Non-invasive screening method and device for Alzheimer's disease proactive period based on expression features

Abstract

The invention provides a non-invasive screening method and a non-invasive screening device for an Alzheimer's disease proactive period based on expression features. The non-invasive screening method comprises the following steps: S1, collecting facial video data of a user; s2, preprocessing the face video data; s3, subject extraction is carried out on the preprocessed video clips based on subject modeling, so that positive emotion clips are eliminated, and negative emotion clips and neutral emotion clips are reserved; s4, extracting a key frame from the reserved video clip; s5, extracting micro-expression features of the user from the key frame; s6, based on the key frame, analyzing lip motion features of the user to perform dysarthria assessment; s7, carrying out multi-modal feature fusion on the micro expression features and the dysarthria features; and S8, inputting the fused features into a pre-trained pAD risk assessment model to obtain a pAD risk assessment result. Noninvasive, low-cost and high-efficiency Alzheimer's disease prophase screening is realized for the first time, and a feasible technical path is provided for early screening of large-scale crowds.

Description

Technical Field

[0001] This invention belongs to the field of medical testing technology and relates to a non-invasive screening method and device for the prodromal stage of Alzheimer's disease based on facial expression features. Background Technology

[0002] Alzheimer's disease (AD) is a degenerative disease of the central nervous system with a long and irreversible course. Before the onset of clinically significant dementia symptoms, there exists a crucial prodromal period (pAD), which typically remains dormant for decades. Although pathophysiological changes occur during this stage, there are no obvious clinical symptoms. Progressive cognitive impairment and behavioral disturbances that appear after age 50 are the main clinical manifestations. Studies have shown that intervention during the pAD stage can effectively slow disease progression; therefore, early screening and diagnosis of pAD have significant clinical and social implications.

[0003] Currently, clinical screening and diagnosis of pAD mainly rely on the following methods, but all of them have significant limitations: 1. Cerebrospinal Fluid Testing (CSF Testing): This method involves collecting cerebrospinal fluid via lumbar puncture to detect biomarkers such as Aβ42 and tau protein. While highly accurate, it is an invasive procedure that causes pain and fear for patients and carries potential risks such as infection and bleeding. The complex procedure requires specialized physicians in a specific medical environment, significantly limiting its application in large-scale population screening.

[0004] 2. Positron Emission Tomography (PET): This method uses the accumulation of radioactive tracers (such as Amyvid) in the brain to visualize Aβ plaques or tau protein tangles. While PET imaging is an important gold standard for diagnosis, it is extremely expensive (approximately 10,000 RMB per examination) and subjects are exposed to ionizing radiation. Furthermore, PET equipment is expensive and scarce, typically found only in large medical centers, making it unsuitable for routine screening or dynamic monitoring.

[0005] 3. Neuropsychological scale assessments (such as MMSE, MoCA, etc.): These are currently the most commonly used initial screening tools. However, scale assessments are highly dependent on the subjects' education level, cultural background, and cooperation, resulting in significant individual differences and potential bias. Furthermore, the assessments require professional assistance and interpretation, are time-consuming (usually exceeding 30 minutes), and lack sensitivity to very early, subtle cognitive changes, as well as objective quantitative indicators.

[0006] In summary, existing pAD screening technologies face a "three-high" dilemma—high risk (invasive), high cost (expensive), and high resource consumption (dependent on professional personnel and equipment). This results in a large number of potential patients not being identified in the prodromal stage of the disease, missing the optimal intervention window.

[0007] Therefore, there is an urgent need in this field for a new screening technology that can overcome the above-mentioned shortcomings: it should be a non-invasive, rapid, low-cost, and easily deployed solution that can serve as an effective "initial screening gateway" to efficiently screen high-risk individuals for pAD from a large population and then recommend them for gold standard diagnosis, thereby optimizing the allocation of medical resources and improving the early diagnosis rate.

