Sub-health state evaluation modeling method for terahertz multi-modal data

Abstract

The invention provides a modeling method for evaluating the sub-health state of terahertz multi-modal data, and belongs to the technical field of terahertz multi-modal data processing and sub-health state evaluation.The method comprises the steps that terahertz time-domain spectroscopy, terahertz imaging and physiological sign data are synchronously collected, and targeted denoising, normalization and time alignment preprocessing are conducted on the terahertz time-domain spectroscopy, the terahertz imaging and the physiological sign data; key features of each modal are extracted, and feature vectors are generated through attention mechanism fusion; an evaluation model is constructed based on an improved algorithm, evaluation results such as sub-health risk levels are output after training verification of a classification sample set, and the model is periodically and incrementally updated. The method is applied to special evaluation equipment, and all modules are linked according to data logic. The method makes up for the limitation of single data evaluation, improves the data quality and evaluation accuracy, guarantees the stable evaluation effect, achieves non-invasive rapid evaluation, promotes technical engineering landing, and provides support for sub-health recognition and intervention.

Description

Technical Field

[0001] This invention relates to the field of terahertz multimodal data processing and sub-health status assessment technology, specifically to a sub-health status assessment modeling method based on terahertz multimodal data. Background Technology

[0002] Current assessments of sub-health status often rely on single detection methods or limited data types, making it difficult to comprehensively reflect the overall state of the human body. Some assessment methods only use physiological signs data, lacking the capture of microscopic features of human tissues; others rely solely on a single terahertz data mode, failing to incorporate the dynamic changes in physiological functions. Furthermore, existing technologies lack unified standards for data collection, resulting in poor handling of heterogeneous data from different sources. Cross-modal feature fusion is often superficial, failing to deeply explore the complementary relationships between data. In addition, assessment models often employ general algorithms, making it difficult to adapt to the characteristics of multimodal data, and lacking continuous optimization mechanisms, leading to a decline in assessment effectiveness over long-term use. At the engineering application level, the lack of integrated equipment specifically adapted to these assessment methods hinders the effective implementation and promotion of the technology, failing to meet the needs of rapid, non-invasive, and accurate assessment of sub-health status in real-world scenarios. Summary of the Invention

[0003] The present invention aims to solve the problems mentioned in the background art by providing a sub-health status assessment and modeling method based on terahertz multimodal data.

[0004] The specific technical solution is as follows: A sub-health state assessment modeling method based on terahertz multimodal data includes the following steps: (1) Terahertz multimodal data acquisition: Acquire terahertz time-domain spectral data, terahertz imaging data, and synchronously acquired physiological data of the subject; (2) Multi-source data preprocessing: The terahertz time-domain spectral data, terahertz imaging data and physiological sign data are respectively denoised, normalized and time-aligned to obtain a standardized multimodal dataset; (3) Cross-modal feature extraction: Extract terahertz time-domain features, terahertz frequency-domain features, terahertz imaging texture features, and physiological sign time-series features from the standardized multimodal dataset, and generate a fused feature vector through a cross-modal fusion algorithm; (4) Evaluation model construction: Based on the fused feature vector, a sub-health status evaluation model is constructed using a deep learning algorithm, and the evaluation model is trained and validated using a set of labeled sub-health status samples; (5) Sub-health status assessment: Input the terahertz multimodal data of the object to be assessed into the trained sub-health status assessment model, and output the corresponding sub-health status assessment results.

[0005] As a preferred embodiment of the present invention, in step (1), the terahertz time-domain spectral data is acquired by a terahertz time-domain spectral acquisition device, with an acquisition frequency band of 0.1THz to 10THz; the terahertz imaging data is acquired by a terahertz area array imaging device, with a resolution of not less than 320×240 pixels; the physiological signs data include at least one of electrocardiogram data, blood oxygen saturation data, and heart rate variability data.

[0006] As a preferred embodiment of the present invention, in step (2), the denoising process uses a wavelet threshold denoising algorithm for terahertz time-domain spectral data, a non-local mean denoising algorithm for terahertz imaging data, and an adaptive filtering denoising algorithm for physiological sign data; the normalization process uses a minimum-maximum normalization method to map each modal data to the [0,1] interval; the time alignment process synchronizes the time axis of the multimodal data based on the acquisition timestamp.

[0007] As a preferred embodiment of the present invention, in step (3), the terahertz time-domain features include peak amplitude, rise time, and delay time; the terahertz frequency-domain features include characteristic absorption peak frequency, absorption peak half width at half maximum, and frequency domain energy spectrum; the terahertz imaging texture features include contrast, correlation, and entropy value of the gray-level co-occurrence matrix; and the physiological sign time-series features include RR interval standard deviation and blood oxygen fluctuation amplitude.

[0008] As a preferred embodiment of the present invention, in step (3), the cross-modal fusion algorithm is a feature fusion algorithm based on the attention mechanism. The multi-head attention module performs weight allocation and correlation modeling on the features of different modalities to generate a fusion feature vector containing multimodal complementary information.

