Alzheimer's disease prediction method and system based on millimeter wave radar

Abstract

The application discloses an Alzheimer's disease prediction method and system based on a millimeter wave radar. The method comprises the following steps: reorganizing original data of human gait collected by the millimeter wave radar to obtain a three-dimensional matrix block, and operating the three-dimensional matrix block to obtain multi-modal data; performing feature extraction, alignment and fusion on the multi-modal data and text data to obtain reweighted multi-modal data; training a GLM by using the multi-modal data and the text data to obtain a trained GLM, and inputting the reweighted multi-modal data into the model to obtain signal semantic features; inputting the signal semantic features and random noise into a diffusion model and training the diffusion model to obtain a trained diffusion model; and generating an image result by using the trained diffusion model, and combining a text prediction result to assist a user in predicting Alzheimer's disease. The application can combine the advantages of millimeter waves and multi-modal large models, improve the robustness and generalization ability of the model, and improve the early prediction ability of Alzheimer's disease.

Description

Technical Field

[0001] This invention relates to the fields of wireless sensing, multimodal large models, and disease prediction, specifically to an Alzheimer's disease prediction method and system based on millimeter-wave radar. Background Technology

[0002] Clinical reports generally indicate that gait analysis serves as an important clinical biomarker for cognitive decline, demonstrating significant clinical value in the early identification and predictive assessment of Alzheimer's disease. Currently, numerous clinical gait and balance analysis tests are available to assess balance ability and detect Alzheimer's disease. However, these analyses lack reliable external devices capable of accurately measuring gait and quantifying balance parameters. Traditional methods using cameras, foot switches, or electronic pads are limited to laboratory settings. Existing analyses are often constrained by accuracy issues and are difficult to deploy and widely implement, hindering the expansion of clinically meaningful gait parameters through more precise gait kinematic analysis.

[0003] With the development of artificial intelligence, deep learning combined with wearable devices has shown a recognition ability far exceeding that of traditional methods. By collecting patients' gait data through inertial measurement units and inputting it into artificial intelligence models for automatic prediction, this method requires patients to correctly install wearable devices, which is a cumbersome and inefficient process. After use, it requires disinfection and maintenance, which is costly. It is not suitable for the complex physical conditions of different patients and cannot meet the large number of medical needs.

[0004] One current technology involves a method described in the paper "Early Alzheimer's disease diagnosis using wearable sensors and multilevel gait assessment: a machine learning ensemble approach," which uses wearable devices based on inertial sensors combined with artificial intelligence to analyze the gait and balance of Alzheimer's patients. This method designs four patient movement modes: simple walking, obstacle-crossing walking, simultaneous memory and cognitive tasks with walking, and simultaneous complex tasks with walking. It collects acceleration and rotational direction parameters during patient movement using inertial sensors, which are then input into a machine learning classifier for automatic classification and prediction. The drawbacks of this method are that it only focuses on walking and does not reflect balance information; the inertial measurement unit can only acquire limited data; and the wearable device is time-consuming, labor-intensive, and inefficient.

[0005] The second existing technology is a method for identifying Alzheimer's disease based on depth cameras combined with a deep learning model, as described in the paper "Alzheimer's Disease Distinction Based OnGait Feature Analysis". This method places multiple depth cameras at different angles to capture human movement, obtaining real-time skeletal information. The combined data from the multiple cameras is then arranged to preserve spatial information and input into a temporal deep learning framework for analysis. The drawbacks of this method are that it requires multiple cameras at different angles, resulting in a large footprint, high equipment and site costs, susceptibility to lighting interference, and the cameras used in this method are no longer in production. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing methods and propose an Alzheimer's disease prediction method and system based on millimeter-wave radar. The main problem addressed by this invention is how to efficiently, conveniently, and accurately perform Alzheimer's disease screening, while improving data accuracy, reducing site requirements, enhancing the model's learning ability to avoid the intervention of manual feature engineering, and improving system robustness.

[0007] To address the aforementioned problems, this invention proposes an Alzheimer's disease prediction method based on millimeter-wave radar, the method comprising:

[0008] The raw data of human gait collected by millimeter-wave radar is reorganized into a three-dimensional matrix block. The three-dimensional matrix block is then manipulated to obtain multimodal data, including range-velocity spectrum, angle spectrum, and point cloud data.

[0009] Feature extraction, alignment, and fusion are performed on the multimodal data and text data, including text data describing velocity spectrum, text data describing angle spectrum, and text data describing point cloud, to obtain reweighted multimodal data, including reweighted velocity spectrum features, reweighted angle spectrum features, and reweighted point cloud features.