[0008] Although recent studies have explored the use of behavioral biomarkers such as facial expression analysis for the assessment of neurological diseases, existing technologies are mostly limited to a single modality (e.g., analyzing only facial expressions or only speech) and lack a deep connection with clinical pathological mechanisms. Their accuracy and reliability remain far from clinical applicability. How to effectively integrate multimodal information (such as micro-expressions and dysarthria) and establish a complete, AI-based automated screening system remains a challenging technical problem to be solved. Summary of the Invention

[0009] The purpose of this invention is to provide a non-invasive screening method and apparatus for the prodromal stage of Alzheimer's disease based on facial expression features, in order to solve the problem.

[0010] To address the aforementioned technical problems, this invention provides a non-invasive screening method for the prodromal stage of Alzheimer's disease based on facial expression features, comprising the following steps: S1. Collect user's facial video data; S2. Preprocess the facial video data, the preprocessing including face detection, alignment and illumination normalization; S3. Based on topic modeling, extract topics from the preprocessed video clips to exclude video clips containing positive emotional topics and retain video clips containing negative and neutral emotional topics. S4. Extract keyframes from the retained video clips; S5. Extract the user's micro-expression features from the keyframes; S6. Based on the keyframes, analyze the user's lip movement features to assess articulation disorders and obtain articulation disorder features; S7. The micro-expression features and the articulation disorder features are fused using multimodal features to obtain the fused features; S8. Input the fused features into the pre-trained pAD risk assessment model to obtain the pAD risk assessment result.

[0011] More preferably, the face detection and alignment process in step S2 specifically includes: Face detection and 68 key point localization were performed using a multi-task convolutional neural network algorithm. The face image is normalized to 256×256 pixels using affine transformation; An adaptive histogram equalization technique was used for illumination normalization.

[0012] More preferably, the topic modeling algorithm described in step S3 adopts the latent Dirichlet distribution model.

[0013] More preferably, the keyframe extraction described in step S4 is achieved in the following manner: Calculate the optical flow energy value between consecutive frames in a video sequence; The optical flow energy value is calculated using the following formula: in, and Each pixel exist Always and Optical flow component in the direction; The frame corresponding to the local maximum value of the optical flow energy is selected as the key frame.

[0014] More preferably, the micro-expression feature extraction in step S5 is implemented using a deep convolutional neural network; The deep convolutional neural network is a ResNet-50 architecture; Output a 512-dimensional micro-expression feature vector.

[0015] More preferably, the articulation disorder feature analysis in step S6 specifically includes: The lip region was extracted using a facial key point detection algorithm; A three-dimensional convolutional neural network was used to extract the spatiotemporal features of lip movements; Modeling temporal dynamic features using long short-term memory networks; Output a score indicating the severity of articulation disorder.

[0016] More preferably, the feature fusion method described in step S7 specifically includes: The dynamic time warping algorithm is used to perform temporal alignment of micro-expression feature sequences and lip movement feature sequences; The attention weights of each modality feature are calculated using an attention mechanism; The attention weights are used to perform weighted fusion to obtain the fused features.

[0017] Further preferably, the pAD risk assessment model described in step S8 is pre-trained using a contrastive learning framework; The loss function of the contrastive learning framework is: in, These are feature representations of different enhanced versions from the same original sample.

[0018] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described above.

