in limited dataset size. Under these conditions, training deep neural networks makes it easy for models to learn random noise in the data rather than true pathological features, leading to poor generalization ability. Second, the lack of meridian topology. Most existing models treat the twelve meridians as independent statistical variables, completely ignoring core association rules in TCM theory such as "the lung and large intestine are internally and externally related" and "wood generates fire". This results in models only learning superficial statistical correlations and failing to establish a systematic understanding of meridian connections. Third, the black-box crisis of interpretability. In clinical decision-making, the basis for the model's health assessment is more important than the diagnosis itself. Existing end-to-end deep learning models lack transparent reasoning paths, making it impossible for doctors to determine whether the model's decisions are based on abnormalities in core meridians or irrelevant statistical pseudo-correlation. This seriously hinders the trust and application of intelligent health assessment systems in clinical practice.

[0004] Therefore, there is an urgent need to build a new type of intelligent health assessment model that can effectively integrate the core theories of traditional Chinese medicine as inductive bias into a deep learning architecture. This model can achieve accurate identification and quantification of syndromes while providing interpretable evidence that conforms to the logic of traditional Chinese medicine theory, thereby breaking through the performance bottleneck and trust barrier of existing pure data-driven methods in traditional Chinese medicine diagnosis and treatment scenarios. Summary of the Invention

[0005] This invention addresses the problems of model overfitting, lack of TCM topological correlation, and insufficient interpretability in existing technologies by proposing a syndrome-based quantitative health assessment model and construction method based on deep learning. By structuring and encoding TCM prior knowledge and embedding it into the core layer of the model, it achieves accurate, robust, and interpretable syndrome differentiation through the synergy of theoretical guidance and data-driven approaches.

[0006] One objective of this invention is to provide a deep learning-based syndrome-based health assessment model, characterized by the following modules: a multimodal data encoding module, used to map meridian features and spatiotemporal context features including demographic features, geographical environment, innate five elements, and meridian flow information to a unified high-dimensional latent space through independent linear projection layers, and embed learnable positional codes to form an input sequence; and a Deep TCM deep feature extraction module, which is based on a Transformer encoder and introduces an adaptive soft-gated multi-head attention mechanism, using learnable parameters... The weights between the TCM prior knowledge matrix and data-driven attention are dynamically adjusted to extract deep meridian features that integrate TCM theory. The XGBoost fusion decision module is used to concatenate the global fusion features output by the Deep TCM deep feature extraction module with the original meridian features and multimodal context features to generate enhanced features. Based on these enhanced features, the XGBoost multi-output classifier and regressor are trained. Finally, the prediction results of the regressor are equally weighted and fused with the regression prediction results of the Deep TCM module to obtain the final quantitative value of the severity of the syndrome.

[0007] The second objective of this invention is to provide a method for constructing a TCM a priori knowledge matrix. This matrix is ​​a 13×13 relational matrix containing one global context node and twelve meridian nodes. Its assignment rules are based on the core theoretical system of TCM, specifically including: classifying meridians according to the five elements (Wu Xing) attributes of the twelve primary meridians: Gallbladder and Liver meridians belong to Wood; Lung and Large Intestine meridians belong to Metal; Stomach and Spleen meridians belong to Earth; Bladder and Kidney meridians belong to Water; Heart, Pericardium, Triple Energizer, and Small Intestine meridians belong to Fire; assigning positive weights to meridian pairs with a generating relationship based on the five elements' mutual generation principle; assigning negative weights to meridian pairs with a restraining relationship based on the five elements' mutual restraint principle; assigning synergistic weights to meridians with the same five elements attributes; assigning association weights to meridian pairs with a "same-name meridian" relationship; and setting the connection weights between the global context node and all twelve meridian nodes to fixed values ​​to achieve effective interaction between global spatiotemporal features (such as innate endowment, seasonal timing, and environmental characteristics) and local meridian features, reflecting the holistic view of "harmony between man and nature" in TCM.

[0008] The third objective of this invention is to provide a computational method for an adaptive soft-gated multi-head attention mechanism. This mechanism improves upon the self-attention computation of the standard Transformer, characterized by using a learnable gating parameter. The above-mentioned matrix of a priori knowledge of traditional Chinese medicine Attention weights are introduced as a bias term in the calculation, and their calculation formula is as follows: ,in, For querying key-value matrices, For single-head feature dimensions, This is the aforementioned matrix of prior knowledge in Traditional Chinese Medicine. This is a learnable scalar initialized during model training. As independent parameters, they are co-optimized with network weights through standard backpropagation to dynamically balance data-driven feature associations with prior constraints guided by traditional Chinese medicine theory. This allows the model to extract real patterns from data with small samples while avoiding deviations from core medical principles.