[0009] As a preferred embodiment of the present invention, in step (4), the deep learning algorithm is an improved Transformer algorithm, which introduces modality-aware position coding in the Transformer encoder, and distinguishes the temporal differences between terahertz data and physiological characteristic data through position coding of different modalities, thereby improving the model's ability to model cross-modal features.

[0010] As a preferred embodiment of the present invention, in step (4), the labeled sub-health state sample set includes healthy state samples, mild sub-health samples, moderate sub-health samples, and severe sub-health samples. Each sample is labeled with a TCM constitution type or a Western medicine sub-health symptom label. The training and verification process adopts the five-fold cross-validation method and optimizes the model parameters through a loss function. The loss function is a weighted combination of cross-entropy loss and cosine similarity loss.

[0011] As a preferred embodiment of the present invention, in step (5), the sub-health status assessment results include sub-health risk level, associated abnormal physiological indicators, and personalized intervention suggestions, wherein the sub-health risk level is divided into three levels: low risk, medium risk, and high risk, and the personalized intervention suggestions are generated based on the assessment results combined with the age, gender, and lifestyle of the tested subject.

[0012] As a preferred embodiment of the present invention, the method further includes a model dynamic update step: periodically collecting new terahertz multimodal data and corresponding sub-health status annotation information, incrementally training the sub-health status assessment model, and updating the model parameters to improve the assessment accuracy.

[0013] As a preferred embodiment of the present invention, the method is applied to a terahertz sub-health assessment device, which includes a terahertz data acquisition module, a physiological sign acquisition module, a data processing module, a model calculation module, and a result display module. The method enables non-invasive and rapid assessment of the sub-health status of the tested subject.

[0014] The present invention has the following beneficial effects: The sub-health status assessment and modeling method using terahertz multimodal data provided in this invention comprehensively addresses the shortcomings of existing sub-health assessment technologies by integrating multiple data types, standardizing processing procedures, optimizing fusion and modeling algorithms, and combining them with specialized equipment. The collaborative application of multimodal data overcomes the information limitations of single data types, enabling a more comprehensive reflection of human health and improving the completeness and reliability of assessment results. Targeted preprocessing and feature extraction methods effectively improve the quality of data and features, providing solid support for subsequent modeling. Attention-based fusion algorithms and improved modeling algorithms can deeply mine data correlations, adapt to the characteristics of multimodal data, and improve the accuracy and adaptability of the assessment. The dynamic model update mechanism avoids performance degradation after long-term use, ensuring the stability of the assessment results. The integration of specialized equipment and modular collaborative design realizes the engineering implementation of the technology, meeting the needs of non-invasive and rapid practical applications, breaking through the scenario limitations of traditional assessment methods, and improving the practicality and widespread applicability of the technology. Overall, this method constructs a complete, efficient, and stable sub-health status assessment system, providing strong support for the accurate identification and intervention guidance of sub-health status. Attached Figure Description

[0015] Figure 1 A flowchart illustrating the sub-health state assessment modeling method using terahertz multimodal data provided in this embodiment of the invention; Figure 2 The curve showing the change in assessment accuracy over time for the sub-health status assessment modeling method based on terahertz multimodal data provided in this embodiment of the invention; Figure 3A multimodal data contribution weight distribution diagram for the sub-health status assessment modeling method using terahertz multimodal data provided in this embodiment of the invention; Figure 4 A bar chart comparing the feature extraction performance of the sub-health status assessment modeling method for terahertz multimodal data provided in this embodiment of the invention; Figure 5 A comparison chart showing the dynamic update effect of the model for the sub-health status assessment modeling method using terahertz multimodal data provided in this embodiment of the invention. Detailed Implementation

[0016] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0017] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual images. They should not be construed as limiting the scope of this application. To better illustrate the embodiments of the present invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0018] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "inner," and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present application. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0019] In the description of this invention, unless otherwise explicitly specified and limited, the term "connection" or similar designation indicating a connection between components should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0020] Reference Figure 1-5 ,in, Figure 1 The process of modeling a sub-health status assessment method using terahertz multimodal data is demonstrated. Figure 2The model's evaluation accuracy changes at different time points (January / March / June / December), showing how the accuracy improved from 85% to 92% through regular incremental updates; Figure 3 The weights assigned based on the attention mechanism show that physiological signs data (0.45) contribute the most to the evaluation results, followed by terahertz time-domain spectroscopy (0.3). Figure 4 The best performance in displaying the temporal features of physiological signs was achieved with an F1 score of 0.85, followed by terahertz frequency domain features (0.8) and imaging texture features (0.65). Figure 5 The results show that in four typical scenarios, the accuracy of the updated model is improved by an average of 6-7 percentage points compared with the initial model, which verifies the effectiveness of incremental training.