[0010] The open-source pre-trained large language model is trained using the multimodal data and the text data to obtain a trained large language model. The reweighted multimodal data is then input into the trained large language model to obtain signal semantic features, including text prediction results and signal semantic units.

[0011] The signal semantic features and random noise are input into the diffusion model and trained to obtain a trained diffusion model.

[0012] The trained diffusion model is used to generate limb motion analysis diagrams and two-dimensional diagrams of point cloud limb data, which, combined with the text prediction results, assist users in predicting Alzheimer's disease.

[0013] Preferably, the raw human gait data acquired by millimeter-wave radar is reorganized into a three-dimensional matrix block, and operations are performed on the three-dimensional matrix block to obtain multimodal data, including range-velocity spectrum, angle spectrum, and point cloud data, specifically:

[0014] The third dimension of the three-dimensional matrix block is a channel composed of each transmitting and receiving antenna, the second dimension is the intermediate frequency signal received in each channel, and the first dimension is a frequency-modulated wave sample in the intermediate frequency signal.

[0015] A fast Fourier transform is performed on the first and second dimensions of the three-dimensional matrix block to obtain the corresponding spectrum and detection peak value. The relative distance between the target and the millimeter-wave radar is then calculated. Combining the phase difference between the same peaks in the spectrum corresponding to the first dimension and the spectrum corresponding to the second dimension of the three-dimensional matrix block, the range-velocity spectrum is calculated using the following formula:

[0016]

[0017] Where φ is the phase difference, λ is the wavelength of the electromagnetic wave, and T c It is the interval between two linear frequency modulated waves transmitted by the millimeter-wave radar, where π is a constant.

[0018] Using a conventional beamforming algorithm, the third dimension of the three-dimensional matrix block is used to construct the field of view's guiding vector. The guiding vector is multiplied by the third dimension of the three-dimensional matrix block, and its maximum value is taken to determine the direction of the actual echo, thus outputting the angle spectrum.

[0019] By combining the distance-velocity spectrum and the angle spectrum to perform constant false alarm rate detection, the target point is obtained. The three-dimensional spatial distribution of the point is obtained by using the target point and the angle spectrum to obtain point cloud data.

[0020] Preferably, the multimodal data and text data, including text data describing velocity spectra, text data describing angular spectra, and text data describing point clouds, are subjected to feature extraction, alignment, and fusion to obtain reweighted multimodal data, including reweighted velocity spectrum features, reweighted angular spectrum features, and reweighted point cloud features, specifically:

[0021] The distance-velocity spectrum is input into the image encoder ViT to obtain velocity spectrum features; the angle spectrum is input into the sequence word encoder LSTM to obtain angle spectrum features; the point cloud data is input into the point cloud encoder Point-M2AE to obtain point cloud features; the text data describing the velocity spectrum, the text data describing the angle spectrum, and the text data describing the point cloud are respectively input into the text encoder Transformer-encoder to obtain velocity spectrum text features, angle spectrum text features, and point cloud text features. The velocity spectrum feature-velocity spectrum text feature, angle spectrum feature-angle spectrum text feature, and point cloud feature-point cloud text feature sets are used as training sets and input into the projection layer, respectively. Optimization training is performed using the following loss function:

[0022]

[0023] Where τ is a temperature coefficient controlling the smoothness of the normalized exponential function distribution, I is the text data, G is the modal data, i.e., one of the distance-velocity spectrum, angle spectrum, or point cloud data, and q i It is a text feature, k i It is a modal feature corresponding to G, namely, one of the velocity spectrum feature, angle spectrum feature, or point cloud feature;

[0024] By using the velocity spectrum text features, the angle spectrum text features, and the point cloud text features as intermediaries, the alignment of the velocity spectrum features, the angle spectrum features, and the point cloud features is indirectly achieved, resulting in aligned velocity spectrum features, aligned angle spectrum features, and aligned point cloud features.

[0025] A selective data fusion algorithm using a soft fusion strategy is used to process the aligned velocity spectrum features, the aligned angle spectrum features, and the aligned point cloud features to obtain reweighted velocity spectrum features, reweighted angle spectrum features, and reweighted point cloud features.