[0009] Preferably, the gating parameters can be learned in the above method. Methods for explaining the evolutionary characteristics, learnable gating parameters The model exhibits a hierarchical adaptive evolution across its Transformer layers: it approaches zero in the bottom layer network to achieve unbiased data feature extraction, and gradually increases in the middle and upper layers to introduce and strengthen the macroscopic constraints of traditional Chinese medicine theory, thus forming a progressive feature learning mode that is "from bottom to top and from data to theory".

[0010] The fourth objective of this invention is to provide an internal construction method for an XGBoost fusion decision module, characterized by comprising: a feature enhancement unit, used to concatenate the global fusion features (48-dimensional) output by the Deep TCM deep feature extraction module, the original meridian feature data (48-dimensional), and the spatiotemporal context feature data (21-dimensional) to generate 117-dimensional enhanced features; a multi-output XGBoost classifier, used to process the enhanced features to generate classification results for 8 types of syndromes; a multi-output XGBoost regressor, used to process the enhanced features to generate severity quantification results for the 8 types of syndromes; and an equal-weight fusion unit, used to average the regression prediction values ​​of the Deep TCM module with the prediction results of the XGBoost regressor to obtain the final severity quantification value of the syndrome.

[0011] The fifth objective of this invention is to provide a method for constructing spatiotemporal contextual feature data in the aforementioned model, characterized by the following construction method: the subject's gender feature is binary encoded, and the age feature is numerically processed to form demographic features; the latitude and longitude coordinates of the subject's place of residence are encoded to form geospatial features; based on the Heavenly Stem of the subject's birth year, a uniquely hot encoded vector representing the subject's innate Five Elements attribute is generated; based on the subject's detection time, a uniquely hot encoded vector representing the meridian corresponding to the detection time is generated to reflect the meridian flow pattern.

[0012] The sixth objective of this invention is to provide a method for constructing and visualizing meridian association networks based on an attention mechanism, used for quantitative evaluation and intuitive presentation of meridian interaction relationships learned by the model. Its key feature is that a meridian association network is constructed using the twelve meridians as network nodes and the strength of the association between meridians as network edge weights; the calculation method for the strength of the association between meridians is as follows: under the condition of integrating prior knowledge of traditional Chinese medicine, the meridians... With meridians The strength of the correlation between The calculation formula is: Under purely data-driven conditions, meridians With meridians The strength of the correlation between The calculation formula is: Where, in the formula: , The first The query matrix and key matrix of each attention head; It is a single-head feature dimension; For the number of heads; For learnable adaptive soft gating parameters; This is a matrix of prior knowledge in Traditional Chinese Medicine (TCM). By visualizing the correlation strength under two conditions using heatmaps and fully connected topological network diagrams, it is possible to intuitively verify whether the model autonomously discovers meridian interaction patterns that are consistent with classical TCM theories (such as the exterior-interior relationship between the Bladder Meridian and Kidney Meridian, and the synergistic relationship between the Spleen Meridian and Stomach Meridian), thus providing objective evidence for the scientific interpretation of TCM theories.

[0013] The Deep TCM-XGB fusion model proposed in this invention systematically solves the key challenges of sample sparsity, topological missingness, and insufficient interpretability in TCM intelligent health assessment by encoding TCM theory into a prior knowledge matrix, designing an adaptive soft-gated attention mechanism, fusing multimodal spatiotemporal features, introducing Mixup data augmentation, and constructing an XGBoost joint decision module. Experiments demonstrate that this model comprehensively outperforms existing methods in syndrome classification and severity quantification tasks, and can autonomously learn and present meridian connections consistent with medical principles, providing a practical and feasible technical solution for the objectification and intelligentization of TCM syndromes. Attached Figure Description

[0014] Figure 1 This is the overall architecture diagram of the Deep TCM-XGB fusion model in Example 1; Figure 2 This is a schematic diagram illustrating the generating and restraining relationships of the Five Elements in Example 1 and their correspondence with the Twelve Cardinal Meridians; Figure 3 This is a graph showing the training loss versus performance variation of the model described in Example 1 during 5-fold cross-validation. Figure 4 This refers to the meridian interaction heatmap and fully connected topology network diagram learned by the model under the condition of integrating traditional Chinese medicine theory in Example 2. Figure 5 Example 2 shows the meridian interaction heatmap and fully connected topology network diagram learned by the model under purely data-driven conditions, without the removal of traditional Chinese medicine theory. Figure 6 Example 3 shows the adaptive soft gating parameters for each Transformer layer during model training. Evolutionary trajectory diagram. Detailed Implementation

[0015] The present invention will now be described in detail with reference to specific implementation examples. These examples are intended to help those skilled in the art to further understand the invention, but do not limit the invention in any way.