[0021] The sub-health state assessment modeling method based on terahertz multimodal data provided in this embodiment, such as... Figures 1-5 As shown, it includes the following steps: (1) Terahertz multimodal data acquisition: Acquire terahertz time-domain spectral data, terahertz imaging data, and synchronously acquired physiological data of the subject; (2) Multi-source data preprocessing: Denoising, normalization and time alignment are performed on terahertz time-domain spectral data, terahertz imaging data and physiological sign data respectively to obtain a standardized multimodal dataset; (3) Cross-modal feature extraction: Extract terahertz time-domain features, terahertz frequency-domain features, terahertz imaging texture features, and physiological sign time-series features from the standardized multimodal dataset, and generate a fused feature vector through a cross-modal fusion algorithm; (4) Assessment model construction: Based on the fusion feature vector, a sub-health status assessment model is constructed using a deep learning algorithm. The assessment model is trained and validated using a set of labeled sub-health status samples. (5) Sub-health status assessment: Input the terahertz multimodal data of the subject to be assessed into the trained sub-health status assessment model, and output the corresponding sub-health status assessment results.

[0022] This solution integrates multiple data sources, including terahertz time-domain spectroscopy, terahertz imaging, and physiological characteristics. Through a complete process of preprocessing, feature extraction and fusion, model construction, and state assessment, it provides multi-dimensional support for sub-health state assessment. Compared to assessment methods based on single data types, it effectively compensates for the limitations of single-data-information approaches, providing a more comprehensive reflection of the subject's physical condition and improving the reliability and completeness of the assessment results.

[0023] Specifically, in this embodiment, in step (1), terahertz time-domain spectral data is acquired through a terahertz time-domain spectral acquisition device with a acquisition frequency band of 0.1THz to 10THz; terahertz imaging data is acquired through a terahertz area array imaging device with a resolution of not less than 320×240 pixels; physiological sign data includes at least one of electrocardiogram data, blood oxygen saturation data, and heart rate variability data. This scheme clarifies the acquisition devices and key parameters for terahertz time-domain spectral data and terahertz imaging data, and defines the scope of physiological sign data. By standardizing the selection of equipment and parameter standards for data acquisition, the stability and effectiveness of the acquired data are ensured, providing a high-quality data foundation for subsequent multi-source data processing, feature extraction, and model construction, and ensuring the smooth progress and effective implementation of the entire evaluation process.

[0024] Specifically, in this embodiment, in step (2), the denoising process uses a wavelet threshold denoising algorithm for terahertz time-domain spectral data, a non-local mean denoising algorithm for terahertz imaging data, and an adaptive filtering denoising algorithm for physiological characteristic data; the normalization process uses a minimum-maximum normalization method to map each modal data to the [0,1] interval; and the time alignment process synchronizes and calibrates the time axis of the multimodal data based on the acquisition timestamp. This scheme adopts appropriate denoising algorithms, normalization methods, and time alignment methods according to the characteristics of different modal data. It can effectively reduce the interference of various noises on the data, unify the scale standards of data from different sources, achieve accurate matching of multimodal data in the time dimension, reduce the negative impact of data heterogeneity, significantly improve the quality of standardized datasets, and provide a guarantee for the accuracy of subsequent feature extraction.

[0025] Specifically, in this embodiment, in step (3), the terahertz time-domain features include peak amplitude, rise time, and delay time; the terahertz frequency-domain features include the frequency of the characteristic absorption peak, the full width at half maximum (FWHM) of the absorption peak, and the frequency-domain energy spectrum; the terahertz imaging texture features include the contrast, correlation, and entropy value of the gray-level co-occurrence matrix; and the physiological sign time-series features include the standard deviation of the RR interval and the amplitude of blood oxygen fluctuations. This scheme clarifies the core feature types of terahertz time-domain, frequency-domain, imaging texture, and physiological sign time-series. By systematically extracting key information from each modality of data, redundant interference from invalid features is avoided, and the feature patterns related to sub-health status can be accurately captured. This provides targeted feature support for cross-modal feature fusion and assessment model construction, helping to improve the accuracy of subsequent assessments.

[0026] Specifically, in this embodiment, in step (3), the cross-modal fusion algorithm is a feature fusion algorithm based on an attention mechanism. A multi-head attention module is used to assign weights and model the correlations of features from different modalities, generating a fused feature vector containing complementary information from multiple modalities. This scheme employs a cross-modal fusion algorithm based on an attention mechanism, using a multi-head attention module to achieve weight assignment and correlation modeling of features from different modalities. It can dynamically adjust the weights according to the correlation between each modal feature and the sub-health assessment, deeply mining the complementary information and intrinsic correlations between different modal data, avoiding the limitations of single-modal features, making the generated fused features more suitable for the needs of sub-health status assessment, and improving the feature representation ability.

[0027] Specifically, in this embodiment, in step (4), the deep learning algorithm is an improved Transformer algorithm. Modality-aware position encoding is introduced into the Transformer encoder. By using position encoding of different modalities, the temporal differences between terahertz data and physiological characteristic data are distinguished, thereby improving the model's ability to model cross-modal features. This scheme adopts an improved Transformer algorithm and introduces a modality-aware position encoding mechanism. It can effectively distinguish the differences between terahertz data and physiological characteristic data in the temporal dimension, solve the modeling adaptation problem caused by the inconsistency of temporal characteristics of different modal data, strengthen the model's ability to capture and model cross-modal features, and improve the model's adaptability and processing effect on complex multimodal data.

