Dementia disease recognition method and system based on space-time-frequency multi-dimensional feature fusion

Abstract

The application discloses a dementia disease recognition method and system based on space-time-frequency multi-dimensional feature fusion, and belongs to the technical field of intelligent medical treatment. The method comprises the following steps: preprocessing an electroencephalogram signal of a target user; extracting space-time-frequency multi-dimensional features of the preprocessed electroencephalogram signal; and classifying based on the space-time-frequency multi-dimensional features to obtain a dementia disease recognition result. The application can recognize dementia diseases such as Alzheimer's disease and frontotemporal dementia with high precision and automation.

Description

Technical Field

[0001] This invention belongs to the field of smart medical technology, specifically relating to a method and system for identifying dementia-like diseases based on spatiotemporal frequency multidimensional feature fusion. Background Technology

[0002] With the increasing aging of the population, the incidence of neurodegenerative diseases, especially dementia, is rising year by year. Alzheimer's disease (AD) and frontotemporal dementia (FTD) are the two most common types, seriously threatening the cognitive health and quality of life of the elderly. Currently, the clinical diagnosis of dementia mainly relies on neuropsychological scale assessments, cerebrospinal fluid biomarker detection, and imaging techniques such as positron emission tomography (PET) or magnetic resonance imaging (MRI). However, these methods generally suffer from high costs, complex operation, poor accessibility, or invasiveness, making it difficult to meet the needs of large-scale screening and long-term dynamic monitoring.

[0003] Electroencephalography (EEG), as a non-invasive neurophysiological detection technique, boasts advantages such as millisecond-level temporal resolution, portable equipment, and low operating costs, showing broad application prospects in cognitive function assessment and neurological disease detection. Studies have shown that Alzheimer's disease (AD) patients often exhibit characteristics such as decreased posterior alpha rhythm power, enhanced slow wave (δ and θ bands) activity, and abnormal brain network functional connectivity; while Freezing Toxic Disease (FTD) patients often show abnormal electrical activity in the frontal lobe region. In recent years, machine learning-based EEG analysis methods have been widely used in the automatic identification of dementia. Existing techniques typically employ traditional spectral analysis combined with classifiers such as Support Vector Machines (SVM) or Random Forests for modeling, but feature extraction is often limited to a single dimension (e.g., only the frequency domain or only the time domain), failing to fully explore the joint information of EEG signals in the three dimensions of time, space, and frequency, resulting in limited feature representation ability and affecting the stability and generalization ability of classification performance. In addition, most studies lack a systematic data preprocessing workflow, relying on manual intervention for key steps such as bad derivative interpolation, rereference, and artifact removal, resulting in low automation and affecting the reproducibility of the results.

[0004] While existing technologies have made some progress in EEG-based dementia identification, many shortcomings remain. Methods exemplified by patent CN120277489A, which use wavelet transform to generate time-frequency maps and combine PSO for channel selection to improve classification performance, rely on computationally complex optimization algorithms. Their preprocessing and feature extraction processes have low automation and lack systematic integration of temporal dynamic features, resulting in a single feature representation dimension. Furthermore, this method converts signals into images for processing, potentially losing the fine temporal structure information of the original signal. Patent CN113558636A, on the other hand, relies solely on permutation entropy, a single nonlinear indicator, for state analysis under musical stimulation. While it boasts ease of operation, it doesn't fully utilize feature information, making it difficult to comprehensively characterize the complex EEG pattern differences between different dementia types such as AD and FTD. Moreover, it lacks a complete end-to-end analysis process and accuracy analysis and system integration capabilities. Overall, existing technologies generally suffer from the following deficiencies: (1) The preprocessing process of EEG signals lacks standardization. Most methods do not effectively integrate modules for removing physiological artifacts such as electrooculogram and electrocardiogram, resulting in severe noise interference in the original signal and a low signal-to-noise ratio, which affects the accuracy of subsequent feature extraction and classification. (2) Feature extraction methods are mostly limited to a single dimension, such as using only frequency domain power features or nonlinear dynamic indicators. They fail to fully integrate the multidimensional information of EEG signals in the time domain, spatial domain (brain region distribution) and frequency domain, resulting in insufficient discrimination ability of the extracted features and difficulty in comprehensively representing the dynamic changes of brain functional state. (3) In the classification application of Alzheimer’s disease and frontotemporal dementia, the existing technology lacks a systematic evaluation of classification performance and does not provide multi-dimensional evaluation indicators including AUC curve, precision, recall and F1 score, resulting in insufficient model performance validation and insufficient reliability and clinical interpretability of the identification results. Summary of the Invention

[0005] This invention proposes a method and system for identifying dementia-like diseases based on spatiotemporal frequency multidimensional feature fusion. By constructing a complete technology chain of preprocessing, feature extraction, model classification, and system integration, it achieves high-precision and automated identification of AD and FTD dementia types, providing a highly efficient, reliable, and easily promoted auxiliary disease identification tool for clinical practice.

