Auxiliary screening system for Alzheimer's disease

Abstract

The invention discloses an auxiliary screening system for Alzheimer's disease. The auxiliary screening system comprises a physiological data acquisition unit, an electric signal preprocessing unit, an expression image processing unit, an expression feature acquisition unit, a time alignment unit, a feature fusion unit and a screening result acquisition unit. Firstly, multi-channel electroencephalogram signal data of a user and expression image data containing the face of the user are synchronously collected, and electroencephalogram features and expression features are extracted respectively; carrying out feature splicing on the electroencephalogram features and the expression features; and finally, inputting a fusion feature vector obtained by feature splicing into an emotional memory disorder recognition model to obtain an Alzheimer's disease early-stage emotional memory risk index and a classification result output by the emotional memory disorder recognition model. Due to the fact that the EEG and the micro-expression are subjected to multi-mode fusion, high-sensitivity recognition and early warning of AD early emotional dysmnesia are achieved through dynamic fusion analysis of neural electrophysiology and facial behavior signals.

Description

Technical Field

[0001] This invention relates to the field of health screening technology, and more specifically to an auxiliary screening system for Alzheimer's disease. Background Technology

[0002] Alzheimer's disease (AD) is an insidious, progressive neurodegenerative disease. Clinically, it is characterized by comprehensive dementia manifestations such as memory impairment, aphasia, agnosia, executive dysfunction, visuospatial skill impairment, and personality and behavioral changes, and it primarily affects the elderly. Early-stage Alzheimer's disease can be treated and controlled with medication, which can help improve cognitive function and delay the clinical progression by 10 to 15 years. Therefore, early screening for Alzheimer's disease is of great significance in controlling its development. With the development of machine learning and deep learning technologies in image recognition, preliminary results have been achieved in using clinical data from methods such as magnetic resonance imaging (MRI) and positron emission tomography (PET) to predict the occurrence and evolution of Alzheimer's disease. Notably, the Montreal Cognitive Assessment (MoCA) is also an important auxiliary screening tool for Alzheimer's disease, and compared to methods such as MRI and PET, MoCA is more cost-effective and therefore more readily accepted by potential patients. Traditional MoCA testing typically involves a face-to-face, one-on-one questionnaire between the patient and doctor. During the test, healthcare professionals guide participants to complete and answer specific questions on a standardized paper questionnaire, thereby assessing their behavioral abilities and disease progression. However, this screening method suffers from several drawbacks, including a heavy workload for healthcare workers, inconsistent evaluation standards, a limited number of participants screened per unit of time, and difficulties in digitizing and displaying paper questionnaires. These limitations significantly restrict screening efficiency and results, as well as the digital management of clinical data. The question then arises: how can new artificial intelligence technologies be utilized to more efficiently and accurately achieve early warning and preliminary screening for Alzheimer's disease, thereby mitigating its adverse impact on current societal development? Summary of the Invention

[0003] The main technical problem addressed by this invention is how to improve and optimize early screening for Alzheimer's disease.

[0004] According to a first aspect, one embodiment provides an auxiliary screening system for Alzheimer's disease, comprising: The physiological data acquisition unit is used to simultaneously acquire multi-channel electroencephalogram (EEG) signal data and facial expression image data of the user; wherein, the time correspondence between the EEG signal data and the facial expression image data is established by identifying the acquisition time of the EEG signal sampling points and the acquisition time of the facial expression image frames; An electrical signal preprocessing unit is used to extract the EEG features of the EEG signal data, the EEG features including θ inhibition depth, θ oscillation recovery rate and / or θ–γ cross-frequency coupling strength; An expression image processing unit is used to convert the video stream of the expression image data into facial behavior data that is quantized over time; wherein, the facial behavior includes instantaneous micro-expression events; The facial expression feature acquisition unit is used to construct a micro-expression co-activation network based on the facial behavior data, and input the micro-expression co-activation network into a preset graph attention network to obtain the facial expression features output by the graph attention network. The facial expression features include dynamic topology features of network density, dynamic topology features of clustering coefficients, and / or dynamic topology features of emotion entropy. A time alignment unit is used to perform time-series alignment of the electroencephalogram (EEG) signal data and the facial behavior data; The feature fusion unit is used to concatenate the EEG features and the facial expression features to obtain a fused feature vector. The screening result acquisition unit is used to input the fused feature vector into a preset emotional memory disorder recognition model, and to acquire the early emotional memory risk index and classification results of Alzheimer's disease output by the emotional memory disorder recognition model.