[0026] Preferably, the open-source pre-trained large language model is trained using the multimodal data and the text data to obtain a trained large language model. The reweighted multimodal data is then input into the trained large language model to obtain signal semantic features, including text prediction results and signal semantic units, specifically:

[0027] The distance velocity spectrum and the text data describing the velocity spectrum, the angle spectrum and the text data describing the angle spectrum, and the point cloud data and the text data describing the point cloud are combined into three datasets respectively. These three datasets are sampled, feature extracted and labeled, and used as learning examples. They are then input into an open-source pre-trained large language model for learning. The model parameters are optimized and adjusted to obtain a trained large language model.

[0028] The reweighted velocity spectrum features, the reweighted angle spectrum features, and the reweighted point cloud features are input into the trained large language model to obtain signal semantic features, including text prediction results and signal semantic units.

[0029] Preferably, the signal semantic features and random noise are input into the diffusion model and trained to obtain a trained diffusion model, specifically as follows:

[0030] The semantic features of the signal and a random noise are input into the Stable Diffusion model, and the model is trained by forward diffusion until the output of the model is close to the center value of the original data, and the output is a data sample with similar features to the original data.

[0031] Backdiffusion training is performed on data samples with similar characteristics to the original data until the random noise is adjusted to match the text prediction results, thus obtaining a trained diffusion model.

[0032] Accordingly, the present invention also provides an Alzheimer's disease prediction system based on millimeter-wave radar, comprising:

[0033] The millimeter-wave radar data acquisition and preprocessing unit is used to reorganize the raw human gait data acquired by the millimeter-wave radar into a three-dimensional matrix block, and to operate on the three-dimensional matrix block to obtain multimodal data, including range-velocity spectrum, angle spectrum and point cloud data.

[0034] The feature extraction, alignment, and fusion unit is used to extract, align, and fuse features from the multimodal data and text data, including text data describing velocity spectrum, text data describing angle spectrum, and text data describing point cloud, to obtain reweighted multimodal data, including reweighted velocity spectrum features, reweighted angle spectrum features, and reweighted point cloud features.

[0035] The large model inference unit is used to train the open-source pre-trained large language model using the multimodal data and the text data to obtain a trained large language model, and input the reweighted multimodal data into the trained large language model to obtain signal semantic features, including text prediction results and signal semantic units.

[0036] The diffusion training unit is used to input the signal semantic features and random noise into the diffusion model and train it to obtain a trained diffusion model.

[0037] The application unit is used to generate limb motion analysis diagrams and two-dimensional diagrams of point cloud limb data using the trained diffusion model, and to assist users in predicting Alzheimer's disease by combining the text prediction results.

[0038] Implementing this invention has the following beneficial effects:

[0039] This invention employs a combination of millimeter-wave radar and a large-scale model. Millimeter-wave radar offers advantages such as being contactless, highly efficient, adaptable, and accurate. By processing the raw human motion information data sensed by the millimeter-wave radar, a series of high-resolution multimodal data, including distance Doppler maps, angle maps, and human point cloud data containing limb movement information, are obtained. The participation of multimodal data enables the system to achieve higher accuracy, stronger robustness, and better generalization ability in early Alzheimer's disease prediction. Attached Figure Description

[0040] Figure 1 This is a flowchart of an Alzheimer's disease prediction method based on millimeter-wave radar according to an embodiment of the present invention;

[0041] Figure 2 This is a structural diagram of an Alzheimer's disease prediction system based on millimeter-wave radar according to an embodiment of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Figure 1 This is a flowchart of an Alzheimer's disease prediction method based on millimeter-wave radar according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes:

[0044] S1 reorganizes the raw human gait data collected by millimeter-wave radar into a three-dimensional matrix block, and operates on the three-dimensional matrix block to obtain multimodal data, including range-velocity spectrum, angle spectrum and point cloud data;

[0045] S2, the multimodal data and text data, including text data describing velocity spectrum, text data describing angle spectrum and text data describing point cloud, are subjected to feature extraction, alignment and fusion to obtain reweighted multimodal data, including reweighted velocity spectrum features, reweighted angle spectrum features and reweighted point cloud features.

[0046] S3, using the multimodal data and the text data, train the open-source pre-trained large language model to obtain a trained large language model, and input the reweighted multimodal data into the trained large language model to obtain signal semantic features, including text prediction results and signal semantic units;

[0047] S4, Input the signal semantic features and random noise into the diffusion model and train it to obtain a trained diffusion model;

[0048] S5. Using the trained diffusion model, generate a limb motion analysis diagram and a two-dimensional diagram of point cloud limb data, and combine the text prediction results to assist the user in predicting Alzheimer's disease.