[0006] To achieve the above objectives, the technical solution of the present invention includes the following:

[0007] A method for identifying dementia-like diseases based on spatiotemporal frequency multidimensional feature fusion, the method comprising: Preprocessing the target user's EEG signals; Extracting spatiotemporal frequency multidimensional features from preprocessed EEG signals; Based on this spatiotemporal frequency multidimensional feature, classification was performed to obtain the identification results of dementia-related diseases.

[0008] Furthermore, the target user's EEG signals are preprocessed, including: Use a sliding window to segment the EEG signal and construct a Raw object based on the segmentation results; Electrode location is performed on the Raw object by combining electrode location information to obtain a Raw object with electrode location information; among them, for the Raw object with electrode location information, bad conductors are marked by interactive visualization and the bad conductors are repaired by interpolation using surrounding normal electrode data. The rereferenced Raw object is obtained based on the average electrode value of all Raw objects with electrode position information. Filter the rereferenced Raw object; Artifact removal is performed on the filtered Raw object to obtain the preprocessed EEG signal.

[0009] Furthermore, the rereferenced Raw object is filtered, including: Set the bandpass filter range; Capture within the bandpass filter range frequency band frequency band frequency band frequency band and The EEG data in the frequency band was filtered using a notch filter.

[0010] Furthermore, artifact removal is performed on the filtered Raw object to obtain the preprocessed EEG signal, including: Perform independent component analysis on the filtered Raw object to obtain the separation matrix; The separation matrix and the filtered Raw object are multiplied together to obtain the preprocessed EEG signal.

[0011] Furthermore, the spatiotemporal frequency multidimensional features of the preprocessed EEG signals are extracted, including: Based on the brain regions where the lead channels are located, the EEG signals are divided into several brain region datasets; wherein, the brain region datasets include the prefrontal region dataset, frontal region dataset, central region dataset, left temporal region dataset, right temporal region dataset, parietal region dataset, and occipital region dataset; For each brain region dataset, EEG data in the delta, theta, alpha, beta, and gamma bands are extracted respectively. The temporal and frequency domain features of the EEG data for each frequency band are extracted separately; wherein, the frequency domain features include spectral energy and differential entropy, and the temporal features include zero-crossing rate, peak-to-peak value, and variance; Based on the brain location of the lead channels, the EEG signals are divided into left hemisphere datasets and right hemisphere datasets. Based on the EEG signals in the left and right hemisphere datasets, hemisphere functional asymmetry features are extracted. The hemisphere functional asymmetry features include power asymmetry index and complexity asymmetry index. Based on the brain location of the lead channels, the EEG signals are divided into anterior brain region datasets and posterior brain region datasets. Based on the EEG signals in the anterior and posterior brain region datasets, functional gradient features of the anterior and posterior brain regions are extracted. The functional gradient features of the anterior and posterior brain regions include the fronto-occipital power gradient ratio and the fronto-occipital complexity gradient ratio. Based on the temporal and frequency domain features of EEG data in each frequency band of the brain region dataset, as well as hemispheric functional asymmetry features and anterior and posterior brain region functional gradient features, the spatiotemporal-frequency multidimensional features of the preprocessed EEG signal are obtained.

[0012] Furthermore, based on these spatiotemporal frequency multidimensional features, classification was performed to obtain dementia disease identification results, including: A hybrid classification model is constructed, comprising a CNN branch, an MLP branch, a concatenation layer, and an output layer. The CNN branch is used to extract local spatiotemporal features of EEG signals, the MLP branch is used to extract global abstract features of EEG signals, the concatenation layer is used to fuse local spatiotemporal features and global abstract features, and the output layer is used to classify the fused results. By inputting spatiotemporal frequency multidimensional features into a hybrid classification model, the identification results of dementia-related diseases are obtained.