[0019] The present invention also provides a non-invasive screening device for the prodromal stage of Alzheimer's disease based on facial expression features, including an electronic device and two opposing and vertically arranged support columns. Each support column is hollow, and a worm gear is rotatably connected to the upper end of each support column. Each support column has a mounting cavity on one side of the corresponding worm gear. A vertically arranged drive motor is installed in each mounting cavity, and the power output shaft of each drive motor is connected to a worm that meshes with the corresponding worm gear. Each support column is connected to a vertically arranged collection column. Each collection column has a rotating groove on its lower side facing the corresponding support column. Each support column has a rotating notch on its upper part facing the corresponding collection column. Each worm gear is connected to a rotating connecting arm that passes through the corresponding rotating notch and rotates in connection with the corresponding rotating groove. Each connecting column has a parallel connecting arm that passes through the corresponding rotating notch and rotates in connection with the corresponding rotating groove at its upper end. Each parallel connecting arm is always parallel to the corresponding rotating connecting arm, so that each collection column and the corresponding support column are always parallel. Each rotating connecting arm has a wire-passing groove on its lower side that is open at both ends and runs along its length. Each parallel connecting arm has a wire-passing port at one end near the corresponding support column. A transmission cable is connected to the top of each rotating groove. Each transmission cable passes through the corresponding wire-passing groove and the corresponding wire-passing port in sequence and extends into the corresponding support column, and is electrically connected to the electronic device. Each acquisition column has an acquisition camera installed on the upper part of the side away from the corresponding support column, and an infrared sensor is installed on the lower part of the side away from the corresponding support column. The acquisition camera and the infrared sensor are electrically connected to the electronic device through corresponding transmission cables. The two acquisition columns are arranged in a figure-eight shape so that the two acquisition cameras can face the statistical target.

[0020] Compared with the prior art, the present invention has the following beneficial effects: Firstly, this invention relies entirely on external video analysis technology, completely eliminating the physical trauma, infection risk, and radiation exposure associated with traditional spinal cord puncture and PET scans, significantly improving user acceptance and the safety of the screening process.

[0021] Secondly, the cost of a single screening using this invention is less than 100 yuan, while the testing time is significantly reduced from several days to several minutes, effectively breaking through the economic and feasibility bottlenecks of large-scale pAD screening.

[0022] Thirdly, this invention can be easily integrated into portable devices such as smartphones, making it simple to operate and enabling screening capabilities to reach communities and families, effectively solving the problem of uneven distribution of high-quality medical resources.

[0023] Fourth, although this invention is not intended as a final diagnostic tool, its high sensitivity makes it an efficient "initial screening gateway" that can quickly identify high-risk individuals from the population and recommend further gold standard diagnosis, thereby significantly saving overall social medical costs.

[0024] Fifth, the present invention also provides a detection device capable of capturing video from the left and right sides of the patient. By shooting the face from the left and right sides, it can more accurately capture facial muscle details, which is more helpful for analysis and judgment. At the same time, it can intelligently adjust the height according to the patient's height, so that the patient's face is always in the center of the video screen, and more complete facial details of the patient are captured. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall structure of Example 2; Figure 2 Used to demonstrate the connection between the rotating connecting arm and the parallel connecting arm and the acquisition column; Figure 3 It is a partial sectional view used to show the internal structure of the support column; Figure 4 Used to demonstrate the threading groove inside the rotating connecting arm.

[0026] The components include: 1. Electronic equipment; 2. Support column; 3. Worm gear; 4. Mounting cavity; 5. Drive motor; 6. Worm; 7. Acquisition column; 8. Rotation groove; 9. Rotation notch; 10. Rotation connecting arm; 11. Parallel connecting arm; 12. Wire groove; 13. Wire passage; 14. Transmission cable; 15. Acquisition camera; and 16. Infrared sensor. Detailed Implementation

[0027] The following detailed description, in conjunction with the accompanying drawings and specific embodiments, provides a non-invasive screening method and apparatus for the prodromal stage of Alzheimer's disease based on facial expression features, as proposed in this invention. The advantages and features of this invention will become clearer from the following description. It should be noted that the accompanying drawings are all in a very simplified form and use non-precise proportions, used only to facilitate and clarify the illustration of the embodiments of this invention. The same or similar reference numerals in the drawings represent the same or similar parts.

