Traditional Chinese medicine intelligent diagnosis system based on pulse condition characteristics

Abstract

The invention relates to the technical field of traditional Chinese medicine intelligent diagnosis, in particular to a pulse condition feature-based traditional Chinese medicine intelligent diagnosis system, which comprises a pulse signal acquisition module, a depth feature extraction and characterization module, a pulse condition semantic mapping and identification module, a personalized adaptation and robustness enhancement module and a diagnosis report module. According to the invention, comprehensive, accurate and robust analysis of pulse condition information is realized, and the accuracy of traditional Chinese medicine intelligent diagnosis is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnostic technology in traditional Chinese medicine, and in particular to an intelligent diagnostic system for traditional Chinese medicine based on pulse characteristics. Background Technology

[0002] Traditional Chinese Medicine (TCM), as a long-standing and profound medical system, plays an irreplaceable role in maintaining human health and treating diseases through its unique holistic concept and syndrome differentiation and treatment principles. Among the four diagnostic methods of TCM—inspection, auscultation and olfaction, inquiry, and palpation—pulse diagnosis occupies a crucial position. Pulse diagnosis involves the doctor palpating the patient's radial artery pulse with their fingertips, perceiving the pulse's location, rhythm, strength, morphology, and other subtle characteristics, thereby judging the state of the internal organs' Qi and blood, the nature of pathogenic factors, and the progression of the disease. It is one of the key links in TCM syndrome differentiation and treatment. The accuracy of its diagnostic results highly depends on the physician's clinical experience, tactile sensitivity, and diagnostic thinking ability, and has always been regarded as the core essence of TCM diagnostic and treatment techniques.

[0003] As research into the modernization, standardization, and intelligent diagnostic techniques of Traditional Chinese Medicine (TCM) deepens, early-stage intelligent pulse diagnosis systems have gradually revealed inherent limitations in their underlying principles. This is because traditional TCM pulse diagnosis perceives pulse characteristics far beyond simple physical waveform parameters. For example, while pulse characteristics such as "floating, deep, slow, and rapid" are related to physical quantities like pulse depth and frequency, pulse characteristics like "slippery, rough, wiry, and tense" focus more on the more abstract, multi-dimensional, and often difficult-to-represent "texture" descriptions of pulse flow, smoothness, and tension—qualities that are difficult to characterize directly with a single physical quantity. Early intelligent pulse diagnosis systems extracted features largely limited to explicit, low-level quantitative parameters of the pulse waveform in the time or frequency domain, such as peak value, trough value, waveform width, and the main frequency components after Fourier transform. While these features are objective, they fail to fully capture the unique, high-level semantic information of "pulse characteristics" that is crucial for clinical diagnosis in TCM pulse diagnosis. Summary of the Invention

[0004] The inventors discovered through research that due to physiological differences such as physique, age, vascular elasticity, and skin thickness, there are significant differences in pulse waveforms among different individuals, even when they are in the same physiological state. At the same time, the pulse waveform of the same patient will also change dynamically under different physiological or pathological states, and even under different pressure and location. Existing intelligent pulse diagnosis systems often lack sufficient adaptability and robustness when dealing with such extensive individual differences and dynamic changes, resulting in insufficient generalization ability and stability of their diagnostic results.

[0005] This invention provides a TCM intelligent diagnostic system based on pulse characteristics, which solves the technical problem that existing intelligent pulse diagnosis systems are limited in capturing multi-dimensional, high-level semantic information of TCM pulse characteristics.

[0006] This invention provides a traditional Chinese medicine intelligent diagnostic system based on pulse characteristics, comprising:

[0007] The pulse signal acquisition module is used to simultaneously acquire pulse signals from the radial artery at the cun, guan, and chi positions of the subject at multiple points, depths, and high resolution.

[0008] The pulse wave depth feature extraction and characterization module is connected to the pulse signal acquisition module and is used to perform deep learning and extract multi-dimensional, high-level features that are highly related to the semantics of traditional Chinese medicine pulse patterns from the acquired pulse signal data.

[0009] The TCM pulse semantic mapping and identification module is connected to the pulse wave depth feature extraction and characterization module. It is used to map the extracted high-dimensional pulse feature vector to a standardized TCM pulse category and provide corresponding identification results and confidence assessment.

[0010] The personalized adaptation and robustness enhancement module is used to improve the system's adaptability and generalization performance to individual physiological differences, environmental disturbances, and complex pulse patterns.

[0011] The diagnostic report module is used to present the system's analysis results to users in a structured, intuitive, and easy-to-understand format, and to provide necessary decision support.

[0012] In some embodiments, the pulse signal acquisition module includes at least:

[0013] An array-type micro pressure sensor unit is composed of multiple microelectromechanical system pressure sensors arranged in a row. The array-type micro pressure sensor unit has at least three independent sensing areas, which correspond to the three pulse positions of Cun, Guan, and Chi respectively.

[0014] A dynamic pressing device is positioned above the array-type micro pressure sensor unit and maintains adjustable physical contact with the sensing surface of the array-type micro pressure sensor unit.

[0015] In some embodiments, the array-type micro pressure sensor unit uses a piezoresistive microelectromechanical system (MEMS) pressure sensor; within each sensing area, the MEMS pressure sensors are distributed at a preset interval and in an array to capture fine spatial distribution information of pulse waves in a local area; the MEMS pressure sensor integrates a temperature compensation circuit.

[0016] In some embodiments, the dynamic pressing device includes a precision stepper motor drive unit for providing precisely controllable vertical pressing force; the drive unit is coupled to a force control feedback unit to achieve closed-loop pressing force control; the pressing device also integrates an infrared ranging sensor for real-time monitoring of pressing depth.