[0013] Furthermore, the CNN branch includes several convolutional blocks and a global average pooling layer, and the several convolutional blocks form a pyramid structure with increasing feature depth; wherein, each convolutional block is composed of a convolutional layer, a batch normalization layer, an activation function layer, a pooling layer and a Dropout layer connected in sequence; a lightweight channel attention module is introduced after the first convolutional block, which generates channel weights based on global average pooling and shared MLP.

[0014] Furthermore, the MLP branch includes a feature attention module and several fully connected layers. The feature attention module is used to generate a feature weight matrix based on the fully connected layers and Sigmoid activation to weight the spatiotemporal frequency multidimensional features according to their importance. The several fully connected layers form a refined structure for feature compression, and each fully connected layer is followed by a batch normalization layer and a Dropout layer.

[0015] Furthermore, the hybrid classification model is trained based on a multi-level comprehensive regularization strategy; wherein the multi-level comprehensive regularization strategy includes structural regularization, training process regularization, and data-level regularization.

[0016] A dementia-related disease identification system based on spatiotemporal frequency multidimensional feature fusion, the system comprising: The preprocessing module is used to preprocess the EEG signals of the target user; The feature extraction module is used to extract the spatiotemporal frequency multidimensional features of the preprocessed EEG signal; The disease identification module is used to classify based on the spatiotemporal frequency multidimensional features to obtain the identification results of dementia-related diseases.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects.

[0018] 1) This invention achieves systematic removal of various artifacts such as electrooculography and electromyography by constructing a standardized 6-stage EEG preprocessing process, which greatly improves signal quality; at the same time, it adopts a spatiotemporal frequency multidimensional feature fusion strategy to overcome the limitations of single frequency domain features and can more comprehensively capture the EEG pattern features of different dementia types.

[0019] 2) The SFNet hybrid deep learning model designed in this invention fully utilizes the complementary advantages of convolutional neural networks in local feature extraction and multilayer perceptrons in global nonlinear modeling, achieving higher accuracy in multi-class classification performance.

[0020] 3) This invention achieves significant results in dementia classification by combining standardized EEG signal preprocessing with spatiotemporal-frequency multidimensional feature fusion technology and an innovative SFNet hybrid deep learning model. This method can achieve high-precision differentiation between Alzheimer's disease, frontotemporal dementia, and healthy controls, with an F1 score exceeding 93%, effectively solving the problems of insufficient feature extraction and limited classification accuracy in traditional methods. These technological advancements provide reliable technical support for the early screening and identification of dementia, helping to promote the transformation of dementia identification from reliance on subjective experience to objective quantitative analysis, and have significant clinical application value. Attached Figure Description

[0021] Figure 1 This is a flowchart of a dementia disease identification method based on spatiotemporal frequency multidimensional feature fusion.

[0022] Figure 2 This invention uses the set_montage() method to locate electrodes in EEG signals, and prints an electrode location map with a total of 19 electrode leads.

[0023] Figure 3 The image shows the effect of artifact removal using independent component analysis in this invention. It can be seen that the electrooculography artifacts are significantly removed.

[0024] Figure 4The graph shows the loss and accuracy changes of the SFNet model designed in this invention on the training and validation sets during the training process.

[0025] Figure 5 This is a visualization of the three-class confusion matrix results of this invention on the test set.

[0026] Figure 6 This invention presents the three-category ROC curve and corresponding AUC value for dementia.

[0027] Figure 7 This is a block diagram of a dementia-related disease identification system based on spatiotemporal frequency multidimensional feature fusion.

[0028] Figure 8 A block diagram of a computer device. Detailed Implementation

[0029] The present invention will now be described in further detail with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0030] This embodiment aims to achieve high-precision, automated identification of individuals with Alzheimer's disease (AD), frontotemporal dementia (FTD), and healthy controls (CN). For example... Figure 1 As shown, the dementia disease identification method based on spatiotemporal frequency multidimensional feature fusion of the present invention includes the following steps.

[0031] Step S1: Perform data preprocessing on the target user's EEG signals.

[0032] First, this invention uses an EEG device to collect the user's electroencephalogram (EEG) signals in a resting state, and transmits the collected data in a standard format (such as .set or .mat files) to the system in real time for recording and analysis.