[0016] Example 1

[0017] Step 1: Multimodal data acquisition and preprocessing.

[0018] Based on 2090 clinical samples, the bioelectrical impedance values ​​of 48 key acupoints (Jing-well points, Yuan-source points, and confluence points of the eight extraordinary meridians) of the twelve regular meridians and eight extraordinary meridians were collected from the subjects using a "multi-channel meridian dynamic intelligent detection system" as the original meridian features. Simultaneously, the subjects' gender, age, latitude and longitude coordinates of their permanent residence, the Heavenly Stem of their birth year, and the detection time were collected. The original meridian impedance values ​​were subjected to a quantile normalization transformation based on ranking to ensure they followed a standard normal distribution, resulting in a 48-dimensional meridian feature vector. The spatiotemporal data were processed according to the method described in this invention: gender and age were encoded as 2-dimensional demographic features; latitude and longitude were encoded as 2-dimensional geospatial features; a 5-dimensional innate five-element unique-heat code was generated based on the Heavenly Stem of the birth year; and a 12-dimensional meridian flow unique-heat code was generated based on the detection time. These were then concatenated and normalized to obtain a 21-dimensional spatiotemporal context feature.

[0019] Step 2: Model architecture setup.

[0020] according to Figure 1 The overall architecture model is shown. In the multimodal data encoding module, two independent linear projection layers are used to map the 48-dimensional meridian features and 21-dimensional spatiotemporal context features to a 48-dimensional high-dimensional latent space. The context embedding is concatenated as a [CLS]-like token to the front end of the meridian embedding sequence, and a learnable positional encoding is added to form a 13×48 input sequence.

[0021] In the Deep TCM deep feature extraction module, according to Figure 2 The correspondence between the Five Elements' generating and restraining relationships and the Twelve Cardinal Meridians, as shown, is used to construct a 13×13 TCM a priori knowledge matrix based on the method described in this invention. A three-layer Transformer encoder was constructed, with each layer employing the adaptive soft-gating multi-head attention mechanism described in this invention, and the gating parameters were initialized. The value is 0.5, and the number of heads for multi-head attention is set to 8.

[0022] In the Deep TCM deep feature extraction module, according to Figure 2 The correspondence between the Five Elements' generating and restraining relationships and the Twelve Cardinal Meridians, as shown, is used to construct a 13×13 TCM a priori knowledge matrix based on the method described in this invention. A three-layer Transformer encoder is constructed, with each layer employing the adaptive soft-gating multi-head attention mechanism described in this invention.

[0023] Step 3: Model training and evaluation.

[0024] The model training employed a 5-fold cross-validation strategy. In each fold, the DeepTCM feature extractor was trained using only the training set data from that fold, and Mixup data augmentation was applied to mitigate overfitting. The optimizer used was AdamW with an initial learning rate of 1e-3, a batch size of 64, and 80 training epochs. After training, the Deep TCM parameters were frozen, and global fusion features were extracted from the training and validation sets for training and evaluating the XGBoost fusion module. No test set information was used for parameter updates in any model.

[0025] Experimental results are as follows Figure 3 As shown, the model achieves excellent performance in both syndrome classification and severity quantification tasks.

[0026] Example 2

[0027] To verify the consistency between the model's decision-making logic and traditional Chinese medicine theory, the model trained in Example 1 was used to perform inference on the test set. Attention weights were extracted according to the method described in this invention, the correlation strength between meridians was calculated, and plotted as shown below. Figure 4 , Figure 5 The heatmap and topological network diagram shown are illustrated. The visualization results demonstrate that the model can automatically discover and present meridian interaction patterns that highly align with classic TCM theories such as the "bladder meridian-kidney meridian" exterior-interior relationship and the "spleen meridian-stomach meridian" synergistic relationship. Notably, even under purely data-driven conditions without TCM priors, the model can still autonomously learn similar core relational structures, providing objective data-driven evidence for TCM meridian theory.

[0028] Example 3

[0029] During the training process in Example 1, the learnable parameters in each Transformer layer are recorded and visualized. The evolutionary trajectory, the result is as follows Figure 6 As shown. The underlying network The value rapidly decays from 0.5 to near 0, achieving unbiased data feature mining; the middle layer network The value stabilizes at around 0.4, indicating the beginning of data and knowledge integration; high-level networks... The value is maintained above 0.5, allowing for full application of traditional Chinese medicine theory in global decision-making. This bottom-up, data-to-theory progressive learning model is key to the model's high generalization performance and interpretability with small sample sizes.