[0028] Specifically, in this embodiment, in step (4), the labeled sub-health state sample set includes healthy state samples, mild sub-health samples, moderate sub-health samples, and severe sub-health samples. Each sample is labeled with a TCM constitution type or a Western medicine sub-health symptom label. The training and validation process adopts the five-fold cross-validation method, and the model parameters are optimized through a loss function, which is a weighted combination of cross-entropy loss and cosine similarity loss. This scheme clarifies the classification criteria of the labeled sample set, adopts the five-fold cross-validation method, and optimizes the model parameters through a weighted combination of loss functions. It can provide comprehensive and targeted sample support for model training, reduce overfitting during training, improve the stability and reliability of model training, enable the model to better adapt to different sub-health state differentiation scenarios, and enhance the model's generalization ability.

[0029] Specifically, in this embodiment, step (5) includes the sub-health status assessment results, including the sub-health risk level, associated abnormal physiological indicators, and personalized intervention suggestions. The sub-health risk level is divided into three levels: low risk, medium risk, and high risk. Personalized intervention suggestions are generated based on the assessment results combined with the subject's age, gender, and lifestyle habits. This scheme clarifies the content covered by the sub-health status assessment results, including risk level, abnormal physiological indicators, and personalized intervention suggestions. It can provide comprehensive and intuitive assessment information for the subjects and related users, not only clearly understanding their sub-health status but also enabling targeted interventions based on the given suggestions, thus improving the practicality and guiding value of the assessment results.

[0030] Specifically, this embodiment also includes a model dynamic update step: periodically collecting new terahertz multimodal data and corresponding sub-health state annotation information, incrementally training the sub-health state assessment model, and updating model parameters to improve assessment accuracy. This solution adds a model dynamic update step by periodically collecting new data and annotation information to incrementally train the model. This enables the model to adapt to new sample data features and changes in actual application scenarios in a timely manner, continuously optimize model parameters, avoid performance degradation after long-term use, and ensure that the model maintains stable assessment results at different times.

[0031] Specifically, in this embodiment, the sub-health status assessment modeling method based on terahertz multimodal data is applied to a terahertz sub-health assessment device. The device includes a terahertz data acquisition module, a physiological sign acquisition module, a data processing module, a model calculation module, and a result display module. This method enables non-invasive and rapid sub-health status assessment of the tested subject. The scheme applies the method to a dedicated terahertz sub-health assessment device, specifying that the device includes a terahertz data acquisition module, a physiological sign acquisition module, a data processing module, a model calculation module, and a result display module. Each module is connected sequentially according to data transmission and processing logic: the outputs of the terahertz data acquisition module and the physiological sign acquisition module are connected to the input of the data processing module; the output of the data processing module is connected to the input of the model calculation module; and the output of the model calculation module is connected to the input of the result display module. This modular connection design is deeply adapted to the methodology, enabling orderly data transmission, efficient processing, and accurate result presentation. It promotes the engineering implementation and practical application of the method. Through the integrated design of the equipment, it achieves non-invasive and rapid evaluation of the tested object, breaking through the scenario limitations of traditional evaluation methods and improving the practicality and popularity of the method.

[0032] Specifically, in this embodiment, the cross-modal fusion algorithm adopts the terahertz-physiological cross-modal attention fusion equation (THz-PhysioCross-ModalAttentionFusion, THz-CMAF), the expression of which is as follows: ; ; ; in: The fused feature vector; For the i-th terahertz mode feature (time domain, frequency domain, imaging texture, etc.); g is the temporal feature vector of physiological signs; and These are nonlinear mapping functions for terahertz characteristics and physiological characteristics, respectively, implemented through a fully connected layer; The attention weight for the i-th terahertz feature; β is the fusion weight of physiological features, which can be set to a learnable parameter or a fixed value (e.g., 0.2–0.4). γ is a temperature coefficient that controls the degree of sharpening of the weight distribution; The feature similarity function is expressed as the complement of the normalized Euclidean distance. For a minimal constant (e.g.) Prevent division by zero; N represents the number of terahertz features.

[0033] The derivation of the equation is as follows: 1. Basic attention mechanism: Starting from the standard attention formula, the weights are determined by the relevance of features: ; However, direct application of dot product attention is not suitable for heterogeneous modalities.

[0034] 2. Introduce intermodal similarity: Use a distance-based similarity metric to accommodate the dimensional differences between terahertz and physiological data. ; The function takes values ​​in the range [0, 1], with values ​​closer to 1 indicating greater similarity.

[0035] 3. Temperature coefficient adjustment: Introducing γ to control the concentration of the weight distribution and avoid the weights being too evenly distributed. .

[0036] 4. Construction of fusion equation: Weighted summation of features after modality mapping, while retaining the independent contribution terms of physiological features. This enhances the model's sensitivity to physiological changes.

[0037] For example, suppose the following features are extracted: Terahertz time-domain characteristics Frequency domain characteristics Imaging texture features Physiological temporal characteristics g (such as statistics of heart rate variability sequences).

[0038] 1. Calculate similarity: , , .

[0039] 2. Set γ=2.0 and calculate the weights: ≈0.52, ≈0.30, ≈0.18.

[0040] 3. Set β=0.3, and perform weighted fusion after mapping through a fully connected layer: .