[0033] Next, the present invention performs data preprocessing on the acquired EEG signals, which can efficiently remove artifacts such as electrooculogram (EOG) and electrocardiogram (ECG) data from the original EEG signals, thereby improving data quality. The data preprocessing includes six stages: format conversion, electrode localization, interpolation bad derivatives, rereference, filtering and noise reduction, and artifact removal.

[0034] (1) Format conversion.

[0035] The EEG dataset used in this embodiment includes 88 subjects, including 36 Alzheimer's patients, 23 frontotemporal dementia patients, and 29 healthy subjects. The data for each subject was segmented using methods of the `mne.io.RawArray` class to construct a `Raw` object. Here, a 10-second sliding window was used to obtain the dataset for each subject, with the shape being (number of samples, number of channels, duration).

[0036] (2) Electrode positioning.

[0037] When using the MNE library for EEG signal visualization, to accurately map electrode-recorded data to corresponding brain regions, it's typically necessary to import location information matching the electrodes. The implementation references the "10-20" international standard lead system. First, the acquired electrode location information is modified and missing electrode locations are supplemented, totaling 19 channels. Next, the electrode channel names and corresponding 3D coordinates are stored as a dictionary. Due to some discrepancies, the three key reference points—nasion (nasal root), lpa (left preauricular point), and rpa (right preauricular point)—are adjusted and encapsulated into a custom montage_1020 system. Finally, the montage_1020 system and the set_montage() method are used to locate the electrodes in the raw object, and the electrode location map is printed as shown below. Figure 2 As shown.

[0038] (3) Interpolation bad derivative.

[0039] When collecting EEG data in the laboratory, some electrodes on the scalp may malfunction due to equipment failure, poor contact, or other factors, leading to data corruption or loss, known as "bad leads." In such cases, a common approach is to use data from normal electrodes and interpolation techniques to estimate and repair these damaged data, thereby ensuring data continuity and reliability. First, the EEG data image is output in an interactive window, allowing the user to identify and mark the bad leads based on the EEG waveform. Next, the program prints the names of the selected bad leads, which are then confirmed by the user. After confirmation, the system selects nearby channels based on electrode location for interpolation of the bad leads. Finally, the repaired data is output for confirmation.

[0040] (4) Re-reference.

[0041] In EEG data analysis, common modal signals introduced by the reference electrode are typically removed, i.e., dereference processing is performed using averaging or other suitable reference methods. In this invention, EEG data rereference employs averaging, creating a new virtual reference electrode based on the average value of all electrodes. If erroneous EEG channels existed in the previous stage, they are automatically excluded.

[0042] (5) Filtering and noise reduction.

[0043] EEG signal acquisition is highly demanding in terms of environmental requirements. In actual acquisition processes, it is always difficult to completely avoid the influence of various interference factors, such as 50Hz AC noise, human electromyography (EMG) artifacts, and eye movement artifacts. To reduce the impact of these interferences on EEG signal quality, this invention employs a bandpass filter based on Finite Impulse Response (FIR) to achieve this goal.

[0044] The output formula of an Nth-order FIR filter As follows, among which The correlation coefficient of the filter. Represented as Delayed One cycle.

[0045] It is worth noting that the phase frequency characteristics of FIR filters exhibit linearity, which means that the phase information of the signal can be effectively preserved after the signal is processed by the filter. This is especially important for processing EEG signals that require precise phase information.

[0046] Considering that the main emotional information in induced EEG signals is concentrated in the lower frequency band, a bandpass filter range of 0.5Hz to 45Hz was set to capture and retain EEG data information in the five key frequency bands: δ, θ, α, β, and γ. Furthermore, to minimize power frequency interference from 50Hz AC power, this invention also employs a notch filter for additional filtering to ensure the purity and reliability of the EEG data.

[0047] (6) Artifact removal.

[0048] A key challenge in EEG signal research is that its microvolt-level amplitude is easily interfered with by other physiological signals. These interfering signals, such as electrooculography (EOG), electrocardiography (ECG), and electromyography (EMG), are called artifacts in EEG recordings. Their presence can seriously affect the interpretation of EEG signals, so special attention needs to be paid to removing these artifacts when processing EEG data.

[0049] Blind Source Separation (BSS) is used to separate individual source signals from mixed signals and can effectively remove artifact components from the source signals. In BSS, this invention employs Independent Components Analysis (ICA). The ICA algorithm, based on the non-Gaussianity and independent distribution of the source signals, can quickly separate artifact components from multi-channel data.