[0005] In one embodiment, the emotional memory disorder recognition model is established using the XGBoost classification model.

[0006] In one embodiment, the screening result acquisition unit is further configured to apply the LASSO regression algorithm to perform feature screening on the fused feature vector, and input the screened fused feature vector into the emotional memory disorder identification model to obtain the early emotional memory risk index and classification results of Alzheimer's disease; wherein, the feature screening of the fused feature vector is used to select non-redundant fused feature vectors that have discriminative power for identifying emotional memory disorders from high-dimensional features.

[0007] In one embodiment, temporal alignment of the electroencephalogram (EEG) signal data and the facial behavior data includes: A hierarchical dynamic time warping algorithm is applied in combination with optical flow motion compensation to perform temporal alignment on the EEG signal data and the facial behavior data, so as to ensure that the synchronization deviation of the EEG signal data and the facial behavior data on the time axis is less than a preset time value.

[0008] In one embodiment, the optical flow motion compensation algorithm uses the Farneback dense optical flow model to perform facial motion correction through inter-frame pixel displacement estimation, so as to ensure that the displacement error does not exceed a preset number of pixels.

[0009] In one embodiment, converting the video stream of the facial expression image data into facial behavior data that is quantized over time includes: An action unit recognition algorithm is applied to extract the AU intensity time series, and instantaneous micro-expression events are detected through continuous wavelet transform; The construction of the micro-expression co-activation network based on the facial behavior data includes: A micro-expression co-activation network is constructed based on the correlation between AUs in the facial behavior data.

[0010] In one embodiment, the electroencephalogram (EEG) features and the facial expression features are spliced ​​together, including: The phase-behavior coupling index is calculated to quantify the modulation relationship between the EEG theta phase and AU intensity, in order to form the fusion feature vector.

[0011] In one embodiment, the electrical signal preprocessing unit is further configured to apply bandpass filtering and independent component analysis methods to filter and remove artifacts from the EEG signal data.

[0012] In one embodiment, the feature fusion unit is further used to calculate the Shannon entropy distribution of AU intensity in different θ phase intervals by referencing the phase-behavior modulation index, in order to measure the emotional coupling degree of the fused feature vector.

[0013] In one embodiment, the screening result acquisition unit is further configured to convert the early emotional memory risk index and classification results of Alzheimer's disease into individualized emotional memory risk curves and classification confidence distribution maps for visualization display according to the auxiliary screening system of the above embodiment. Based on the auxiliary screening system in the above embodiments, EEG and micro-expression are fused in a multimodal manner to achieve highly sensitive identification and early warning of early emotional memory disorders in AD through dynamic fusion analysis of neurophysiological and facial behavioral signals. It can quickly and objectively assess cognitive function status and has the advantages of high accuracy, strong interpretability and good applicability. Attached Figure Description

[0014] Figure 1 This is a functional structure block diagram of an auxiliary screening system in one embodiment; Figure 2 This is a schematic diagram of the workflow of an auxiliary screening system in one embodiment; Figure 3 This is a schematic diagram of multimodal acquisition in one embodiment; Figure 4 This is a schematic diagram of cross-modal feature fusion in one embodiment. Detailed Implementation

[0015] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of this application. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to this application are not shown or described in the specification. This is to avoid obscuring the core parts of this application with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0016] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.