[0017] In some embodiments, the pulse wave depth feature extraction and representation module is a feature learning architecture based on a deep neural network, the feature learning architecture including:

[0018] Multi-scale spatiotemporal convolutional networks, consisting of multiple parallel or cascaded convolutional layers, are used to capture local patterns of pulse waveforms at different time scales and spatial locations.

[0019] Long Short-Term Memory (LSTM) network layers are set after the multi-scale spatiotemporal convolutional network to capture long-term dependencies and dynamic evolution patterns in the pulse waveform time series.

[0020] A self-attention mechanism, integrated into the middle or top layer of the feature learning architecture, is used to dynamically evaluate and weight the contribution of different spatiotemporal points and different compression depth channels in the pulse waveform data to the final pulse characteristics.

[0021] The feature fusion layer, set after the multi-scale spatiotemporal convolutional network, long short-term memory network layers and self-attention mechanism, is used to cascade feature representations from different sub-networks to generate fixed-dimensional, high-dimensional, semantically rich pulse feature vectors.

[0022] In some embodiments, the TCM pulse semantic mapping and identification module includes at least:

[0023] A multi-task learning classifier receives a high-dimensional pulse feature vector as input; the multi-task learning classifier is a multi-layer fully connected neural network.

[0024] The TCM knowledge graph embedding unit works in collaboration with the multi-task learning classifier. The knowledge graph is pre-built and stores rich TCM theoretical knowledge. The knowledge graph embedding unit uses a knowledge graph reasoning engine to fuse and verify the preliminary pulse recognition results output by the multi-task learning classifier with the semantic information in the knowledge graph.

[0025] In some embodiments, the personalized adaptation and robustness enhancement module includes at least: a transfer learning and domain adaptation unit, an adversarial training and data augmentation unit, and an uncertainty quantification and anomaly detection unit; wherein...

[0026] The data augmentation in the adversarial training and data augmentation unit includes: time stretching and compression of the pulse waveform, random scaling of amplitude, random shifting, and generating synthetic pulse data through a generative adversarial network;

[0027] The uncertainty quantification and anomaly detection unit is based on Bayesian deep learning and provides a quantified uncertainty measure for pulse recognition results.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] This invention utilizes a pulse signal acquisition module that combines an array of micro-pressure sensor units with a dynamic compression device to achieve comprehensive and refined acquisition of pulse waves in three dimensions: space, depth, and time. This enriches the data information of the pulse signal and overcomes the limitations of existing technologies that rely solely on a single physical quantity or low-level parameters.

[0030] This invention proposes a deep neural network architecture based on multi-scale spatiotemporal convolutional networks, long short-term memory networks, and self-attention mechanisms to construct a pulse wave deep feature extraction and representation module. This module can automatically learn and mine the multi-dimensional, high-level, and nonlinear semantic features hidden in the pulse waveform, such as the smoothness, tension, and subtle fluctuations of the pulse, rather than being limited to explicit physical parameters. The system can better simulate the perception and judgment of the "texture" of the pulse by senior TCM experts, solving the problem that existing technologies cannot fully capture the semantic information of TCM pulse characteristics.

[0031] The multi-task learning classifier and TCM knowledge graph embedding unit of this invention can achieve accurate identification of standard pulse types. Furthermore, by introducing TCM knowledge graph, the identification results are deeply integrated and verified with the TCM theoretical system, which greatly improves the accuracy, consistency and clinical interpretability of the diagnostic results.

[0032] The personalized adaptation and robustness enhancement module of this invention enables the system to effectively cope with individual physiological differences, environmental noise, and complex and variable pulse signals, significantly improving the system's generalization ability and diagnostic stability, and overcoming the inherent defects of existing systems in terms of insufficient robustness when facing a wide range of individual differences.

[0033] The diagnostic report module of this invention provides structured and graphical diagnostic reports and a variety of intuitive visualization tools, clearly displaying pulse waveforms, depth features, recognition results, and confidence levels. This facilitates doctors' understanding and adoption of diagnostic suggestions, provides doctors with objective auxiliary decision-making basis, and accelerates the clinical application and promotion of TCM intelligent diagnostic technology. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a schematic diagram of the structure of the TCM intelligent diagnostic system based on pulse characteristics according to the present invention. Detailed Implementation

[0036] The following will refer to the appendices in the embodiments of the present invention. Figure 1 The technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0037] Example

[0038] Figure 1 As shown, the TCM intelligent diagnostic system based on pulse characteristics of the present invention has a core structure including a pulse signal acquisition module, a pulse wave depth feature extraction and characterization module, a TCM pulse semantic mapping and identification module, a personalized adaptation and robustness enhancement module, and a diagnostic report module. The modules achieve efficient data transmission and functional collaboration through standardized data interfaces and communication protocols, forming an intelligent diagnostic closed loop.

[0039] The pulse signal acquisition module is used to simultaneously acquire pulse signals from the radial artery at the cun, guan, and chi positions of the subject at multiple points, depths, and high resolutions. The module includes an array-type micro-pressure sensor unit, a dynamic pressing device, and may also include a data acquisition and preprocessing unit.

[0040] Specifically, the array-type micro-pressure sensor unit consists of multiple microelectromechanical system (MEMS) pressure sensors arranged in a matrix or linear pattern. More specifically, the array-type micro-pressure sensor unit employs a silicon-based MEMS piezoresistive pressure sensor array with a sensing surface size of approximately 30 mm × 15 mm, capable of completely covering the radial artery's cun, guan, and chi pulse positions. The sensor unit has three independent sensing areas, each containing a 3x3 sensor matrix, totaling 27 MEMS pressure sensors. For example, the sensor matrix in the cun position area covers approximately 1 cm above the radial styloid process, the guan position area covers the radial styloid process, and the chi position area is located approximately 1 cm below the guan position. Each MEMS pressure sensor measures 2 mm × 2 mm with a 1 mm spacing to capture fine spatial distribution information of the pulse wave in a localized area.