[0049] Step S1 is as follows:

[0050] S1-1, the third dimension of the three-dimensional matrix block is a channel composed of each transmitting and receiving antenna, the second dimension is the intermediate frequency signal received in each channel, and the first dimension is a frequency-modulated wave sample in the intermediate frequency signal;

[0051] S1-2, perform a Fast Fourier Transform on the first and second dimensions of the three-dimensional matrix block to obtain the corresponding spectrum and detection peak value, and calculate the relative distance between the target and the millimeter-wave radar. Combining the phase difference of the same peaks in the spectrum corresponding to the first dimension of the three-dimensional matrix block and the spectrum corresponding to the second dimension of the three-dimensional matrix block, calculate the range-velocity spectrum, the formula of which is as follows:

[0052]

[0053] Where φ is the phase difference, λ is the wavelength of the electromagnetic wave, and T c It is the interval between two linear frequency modulated waves transmitted by the millimeter-wave radar, where π is a constant.

[0054] S1-3, using a conventional beamforming algorithm, the third dimension of the three-dimensional matrix block is used to construct the field of view's guiding vector. The guiding vector is multiplied by the third dimension of the three-dimensional matrix block, and its maximum value is taken to determine the direction of the actual echo, and the angle spectrum is output.

[0055] S1-4, combine the distance-velocity spectrum and the angle spectrum to perform constant false alarm rate detection, obtain the target point, and use the target point and the angle spectrum to obtain the three-dimensional spatial distribution of the point, thus obtaining point cloud data.

[0056] Step S2 is as follows:

[0057] S2-1, the distance-velocity spectrum is input into the image encoder ViT to obtain velocity spectrum features; the angle spectrum is input into the sequence word encoder LSTM to obtain angle spectrum features; the point cloud data is input into the point cloud encoder Point-M2AE to obtain point cloud features; the text data describing the velocity spectrum, the text data describing the angle spectrum, and the text data describing the point cloud are respectively input into the text encoder Transformer-encoder to obtain velocity spectrum text features, angle spectrum text features, and point cloud text features; the velocity spectrum feature-velocity spectrum text features, angle spectrum feature-angle spectrum text features, and point cloud feature-point cloud text features are used as training sets and input into the projection layer respectively, and optimized training is performed using the following loss function:

[0058]

[0059] Where τ is a temperature coefficient controlling the smoothness of the normalized exponential function distribution, I is the text data, G is the modal data, i.e., one of the distance-velocity spectrum, angle spectrum, or point cloud data, and q i It is a text feature, k i It is a modal feature corresponding to G, namely, one of the velocity spectrum feature, angle spectrum feature, or point cloud feature;

[0060] By using the velocity spectrum text features, the angle spectrum text features, and the point cloud text features as intermediaries, the alignment of the velocity spectrum features, the angle spectrum features, and the point cloud features is indirectly achieved, resulting in aligned velocity spectrum features, aligned angle spectrum features, and aligned point cloud features.

[0061] S2-2, The aligned velocity spectrum features, the aligned angle spectrum features, and the aligned point cloud features are processed using a selective data fusion algorithm with a soft fusion strategy to obtain reweighted velocity spectrum features, reweighted angle spectrum features, and reweighted point cloud features.

[0062] Step S3 is as follows:

[0063] S3-1, combine the distance velocity spectrum and the text data describing the velocity spectrum, the angle spectrum and the text data describing the angle spectrum, and the point cloud data and the text data describing the point cloud into three datasets respectively, and sample, extract features and label these three datasets as learning examples, input them into an open-source pre-trained large language model for learning, optimize and adjust the model parameters to obtain a trained large language model;

[0064] S3-2, Input the reweighted velocity spectrum features, the reweighted angle spectrum features, and the reweighted point cloud features into the trained large language model to obtain signal semantic features, including text prediction results and signal semantic units.

[0065] Step S4 is as follows:

[0066] S4-1, Input the semantic features of the signal and a random noise into the Stable Diffusion model, perform forward diffusion training on the model until the output of the model is close to the center value of the original data, and output data samples with similar features to the original data;

[0067] S4-2, Perform backdiffusion training on data samples with similar characteristics to the original data until the random noise is adjusted to match the text prediction result, and obtain the trained diffusion model.

[0068] Accordingly, the present invention also provides an Alzheimer's disease prediction system based on millimeter-wave radar, such as... Figure 2 As shown, it includes:

[0069] The millimeter-wave radar data acquisition and preprocessing unit 1 is used to reorganize the raw human gait data acquired by the millimeter-wave radar into a three-dimensional matrix block, and to operate on the three-dimensional matrix block to obtain multimodal data, including range-velocity spectrum, angle spectrum and point cloud data.