Claims

1. A syndrome-quantified health assessment model based on deep learning, characterized in that... The model comprises the following modules: a multimodal data encoding module, which maps meridian features and spatiotemporal context features containing demographic features, geographical environment, innate five elements, and meridian flow information to a unified high-dimensional latent space through independent linear projection layers, and embeds learnable positional codes to form an input sequence; a Deep TCM deep feature extraction module, which, based on a Transformer encoder, introduces an adaptive soft-gated multi-head attention mechanism, dynamically adjusting the weights of the TCM prior knowledge matrix and data-driven attention through learnable parameters to extract deep meridian features that integrate TCM theory; and an XGBoost fusion decision module, which concatenates the global fusion features output by the DeepTCM deep feature extraction module with the original meridian features and multimodal context features to generate enhanced features, and trains an XGBoost multi-output classifier and regressor based on these enhanced features. Finally, the regressor prediction results are equally weighted and fused with the regression prediction results of the Deep TCM module to obtain the final quantitative value of syndrome severity.

2. A method for constructing a TCM prior knowledge matrix in the model described in claim 1, characterized in that... The construction method is as follows: Based on traditional Chinese medicine theory, a relationship matrix containing a global context node and twelve meridian nodes is assigned values. The assignment rules include: classifying meridians according to the five elements attributes of the twelve regular meridians; assigning positive weights to meridian pairs with mutual generation relationships based on the five elements' mutual generation law; assigning negative weights to meridian pairs with mutual restraint relationships based on the five elements' mutual restraint law; assigning synergistic weights to meridians with the same five elements attributes; assigning association weights to meridians with "same-name meridians" relationships; and setting the connection weights between the global context node and all meridian nodes to fixed values ​​to achieve the interaction between global spatiotemporal features and local meridian features.

3. A method for calculating feature weights using the adaptive soft-gated multi-head attention mechanism in the model described in claim 1, characterized in that... The method involves dynamically adjusting the weights of the TCM prior knowledge matrix in attention calculation using a learnable gating parameter α, the calculation formula of which is as follows: in, For querying key-value matrices, This is the aforementioned matrix of prior knowledge in Traditional Chinese Medicine. For initialization, a learnable scalar; during model training, As an independent parameter, it is optimized in conjunction with network weights to dynamically balance data-driven feature associations and prior constraints guided by traditional Chinese medicine theory.

4. A method for internally constructing the XGBoost fusion decision module in the model of claim 1, characterized in that... The XGBoost fusion decision module includes: a feature enhancement unit, used to concatenate the global fusion features output by the Deep TCM deep feature extraction module, the meridian feature data, and the spatiotemporal context feature data to generate the enhanced features; a multi-output XGBoost classifier, used to process the enhanced features to generate syndrome classification results; and a multi-output XGBoost regressor, used to process the enhanced features to generate severity quantification results.

5. A method for constructing spatiotemporal contextual feature data in the model of claim 1, characterized in that... The construction method is as follows: the gender characteristics of the subjects are binary encoded, the age characteristics are numerically processed to form demographic characteristics; and the latitude and longitude coordinates of the subjects' permanent residence are encoded to form geospatial characteristics. Based on the Heavenly Stem of the subject's birth year, a unique hot coding vector representing their innate Five Elements attribute is generated; based on the subject's detection time, a unique hot coding vector representing the meridian corresponding to the detection time is generated to reflect the meridian flow pattern.

6. A learnable gating parameter for the method of claim 3 A method for describing the evolutionary characteristics, characterized in that... The learnable gating parameters The model exhibits a hierarchical adaptive evolution across its Transformer layers: it approaches zero in the bottom layer to achieve unbiased data feature extraction, while gradually increasing in the middle and upper layers to introduce and strengthen the macroscopic constraints of traditional Chinese medicine theory, thus forming a progressive feature learning mode that is "from bottom to top and from data to theory".

7. A method for constructing a meridian connection network based on an attention mechanism, used for constructing the deep learning-based syndrome quantitative health assessment model as described in claim 1, characterized in that: Under the condition of integrating the prior knowledge of traditional Chinese medicine, meridians With meridians The strength of the correlation between The calculation formula is: ; Under purely data-driven conditions, meridians With meridians The strength of the correlation between The calculation formula is: ,in, , The first The query matrix and key matrix of each attention head; It is a single-head feature dimension; For the number of heads; For learnable adaptive soft gating parameters; This is a matrix of prior knowledge in Traditional Chinese Medicine. Using the twelve meridians as network nodes and the calculated inter-meridian correlation strength as network edge weights, a meridian correlation network is constructed, and a heat map and a fully connected topology network diagram are generated for visualization.

Images