[0041] The MEMS pressure sensor uses biocompatible polymer materials, such as medical-grade polyimide film, for encapsulation to ensure comfort and safety when in contact with human skin. The sensor boasts high sensitivity, typically 0.5 mV / kPa, and a wide frequency response range covering 0.5 Hz to 200 Hz, enabling it to capture the fundamental frequency component and higher harmonics of the pulse wave, thus resolving the fine structure of the waveform. The sensor's measurement range is precisely designed from 0 to 50 kPa, sufficient to sense the subtle pressure changes generated by the radial artery pulsation, while avoiding saturation at higher pressures. Its full-range accuracy is better than 0.1%, linearity error is less than 0.05%, and hysteresis error is less than 0.03%, ensuring the accuracy and reliability of pulse pressure measurement. Each MEMS pressure sensor integrates a temperature compensation circuit to eliminate the influence of ambient temperature fluctuations on the measurement results, ensuring measurement consistency under different ambient temperatures.

[0042] The dynamic pressing device is positioned above the array-type micro-pressure sensor unit and maintains adjustable physical contact with the sensing surface of the sensor unit. The dynamic pressing device includes a precision stepper motor drive unit connected to a pressure head via a high-precision ball screw mechanism. The stepper motor drive unit uses a closed-loop stepper motor with micro-stepping capability, such as a resolution of 20,000 steps / revolution, coupled with a planetary gearbox with a reduction ratio of 10:1, theoretically achieving a minimum displacement step size of 0.1 micrometers. The precision stepper motor drive unit provides precisely controllable vertical pressing force, allowing the pressure head to press down vertically at a preset speed (e.g., 0.5 mm / s to 2 mm / s) or remain at a specific depth.

[0043] A precision stepper motor drive unit is coupled to a force control feedback unit, which consists of a high-precision miniature strain gauge-type force sensor integrated into the bottom of the compression head of the compression device. It directly measures the vertical pressure applied by the compression head to the sensor array. The force sensor has a range of 0 to 20 Newtons and an accuracy better than 0.05 Newtons. Through this closed-loop feedback control mechanism, the system can dynamically and continuously acquire pulse signals at multiple preset compression depths or pressure levels. For example, the system can perform three compression techniques: floating, medium, and deep. Floating corresponds to a pressure level of 0.5 Newtons, medium to 2.0 Newtons, and deep to 5.0 Newtons, with each pressure level continuously acquired for 5 to 10 seconds. The pressure adjustment accuracy is better than 0.1 Newtons, and the response time is less than 100 milliseconds, ensuring the stability and repeatability of the compression process, thereby acquiring pulse depth information and its waveform characteristics under different pressure levels.

[0044] The dynamic pressing device also integrates an infrared distance sensor for real-time monitoring of pressing depth, with a measurement accuracy of 0.05 mm, to facilitate calibration and reference with force control feedback data.

[0045] The data acquisition and preprocessing unit is electrically connected to the array-type micro pressure sensor unit and the dynamic pressing device. The data acquisition and preprocessing unit includes a multi-channel synchronous analog-to-digital converter (ADC), a low-noise amplifier array, and a digital signal processor (DSP).

[0046] The multi-channel synchronous ADC uses the Analog Devices AD7175-2 or an equivalent 24-bit high-resolution ADC, which features multiple differential input channels and can synchronously acquire analog electrical signals from all 27 MEMS pressure sensors at a sampling rate of 5000Hz per channel, ensuring high time synchronization and data accuracy. Each ADC channel is pre-amplified with a low-noise amplifier array, which uses ultra-low-noise operational amplifiers (such as the Analog Devices AD797) to provide adjustable gain, for example, from 100x to 1000x, with an input reference noise voltage of less than 1.0nV / √Hz, effectively improving the signal-to-noise ratio of the weak sensor output signals and ensuring signal integrity and purity.

[0047] The digital signal processor (DSP) selected is the TI C6748 series high-performance DSP, with a main frequency of up to 456MHz and floating-point operation capability, used to execute real-time signal preprocessing algorithms. These preprocessing algorithms include: baseline drift elimination, such as a method combining Empirical Mode Decomposition (EMD) and wavelet transform, where EMD decomposes the signal into intrinsic mode functions (IMFs), then identifies and removes low-frequency drift IMF components, or uses cubic spline interpolation for polynomial fitting to remove baseline drift; high-frequency noise filtering, employing adaptive Kalman filtering or moving average filtering algorithms optimized for the non-stationary characteristics of pulse signals to effectively filter out high-frequency random noise, such as a fourth-order Butterworth low-pass filter with a cutoff frequency of 50Hz, which has minimal phase distortion; and power frequency interference suppression, using notch filters or adaptive notch-based digital signal processing techniques to accurately filter out 50Hz or 60Hz power frequency interference and its harmonic components.

[0048] The data acquisition and preprocessing unit ultimately outputs a standardized, denoised, and time-aligned multi-channel, multi-compression-depth (i.e., three-dimensional spatiotemporal pressure matrix) digital pulse waveform dataset. This dataset accurately characterizes the spatial distribution of pulse waves at the cun, guan, and chi positions, their intensity and morphology at different compression depths, and their dynamic characteristics over time. For example, the dataset format is a four-dimensional tensor of N x C x D x T, where N is the number of samples, C is the number of sensor channels (27), D is the number of compression depths (3 or more), and T is the time series length (e.g., for a 10-second acquisition time with a sampling rate of 5000Hz, T = 50000).