[0050] Based on the fundamental assumptions of the ICA algorithm, the brain electrical signal is assumed to be: , Where n is the number of sampling points and k is the number of leads. These are the source signals corresponding to each electrode. The acquired signals are: , These represent the signals actually acquired by each electrode. The core of the ICA algorithm lies in calculating a confusion matrix. This matrix describes the linear relationship between the source signal and the observed signal, making and satisfy: The goal of ICA is to find a separation matrix W such that the observed signal, after being transformed by W, approximates the original independent source signal as closely as possible. There is an inverse relationship between the confusion matrix A and the separation matrix W. .

[0051] Substitute the separation matrix W obtained above into This allows us to obtain the best estimate of the input vector. That is, independent components, the signal waveforms before and after artifact removal, for example... Figure 3 It is evident that the oculomotor interference was significantly reduced.

[0052] Step S2: Extract the spatiotemporal frequency multidimensional features of the preprocessed EEG signal.

[0053] Given the non-stationary, nonlinear, and time-varying characteristics of EEG signals, features extracted solely from the time or frequency domain often fail to fully capture their complex dynamic properties. Therefore, this invention proposes combining time and frequency domain information and considering the interactions between brain regions to achieve the fusion extraction of spatiotemporal-frequency multidimensional features.

[0054] (1) Brain region feature extraction.

[0055] To achieve refined analysis of EEG signals, this invention performs a first-step channel decomposition on the original 19-lead EEG signals based on the brain anatomical region location information corresponding to each lead in the international standard 10–20 system. Specifically, the 19 leads are divided into seven functional subsets according to their respective brain regions, including the prefrontal cortex, frontal cortex, central cortex, left temporal cortex, right temporal cortex, parietal cortex, and occipital cortex, thereby constructing a multi-channel dataset after brain region segmentation. The specific correspondences between leads and brain regions are shown in Table 1. This partitioning method not only effectively reduces the computational complexity of subsequent model processing, but also reduces the adverse effects of individual channel data anomalies or noise interference on the overall analysis results through regional isolation.

[0056] Table 1. 19-Lead Brain Region Division Based on Functional Anatomy (2) Frequency domain feature extraction based on brain region division.

[0057] Frequency domain analysis is a method for observing signal changes from a frequency perspective. In the processing of electroencephalogram (EEG) signals, this invention employs a five-band segmentation method, subdividing the EEG signal into five frequency bands: delta waves, theta waves, alpha waves, beta waves, and gamma waves, completing the second step of decomposition. Each frequency band represents a different state of EEG activity and frequency range, as shown in Table 2. Table 2 Five-band division standard 1) Spectral energy.

[0058] First, based on the five-band segmentation method described above, the EEG signal is converted into spectral information; then, according to Parseval's theorem, the time-domain energy and frequency-domain energy of a signal are equal, let an energy signal be... The spectral density of this energy signal is ,have: Therefore we will The energy spectral density, also known as the energy signal, is expressed at a frequency of [missing information]. Width is The signal energy within the frequency band, that is, the signal energy within a unit frequency band, is used to extract the energy features in the EEG signal.

[0059] 2) Differential Entropy (DE).

[0060] Differential entropy is a generalization of Shannon information entropy to continuous variables, used to measure the complexity or uncertainty of the distribution of continuous random variables. Its expression is usually defined based on the integral of the probability density function (PDF), but in practical applications, it is often approximated using sampled data. Its expression is as follows: in, For the calculation of differential entropy, For random variables, for The probability density function is given. First, the EEG signal is divided into five frequency bands: delta, theta, alpha, beta, and gamma. Since the EEG signals in each frequency band statistically approximately follow a Gaussian distribution, the differential entropy of the EEG signals in each frequency band can be simplified and obtained. Equation

[0061] (3) Temporal feature extraction based on brain region division.

[0062] Time-domain analysis typically uses various statistical features to describe the waveform characteristics of EEG signals over time. Statistical features were the initial method for characterizing EEG data; previous researchers considered EEG signals a type of signal analysis and thus introduced commonly used statistical indicators for description. This invention extracts multiple statistical features from EEG signals, including peak-to-peak value, variance, and zero-crossing rate (ZCR) variation. These features together constitute a multi-dimensional set of statistical indicators describing the waveform characteristics of EEG signals.

[0063] Zero-crossing rate is the number of times an EEG signal crosses zero (from positive to negative or from negative to positive) per unit time, i.e., the frequency at which the signal flips along the zero value.