[0028] Example 1: A non-invasive, rapid, and low-cost screening method for the prodromal stage of Alzheimer's disease based on facial expression features. System Architecture Overview This innovative technical solution employs a multimodal fusion analysis method to establish a complete early screening system for pAD (painful articulation disorder) by capturing subtle changes in facial micro-expression features and lip-and-tooth articulation characteristics. This system integrates cutting-edge technologies from multiple fields, including computer vision, deep learning, and medical diagnostics, and comprises five core modules: data acquisition, preprocessing, feature extraction, multimodal fusion, and risk assessment, forming a complete technical closed loop from data input to risk assessment. This solution not only solves the problems of invasiveness, high cost, and long cycle time associated with traditional detection methods, but also provides a feasible technical path for large-scale population screening.

[0029] I. Data Acquisition and Preprocessing Stage Step 1.1: Standardized Video Acquisition This system employs a standardized video capture scheme, specifically designed for high-risk individuals aged 50 and above. The capture device uses a standard smartphone camera or dedicated video equipment, requiring a video resolution of at least 1080p and a frame rate of no less than 30fps to ensure the capture of subtle facial muscle movements. The capture environment requires strict control of lighting conditions between 300-800 lux to avoid excessively dark or bright environments affecting image quality. During capture, the subject's face must be unobstructed, facing the camera directly. A multi-angle synchronous capture scheme is used to obtain more comprehensive three-dimensional facial motion information. Each subject's capture time is approximately 3-5 minutes, including multiple scenarios such as natural dialogue, emotion-evoked tasks, and free speech, with 3-5 seconds of abnormal segments extracted for analysis.

[0030] Quality control formula: This formula ensures that the quality of the acquired video meets the requirements of subsequent analysis by weighting the resolution, frame rate and lighting conditions. Videos with a total score below 0.8 need to be re-acquired.

[0031] Step 1.2: Face Detection and Alignment An improved MTCNN (Multi-Task Convolutional Neural Network) algorithm is employed for accurate face detection and keypoint localization. This algorithm refines the detection results step-by-step through three cascaded networks: first, P-Net rapidly generates candidate face regions; then, R-Net further refines them; and finally, O-Net outputs precise face bounding boxes and coordinates of 68 keypoints. Based on these keypoints, the system standardizes the face image to a uniform 256×256 pixel format using affine transformation, ensuring the accuracy and consistency of subsequent feature extraction.

[0032] Face alignment formula: in The transformation uses coordinates for 68 key facial points to eliminate the effects of pose and distance variations.

[0033] Step 1.3: Illumination Normalization Process To address the issue of significant variations in illumination conditions under different acquisition environments, the system employs Adaptive Histogram Equalization (CLAHE) for illumination normalization. This method divides the image into multiple small regions, performs histogram equalization on each region individually, and then eliminates block artifacts through bilinear interpolation, thereby enhancing local contrast while avoiding excessive noise amplification. This process ensures stable analysis results under varying illumination conditions, improving the system's robustness.

[0034] II. Feature Extraction and Analysis Stage Step 2.1: Video Segmentation Based on Topic Modeling Deep semantic analysis of video content is performed using the Latent Dirichlet Allocation (LDA) topic model. The system treats the visual features of video frames as "visual words" and automatically identifies the emotional topic distribution of video segments through topic modeling. This process effectively distinguishes between positive emotional segments (such as laughter and smiles) and neutral and negative emotional segments, automatically excluding the former and retaining the latter for subsequent in-depth analysis. This method significantly improves analysis efficiency and avoids interference from irrelevant data in the analysis results.

[0035] Theme extraction formula: By calculating the probability distribution of topics, the system enables automatic classification and filtering of video clips.

[0036] Step 2.2: Keyframe Extraction Algorithm A keyframe extraction method based on optical flow energy is employed. By calculating the motion changes between consecutive frames, the most representative keyframes are selected for analysis. The optical flow algorithm calculates the motion energy distribution of the entire face by comparing the motion vectors of pixels in adjacent frames. The system selects frames corresponding to local maxima of optical flow energy as keyframes, as these frames typically contain the richest information on facial expression changes.

[0037] Optical flow energy calculation: in, and Each pixel exist Always and Optical flow component in the direction; This formula calculates the overall motion energy of each frame, providing a quantitative basis for keyframe selection.