[0017] The serial numbers assigned to components in this document, such as "first" and "second," are used only to distinguish the described objects and have no sequential or technical meaning. The terms "connection" and "linkage" used in this application, unless otherwise specified, include both direct and indirect connections (linkages).

[0018] Alzheimer's disease is a neurodegenerative disease characterized primarily by progressive memory decline, often accompanied in its early stages by significant mood disturbances such as anxiety, apathy, and depression. Numerous studies have confirmed that these non-cognitive symptoms often appear before memory decline and accelerate disease progression. However, currently, there is a lack of objective and quantitative methods for detecting abnormal emotional functioning in clinical practice. To reduce the risk of misdiagnosis and improve the accuracy of assessment, simpler and more sensitive methods need to be introduced to supplement consciousness assessment. Patients with unresponsive wakefulness syndrome (UWS) are unaware of their surroundings and themselves, while patients with minimally conscious states (MCS) have weaker consciousness. Therefore, it can be hypothesized that patients experiencing emotional changes in a passive paradigm have a better chance of recovery. In recent years, due to the rapid development of computer vision and brain-computer interfaces (BCIs), the accuracy of emotion recognition based on facial images and electroencephalography (EEG) has improved. Compared to assessment methods based on motor responses and fMRI, multimodal emotion recognition based on EEG and microexpressions can more easily diagnose the residual state of consciousness in patients with impaired consciousness and serve as an auxiliary diagnostic tool.

[0019] At the neural mechanism level, the consolidation of emotional memories depends on the brain's offline reactivation process, which is regulated by theta (4–8 Hz) oscillations in the prefrontal-hippocampal-limbic system. Theta oscillation recovery rate (TORR) and theta-γ cross-frequency coupling (CFC) reflect the dynamic integration capacity of neural networks. AD patients show significantly decreased TORR and CFC, suggesting impaired synchronicity of neural oscillations. On the other hand, microexpressions are extremely brief (40–500 ms), unconscious facial muscle activities that can truly reflect the dynamic balance between emotional impulses and prefrontal inhibition. Prefrontal cortex dysfunction in AD patients may lead to altered microexpression patterns. Therefore, joint analysis of EEG signals and microexpressions can provide integrated central-peripheral information on emotion regulation. However, existing EEG and video data struggle to achieve high-precision temporal alignment and multimodal fusion, making it impossible to quantify the dynamic coupling relationship between the two signals. This application provides a systematic solution for signal acquisition, processing, and risk assessment of emotional memory disorders. Example