[0041] 4. Output Input into the subsequent Transformer evaluation model.

[0042] Parameter description table: .

[0043] Technical effect 1. Improve fusion quality: By using similarity weighting, terahertz features that are more relevant to physiological states are given higher weights, thereby enhancing feature complementarity; 2. Solving the heterogeneity problem: Normalized distance metric avoids dimensional differences and is suitable for direct comparison between terahertz and physiological data; 3. Enhance interpretability: weighting It can reflect the contribution of each terahertz mode to the current physiological state; 4. Strong compatibility: It can be flexibly extended to more terahertz sub-modes (such as polarization and phase imaging features).

[0044] Working principle and process 1. Input: Standardized multimodal feature set ; 2. Similarity calculation: For each Calculate the Sim value of g; 3. Weighting: Attention weights are obtained by using softmax plus a temperature coefficient γ. ; 4. Feature mapping: respectively through ψ maps terahertz and physiological characteristics to the same latent space; 5. Weighted fusion: based on weights A linear combination with β yields the output. ; 6. Output: The fused vector is fed into the evaluation model for sub-health status classification.

[0045] Working Principle: This method achieves accurate assessment of sub-health status through the integrated processing and modeling analysis of multimodal data, relying on a collaborative mechanism of data flow and model computation. First, a dedicated module simultaneously collects three types of data: terahertz time-domain spectroscopy, terahertz imaging, and physiological signs, providing multi-dimensional information support for the assessment. The collected multi-source data undergoes targeted preprocessing to remove various noise interferences, standardize data scales, and achieve precise alignment in the time dimension, eliminating the impact of data heterogeneity. Subsequently, key features related to sub-health status for each modality are extracted from the standardized data. Then, an attention-driven fusion algorithm dynamically allocates the feature weights for each modality, uncovers the intrinsic correlations between features, and generates fusion features that meet the assessment requirements. Based on the fusion features, an improved modeling algorithm is constructed, trained and validated using a well-defined sample set, and the model parameters are optimized to improve the model's ability to distinguish sub-health status. Finally, the data to be assessed is input into the trained model, and the assessment results are output and displayed through a dedicated module. Simultaneously, new data is collected periodically to incrementally update the model, ensuring its long-term stability. At the device level, the modules are connected according to the data transmission logic. The acquisition module transmits data to the processing module, the processed feature data is sent to the modeling module for calculation, and the final result is output by the display module, realizing efficient collaboration throughout the entire process.

[0046] How to use: 1. Equipment Deployment and Preparation: Set up the terahertz sub-health assessment equipment, ensuring that the terahertz data acquisition module, physiological sign acquisition module, data processing module, model calculation module, and result display module are assembled in place according to their connection relationships, complete the equipment debugging and initialization, and ensure that each module works normally and collaboratively.

[0047] 2. Multimodal data acquisition: Start the device and acquire terahertz time-domain spectral data and terahertz imaging data of the subject through the terahertz time-domain spectral acquisition device and the terahertz area array imaging device, respectively. At the same time, acquire physiological sign data such as electrocardiogram and blood oxygen saturation through the physiological sign acquisition component to ensure that the data acquisition process meets the preset parameter standards.

[0048] 3. Multi-source data preprocessing: The device automatically performs targeted processing on the three types of data collected. Wavelet thresholding is used to denoise the terahertz time-domain spectral data, non-local mean denoising is used for the terahertz imaging data, and adaptive filtering denoising is used for the physiological sign data. The data scale is unified by the min-max normalization method, and the time alignment of the multimodal data is completed based on the collection timestamp to generate a standardized multimodal dataset.

[0049] 4. Feature Extraction and Fusion: The system extracts terahertz time-domain, frequency-domain, imaging texture, and physiological characteristics temporal features from a standardized dataset. Through a cross-modal fusion algorithm based on an attention mechanism, it completes the weight allocation and correlation modeling of each modality feature and generates a fused feature vector.

[0050] 5. Model Training and Validation (for first use or updates): Import sample sets labeled with healthy and sub-healthy states of varying degrees, construct an evaluation model using an improved modeling algorithm, optimize model parameters using a weighted combination loss function through five-fold cross-validation, and complete model training and validation.

[0051] 6. Sub-health status assessment: Input the preprocessed data of the subject to be assessed and the extracted and fused feature vectors into the trained model. The model automatically calculates and outputs the sub-health risk level, abnormal physiological indicators and personalized intervention suggestions, which are presented to the user through the results display module.

[0052] 7. Dynamic Model Update: New terahertz multimodal data and corresponding sub-health status annotation information are collected periodically and input into the existing model for incremental training to update model parameters and ensure the continuity and stability of model evaluation results.