[0049] The pulse wave depth feature extraction and representation module is used to automatically and deeply learn and extract multi-dimensional, high-level features highly correlated with the semantics of traditional Chinese medicine pulse diagnosis from the digital pulse waveform data output by the pulse signal acquisition module. Its core is a feature learning architecture based on a deep neural network, with preprocessed multi-channel, multi-pressure depth pulse wave time series data as input. The deep neural network architecture includes a multi-scale spatiotemporal convolutional network (MSTSCN), which consists of multiple parallel or cascaded convolutional layers, designed to capture local patterns of the pulse waveform at different time scales and spatial locations.

[0050] MSTSCN integrates multiple 1D convolutional layers, primarily used to extract the rhythm, morphology, rising edge, and falling edge features of the pulse wave in the temporal dimension. For example, the initial layers can use 1D convolutional kernels with a kernel size of 3 or 5 and a stride of 1 to capture the instantaneous changes and fine texture features of the pulse wave, such as peaks, troughs, and diphtheria notches. As the network depth increases, the kernel size can be increased to 15 or 30, or small-sized convolutional kernels can be stacked (e.g., three consecutive convolutional layers with a kernel size of 3 are equivalent to one convolutional layer with a kernel size of 7) to expand the receptive field and capture the macroscopic morphological features of the pulse wave, such as waveform width and symmetry.

[0051] Meanwhile, multiple 2D or 3D convolutional layers are used to capture the correlation and variation patterns of pulse waveforms between different sensor channels (spatial dimension) and between different compression depths (depth dimension). For example, 2D convolutional layers can operate on a sensor matrix (e.g., 3x3) at a certain compression depth to capture spatial pressure gradient features within or between the Cun, Guan, and Chi pulse positions. 3D convolutional layers can perform convolutions simultaneously in the temporal and spatial depth dimensions, for example, using 3x3x3 convolutional kernels to capture the dynamic evolution patterns of pulse waves in a three-dimensional spatiotemporal pressure matrix. The MSTSCN extensively integrates residual connections, such as the identity mapping in the ResNet architecture, or dense connections, such as the feature reuse mechanism in the DenseNet architecture, to solve the gradient vanishing problem in deep network training and promote effective feature propagation and reuse. Each convolutional layer is typically followed by a batch normalization layer and a ReLU activation function to accelerate training and enhance nonlinear expressive power. Pooling layers (such as max pooling or average pooling) are used to reduce feature dimensionality and improve feature scale invariance.

[0052] Furthermore, Long Short-Term Memory (LSTM) layers or Gated Recurrent Unit (GRU) layers are placed after the MSTSCN to capture long-term dependencies and dynamic evolution patterns in the pulse waveform time series. Through its unique gating mechanism (input gate, forget gate, output gate or update gate, reset gate), this LSTM or GRU layer can memorize and process longer sequence information, effectively characterizing the continuity, smoothness, tension, and rhythm of the pulse wave. These characteristics are key to identifying pulse patterns such as "slippery pulse," "hesitant pulse," "wiry pulse," and "intermittent pulse."

[0053] In one embodiment, the layer group employs a two-layer stacked bidirectional LSTM network, with each layer containing 256 hidden units. The bidirectional structure allows the network to simultaneously consider past and future information of the sequence, thereby providing a more comprehensive model of the dynamic characteristics of the pulse wave. The output of this LSTM layer group, namely the hidden state vector, effectively encodes the time-series characteristics of the pulse wave, such as the stability and variability of the pulse rate, the regularity of the waveform, and anomalous rhythms.

[0054] Furthermore, a self-attention mechanism is integrated into the middle or top layer of the deep neural network architecture to dynamically evaluate and weight the contribution of different spatiotemporal points and compression depths in the pulse waveform data to the final pulse characteristics. This self-attention mechanism identifies key regions or feature points in the pulse waveform that are discriminative for specific pulse types by calculating the similarity between query, key, and value vectors. Examples include the morphology and location of the dicrotic notch, the width and height of the main peak, and the amplitude variation trend at different compression depths, and assigns them higher weights.

[0055] In a preferred embodiment, a multi-head self-attention mechanism is employed, with eight attention heads. Each attention head independently learns different feature associations, and the outputs of the multiple heads are concatenated to provide a richer attention representation. Through this attention mechanism, the model can better focus on key information, improving the accuracy and robustness of feature extraction.

[0056] Furthermore, a feature fusion layer is placed after the MSTSCN, LSTM / GRU layers, and self-attention mechanism. The feature fusion layer concatenates or weights the feature representations from different sub-networks. For example, the spatiotemporal feature map output by MSTSCN is flattened into a vector, the final hidden state vector output by the LSTM / GRU layers, and the weighted feature representation output by the self-attention mechanism are merged through concatenation into a unified, fixed-dimensional (e.g., 1024-dimensional), high-dimensional, semantically rich pulse feature vector. This feature vector comprehensively encodes deep information about the pulse wave across multiple dimensions, including spatiotemporal, depth, rhythm, and morphology.

[0057] Deep neural network architectures learn robust and generalizable feature representations through pre-training on large-scale, diverse pulse datasets using self-supervised learning. For example, masked pulse waveform prediction randomly masks a portion of the pulse waveform's time points or sensor channel data, then trains the model to predict the masked data; or cross-modal contrastive learning maximizes the similarity of the same pulse waveform features at different compression depths while minimizing the similarity with features from different pulse waveforms, thereby learning deep semantic features.