[0020] Please refer to Figure 1This is a functional block diagram of an auxiliary screening system in one embodiment. The auxiliary screening system is used for auxiliary screening of Alzheimer's disease and specifically includes a physiological data acquisition unit 100, an electrical signal preprocessing unit 200, an facial expression image processing unit 300, an facial expression feature acquisition unit 400, a time alignment unit 500, a feature fusion unit 600, and a screening result acquisition unit 700. The physiological data acquisition unit 100 is used to simultaneously acquire multi-channel EEG signal data and facial expression image data containing the user's face. The temporal correspondence between the EEG signal data and the facial expression image data is established by identifying the acquisition time of the EEG signal sampling points and the acquisition time of the facial expression image frames. The electrical signal preprocessing unit 200 is used to extract EEG features from the EEG signal data, including theta inhibition depth, theta oscillation recovery rate, and / or theta-γ cross-frequency coupling strength. In one embodiment, the electrical signal preprocessing unit 200 is also used to apply bandpass filtering and independent component analysis methods to filter and remove artifacts from the EEG signal data. The facial expression image processing unit 300 converts the video stream of facial expression image data into facial behavior data that is quantized over time, wherein facial behavior includes instantaneous micro-expression events. In one embodiment, an action unit recognition algorithm is applied to extract the AU intensity time series, and instantaneous micro-expression events are detected through continuous wavelet transform. The facial expression feature acquisition unit 400 constructs a micro-expression co-activation network based on the facial behavior data, and inputs the micro-expression co-activation network into a preset graph attention network to obtain the facial expression features output by the graph attention network. The facial expression features include dynamic topological features of network density, dynamic topological features of clustering coefficients, and / or dynamic topological features of emotion entropy. In one embodiment, the micro-expression co-activation network is constructed based on the correlation between AUs in the facial behavior data. The time alignment unit 500 performs time alignment on the EEG signal data and facial behavior data. In one embodiment, a hierarchical dynamic time warping algorithm combined with optical flow motion compensation is applied to perform time alignment on the EEG signal data and facial behavior data to ensure that the synchronization deviation between the EEG signal data and facial behavior data on the time axis is less than a preset time value. The optical flow motion compensation algorithm employs the Farneback dense optical flow model and performs facial motion correction through inter-frame pixel displacement estimation to ensure that the displacement error does not exceed a preset number of pixels. The feature fusion unit 600 concatenates EEG features and facial expression features to obtain a fused feature vector. In one embodiment, a phase-behavior coupling index is calculated to quantify the modulation relationship between the EEG theta phase and AU intensity, forming the fused feature vector. In another embodiment, the feature fusion unit 600 also uses the phase-behavior modulation index to calculate the Shannon entropy distribution of AU intensity within different theta phase intervals to measure the emotional coupling degree of the fused feature vector. The screening result acquisition unit 700 inputs the fused feature vector into a preset emotional memory disorder recognition model and obtains the early emotional memory risk index and classification results of Alzheimer's disease output by the emotional memory disorder recognition model.In one embodiment, the emotional memory disorder identification model is established using an XGBoost classification model. In another embodiment, the screening result acquisition unit 700 is further configured to apply a LASSO regression algorithm to perform feature filtering on the fused feature vector, and input the filtered fused feature vector into the emotional memory disorder identification model to obtain the early emotional memory risk index and classification results for Alzheimer's disease. The feature filtering of the fused feature vector is used to select non-redundant fused feature vectors with discriminative power for identifying emotional memory disorders from high-dimensional features. In another embodiment, the screening result acquisition unit 700 is further configured to convert the early emotional memory risk index and classification results for Alzheimer's disease into individualized emotional memory risk curves and classification confidence distribution maps for visualization.

[0021] Please refer to Figure 2 This is a schematic diagram of the workflow of an auxiliary screening system in one embodiment. The workflow of the auxiliary screening system specifically includes: Step 101, Multimodal Acquisition.

[0022] Simultaneously collect multi-channel EEG signal data and facial expression image data of the user, including the acquisition time of EEG signal sampling points and the acquisition time of facial expression image frames.

[0023] Please refer to Figure 3 This diagram illustrates multimodal acquisition in one embodiment. In one embodiment, EEG signal data acquisition uses 16-lead electrodes (Fp1, Fp2, F3, F4, C3, C4, P3, P4, O1, O2, T3, T4, F7, F8, T5, T6), with a sampling rate ≥500Hz. Reference electrodes A1 / A2 are used. Skin impedance is maintained <5 kΩ during sampling. The acquired data includes a resting state (2 minutes), an emotion-evoked phase (8 positive / neutral / negative images each), and a recovery phase (15 minutes). Synchronous trigger signals ensure that event markers are consistent with video time frames. Facial expression images are high-frame-rate facial videos. Facial expression image data acquisition uses a high-definition camera with a frame rate of 120fps and a resolution of 1920×1080, maintaining a frontal angle and D65 standard lighting to avoid shadows and occlusions. The entire facial expression change of the subject is recorded and time-stamped during acquisition. A unified clock synchronization mechanism achieves time alignment between the EEG signal and the facial expression video, with an error of less than 5 ms.

[0024] Step 102, EEG signal preprocessing.