[0038] Step 2.3: Micro-expression feature extraction Facial micro-expression features were extracted using a deep convolutional neural network. The network architecture employed a pre-trained ResNet-50 model with targeted improvements made to the final layer. Feature extraction focused on subtle movement patterns in areas such as the periorbital region, periorbital region, and forehead, regions proven to be closely related to neurodegenerative changes. The network outputs a 512-dimensional micro-expression feature vector, containing rich dynamic information about facial expressions.

[0039] Feature extraction formula: Here, θ represents the network parameters, which are obtained through training with a large amount of data and can effectively capture micro-expression features related to pAD.

[0040] Step 2.4: Analysis of labiodental articulation disorders This system detects articulation disorder features related to pAD (palatational articulation disorder) through high-precision lip movement trajectory analysis and speech feature extraction. First, a facial landmark detection algorithm is used to accurately extract the lip region, then the lip movement trajectory is tracked during pronunciation. An LSTM (Long Short-Term Memory) network is employed to model temporal dynamic features, capturing abnormal patterns in lip movement that often reflect degeneration of facial neuromuscular control.

[0041] Articulation disorder score: Where φ is the parameter of the LSTM network, and the output is a score of the severity of articulation disorder between 0 and 1. The higher the score, the more severe the disorder.

[0042] III. Multimodal Fusion and Risk Assessment Step 3.1: Feature Alignment and Standardization Because micro-expression features and lip movement features may be asynchronous in time, the system uses the Dynamic Time Warping (DTW) algorithm for feature alignment. DTW eliminates the impact of temporal differences on the fusion effect by finding the optimal matching path between the two sequences, ensuring that the two modal features can be fused at the correct time points.

[0043] DTW distance calculation: Where π is the optimal alignment path, the algorithm guarantees the temporal consistency of feature fusion.

[0044] Step 3.2: Attention Mechanism Fusion We design a multimodal feature fusion module based on an attention mechanism. This module can automatically learn the importance of each modality feature and dynamically adjust the weights according to different input samples. The attention mechanism first calculates the similarity between the query vector and the key vector, then generates attention weights through a softmax function, and finally applies these weights to the value vector to achieve weighted fusion.

[0045] Attention weight calculation: This fusion method can highlight important features, suppress noise, and significantly improve the fusion effect.

[0046] Step 3.3: pAD Risk Assessment Model A risk assessment model based on a multilayer perceptron was established, consisting of two fully connected layers and a sigmoid output layer. The model takes a fused feature vector as input and outputs a pAD risk probability between 0 and 1 after a nonlinear transformation. Through training with a large amount of clinical data, the model learned to identify feature patterns associated with pAD and is able to provide accurate risk assessments.

[0047] Risk assessment formula: Where σ is the sigmoid function, which compresses the output value to between 0 and 1, representing the risk probability of pAD.

[0048] IV. Model Training and Optimization Step 4.1: Comparative Learning Pre-training Self-supervised pre-training is performed using the SimCLR framework. This framework generates positive sample pairs through data augmentation and uses contrastive learning to make the feature representations of similar samples more similar and the feature representations of dissimilar samples more dissimilar. This method makes full use of unlabeled data and enhances the model's feature representation capabilities.

[0049] Comparison loss function: in, These are feature representations of different enhanced versions from the same original sample; This loss function optimizes the model by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs.

[0050] Step 4.2: Multi-task joint training A multi-task learning framework is designed, simultaneously optimizing classification loss and reconstruction loss. The classification loss ensures the model can accurately distinguish pAD risk levels, while the reconstruction loss, through the autoencoder structure, guarantees the model learns meaningful feature representations. Regularization terms prevent overfitting. This multi-task learning strategy significantly improves the model's generalization ability.

[0051] Total loss function: By adjusting the weights, the importance of different tasks can be balanced.