[0058] In addition, the pulse wave depth feature extraction and characterization module also applies data augmentation techniques, including time stretching (randomly stretching or compressing the pulse waveform time axis by 5% to 15%), amplitude scaling (randomly scaling the waveform amplitude by 80% to 120%), random noise injection (adding Gaussian white noise with a signal-to-noise ratio of 20dB to 30dB), and channel perturbation (randomly selecting some sensor channels to inject high-frequency noise or make small amplitude shifts) to expand the diversity of the training dataset, reduce the risk of model overfitting, and improve the robustness of the feature extraction model to the variability of pulse signals.

[0059] The Traditional Chinese Medicine (TCM) pulse semantic mapping and identification module maps the high-dimensional pulse feature vector output by the pulse wave depth feature extraction and representation module to standardized TCM pulse categories, and provides corresponding identification results and confidence assessments. The TCM pulse semantic mapping and identification module includes a multi-task learning classifier, a TCM knowledge graph embedding unit, and a confidence assessment unit.

[0060] Specifically, the multi-task learning classifier receives a high-dimensional pulse feature vector as input. The classifier is a multilayer perceptron consisting of three fully connected neural networks. The first layer contains 512 neurons, the second layer contains 256 neurons, and the third layer contains 128 neurons. The ReLU activation function is used between each layer, supplemented by a dropout layer (e.g., a dropout rate of 0.3) to prevent overfitting. Its output layer contains multiple independent prediction branches.

[0061] A primary classification branch is used for multi-label or multi-classification identification of 28 or more predefined standard TCM pulse types. This branch contains 28 sigmoid activation function output neurons, outputting the probability distribution or confidence score for each pulse type. For example, the identification of "floating pulse" depends on the pulse amplitude and morphological characteristics at a superficial pressure depth; the identification of "deep pulse" focuses on the amplitude and morphological characteristics at a deep pressure depth; the identification of "slow pulse" and "rapid pulse" focuses more on the stability of pulse rate and rhythm; "slippery pulse" is characterized by a smooth and rounded pulse waveform with a large amplitude and obvious dicrotic wave; "hesitant pulse" has a thin and irregular waveform with a sense of abruptness. In addition, four auxiliary classification branches are included, each used to predict the independent core attributes of the pulse: depth (floating / deep binary classification), speed (slow / rapid / slow triad classification), morphology (slippery / hesitant / wiry / tight quadruple classification), and intensity (weak / full binary classification). By employing a multi-task learning paradigm—that is, including a weighted sum of the cross-entropy losses of the main classification task and the auxiliary classification task in the total loss function—the model is encouraged to better learn the discriminative features of these key attributes, thereby improving the accuracy and generalization ability of the main classification task. For example, the total loss function L = w1L_main + w2L_depth + w3L_rate + w4L_shape + w5*L_intensity, where w1-w5 are weight coefficients.

[0062] The TCM knowledge graph embedding unit works collaboratively with a multi-task learning classifier. The knowledge graph (KG) is pre-built and stores rich TCM theoretical knowledge, and its nodes include: pulse types (e.g., 28 standard pulse types and their composite pulse types, such as floating pulse, deep pulse, slow pulse, rapid pulse, slippery pulse, hesitant pulse, wiry pulse, tense pulse, weak pulse, and strong pulse), viscera (liver, heart, spleen, lung, and kidney), pathogenesis (liver qi stagnation, spleen deficiency with dampness, kidney essence deficiency, qi stagnation and blood stasis, etc.), symptoms (dizziness, fatigue, chest tightness, abdominal distension, etc.), prescriptions, and acupoints.

[0063] The edges of the knowledge graph represent the relationships between different entities. For example, there is a "correspondence" relationship between "floating pulse" and "exterior syndrome," a "hint" relationship between "wiry pulse" and "abnormal liver function in regulating qi," or a "combination" relationship where "slippery pulse" often co-occurs with "rapid pulse." The knowledge graph embedding unit uses a graph neural network (GNN), such as Graph Attention Networks (GAT) or Relational Graph Convolutional Networks (R-GCN), to fuse and verify the preliminary pulse identification results (e.g., probability distribution) output by the multi-task learning classifier with the semantic information in the knowledge graph.

[0064] Specifically, the Generative Neural Network (GNN) uses pulse diagnosis results as initial node features. Through multi-layer message passing and aggregation on the knowledge graph, it encodes the associated information in the graph structure into the pulse diagnosis results. For example, if the classifier identifies "wiry pulse" with a probability of 0.8 and "rapid pulse" with a probability of 0.7, and "wiry and rapid pulse" has a strong correlation with "liver stagnation transforming into fire" in the knowledge graph, the GNN will adjust the final confidence of the pulse diagnosis based on these relationships and deduce its possible association with pathogenesis such as liver stagnation transforming into fire. This mechanism not only improves the accuracy and consistency of pulse diagnosis results, but more importantly, it provides the system with an internal logical explanation and TCM theoretical basis for the pulse diagnosis results, enhancing the system's interpretability and clinical acceptability.

[0065] The confidence assessment unit provides a quantified confidence score or uncertainty interval for each identified pulse type. This can be achieved using Bayesian deep learning methods, such as applying Monte Carlo Dropout after each fully connected layer of a deep neural network. This means the Dropout layer remains active during the testing phase, and multiple predictions are obtained through multiple forward propagations (e.g., 100 times). The mean of these predictions is then calculated as the final prediction, and the standard deviation is used as a measure of uncertainty. Alternatively, ensemble learning can be employed, constructing multiple base classifiers (e.g., training multiple deep neural networks with different initialization parameters or subsets). The predictions from these base classifiers are then voted on or weighted to obtain a more stable and reliable confidence score.