[0025] The raw EEG signal is denoised to obtain EEG signal data, followed by feature extraction. In one embodiment, the denoising and artifact removal process uses a 0.5–50 Hz bandpass filter and a 50 Hz notch filter to remove power frequency interference, and then uses independent component analysis (ICA) to remove eye movement, blinking, and electromyography artifacts. The signal is segmented and artifacts are labeled. The feature extraction process includes: Calculate the theta wave power (4–8 Hz) and gamma wave power (30–80 Hz), and extract three types of indicators: 1. Theta suppression depth (TID): The percentage decrease in θ power relative to the resting state during the mission phase is calculated using the following formula: TID=(P base - P task ) / P base ×100%; Among them, P base It is the baseline power, P task It is the power during the mission period.

[0026] 2. Theta oscillation recovery rate (TORR): The time constant for recovering to the baseline is obtained by fitting an exponential function, and the calculation formula is as follows: P(t) = P0·e -t / τ +C; Where P(t) is the instantaneous power of θ power during the resting state recovery process, and τ represents the recovery time constant (seconds).

[0027] 3. θ–γ interfrequency coupling strength (CFC): The phase-amplitude coupling strength is calculated using the Tort modulation index.

[0028] These features are used to quantify the mechanisms of prefrontal cortex functional inhibition and mood recovery, and are important indicators of early neurophysiological changes in Alzheimer's disease (AD).

[0029] Step 103, Micro-expression feature extraction.

[0030] This includes the micro-expression extraction process and the feature extraction process. The micro-expression extraction process includes: Action Unit (AU) recognition algorithms (such as OpenFace) are used to extract AU intensity sequences (values ​​0–5). AUs associated with emotion (AU1, AU4, AU6, AU12, AU15, AU17, etc.) are selected. Brief micro-expression events lasting 40–500 ms are detected using continuous wavelet transform (CWT); time-frequency analysis is performed on the AU intensity sequences to detect instantaneous AU activation events, yielding the following formula: ; Where x(t) represents the AU intensity signal, Ψ is the Morlet mother wavelet, s is the scale parameter, τ is the time shift parameter, and each detected micro-expression event records the occurrence time, duration, and peak amplitude.

[0031] In one embodiment, the automatic identification results are manually checked and calibrated to ensure that the AU identification is accurate and reliable.

[0032] Step 104, Graph Attention Network Analysis.

[0033] The feature extraction process involves the construction and analysis of co-activation networks, specifically including: Calculate the pairwise Pearson correlation coefficients of AUs within a 1-second sliding window (100 ms step). A co-activation edge is considered to exist when r > 0.6, where nodes represent AUs and edge weights represent co-activation strengths. The input is a Graph Attention Network (GAT), specifically: First, the nodes are calculated using a self-attention mechanism. Its neighboring nodes Attention coefficient between: ; in, The feature mapping weight matrix, Here, || represents the attention parameters, and || denotes feature concatenation. The attention coefficients are then normalized using the softmax function. ; in, For nodes The set of neighbors.

[0034] Node features are updated to a weighted sum of neighboring nodes: ; Where δ(·) is the activation function.

[0035] GAT achieves efficient extraction of dynamic network features through the above mechanism, thereby improving the performance of emotion or disease classification models.

[0036] Step 105, timing alignment.

[0037] To achieve precise synchronization between EEG and micro-expression signals, a layered dynamic time warping (HDTW) algorithm combined with optical flow compensation is employed. The layered DTW alignment includes a coarse alignment stage and a fine alignment stage. The coarse alignment stage uses a downsampled signal (10 Hz) for cosine similarity matching, while the fine alignment stage performs high-resolution DTW matching within a ±50 ms window. Finally, an optimal time mapping path is output, ensuring that the time deviation between the two modalities is ≤ ±30 ms. Optical flow compensation uses the Farneback dense optical flow algorithm to calculate the pixel displacement of adjacent frames and performs motion correction on the AU intensity sequence, controlling the average error to within 2 pixels. Through this dual-layer mechanism of HDTW and optical flow compensation, millisecond-level synchronization of different modal signals can be achieved.

[0038] Step 106, cross-modal feature fusion.