[0052] Step 4.3: Model Compression and Acceleration Knowledge distillation is employed to transfer knowledge from a large teacher model to a small student model, while quantization techniques are used to reduce FP32 precision to INT8 precision. These techniques significantly reduce the computational and storage requirements of the model, enabling it to run in real time on mobile devices.

[0053] Distillation loss: By softening labels to convey knowledge from the teacher's model, the performance of the student's model can be improved.

[0054] V. System Deployment and Application Step 5.1: Edge computing optimization The optimized model was deployed to mobile and edge computing devices, and inference was accelerated using TensorRT. Through techniques such as layer fusion, accuracy calibration, and automatic kernel tuning, a significant improvement in inference speed was achieved, ensuring a single-frame processing time of less than 50ms, meeting real-time requirements.

[0055] Step 5.2: Privacy Protection Mechanism A federated learning framework is employed to achieve continuous model optimization while protecting user privacy. User data is always stored locally on the device, and only model gradient updates are uploaded to a central server for aggregation and updating. This approach protects user privacy while enabling the collective intelligence of the model to improve.

[0056] Step 5.3: Clinical Validation Protocol Design a multicenter clinical validation trial to compare and validate the system with gold standard detection methods (PET, CSF testing). Through rigorous statistical analysis, verify the system's accuracy, sensitivity, and specificity to ensure the reliability of screening results.

[0057] Screening accuracy: shows potential, but still needs optimization. In a Phase I clinical validation study involving 300 participants (100 confirmed pAD patients and 200 healthy controls), the core performance data of this system are as follows: Overall accuracy: 67.0% (95% CI: 61.5% - 72.5%) Sensitivity: 72.0% (meaning it can correctly identify 72% of true pAD patients) Specificity: 64.5% (meaning it correctly excluded 64.5% of healthy controls) AUC (Area Under Receiver Operating Characteristic): 0.74 (95% CI: 0.68 - 0.80) Interpretation of Results: The current accuracy rate of 67% indicates that this technology, as a novel screening method, has shown the potential to distinguish pAD patients from healthy individuals (AUC=0.74, superior to random guessing's 0.5). However, its performance has not yet reached the level of replacing existing gold standard diagnostic methods (such as PET and cerebrospinal fluid testing). Its sensitivity (72%) is higher than its specificity (64.5%), meaning it tends to "better to err on the side of caution than to miss a case," which to some extent aligns with the characteristics of early screening tools and helps reduce missed diagnoses.

[0058] Example 2: A non-invasive screening device for the prodromal stage of Alzheimer's disease based on facial expression features, such as... Figures 1 to 4As shown, the device includes an electronic device 1, which contains a memory, a processor, and a computer program (not shown) stored in the memory. When the processor executes the computer program, it implements the method as described in Embodiment 1. The device also includes two opposing and vertically arranged support columns 2. Each support column 2 is hollow, and a worm gear 3 is rotatably connected to the upper end of each support column 2. Each support column 2 has a mounting cavity 4 on one side of the corresponding worm gear 3. Each mounting cavity 4 is equipped with a vertically arranged drive motor 5. The power output shaft of each drive motor 5 is connected to a worm 6 that meshes with the corresponding worm gear 3.

[0059] Each support column 2 is connected to a vertically arranged collection column 7. Each collection column 7 has a rotating groove 8 on its lower side facing the corresponding support column 2. Each support column 2 has a rotating notch 9 on its upper part facing the corresponding collection column 7. Each worm gear 3 is connected to a rotating connecting arm 10 that passes through the corresponding rotating notch 9 and rotates with the corresponding rotating groove 8. Each connecting column has a parallel connecting arm 11 that passes through the corresponding rotating notch 9 and rotates with the corresponding rotating groove 8. Each parallel connecting arm 11 is always parallel to the corresponding rotating connecting arm 10, so that each collection column 7 and the corresponding support column 2 are always parallel.