[0053] In summary, the sub-health state assessment modeling method based on terahertz multimodal data provided in this embodiment has the following advantages: This method comprehensively addresses the shortcomings of existing sub-health assessment technologies by integrating multiple data types, standardizing processing procedures, optimizing fusion and modeling algorithms, and combining them with specialized equipment. The collaborative application of multimodal data overcomes the information limitations of single data types, providing a more comprehensive reflection of the human body's state and improving the completeness and reliability of assessment results. Targeted preprocessing and feature extraction methods effectively improve the quality of data and features, providing solid support for subsequent modeling. Attention-based fusion algorithms and improved modeling algorithms can deeply mine data correlations, adapt to the characteristics of multimodal data, and improve the accuracy and adaptability of the assessment. The dynamic model update mechanism avoids performance degradation after long-term use, ensuring the stability of the assessment results. The integration of specialized equipment and modular collaborative design realizes the engineering implementation of the technology, meeting the needs of non-invasive and rapid practical applications, breaking through the scenario limitations of traditional assessment methods, and improving the practicality and accessibility of the technology. Overall, this method constructs a complete, efficient, and stable sub-health state assessment system, providing strong support for the accurate identification and intervention guidance of sub-health states.

[0054] In addition, this application also provides examples of sub-health state assessment modeling methods using terahertz multimodal data. I. Application Scenarios This example is applied to a sub-health screening scenario for adults in a community health service center. It targets individuals aged 18-65 without serious organic diseases and achieves accurate assessment and intervention guidance for their sub-health status through non-invasive and rapid data collection and analysis.

[0055] II. Technical Solution (a) Implementation preparation 1. Equipment Setup: A customized terahertz sub-health assessment device is used. The device includes a terahertz data acquisition module (integrating a terahertz time-domain spectral acquisition device and a terahertz area array imaging device), a physiological sign acquisition module, a data processing module, a model calculation module, and a results display module. The connections between the modules are as follows: the outputs of the terahertz data acquisition module and the physiological sign acquisition module are connected to the input of the data processing module via a USB 3.0 high-speed interface. The data processing module communicates with the model calculation module via a PCIe bus. The outputs of the model calculation module are connected to the display screen and printing unit of the results display module, ensuring efficient collaboration throughout the entire process of data acquisition, processing, calculation, and display.

[0056] 2. Sample Set Preparation: Sample data were collected from 2000 subjects, including 500 healthy subjects, 600 subjects with mild sub-health, 500 subjects with moderate sub-health, and 400 subjects with severe sub-health. All samples were labeled with their TCM constitution type (balanced, Qi-deficient, etc.) and Western medicine sub-health symptom labels (fatigue, sleep disorders, etc.), excluding patients with serious diseases.

[0057] (II) Terahertz Multimodal Data Acquisition The device is started to collect three types of data simultaneously, with a uniform collection time of 5 seconds: 1. Terahertz time-domain spectral data: Data on the inner wrist skin of the subject was collected in the frequency band of 0.1THz to 10THz using a terahertz time-domain spectral acquisition device, with a sampling frequency of 1kHz. 2. Terahertz imaging data: Images of a 2cm×2cm area on the inside of the wrist were acquired using a terahertz area array imaging device, with a resolution of 320×240 pixels; 3. Physiological data: Simultaneous acquisition of electrocardiogram data (sampling frequency 250Hz), blood oxygen saturation data (sampling frequency 1Hz), and heart rate variability data (sampling frequency 250Hz).

[0058] (III) Multi-source data preprocessing The data processing module automatically performs the following operations: 1. Denoising: Terahertz time-domain spectral data are denoised using a wavelet thresholding algorithm based on the db4 wavelet basis; terahertz imaging data are denoised using a nonlocal mean denoising algorithm with a 7×7 search window and a 5×5 similar window; and physiological sign data are denoised using an adaptive filtering algorithm to filter out various types of noise. 2. Normalization: By using the min-max normalization method, all modal data are mapped to the [0,1] interval to unify the data scale; 3. Time alignment processing: Based on the acquisition timestamp with 1ms accuracy, the time axis of multimodal data is calibrated to achieve accurate matching of data at the same time point and generate a standardized multimodal dataset.

[0059] (iv) Cross-modal feature extraction Extracting key features from a standardized dataset: 1. Terahertz time-domain characteristics: Calculate peak amplitude, rise time, and delay time; 2. Terahertz frequency domain characteristics: The time domain data is converted into the frequency domain through Fourier transform, and the characteristic absorption peak frequency, absorption peak half width at half maximum (FWHM), and frequency domain energy spectrum are extracted. 3. Terahertz imaging texture features: Based on the gray-level co-occurrence matrix, calculate contrast, correlation, and entropy. 4. Physiological signs and temporal characteristics: Extract the standard deviation of the RR interval and the amplitude of blood oxygen fluctuations to form a multi-dimensional feature set.

[0060] (v) Cross-modal feature fusion The THz-CMAF fusion algorithm, based on an attention mechanism, is used to generate the fused feature vector through the following equation: ; ; ; Where f1 represents the terahertz time-domain feature, f2 represents the terahertz frequency-domain feature, f3 represents the terahertz imaging texture feature, and g represents the physiological sign temporal feature. ψ and β are the nonlinear mapping functions implemented in the fully connected layer; β=0.3, γ=2.0, =10−8. This algorithm dynamically allocates the weights for each modality to generate a fused feature vector.