[0064] Where S is the signal and N is the length of the signal. It is a symbolic function: (4) Extraction of cross-brain region interaction features.

[0065] To delve deeper into the functional connectivity and collaborative mechanisms between brain regions, this invention designs a cross-brain region interaction feature extraction module. This module overcomes the limitations of traditional single-channel or isolated brain region analysis by quantifying the dynamic interaction relationships between different brain regions, capturing information integration and functional lateralization features at the brain network level, and significantly enhancing the spatial discriminative ability of the features.

[0066] 1) Hemispherical functional asymmetry characteristics.

[0067] The left and right hemispheres of the brain have specific specializations in cognitive function, and this functional lateralization can be quantified by the asymmetry of electroencephalogram (EEG) signals. This invention calculates the following two key asymmetry indicators: Power Asymmetry Index (PAI): Reflects the difference in energy activity between the left and right hemispheres in a specific frequency band. The calculation formula is as follows: in, and These represent the average power of corresponding brain regions in the left and right hemispheres (e.g., F3 in the left frontal lobe and F4 in the right frontal lobe). This is a very small constant to prevent division by zero errors. Positive values ​​indicate left-hemispheric dominance, and negative values ​​indicate right-hemispheric dominance.

[0068] Complexity Asymmetry Index (CAI): Based on differential entropy calculation, it characterizes the difference in signal complexity between the left and right hemispheres. Its calculation method is similar to PAI, but power is replaced by differential entropy value.

[0069] 2) Functional gradient characteristics of the anterior and posterior brain regions.

[0070] The functional gradient of the brain from front to back (from higher cognition in the prefrontal cortex to sensory processing in the occipital lobe) is a crucial principle in brain functional organization. This invention quantifies this gradient by calculating the ratio of features between the anterior and posterior brain regions: Frontal-Occipital Power Ratio (FOPR): The formula is as follows: in, The average power of the prefrontal region (including the prefrontal region and the frontal region). This represents the average power in the occipital region. This ratio reflects the energy distribution between cognitive control and sensory processing.

[0071] Frontal-Occipital Complexity Ratio (FOCR): Using differential entropy instead of power, and calculated in the same way as above, it is used to assess the difference in information processing complexity between the frontal and occipital brain regions.

[0072] Finally, by integrating temporal, frequency, and cross-brain region interaction features, this invention constructs a comprehensive spatiotemporal-frequency multidimensional feature vector. Specifically, based on brain region segmentation (raw_pick), for each frequency band of each brain region, the system extracts multiple sets of basic features, including zero-crossing rate, peak-to-peak value, differential entropy, variance, and energy, and further incorporates brain region asymmetry and functional gradient interaction features. These features are systematically fused according to a fixed hierarchical structure to form a high-dimensional feature vector with a total dimension of 195 (i.e., 5 frequency bands × 7 brain regions × 5 basic features + 5 frequency bands × 4 cross-brain region interaction features), serving as a comprehensive representation of each EEG sample. This feature matrix deeply integrates the temporal dynamics, spatial distribution patterns, and frequency rhythm characteristics of EEG signals, thereby significantly enhancing the discriminative ability and generalization performance of subsequent classification models.

[0073] Step S3: Classify based on the spatiotemporal frequency multidimensional features to obtain the identification results of dementia-related diseases.

[0074] To effectively learn high-dimensional nonlinear EEG features, this invention addresses the limitations of traditional CNN-MLP hybrid models, such as limited feature representation capabilities and insufficient extraction of key information. It innovatively designs a CNN-MLP hybrid neural network model (SFNet) that integrates a lightweight attention mechanism. While maintaining a lightweight architecture, this model significantly improves feature selection and representation capabilities by introducing a dual enhancement mechanism of channel attention and feature attention, while reducing computational complexity. This makes it particularly suitable for scenarios such as dementia disease recognition.

[0075] (1) Model structure design.