[0070] Specifically, the third dimension of the three-dimensional matrix block is a channel composed of each transceiver antenna, the second dimension is the intermediate frequency signal received in each channel, and the first dimension is a frequency-modulated wave sample in the intermediate frequency signal;

[0071] A fast Fourier transform is performed on the first and second dimensions of the three-dimensional matrix block to obtain the corresponding spectrum and detection peak value. The relative distance between the target and the millimeter-wave radar is then calculated. Combining the phase difference between the same peaks in the spectrum corresponding to the first dimension and the spectrum corresponding to the second dimension of the three-dimensional matrix block, the range-velocity spectrum is calculated using the following formula:

[0072]

[0073] Where φ is the phase difference, λ is the wavelength of the electromagnetic wave, and T c It is the interval between two linear frequency modulated waves transmitted by the millimeter-wave radar, where π is a constant.

[0074] Using a conventional beamforming algorithm, the third dimension of the three-dimensional matrix block is used to construct the field of view's guiding vector. The guiding vector is multiplied by the third dimension of the three-dimensional matrix block, and its maximum value is taken to determine the direction of the actual echo, thus outputting the angle spectrum.

[0075] By combining the distance-velocity spectrum and the angle spectrum to perform constant false alarm rate detection, the target point is obtained. The three-dimensional spatial distribution of the point is obtained by using the target point and the angle spectrum to obtain point cloud data.

[0076] The feature extraction, alignment and fusion unit 2 is used to extract, align and fuse features from the multimodal data and text data, including text data describing velocity spectrum, text data describing angle spectrum and text data describing point cloud, to obtain reweighted multimodal data, including reweighted velocity spectrum features, reweighted angle spectrum features and reweighted point cloud features.

[0077] Specifically, the distance-velocity spectrum is input into the image encoder ViT to obtain velocity spectrum features, the angle spectrum is input into the sequence word encoder LSTM to obtain angle spectrum features, and the point cloud data is input into the point cloud encoder Point-M2AE to obtain point cloud features. Text data describing the velocity spectrum, the angle spectrum, and the point cloud are respectively input into the text encoder Transformer-encoder to obtain velocity spectrum text features, angle spectrum text features, and point cloud text features. The velocity spectrum feature-velocity spectrum text feature, angle spectrum feature-angle spectrum text feature, and point cloud feature-point cloud text feature sets are used as training sets and input into the projection layer, where optimization training is performed using the following loss function:

[0078]

[0079] Where τ is a temperature coefficient controlling the smoothness of the normalized exponential function distribution, I is the text data, G is the modal data, i.e., one of the distance-velocity spectrum, angle spectrum, or point cloud data, and q i It is a text feature, k i It is a modal feature corresponding to G, namely, one of the velocity spectrum feature, angle spectrum feature, or point cloud feature;

[0080] By using the velocity spectrum text features, the angle spectrum text features, and the point cloud text features as intermediaries, the alignment of the velocity spectrum features, the angle spectrum features, and the point cloud features is indirectly achieved, resulting in aligned velocity spectrum features, aligned angle spectrum features, and aligned point cloud features.

[0081] A selective data fusion algorithm using a soft fusion strategy is used to process the aligned velocity spectrum features, the aligned angle spectrum features, and the aligned point cloud features to obtain reweighted velocity spectrum features, reweighted angle spectrum features, and reweighted point cloud features.

[0082] The large model inference unit 3 is used to train the open-source pre-trained large language model using the multimodal data and the text data to obtain a trained large language model, and input the reweighted multimodal data into the trained large language model to obtain signal semantic features, including text prediction results and signal semantic units.

[0083] Specifically, the distance velocity spectrum and the text data describing the velocity spectrum, the angle spectrum and the text data describing the angle spectrum, and the point cloud data and the text data describing the point cloud are combined into three datasets respectively. These three datasets are sampled, feature extracted and labeled, and used as learning examples. They are then input into an open-source pre-trained large language model for learning. The model parameters are optimized and adjusted to obtain a trained large language model.

[0084] The reweighted velocity spectrum features, the reweighted angle spectrum features, and the reweighted point cloud features are input into the trained large language model to obtain signal semantic features, including text prediction results and signal semantic units.

[0085] Diffusion training unit 4 is used to input the signal semantic features and random noise into the diffusion model and train it to obtain a trained diffusion model.