[0066] When the confidence level of the recognition result is lower than the preset threshold (e.g., 0.7 or 0.8), the system can indicate that the result has high uncertainty and can activate the personalized adaptation mechanism or request manual review to avoid misdiagnosis or missed diagnosis. For example, for complex or mixed pulses with low confidence, the system will suggest that doctors make a comprehensive judgment by combining tongue appearance, medical history and other information.

[0067] The personalized adaptation and robustness enhancement module is used to improve the system's adaptability and generalization performance to individual physiological differences, environmental disturbances and complex pulse patterns. Specifically, it includes a transfer learning and domain adaptation unit, an adversarial training and data augmentation unit, and an uncertainty quantification and anomaly detection unit.

[0068] The transfer learning and domain adaptation unit utilizes a deep neural network model pre-trained on a large general pulse dataset as its base model. This general pulse dataset can contain pulse data from tens of thousands of healthy individuals and patients with common diseases, covering different ages, genders, and body types. When faced with pulse data from new individuals or specific subgroups (e.g., the elderly, diabetic patients, or people from a specific region), the transfer learning and domain adaptation unit uses a small amount of labeled data from that individual or subgroup (e.g., 50 to 200 samples) and employs strategies such as fine-tuning or domain adversarial training to optimize and adjust the parameters of the base model.

[0069] The specific implementation of fine-tuning strategies includes: retraining all layers of the pre-trained model with a small learning rate (e.g., 1e-5); or freezing the bottom feature extraction layers of the pre-trained model (e.g., the MSTSCN layer) and training only the top classifier; or adopting a learning rate decay strategy, using a larger learning rate in the early stage of training and gradually reducing it in the later stage.

[0070] Domain adversarial training introduces a domain discriminator, forcing the feature extractor to generate domain-invariant features, thus enabling the model to better generalize to new target domains. This mechanism allows the system to quickly and effectively adapt to physiological differences between individuals, such as the influence of vascular elasticity, subcutaneous fat thickness, body type, and even psychological state on pulse, thereby overcoming the problem of insufficient adaptability to individual specificities in traditional methods.

[0071] The adversarial training and data augmentation unit enhances the robustness of deep neural network models to input signal perturbations and noise by generating adversarial examples and adding them to the training dataset. For example, by using methods such as the Fast Gradient Sign Method (FGSM) or Projected Gradient Descent (PGD), subtle perturbations that are imperceptible to the naked eye but can mislead the model are injected into the original pulse waveform data. This simulates real physiological noise (such as breathing artifacts, body movement artifacts, and arrhythmia) or environmental noise (such as electromagnetic interference and slight sensor vibrations). These perturbed samples and their correct labels are then added to the training set, forcing the model to correctly classify these perturbed samples. In this way, the model's generalization ability and anti-interference ability in real and complex environments are enhanced.

[0072] Data augmentation techniques further include: time stretching and compression of pulse waveforms (randomly stretching or compressing the time axis of the waveform by 5% to 15%), random amplitude scaling (randomly scaling the waveform amplitude by 80% to 120%), random shifting (randomly shifting left and right on the time axis by 10 milliseconds to 50 milliseconds), and generating synthetic pulse data through generative adversarial networks (GANs), such as using WGAN-GP or StyleGAN2 to generate synthetic waveforms with specific pulse characteristics, in order to expand the diversity of the training dataset, reduce the risk of model overfitting, and improve the ability to identify atypical, rare, or complex mixed pulses.

[0073] The uncertainty quantification and anomaly detection unit employs Bayesian deep learning or ensemble learning methods to provide a quantified measure of uncertainty for pulse recognition results. For example, the system not only outputs the recognition result of "string pulse" but also its confidence level of 92% or prediction interval of [88%, 96%]. When the feature vector of the input pulse signal deviates from the distribution of the trained data (i.e., it belongs to an abnormal or unseen new pulse type), the uncertainty quantification and anomaly detection unit can identify potential abnormal samples through outlier detection algorithms. For example, it can use an autoencoder to reconstruct the pulse feature vector and calculate the reconstruction error. If the reconstruction error exceeds a preset threshold, the sample is considered abnormal; or it can use density estimation-based algorithms, such as Isolation Forest or One-Class SVM, to determine the sparsity of the input feature vector in the training data distribution. Once a potential abnormal sample is identified, the system will mark the result as "uncertain" or "requires manual intervention," avoiding misidentification of rare or complex pulses and ensuring the safety of diagnosis.

[0074] The diagnostic report module presents the analysis results of the TCM intelligent diagnostic system to users in a structured, intuitive, and easy-to-understand format, providing necessary decision support. This module includes a structured diagnostic report generation unit, a pulse characteristic visualization unit, and a clinical decision support and human-computer interaction unit.

[0075] Specifically, the diagnostic report automatically generates a standardized electronic diagnostic report based on the raw data from the pulse signal acquisition module, the intermediate features from the deep feature extraction and characterization module, and the final recognition results and confidence level from the traditional Chinese medicine pulse semantic mapping and identification module.

[0076] The report includes: basic information of the examinee (name, age, gender, height, weight, chief complaint, etc.), time-domain and frequency-domain analysis of pulse waveforms (e.g., original multi-channel pulse waveform, Fourier transform spectrum, wavelet transform time-frequency graph), the main pulse types identified and their distribution in the cun, guan, and chi positions (e.g., floating and rapid cun pulse, wiry and slippery guan pulse, deep and slow chi pulse), quantitative characteristic descriptions of each pulse type (e.g., pulse rate, amplitude, dicrotic wave index, pulse wave propagation velocity, PWV, etc.), confidence score of the identification results, suggestions for potential pathogenesis or organ dysfunction derived based on the TCM knowledge graph (e.g., liver qi stagnation, spleen and stomach weakness, kidney essence deficiency), and corresponding differential diagnosis suggestions.