[0076] The proposed SFNet employs a lightweight dual-branch parallel architecture. The CNN branch is dedicated to extracting local spatiotemporal features of EEG signals, using a 5-3-3 progressive convolutional kernel design: the first convolutional layer uses a 5×1 kernel with 32 filters to capture long-term EEG signal dependencies; the second convolutional layer uses a 3×1 kernel with 64 filters to extract medium-granularity features; and the third convolutional layer uses a 3×1 kernel with 128 filters for fine-granular feature refinement. Each convolutional block is sequentially connected to a batch normalization layer, a ReLU activation function layer, a max-pooling layer (pooling size 2), and a Dropout layer, with the Dropout rate progressively increasing. Notably, a lightweight channel attention module is innovatively introduced after the first convolutional block. This module uses global average pooling and shared MLP to generate channel weights, dynamically weighting important temporal channels and ultimately outputting feature vectors through a global average pooling layer.

[0077] The MLP branch is responsible for learning the global statistical features of EEG signals and adopts a feature attention enhancement design: First, through the feature attention module, a feature weight matrix is ​​generated based on fully connected layers and Sigmoid activation to weight the input features according to their importance; then, through two fully connected layers (with 128 and 64 neurons respectively), each layer is followed by a batch normalization layer, a ReLU activation function, and a Dropout layer (with dropout rates of 0.3 and 0.2 respectively), forming a refined feature compression process.

[0078] The feature fusion module deeply fuses the output features of the two branches at the concatenation layer, forming a comprehensive representation that includes local spatiotemporal features and global statistical features. The fused features are further refined through two fully connected layers with 64 and 32 neurons respectively, and finally complete the three-class classification task (AD, FTD, and CN) through the softmax output layer. This dual-branch fusion architecture fully leverages the complementary advantages of CNN in local feature extraction and MLP in global feature learning, and achieves effective selection and enhancement of key features in a lightweight architecture through a dual attention mechanism.

[0079] (2) Regularization optimization strategy.

[0080] This invention employs a multi-level comprehensive regularization strategy to ensure the model's generalization ability. For structural regularization, L2 weight regularization is introduced into all convolutional and fully connected layers, with a regularization coefficient set to 0.001, effectively constraining the weight value range to prevent overfitting. A progressive Dropout strategy is adopted, with the Dropout rate gradually increasing from 0.2 to 0.4 as the network depth increases, preserving more feature information in shallow layers and enhancing the random deactivation effect in deeper layers.

[0081] Regarding regularization during training, an early stopping monitoring mechanism is implemented. Training is automatically terminated and the optimal weights are restored when the validation loss fails to improve for 15 consecutive rounds. A dynamic learning rate decay strategy is employed, combining exponential decay scheduling and platform-based dynamic adjustment: the initial learning rate is set to 0.001, decaying by a factor of 0.9 every 1000 steps; the learning rate is automatically halved when the validation loss fails to decrease for 8 consecutive rounds, finely adjusting the model parameters. To address the class imbalance problem, a class weight balancing mechanism is introduced, dynamically adjusting the loss function weights based on the number of samples in each class.

[0082] Data-level regularization is achieved through input data standardization, eliminating scale differences across feature dimensions. Global average pooling replaces traditional flattening operations, significantly reducing the number of model parameters. The comprehensive application of batch normalization layers stabilizes the input distribution of each layer, accelerating training convergence while also providing a slight regularization effect.

[0083] (3) Training configuration.

[0084] This invention employs an optimized training configuration. The optimizer chosen is the Adam algorithm, with an initial learning rate of 0.001, momentum parameters β1 of 0.9, β2 of 0.999, and a numerical stability constant ε of 1e-07. The batch size is set to 32, striking a balance between training efficiency and memory consumption. The maximum number of training epochs is set to 150, coupled with an early stopping mechanism to achieve dynamic control of the training process.

[0085] The training employs a five-fold cross-validation strategy to evaluate model stability and ensure reliable results. The loss function chosen is classification cross-entropy, combined with class weights to address the issue of imbalanced data distribution. The callback mechanism includes early stopping callbacks and learning rate scheduling callbacks. Early stopping callbacks monitor changes in validation loss, while learning rate scheduling callbacks dynamically adjust the learning rate based on the plateauing period of the validation loss. Specific model parameter settings are shown in Table 3.

[0086] Table 3 SFNet Model Parameter Settings To comprehensively evaluate model performance, this invention employs a stratified five-fold cross-validation method to robustly validate and evaluate the classification model. The specific steps are as follows: First, the labeled dataset is stratified and randomly divided according to class proportions, generating five equal-sized subsets (folds). One subset is selected sequentially as the test set, and the remaining four are used as training sets to construct and train five independent classification models. Each model performs predictions on its corresponding test set, and its classification performance metrics are recorded. Finally, the results of the five validation runs are summarized, and the macro-average precision, recall, and F1 score are calculated as the core evaluation metrics for the three-class classification task.