[0039] At each time point t, EEG and micro-expression features are concatenated to complete the initial fusion. The EEG features (FEEG) include TID, RR, and MI, while the micro-expression features (FAU) include network density, clustering coefficient, and emotion entropy.

[0040] Please refer to Figure 4 This is a schematic diagram of cross-modal feature fusion in one embodiment. All features are normalized, specifically including feature concatenation and phase-behavior coupling calculation. Feature concatenation involves concatenating F1 and F2 within the same time window to form a joint feature matrix. ; Z-score standardization is applied to each feature.

[0041] Phase-behavior coupling calculation (PBC) extracts the instantaneous phase by performing a Hilbert transform on the EEG theta wave. The phase is divided into 18 equally spaced intervals, and the modulation index (MI) between the EEG theta phase and the AU intensity of the micro-expression is calculated as the PBC. Specifically: ; in, This represents the AU intensity distribution across different θ phase intervals. For Shannon entropy, =18.

[0042] Step 107: Feature selection and classification prediction.

[0043] Feature selection used LASSO regression (λ=0.01) to select the most discriminative fused features. The XGBoost ensemble learning algorithm was used for model training, with hyperparameters of 200 trees, a maximum depth of 6, and a learning rate of 0.1. The model input was the fused feature vector, and the output was the classification probability. The Emotion–Memory Risk Index (EMI) was calculated to generate an individualized assessment report, including risk level, contribution of key features, and confidence distribution.

[0044] In the embodiments of this application, the training set for the XGBoost ensemble learning algorithm consists of 150 AD patients and 150 controls. The training results show that the multimodal model has an AUC of 0.91, with a 12% improvement in accuracy; the PBC index and emotional memory score are r=0.63 (p<0.001), indicating that the auxiliary screening system in this embodiment can sensitively detect early emotional memory impairment in AD. In one embodiment, the auxiliary screening system is embedded in a medical auxiliary diagnostic platform to achieve real-time emotional memory assessment.

[0045] The embodiments of this application disclose an auxiliary screening system that detects early emotional memory impairment in Alzheimer's disease based on the multimodal fusion of EEG and microexpression. Specifically, it includes a physiological data acquisition unit, an electrical signal preprocessing unit, an facial expression image processing unit, an facial expression feature acquisition unit, a time alignment unit, a feature fusion unit, and a screening result acquisition unit.

[0046] The physiological data acquisition unit simultaneously acquires the user's electroencephalogram (EEG) signals and high-frame-rate facial video signals to achieve a temporal correspondence between central nervous system activity and peripheral emotional behavior. In one embodiment, the acquisition process includes a resting state, an emotion induction phase, and a recovery phase. The emotional stimulus materials use positive, neutral, and negative image stimuli, and the data from each phase are synchronously time-stamped for subsequent alignment. A unified time reference for multimodal signals is achieved through a synchronous triggering mechanism.

[0047] The electrical signal preprocessing unit preprocesses the acquired raw EEG signals to remove artifacts and extract dynamic oscillation features related to emotional memory. In one embodiment, a 0.5–50 Hz bandpass filter is used, with a 50 Hz notch filter superimposed to remove power line interference. Eye movement, blinking, and electromyography artifacts are eliminated using the Independent Component Analysis (ICA) algorithm. Theta inhibition depth (TID), theta oscillation recovery rate (TORR), and theta-gamma cross-frequency coupling (CFC) are extracted. TID is the percentage decrease in task-state theta power compared to the resting state; TORR reflects the speed at which theta power recovers to baseline levels after the task; and CFC is calculated using the Tort modulation index to determine the degree of coupling between theta phase and gamma amplitude. These features are used to characterize the neural basis of prefrontal dysfunction and emotion regulation disorders in AD patients.