[0086] Specifically, the semantic features of the signal and a random noise are input into the Stable Diffusion model, and the model is trained by forward diffusion until the output of the model is close to the center value of the original data, and the output is a data sample with similar features to the original data.

[0087] Backdiffusion training is performed on data samples with similar characteristics to the original data until the random noise is adjusted to match the text prediction results, thus obtaining a trained diffusion model.

[0088] Application unit 5 is used to generate limb motion analysis diagrams and two-dimensional diagrams of point cloud limb data using the trained diffusion model, and to assist users in predicting Alzheimer's disease by combining the text prediction results.

[0089] Therefore, this invention employs a combination of millimeter-wave radar and a large-scale model. Millimeter-wave radar offers advantages such as being contactless, highly efficient, adaptable, and accurate. By processing the raw human motion information data sensed by the millimeter-wave radar, a series of high-resolution multimodal data, including distance Doppler maps, angle maps, and human point cloud data containing limb motion information, are obtained. The participation of multimodal data enables the system to achieve higher accuracy, stronger robustness, and better generalization ability in early Alzheimer's disease prediction.

[0090] The above provides a detailed description of an Alzheimer's disease prediction method and system based on millimeter-wave radar provided by the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method of predicting Alzheimer's disease based on millimeter wave radar, characterized by, The method comprises: The original data of human gait collected by the millimeter wave radar is reorganized to obtain a three-dimensional matrix block, and the three-dimensional matrix block is operated to obtain multi-modal data, including distance speed spectrum, angle spectrum and point cloud data; Feature extraction, alignment and fusion are performed on the multi-modal data and text data, including text data describing the speed spectrum, text data describing the angle spectrum and text data describing the point cloud, to obtain reweighted multi-modal data, including reweighted speed spectrum features, reweighted angle spectrum features and reweighted point cloud features; The multi-modal data and the text data are used to train an open-source pre-trained large language model to obtain a trained large language model, and the reweighted multi-modal data is input into the trained large language model to obtain signal semantic features, including text prediction results and signal semantic units; The signal semantic features and random noise are input into a diffusion model and trained to obtain a trained diffusion model; The trained diffusion model is used to generate a limb movement analysis graph and a point cloud limb data two-dimensional graph, and the text prediction results are combined to assist users in predicting Alzheimer's disease; Specifically, the original data of human gait collected by the millimeter wave radar is reorganized to obtain a three-dimensional matrix block, and the three-dimensional matrix block is operated to obtain multi-modal data, including distance speed spectrum, angle spectrum and point cloud data, specifically: The third dimension of the three-dimensional matrix block is a channel composed of each transceiver antenna, the second dimension is the intermediate frequency signal received in each channel, and the first dimension is a frequency modulation wave sampling in the intermediate frequency signal; Fast Fourier transform is performed on the first dimension and the second dimension of the three-dimensional matrix block to obtain corresponding frequency spectrum, detect peak value, and calculate the relative distance between the target and the millimeter wave radar, and the phase difference of the same wave peak of the first dimension corresponding frequency spectrum of the three-dimensional matrix block and the second dimension corresponding frequency spectrum of the three-dimensional matrix block is combined to calculate the distance speed spectrum, and the formula is as follows: wherein φ is the phase difference, λ is the wavelength of the electromagnetic wave, T c is the interval between two linear frequency modulation waves transmitted by the millimeter wave radar, and π is a constant; By using the third dimension of the three-dimensional matrix block to construct a guide vector of the field of view through a conventional beam forming algorithm, the guide vector is multiplied by the third dimension of the three-dimensional matrix block and the maximum value is taken to determine the direction of the actual echo, and the angle spectrum is output; Constant false alarm rate detection is performed on the distance speed spectrum and the angle spectrum to obtain a target point, and the three-dimensional spatial distribution of the target point and the angle spectrum acquisition point is obtained to obtain point cloud data.