[0061] (vi) Evaluation Model Construction 1. Model Structure: The model is built based on the improved Transformer algorithm, with 6 encoder layers and 4 attention heads. Modality-aware position coding is introduced (0.6 coding coefficient for terahertz data and 0.4 coding coefficient for physiological sign data) to distinguish the temporal differences between the two types of data. 2. Training and Validation: The samples were divided into a training set (1400 samples) and a validation set (600 samples) in a 7:3 ratio. Five-fold cross-validation was used. The parameters were optimized by a weighted combination of cross-entropy loss (weight 0.7) and cosine similarity loss (weight 0.3). The Adam optimizer (learning rate 0.001, decay coefficient 0.99) was selected and trained for 100 rounds until the indicators stabilized. 3. Model solidification: After verification and compliance, the parameters are solidified for subsequent evaluation.

[0062] (vii) Assessment of sub-health status After the multimodal data of the subjects to be evaluated are preprocessed, feature extracted and fused as described above, they are input into the model. The output includes three parts: sub-health risk level (low, medium and high risk), associated abnormal physiological indicators (such as abnormal blood oxygen fluctuations, characteristic absorption peak shifts, etc.), and personalized intervention suggestions (generated in combination with age, gender and lifestyle habits), which are presented through the display screen and printing unit.

[0063] (viii) Dynamic Model Update Every 3 months, multimodal data and annotation information of 100 new subjects are collected and added to the original sample set to incrementally train the model (20 iterations, learning rate 0.0005), update the model parameters, and continuously optimize and evaluate the performance.

[0064] III. Working Principle The core logic of this example is to achieve accurate assessment through the synergistic complementarity of multimodal data and standardized processing throughout the entire process: 1. Data Acquisition Phase: Relying on specialized equipment, terahertz data (reflecting tissue microstructure) and physiological sign data (reflecting dynamic changes in function) are acquired simultaneously to provide comprehensive information support for the assessment. The various modules of the equipment are connected according to the data flow logic to ensure efficient and accurate data transmission. 2. Preprocessing stage: Adaptive methods are used to eliminate noise, unify scale, and calibrate time for different modal data characteristics, thereby reducing data heterogeneity interference and providing a high-quality data foundation for subsequent processing; 3. Feature extraction and fusion stage: Extract key features of each modality, dynamically assign weights based on similarity using the THz-CMAF algorithm, explore the intrinsic correlation between modalities, enhance feature representation capabilities, and make the fused features more in line with the needs of sub-health assessment; 4. Model building stage: The improved Transformer adapts to multimodal temporal differences through modality-aware positional encoding, and combines a well-defined sample set with a combined loss function to improve the model's ability to distinguish different sub-health states. Five-fold cross-validation reduces overfitting. 5. Evaluation and Update Phase: The trained model outputs comprehensive evaluation results based on the learned feature patterns. Regular incremental updates allow the model to adapt to new sample features, avoiding performance degradation after long-term use and ensuring evaluation stability.

[0065] IV. Experimental Data Evaluation results: In the validation set test, the recognition effect of healthy status, mild sub-health, moderate sub-health and severe sub-health was good. The overall evaluation accuracy, recall and F1 score were better than the single modality evaluation method. Comparative experimental data: Compared with the method using only terahertz time-domain spectral data, the overall assessment accuracy is significantly improved; compared with the method using only physiological sign data, the overall assessment accuracy and F1 score are greatly improved. Stability data: In the five-fold cross-validation, the overall evaluation accuracy of each fold fluctuated little, and the model training was stable; after dynamic updates, the evaluation accuracy of new samples improved compared with before the update, and after 6 months of continuous tracking tests, the accuracy did not decrease significantly. Engineering application data: The complete evaluation process for a single case takes no more than 2 minutes, meeting the needs of rapid screening; the data collection process is non-invasive, the tested subjects are comfortable, and there are no adverse reactions.

[0066] V. Technical Effects Equipment setup and standardized data collection: By clarifying the connection relationships of equipment modules and the standards for data collection parameters, the stability and validity of the data were ensured, providing reliable hardware support and a high-quality data foundation for the entire evaluation process, and guaranteeing the smooth progress of the evaluation process; Targeted preprocessing: Adaptive denoising, normalization, and time alignment methods are adopted for different modal data to effectively reduce noise interference, unify data scale, achieve accurate time matching, reduce the impact of heterogeneity, significantly improve the quality of standardized datasets, and ensure the accuracy of feature extraction. Feature extraction and fusion: The system extracts key features of each modality to avoid redundant interference. It dynamically allocates weights and mines modal correlations through the THz-CMAF fusion algorithm, breaking through the limitations of a single modality and enhancing the feature representation capability, thus laying the foundation for accurate model evaluation. Improved model construction: The Transformer algorithm with modality-aware position encoding is introduced to adapt to multimodal temporal differences. Combined with classification sample sets, combined loss functions and five-fold cross-validation, overfitting is reduced and the model's generalization ability and evaluation accuracy are improved. Comprehensive assessment results: The output includes risk level, abnormal indicators and personalized suggestions, providing users with intuitive and practical guidance information and enhancing the application value of the assessment results; Dynamic update mechanism: Regular incremental training enables the model to adapt to new sample features and scene changes in a timely manner, avoids performance degradation, and ensures the long-term stability of evaluation results; Engineering implementation: The integrated design of specialized equipment and efficient processes enable non-invasive and rapid assessment, breaking through the limitations of traditional assessment scenarios, improving the practicality and accessibility of the technology, and building a complete, efficient, and stable sub-health assessment system.