[0048] The facial expression processing unit performs frame-by-frame analysis of high frame rate videos based on Action Unit (AU) recognition algorithms to extract micro-expression dynamic features. It identifies major AUs related to emotion (such as AU1, AU4, AU6, AU12, AU15, AU17, etc.) and extracts their intensity time series. Continuous Wavelet Transform (CWT) is used to analyze the time-frequency changes in AU intensity, detecting instantaneous micro-expression events with durations of 40–500 milliseconds. An AU co-activation matrix is ​​constructed based on the temporal correlation between different AUs to reflect the competitive relationship between emotional impulses and cognitive inhibition.

[0049] The facial expression feature acquisition unit captures the complex dynamic interaction features between AUs. It inputs the AU co-activation network into a Graph Attention Network (GAT) model, calculates the attention weights between nodes, and extracts high-order topological features. The node input is the temporal intensity vector of each AU, and the edge weights are defined by the AU correlation coefficient r > 0.6. The GAT model calculates the weights of neighboring nodes based on a self-attention mechanism and generates node embedding vectors. The output features include metrics such as network density, average clustering coefficient, emotion entropy, and node centrality, which are used to characterize the coordination and complexity of emotional expression.

[0050] To ensure precise temporal matching between EEG signals and micro-expression signals, this invention employs a Hierarchical Dynamic Time Warping (HDTW) algorithm combined with optical flow motion compensation technology for millisecond-level temporal alignment. HDTW consists of two stages: global coarse alignment and local fine alignment. The global stage utilizes downsampled signals for cosine similarity matching; the local stage optimizes the alignment path between EEG and AU signals within a ±50 ms window; optical flow compensation uses the Farneback dense optical flow model to calculate the pixel displacement field of adjacent frames and performs motion correction on the AU intensity sequence, controlling the head movement displacement error to within 2 pixels. After alignment, the time deviation is controlled within ±30 milliseconds, achieving high-precision synchronization between neural signals and facial expression changes.

[0051] The feature fusion unit fuses EEG features with micro-expression features to construct a multimodal feature matrix and extract synergistic features between neural and behavioral signals. At each time point, EEG feature vectors (TID, TORR, CFC) are concatenated with micro-expression network features (density, clustering coefficient, emotion entropy). The Phase-Behavior Coupling (PBC) index is used to calculate the modulation relationship between the EEG theta wave phase and the AU intensity. Dividing the signal into 18 theta phase intervals, the Shannon entropy of the AU intensity distribution is calculated to obtain the cross-modal modulation index (MI), which reflects the dynamic coupling degree between neural signals and external facial expressions.

[0052] The screening result acquisition unit constructs a predictive model for early emotional memory impairment in Alzheimer's disease using fused multimodal features. LASSO regression is employed for feature sparsity reduction and important feature selection; an XGBoost ensemble learning algorithm is used to build a classification model, taking the selected feature vectors as input and outputting the classification result for emotional memory impairment in Alzheimer's disease. The predicted Emotion-Memory Risk Index (EMI) is calculated to quantify individual risk levels.

[0053] The assisted screening system disclosed in this application includes a physiological data acquisition unit, an electrical signal preprocessing unit, an facial expression image processing unit, an facial expression feature acquisition unit, a time alignment unit, a feature fusion unit, and a screening result acquisition unit. First, it simultaneously acquires multi-channel EEG signal data and facial expression image data containing the user's face, and extracts EEG features and facial expression features respectively. Then, it performs feature splicing on the EEG and facial expression features. Finally, it inputs the fused feature vector obtained from the feature splicing into an emotional memory disorder recognition model to obtain the early emotional memory risk index and classification results of Alzheimer's disease output by the emotional memory disorder recognition model. By fusing EEG and micro-expressions in a multimodal manner, it achieves highly sensitive identification and early warning of early emotional memory disorders in AD through dynamic fusion analysis of neurophysiological and facial behavioral signals.

[0054] The above examples illustrate the present invention only to aid in understanding it and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention.