[0060] Each rotating connecting arm 10 has a wire groove 12 on its lower side that runs along its length and is open at both ends. Each parallel connecting arm 11 has a wire passage 13 at one end near the corresponding support column 2. A transmission cable 14 is connected to the top of each rotating groove 8. Each transmission cable 14 passes through the corresponding wire groove 12 and the corresponding wire passage 13 in sequence and extends into the corresponding support column 2, and is electrically connected to the electronic device 1.

[0061] Each acquisition column 7 has an acquisition camera 15 installed on the upper part of the side away from the corresponding support column 2, and an infrared sensor 16 is installed on the lower part of the side away from the corresponding support column 2. The acquisition camera 15 and the infrared sensor 16 are electrically connected to the electronic device 1 through the corresponding transmission cable 14. The two acquisition columns 7 are arranged in a figure-eight shape, so that the two acquisition cameras 15 can face the statistical target.

[0062] Working principle: The device is installed on the doctor's desk with the two acquisition cameras 15 facing the patient in a V-shape. When a patient sits in front, the infrared sensor 16 detects the patient and the two acquisition cameras 15 are activated. The system judges whether the patient's face is in the center of the camera image. If the face is not in the center, the drive motor 5 is activated, which drives the worm gear 6 to rotate. The rotation of the worm gear 6 drives the worm wheel 3 to rotate, which in turn drives the rotating connecting arm 10 to rotate. Since the parallel connecting arm 11 is parallel to the rotating connecting arm 10, the acquisition column 7 can always remain vertically raised or lowered until the face image is in the center of the image. The device captures the face from both sides, which can more accurately capture facial muscle details and is more helpful for analysis and judgment.

[0063] It should also be noted that all terms such as "set up" and similar descriptive words in this application (especially the specification) indicate that two structures have or exist a connection relationship. However, the specific means by which the two are connected are not limited in detail, and are usually conventional connection methods. That is, the means should be understood as prior art and do not need to be elaborated. For example, "m is set up with n" only indicates that structure m has structure n, and whether the two are connected by welding, riveting, adhesive, or integral molding is within the scope of protection of this application. Similarly, "x is rotatably set up with y" only indicates that y and x can rotate relative to each other, and whether the two are connected by a bearing, or whether y directly passes through x and is rotatably connected to x, or other feasible methods, are all within the scope of protection of this application.

[0064] The above description is merely a description of preferred embodiments of the present invention and is not intended to limit the scope of the present invention in any way. Any changes or modifications made by those skilled in the art based on the above disclosure shall fall within the protection scope of the claims.

Claims

1. A non-invasive screening method for the prodromal stage of Alzheimer's disease based on facial expression features, characterized in that, Includes the following steps: S1. Collect user's facial video data; S2. Preprocess the facial video data, the preprocessing including face detection, alignment and illumination normalization; S3. Based on topic modeling, extract topics from the preprocessed video clips to exclude video clips containing positive emotional topics and retain video clips containing negative and neutral emotional topics. S4. Extract keyframes from the retained video clips; S5. Extract the user's micro-expression features from the keyframes; S6. Based on the keyframes, analyze the user's lip movement features to assess articulation disorders and obtain articulation disorder features; S7. The micro-expression features and the articulation disorder features are fused using multimodal features to obtain the fused features; S8. Input the fused features into the pre-trained pAD risk assessment model to obtain the pAD risk assessment result.

2. The non-invasive screening method for the prodromal stage of Alzheimer's disease based on facial expression features according to claim 1, characterized in that, The face detection and alignment process described in step S2 specifically includes: Face detection and 68 key point localization were performed using a multi-task convolutional neural network algorithm. The face image is normalized to 256×256 pixels using affine transformation; An adaptive histogram equalization technique was used for illumination normalization.

3. The non-invasive screening method for the prodromal stage of Alzheimer's disease based on facial expression features according to claim 1, characterized in that, The topic modeling algorithm described in step S3 uses the latent Dirichlet distribution model.