2. The Alzheimer's disease prediction method based on millimeter wave radar according to claim 1, characterized in that, The multi-modal data and text data, including text data describing the speed spectrum, text data describing the angle spectrum and text data describing the point cloud, are subjected to feature extraction, alignment and fusion to obtain reweighted multi-modal data, including reweighted speed spectrum features, reweighted angle spectrum features and reweighted point cloud features, specifically: The distance velocity spectrum is input into an image encoder ViT to obtain velocity spectrum features, the angle spectrum is input into a sequence word encoder LSTM to obtain angle spectrum features, and the point cloud data is input into a point cloud encoder Point-M2AE to obtain point cloud features; text data describing the velocity spectrum, text data describing the angle spectrum, and text data describing the point cloud are input into a text encoder Transformer-encoder to obtain velocity spectrum text features, angle spectrum text features, and point cloud text features, respectively; the velocity spectrum features-velocity spectrum text features, the angle spectrum features-angle spectrum text features, and the point cloud features-point cloud text features are input into projection layers as training sets, and are optimized and trained through the following loss functions: where τ is a temperature coefficient that controls the smoothness of the normalized exponential function distribution, I is the text data, G is the modal data, i.e., one of the distance velocity spectrum, angle spectrum, or point cloud data, q i is the text feature, k i is the modal feature corresponding to G, i.e., one of the velocity spectrum feature, angle spectrum feature, or point cloud feature. The velocity spectrum text features, the angle spectrum text features, and the point cloud text features are used as intermediates to indirectly align the velocity spectrum features, the angle spectrum features, and the point cloud features, to obtain aligned velocity spectrum features, aligned angle spectrum features, and aligned point cloud features; The aligned velocity spectrum features, the aligned angle spectrum features, and the aligned point cloud features are processed using a selective data fusion algorithm with a soft fusion strategy to obtain reweighted velocity spectrum features, reweighted angle spectrum features, and reweighted point cloud features.

3. The millimeter wave radar-based Alzheimer's disease prediction method of claim 1, wherein, The multi-modal data and the text data are used to train an open-source pre-trained large language model to obtain a trained large language model, and the reweighted multi-modal data is input into the trained large language model to obtain signal semantic features, including a text prediction result and a signal semantic unit, specifically: The distance velocity spectrum and the text data describing the velocity spectrum, the angle spectrum and the text data describing the angle spectrum, and the point cloud data and the text data describing the point cloud are combined into three data sets, respectively, and are sampled, feature-extracted, and labeled as learning examples, and are input into an open-source pre-trained large language model for learning, and the model parameters are optimized and adjusted to obtain a trained large language model; The reweighted velocity spectrum features, the reweighted angle spectrum features, and the reweighted point cloud features are input into the trained large language model to obtain signal semantic features, including a text prediction result and a signal semantic unit.

4. The millimeter wave radar-based Alzheimer's disease prediction method of claim 1, wherein, The signal semantic features and random noise are input into a diffusion model and are trained to obtain a trained diffusion model, specifically: The signal semantic features and a random noise are input into a Stable Diffusion diffusion model for forward diffusion training of the model until the output of the model approaches the center value of the original data, and a data sample with similar features to the original data is output; The data sample with similar features to the original data is subjected to reverse diffusion training until the random noise is adjusted to match the text prediction result, and a trained diffusion model is obtained.

5. A millimeter wave radar-based Alzheimer's disease prediction system, characterized by, The system comprises: A millimeter wave radar data acquisition and preprocessing unit is configured to reorganize original data of human gait collected by a millimeter wave radar into a three-dimensional matrix block, and to operate the three-dimensional matrix block to obtain multi-modal data including a distance-velocity spectrum, an angle spectrum, and point cloud data; A feature extraction, alignment, and fusion unit is configured to perform feature extraction, alignment, and fusion on the multi-modal data and text data including text data describing the velocity spectrum, text data describing the angle spectrum, and text data describing the point cloud to obtain reweighted multi-modal data including reweighted velocity spectrum features, reweighted angle spectrum features, and reweighted point cloud features; A large model inference unit is configured to train an open-source pre-trained large language model using the multi-modal data and the text data to obtain a trained large language model, and to input the reweighted multi-modal data into the trained large language model to obtain signal semantic features including a text prediction result and a signal semantic unit; A diffusion training unit is configured to input the signal semantic features and random noise into a diffusion model and train the diffusion model to obtain a trained diffusion model; An application unit is configured to generate a limb movement analysis graph and a point cloud limb data two-dimensional graph using the trained diffusion model, and to assist a user in predicting Alzheimer's disease in combination with the text prediction result. Specifically, the millimeter wave radar data acquisition and preprocessing unit is configured to reorganize original data of human gait collected by a millimeter wave radar into a three-dimensional matrix block, and to operate the three-dimensional matrix block to obtain multi-modal data including a distance-velocity spectrum, an angle spectrum, and point cloud data, specifically as follows: The third dimension of the three-dimensional matrix block is a channel formed by each transceiving antenna, the second dimension is an intermediate frequency signal received in each channel, and the first dimension is a frequency modulation wave sampling in the intermediate frequency signal; Fast Fourier transform is performed on the first dimension and the second dimension of the three-dimensional matrix block to obtain corresponding frequency spectra, detect peak values, and calculate a relative distance between a target and the millimeter wave radar, and a phase difference of the same wave peak of the frequency spectrum corresponding to the first dimension of the three-dimensional matrix block and the frequency spectrum corresponding to the second dimension of the three-dimensional matrix block is calculated to obtain a distance-velocity spectrum, which is calculated according to the following formula: wherein φ is the phase difference, λ is the wavelength of the electromagnetic wave, T c is the interval between two linear frequency modulation waves transmitted by the millimeter wave radar, and π is a constant; A conventional beam forming algorithm is used to construct a guide vector of a field of view by using the third dimension of the three-dimensional matrix block, the guide vector is multiplied by the third dimension of the three-dimensional matrix block, and a maximum value is taken to determine the direction of the actual echo, and an angle spectrum is output. Constant false alarm rate detection is performed in combination with the distance-velocity spectrum and the angle spectrum to obtain a target point, and a three-dimensional spatial distribution is obtained by using the target point and the angle spectrum to obtain point cloud data.