[0077] The report uses standardized coding based on the International Classification of Diseases (ICD-10-CM) or Traditional Chinese Medicine (TCM) disease diagnostic standard codes (e.g., the "Classification and Codes of TCM Diseases and Syndromes" published by the China Association of Traditional Chinese Medicine) to facilitate data sharing and statistical analysis. The report can be exported in PDF, JSON, or XML formats for easy storage, transmission, and integration with other medical information systems.

[0078] The pulse feature visualization unit visually displays pulse signals and key information from the system analysis process. For example, it provides a 3D dynamic view of the original multi-channel, multi-pressure depth pulse waveform. This view uses color coding to encode pressure values ​​at different pressure depths and dynamically displays the spatial distribution and dynamic changes of the pulse waveform across the three regions of Cun, Guan, and Chi on a time axis. Doctors can drag and zoom to view waveform differences at different pulse positions and pulse changes at different pressure depths. Through dimensionality reduction algorithms such as t-distributed random neighbor embedding (t-SNE) or uniform manifold approximation and projection (UMAP), the 1024-dimensional high-dimensional pulse feature vector is projected onto a 2D or 3D space, visually displaying the clustering distribution of different pulse types in the feature space. This helps doctors understand the model's discrimination criteria; for example, "wiry pulse" and "slippery pulse" will form clearly distinguishable clusters in the feature space.

[0079] By using heatmaps or attention weight maps, the regions or feature points in the pulse waveform that contribute the most to the identification of a specific pulse type can be highlighted. For example, when identifying a "string pulse", the heatmap may highlight the slope of the rising branch of the pulse waveform and the sharpness of the peak; when identifying a "slippery pulse", the amplitude and position of the dicrotic wave may be highlighted, thereby enhancing the interpretability of the system.

[0080] The clinical decision support and human-computer interaction unit submits the generated diagnostic reports and visualization results to clinicians as auxiliary diagnostic information. This system does not replace the doctor's final diagnostic authority, but provides multi-dimensional and objective data support and theoretical basis to assist doctors in making more accurate diagnosis and treatment.

[0081] The human-computer interface allows doctors to provide feedback, corrections, or annotations to the system's identification results. For example, doctors can manually adjust the identified pulse type or supplement the system's pathogenesis suggestions. This feedback data can be used to continuously optimize and update the deep learning model. For instance, through active learning strategies, samples with high model uncertainty or those corrected by doctors can be prioritized for annotation and retraining, forming a closed loop of continuous learning and evolution in intelligent diagnosis.

[0082] In addition, the clinical decision support and human-computer interaction unit can also retrieve and recommend corresponding prescriptions (e.g., Xiaoyao San for liver qi stagnation), acupoints (e.g., Zusanli and Pishu for spleen deficiency) or lifestyle regulation suggestions (e.g., a light diet and regular work and rest) from the integrated TCM knowledge base based on the identified pulse characteristics and other clinical information of the patient (if any). However, these suggestions are for reference only, and doctors need to make the final decision based on the patient's actual situation.

[0083] In a specific clinical practice, the pulse-based TCM intelligent diagnostic system of the present invention was applied to the pulse assessment of a 68-year-old male patient with hypertension. The patient complained of dizziness, irritability, bitter taste in the mouth, and a red tongue with a yellow coating.

[0084] First, the patient undergoes testing by the pulse signal acquisition module. An array of micro-pressure sensor units is precisely placed at the radial artery points (cun, guan, chi) of the patient's left hand. The dynamic compression device executes three compression techniques sequentially according to a preset program: superficial (0.5N), medium (2.0N), and deep (5.0N), with each compression depth continuously acquiring data for 10 seconds. Each MEMS pressure sensor synchronously acquires pulse waveform data at a sampling rate of 5000Hz.

[0085] The data acquisition and preprocessing unit performed baseline drift cancellation, high-frequency noise filtering (using a Butterworth low-pass filter with a cutoff frequency of 50Hz), and power frequency interference suppression on the raw signal. This resulted in a three-dimensional spatiotemporal pressure matrix dataset containing 27 channels, 3 compression depths, and 50,000 time points per channel.

[0086] Next, the dataset is input into the pulse wave depth feature extraction and representation module. The module's internal Multi-Scale Spatiotemporal Convolutional Network (MSTSCN) captures the local spatiotemporal patterns of the pulse waveform. For example, the amplitude of the waveform at the cun position is slightly greater than that at the guan position, while the waveform at the chi position is relatively weaker; the waveform is stronger upon superficial pressure and weaker upon deep pressure, indicating a shallow pulse position. Long Short-Term Memory (LSTM) layers further analyze the temporal series features of the pulse wave, finding that the patient's pulse rhythm is regular, but the pulse rate is relatively fast, approximately 90 beats per minute. The self-attention mechanism highlights areas with steep rising limbs, sharp peaks, and indistinct dicrotic notches; these features have high discriminative power for identifying a "wiry pulse." Finally, the module outputs a 1024-dimensional pulse feature vector.

[0087] Subsequently, this feature vector is input into the TCM pulse semantic mapping and identification module. The main classification branch of the multi-task learning classifier identifies a combination of "wiry pulse," "rapid pulse," and "tense pulse" in the patient, with a confidence score of 0.91 for "wiry pulse," 0.88 for "rapid pulse," and 0.75 for "tense pulse." The auxiliary classification branch predicts that the pulse depth is "superficial," the speed is "rapid," the morphology is "wiry and tense," and the intensity is "full." The TCM knowledge graph embedding unit combines these preliminary identification results and uses the GAT model to perform reasoning in the knowledge graph, verifying the possibility of "wiry, rapid, and tense pulse," and further associating it with the pathogenesis of "liver yang hyperactivity" and "liver wind internal movement." The confidence assessment unit gives an overall confidence score of 0.85 for the comprehensive diagnosis.