4. The non-invasive screening method for the prodromal stage of Alzheimer's disease based on facial expression features according to claim 1, characterized in that, The keyframe extraction described in step S4 is achieved in the following way: Calculate the optical flow energy value between consecutive frames in a video sequence; The optical flow energy value is calculated using the following formula: in, and Each pixel exist Always and Optical flow component in the direction; The frame corresponding to the local maximum value of the optical flow energy is selected as the key frame.

5. The non-invasive screening method for the prodromal stage of Alzheimer's disease based on facial expression features according to claim 1, characterized in that, The micro-expression feature extraction described in step S5 is implemented using a deep convolutional neural network; The deep convolutional neural network is a ResNet-50 architecture; Output a 512-dimensional micro-expression feature vector.

6. The non-invasive screening method for the prodromal stage of Alzheimer's disease based on facial expression features according to claim 1, characterized in that, The articulation disorder feature analysis described in step S6 specifically includes: The lip region was extracted using a facial key point detection algorithm; A three-dimensional convolutional neural network was used to extract the spatiotemporal features of lip movements; Modeling temporal dynamic features using long short-term memory networks; Output a score indicating the severity of articulation disorder.

7. The non-invasive screening method for the prodromal stage of Alzheimer's disease based on facial expression features according to claim 1, characterized in that, The feature fusion method described in step S7 specifically includes: The dynamic time warping algorithm is used to perform temporal alignment of micro-expression feature sequences and lip movement feature sequences; The attention weights of each modality feature are calculated using an attention mechanism; The attention weights are used to perform weighted fusion to obtain the fused features.

8. The non-invasive screening method for the prodromal stage of Alzheimer's disease based on facial expression features according to claim 1, characterized in that, The pAD risk assessment model described in step S8 is pre-trained using a contrastive learning framework; The loss function of the contrastive learning framework is: in, These are feature representations of different enhanced versions from the same original sample.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1-8.

10. A non-invasive screening device for the prodromal stage of Alzheimer's disease based on facial expression features, comprising the electronic device (1) as described in claim 9, characterized in that, It also includes two opposing and vertically arranged support columns (2), each support column (2) is hollow, and the upper end of each support column (2) is rotatably connected to a worm gear (3). Each support column (2) has an installation cavity (4) on one side of the corresponding worm gear (3). Each installation cavity (4) is equipped with a vertically arranged drive motor (5), and the power output shaft of each drive motor (5) is connected to a worm (6) that meshes with the corresponding worm gear (3). Each support column (2) is connected to a vertically arranged collection column (7). Each collection column (7) has a rotating groove (8) facing the corresponding support column (2) on its lower side. Each support column (2) has a rotating notch (9) on its upper part facing the corresponding collection column (7). Each worm gear (3) is connected to a rotating connecting arm (10) that passes through the corresponding rotating notch (9) and rotates with the corresponding rotating groove (8). Each connecting column has a parallel connecting arm (11) that passes through the corresponding rotating notch (9) and rotates with the corresponding rotating groove (8) at its upper end. Each parallel connecting arm (11) is always parallel to the corresponding rotating connecting arm (10), so that each collection column (7) and the corresponding support column (2) are always parallel. Each rotating connecting arm (10) has a wire groove (12) on its lower side that is open at both ends and is arranged along its length. Each parallel connecting arm (11) has a wire outlet (13) at one end near the corresponding support column (2). Each rotating groove (8) has a transmission cable (14) connected to its top. Each transmission cable (14) passes through the corresponding wire groove (12) and the corresponding wire outlet (13) in sequence and extends into the corresponding support column (2) and is electrically connected to the electronic device (1). Each acquisition column (7) has an acquisition camera (15) installed on the upper part of the side away from the corresponding support column (2), and an infrared sensor (16) is installed on the lower part of the side away from the corresponding support column (2). The acquisition camera (15) and the infrared sensor (16) are electrically connected to the electronic device (1) through the corresponding transmission cable (14). The two acquisition columns (7) are arranged in a figure-eight shape so that the two acquisition cameras (15) can face the statistical target.

Images