6. The millimeter wave radar-based Alzheimer's disease prediction system of claim 5, wherein, The feature extraction, alignment, and fusion unit is configured to perform feature extraction, alignment, and fusion on the multi-modal data and text data including text data describing the velocity spectrum, text data describing the angle spectrum, and text data describing the point cloud to obtain reweighted multi-modal data including reweighted velocity spectrum features, reweighted angle spectrum features, and reweighted point cloud features, specifically as follows: The distance velocity spectrum is input into an image encoder ViT to obtain velocity spectrum features, the angle spectrum is input into a sequence word encoder LSTM to obtain angle spectrum features, and the point cloud data is input into a point cloud encoder Point-M2AE to obtain point cloud features; text data describing the velocity spectrum, text data describing the angle spectrum, and text data describing the point cloud are input into a text encoder Transformer-encoder to obtain velocity spectrum text features, angle spectrum text features, and point cloud text features, respectively; the velocity spectrum features-velocity spectrum text features, the angle spectrum features-angle spectrum text features, and the point cloud features-point cloud text features are input into projection layers as training sets, and are optimized and trained by the following loss function: where τ is a temperature coefficient that controls the smoothness of the normalized exponential function distribution, I is the text data, G is the modal data, i.e., one of the distance velocity spectrum, angle spectrum, or point cloud data, q i is the text feature, k i is the modal feature corresponding to G, i.e., one of the velocity spectrum feature, angle spectrum feature, or point cloud feature. The velocity spectrum text features, the angle spectrum text features, and the point cloud text features are used as intermediates to indirectly align the velocity spectrum features, the angle spectrum features, and the point cloud features, to obtain aligned velocity spectrum features, aligned angle spectrum features, and aligned point cloud features; The aligned velocity spectrum features, the aligned angle spectrum features, and the aligned point cloud features are processed by a selective data fusion algorithm using a soft fusion strategy to obtain reweighted velocity spectrum features, reweighted angle spectrum features, and reweighted point cloud features.

7. The millimeter wave radar-based Alzheimer's disease prediction system of claim 5, wherein, The large model inference unit is configured to train an open-source pre-trained large language model using the multi-modal data and the text data to obtain a trained large language model, and input the reweighted multi-modal data into the trained large language model to obtain signal semantic features, including a text prediction result and a signal semantic unit, and specifically comprising: The distance velocity spectrum and the text data describing the velocity spectrum, the angle spectrum and the text data describing the angle spectrum, and the point cloud data and the text data describing the point cloud are combined into three data sets, respectively, and are sampled, feature extracted, and labeled as learning examples, and are input into an open-source pre-trained large language model for learning, and the model parameters are optimized and adjusted to obtain a trained large language model; The reweighted velocity spectrum features, the reweighted angle spectrum features, and the reweighted point cloud features are input into the trained large language model to obtain signal semantic features, including a text prediction result and a signal semantic unit.

8. The millimeter wave radar-based Alzheimer's disease prediction system of claim 5, wherein, The diffusion training unit is configured to input the signal semantic features and random noise into a diffusion model and train the diffusion model to obtain a trained diffusion model, and specifically comprising: The signal semantic features and a random noise are input into a Stable Diffusion diffusion model for forward diffusion training of the model until the output of the model approaches a center value of the original data, and a data sample with similar features to the original data is output; The data sample with similar features to the original data is subjected to reverse diffusion training until the random noise is adjusted to match the text prediction result, and a trained diffusion model is obtained.

Images