[0067] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the description and illustrations of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for modeling sub-health status assessment using terahertz multimodal data, characterized in that, Includes the following steps: (1) Terahertz multimodal data acquisition: Acquire terahertz time-domain spectral data, terahertz imaging data, and synchronously acquired physiological data of the subject; (2) Multi-source data preprocessing: The terahertz time-domain spectral data, terahertz imaging data and physiological sign data are respectively denoised, normalized and time-aligned to obtain a standardized multimodal dataset; (3) Cross-modal feature extraction: Extract terahertz time-domain features, terahertz frequency-domain features, terahertz imaging texture features, and physiological sign time-series features from the standardized multimodal dataset, and generate a fused feature vector through a cross-modal fusion algorithm; (4) Evaluation model construction: Based on the fused feature vector, a sub-health status evaluation model is constructed using a deep learning algorithm, and the evaluation model is trained and validated using a set of labeled sub-health status samples; (5) Sub-health status assessment: Input the terahertz multimodal data of the object to be assessed into the trained sub-health status assessment model, and output the corresponding sub-health status assessment results.

2. The sub-health state assessment and modeling method using terahertz multimodal data according to claim 1, characterized in that, In step (1), the terahertz time-domain spectral data is acquired by a terahertz time-domain spectral acquisition device with a acquisition frequency band of 0.1THz to 10THz; the terahertz imaging data is acquired by a terahertz area array imaging device with a resolution of not less than 320×240 pixels; the physiological signs data include at least one of electrocardiogram data, blood oxygen saturation data, and heart rate variability data.

3. The sub-health state assessment and modeling method using terahertz multimodal data according to claim 1, characterized in that, In step (2), the denoising process uses a wavelet threshold denoising algorithm for terahertz time-domain spectral data, a nonlocal mean denoising algorithm for terahertz imaging data, and an adaptive filtering denoising algorithm for physiological sign data; the normalization process uses a minimum-maximum normalization method to map each modal data to the [0,1] interval; and the time alignment process synchronizes and calibrates the time axis of the multimodal data based on the acquisition timestamp.

4. The sub-health state assessment and modeling method using terahertz multimodal data according to claim 1, characterized in that, In step (3), the terahertz time-domain features include peak amplitude, rise time, and delay time; the terahertz frequency-domain features include characteristic absorption peak frequency, absorption peak half width at half maximum (FWHM), and frequency-domain energy spectrum; the terahertz imaging texture features include the contrast, correlation, and entropy value of the gray-level co-occurrence matrix. The temporal characteristics of the physiological signs include the standard deviation of the RR interval and the amplitude of blood oxygen fluctuations.

5. The sub-health state assessment and modeling method using terahertz multimodal data according to claim 1, characterized in that, In step (3), the cross-modal fusion algorithm is a feature fusion algorithm based on the attention mechanism. It uses a multi-head attention module to perform weight allocation and correlation modeling on features of different modalities, and generates a fusion feature vector containing complementary information of multiple modalities.

6. The sub-health state assessment and modeling method based on terahertz multimodal data according to claim 1, characterized in that, In step (4), the deep learning algorithm is an improved Transformer algorithm. Modality-aware position coding is introduced into the Transformer encoder. The temporal differences between terahertz data and physiological characteristic data are distinguished by position coding of different modalities, thereby improving the model's ability to model cross-modal features.

7. The sub-health state assessment and modeling method using terahertz multimodal data according to claim 1, characterized in that, In step (4), the labeled sub-healthy state sample set includes healthy state samples, mild sub-healthy samples, moderate sub-healthy samples, and severe sub-healthy samples. Each sample is labeled with a TCM constitution type or a Western medicine sub-healthy symptom label. The training and verification process adopts the five-fold cross-validation method and optimizes the model parameters through a loss function. The loss function is a weighted combination of cross-entropy loss and cosine similarity loss.

8. The sub-health state assessment and modeling method based on terahertz multimodal data according to claim 1, characterized in that, In step (5), the sub-health status assessment results include sub-health risk level, associated abnormal physiological indicators, and personalized intervention suggestions. The sub-health risk level is divided into three levels: low risk, medium risk, and high risk. The personalized intervention suggestions are generated based on the assessment results combined with the age, gender, and lifestyle of the tested subject.

9. The sub-health state assessment modeling method using terahertz multimodal data according to claim 1, characterized in that, It also includes a model dynamic update step: periodically collecting new terahertz multimodal data and corresponding sub-health status annotation information, incrementally training the sub-health status assessment model, and updating the model parameters to improve the assessment accuracy.

10. The sub-health state assessment and modeling method based on terahertz multimodal data according to any one of claims 1-9, characterized in that, The method is applied to a terahertz sub-health assessment device, which includes a terahertz data acquisition module, a physiological sign acquisition module, a data processing module, a model calculation module, and a result display module. The method enables non-invasive and rapid assessment of the sub-health status of the tested subjects.

Images