Claims

1. An auxiliary screening system for Alzheimer's disease, characterized in that, include: The physiological data acquisition unit is used to simultaneously acquire multi-channel electroencephalogram (EEG) signal data and facial expression image data of the user; wherein, the time correspondence between the EEG signal data and the facial expression image data is established by identifying the acquisition time of the EEG signal sampling points and the acquisition time of the facial expression image frames; An electrical signal preprocessing unit is used to extract the EEG features of the EEG signal data, the EEG features including θ inhibition depth, θ oscillation recovery rate and / or θ–γ cross-frequency coupling strength; An expression image processing unit is used to convert the video stream of the expression image data into facial behavior data that is quantized over time; wherein, the facial behavior includes instantaneous micro-expression events; The facial expression feature acquisition unit is used to construct a micro-expression co-activation network based on the facial behavior data, and input the micro-expression co-activation network into a preset graph attention network to obtain the facial expression features output by the graph attention network. The facial expression features include dynamic topology features of network density, dynamic topology features of clustering coefficients, and / or dynamic topology features of emotion entropy. A time alignment unit is used to perform time-series alignment of the electroencephalogram (EEG) signal data and the facial behavior data; The feature fusion unit is used to concatenate the EEG features and the facial expression features to obtain a fused feature vector. The screening result acquisition unit is used to input the fused feature vector into a preset emotional memory disorder recognition model, and to acquire the early emotional memory risk index and classification results of Alzheimer's disease output by the emotional memory disorder recognition model.

2. The auxiliary screening system as described in claim 1, characterized in that, The emotional memory disorder identification model was built using the XGBoost classification model.

3. The auxiliary screening system as described in claim 2, characterized in that, The screening result acquisition unit is also used to apply the LASSO regression algorithm to perform feature screening on the fusion feature vector, and input the screened fusion feature vector into the emotional memory disorder recognition model to obtain the early emotional memory risk index and classification results of Alzheimer's disease; wherein, the feature screening of the fusion feature vector is used to select non-redundant fusion feature vectors that have discriminative power for identifying emotional memory disorders from high-dimensional features.

4. The auxiliary screening system as described in claim 1, characterized in that, Temporal alignment of the electroencephalogram (EEG) signal data and the facial behavior data includes: A hierarchical dynamic time warping algorithm is applied in combination with optical flow motion compensation to perform temporal alignment on the EEG signal data and the facial behavior data, so as to ensure that the synchronization deviation of the EEG signal data and the facial behavior data on the time axis is less than a preset time value.

5. The auxiliary screening system as described in claim 4, characterized in that, The optical flow motion compensation algorithm uses the Farneback dense optical flow model to perform facial motion correction through inter-frame pixel displacement estimation, so as to ensure that the displacement error does not exceed the preset number of pixels.

6. The auxiliary screening system as described in claim 1, characterized in that, The video stream of the facial expression image data is converted into facial behavior data that is quantized over time, including: An action unit recognition algorithm is applied to extract the AU intensity time series, and instantaneous micro-expression events are detected through continuous wavelet transform; The construction of the micro-expression co-activation network based on the facial behavior data includes: A micro-expression co-activation network is constructed based on the correlation between AUs in the facial behavior data.

7. The auxiliary screening system as described in claim 6, characterized in that, The brainwave features and facial expression features are concatenated, including: The phase-behavior coupling index is calculated to quantify the modulation relationship between the EEG theta phase and AU intensity, in order to form the fusion feature vector.

8. The auxiliary screening system as described in claim 7, characterized in that, The feature fusion unit is also used to calculate the Shannon entropy distribution of AU intensity in different θ phase intervals by referencing the phase-behavior modulation index, in order to measure the emotional coupling degree of the fused feature vector.

9. The auxiliary screening system as described in claim 1, characterized in that, The electrical signal preprocessing unit is also used to filter and remove artifacts from the EEG signal data using bandpass filtering and independent component analysis methods.

10. The auxiliary screening system as described in claim 1, characterized in that, The screening result acquisition unit is also used to convert the early emotional memory risk index and classification results of Alzheimer's disease into individualized emotional memory risk curves and classification confidence distribution maps for visualization.

Images