[0088] Finally, the diagnostic report generation and visualization interaction module generates a structured electronic diagnostic report. The report clearly displays a 3D dynamic view of the original pulse waveform, a dimensionality-reduced feature clustering diagram (showing the patient's pulse point is located in the central region of the "string-rapid pulse" cluster), and an attention heatmap (highlighting the rising limb and peak of the pulse). The report content includes: the main pulse characteristic is a wiry, rapid, and tense pulse, with a wiry quality visible in all three positions (cun, guan, and chi); the pulse rate is relatively fast and forceful upon palpation; the deduced pathogenesis is liver yang hyperactivity, accompanied by internal liver wind; considering the patient's symptoms of dizziness, irritability, and bitter taste in the mouth, it is recommended to consider treatment to clear liver fire and calm liver yang.

[0089] The clinical decision support and human-computer interaction unit recommended the Tianma Gouteng Decoction formula and acupoints such as Taichong and Fengchi, and advised the patient to rest and avoid spicy and irritating foods. After reviewing the report, the doctor confirmed the system's diagnosis based on the patient's tongue appearance and other clinical information, and adopted some of the suggestions.

[0090] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0091] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A TCM intelligent diagnostic system based on pulse characteristics, characterized in that, include: The pulse signal acquisition module is used to simultaneously acquire pulse signals from the radial artery at the cun, guan, and chi positions of the subject at multiple points, depths, and high resolution. The pulse wave depth feature extraction and characterization module is connected to the pulse signal acquisition module and is used to perform deep learning and extract multi-dimensional, high-level features that are highly related to the semantics of traditional Chinese medicine pulse patterns from the acquired pulse signal data. The TCM pulse semantic mapping and identification module is connected to the pulse wave depth feature extraction and characterization module. It is used to map the extracted high-dimensional pulse feature vector to a standardized TCM pulse category and provide corresponding identification results and confidence assessment. The personalized adaptation and robustness enhancement module is used to improve the system's adaptability and generalization performance to individual physiological differences, environmental disturbances, and complex pulse patterns. The diagnostic report module is used to present the system's analysis results to users in a structured, intuitive, and easy-to-understand format, and to provide necessary decision support.

2. The system according to claim 1, characterized in that, The pulse signal acquisition module includes at least: An array-type micro pressure sensor unit is composed of multiple microelectromechanical system pressure sensors arranged in a row. The array-type micro pressure sensor unit has at least three independent sensing areas, which correspond to the three pulse positions of Cun, Guan, and Chi respectively. A dynamic pressing device is positioned above the array-type micro pressure sensor unit and maintains adjustable physical contact with the sensing surface of the array-type micro pressure sensor unit.

3. The system according to claim 2, characterized in that, The array-type micro pressure sensor unit uses a piezoresistive microelectromechanical system (MEMS) pressure sensor. Within each sensing area, the MEMS pressure sensors are distributed at preset intervals and in an array to capture fine spatial distribution information of pulse waves in a local area. The MEMS pressure sensor integrates a temperature compensation circuit.

4. The system according to claim 2, characterized in that, The dynamic pressing device includes a precision stepper motor drive unit for providing precisely controllable vertical pressing force; the drive unit is coupled with a force control feedback unit to achieve closed-loop pressing force control; the pressing device also integrates an infrared ranging sensor for real-time monitoring of pressing depth.

5. The system according to claim 2, characterized in that, The pulse wave depth feature extraction and representation module is a feature learning architecture based on a deep neural network, which includes: Multi-scale spatiotemporal convolutional networks, consisting of multiple parallel or cascaded convolutional layers, are used to capture local patterns of pulse waveforms at different time scales and spatial locations. Long Short-Term Memory (LSTM) network layers are set after the multi-scale spatiotemporal convolutional network to capture long-term dependencies and dynamic evolution patterns in the pulse waveform time series. A self-attention mechanism, integrated into the middle or top layer of the feature learning architecture, is used to dynamically evaluate and weight the contribution of different spatiotemporal points and different compression depth channels in the pulse waveform data to the final pulse characteristics. The feature fusion layer, set after the multi-scale spatiotemporal convolutional network, long short-term memory network layers and self-attention mechanism, is used to cascade feature representations from different sub-networks to generate fixed-dimensional, high-dimensional, semantically rich pulse feature vectors.

6. The system according to claim 1, characterized in that, The TCM pulse semantic mapping and identification module includes at least the following: A multi-task learning classifier receives a high-dimensional pulse feature vector as input; the multi-task learning classifier is a multi-layer fully connected neural network. The TCM knowledge graph embedding unit works in collaboration with the multi-task learning classifier. The knowledge graph is pre-built and stores rich TCM theoretical knowledge. The knowledge graph embedding unit uses a knowledge graph reasoning engine to fuse and verify the preliminary pulse recognition results output by the multi-task learning classifier with the semantic information in the knowledge graph.

7. The system according to claim 1, characterized in that, The personalized adaptation and robustness enhancement module includes at least: a transfer learning and domain adaptation unit, an adversarial training and data augmentation unit, and an uncertainty quantification and anomaly detection unit; wherein... The data augmentation in the adversarial training and data augmentation unit includes: time stretching and compression of the pulse waveform, random scaling of amplitude, random shifting, and generating synthetic pulse data through a generative adversarial network; The uncertainty quantification and anomaly detection unit is based on Bayesian deep learning and provides a quantified uncertainty measure for pulse recognition results.

Images