[0087] The model demonstrated excellent classification performance on the test set, with specific evaluation metrics shown in Table 4. As can be seen from the table, the model achieved high-precision classification for all three EEG signal categories. The Alzheimer's disease (AD) category achieved a precision of 0.92, a recall of 0.93, and an F1 score of 0.93, indicating that the model can accurately identify the EEG characteristics of AD patients. The frontotemporal dementia (FTD) category showed the best classification performance, with precision, recall, and F1 score of 0.94, 0.95, and 0.94, respectively, demonstrating the model's specific ability to identify FTD features. The classification metrics for the normal control group (CN) were 0.92, 0.90, and 0.91, respectively, proving the model's effectiveness in distinguishing between normal and abnormal EEG patterns.

[0088] Table 4. Classification results of dementia under this invention Figure 4 The diagram illustrates the loss and accuracy curves of the model on the training and validation sets during training. The loss curve shows a rapid decrease in loss value in the early stages of training, followed by a gradual convergence to a stable plateau, indicating that the model's learning process is efficient and stable. The loss curves on the training and validation sets maintain a high degree of consistency, with minimal difference between them, demonstrating the model's good generalization ability and the absence of significant overfitting. The accuracy curves also show ideal training results, with both training and validation accuracy steadily improving and eventually converging to a high level. Validation accuracy consistently exceeds 90%, fully demonstrating the rationality of the model architecture design and the effectiveness of the training strategy.

[0089] Figure 5 The confusion matrix visualization results of the model on the test set are presented. Only a small number of samples in the AD category were misclassified as CN, while the FTD category maintained the highest classification purity. Misclassification of the CN category mainly occurred between it and the AD category, which is consistent with the actual situation in clinical practice where the two categories of EEG features have a certain degree of similarity. The overall accuracy of the confusion matrix reached 92.5%, further validating the classification reliability of the model.

[0090] Figure 6 The multi-class ROC curves and corresponding AUC values ​​of the model are shown. The ROC curves for each class exhibit an ideal shape close to the upper left corner, indicating that the model has excellent discriminative ability across all classes. The AUC value for the AD class is 0.98, for the FTD class it reaches 0.99, and for the CN class it is 0.98. These high AUC values ​​demonstrate that the model maintains stable performance under different classification thresholds, possessing high clinical practical value.

[0091] Based on comprehensive evaluation metrics and visualization analysis results, the SFNet hybrid deep learning classification model proposed in this invention demonstrates outstanding performance in EEG signal classification tasks. The model not only achieves high scores on each individual metric, but more importantly, it maintains high accuracy while exhibiting good generalization ability and clinical applicability. These results fully validate the practical application value of this invention in the early identification and classification of dementia, providing a reliable technical foundation for brain-computer interfaces and medical auxiliary identification systems.

[0092] In summary, this invention constructs a standardized EEG signal preprocessing workflow, comprising six sequentially executed standardized steps: format conversion, electrode localization, interpolation bad lead generation, rereference, filtering and noise reduction, and artifact removal. This method effectively improves EEG signal quality, lays the foundation for subsequent feature extraction and classification, and significantly enhances recognition accuracy.

[0093] This invention presents a technical solution for simultaneously extracting and fusing EEG features across three dimensions: time, space, and frequency. This technology comprehensively captures dementia-related EEG pattern changes through multi-dimensional feature complementarity, overcoming the limitations of single-dimensional features and enhancing the richness and discriminative power of feature representation.

[0094] This invention proposes a multi-dimensional collaborative technique for extracting and fusing EEG features across the temporal, spatial, and frequency domains. This technique overcomes the limitations of traditional single-dimensional analysis by extracting and fusing temporal statistical features, spatial distribution features, and frequency rhythm features to construct a highly complementary comprehensive feature vector. This technique can comprehensively and deeply capture the complex multi-dimensional EEG pattern variations induced by dementia, greatly enhancing the feature representation ability and classification discriminative power.

[0095] This invention designs a spatiotemporal-frequency fusion convolutional multilayer perceptron (SFNet) with a dual-branch architecture. This model innovatively employs a parallel dual-stream architecture of CNN and MLP to efficiently extract local spatiotemporal patterns and global abstract features of signals, respectively. It also achieves adaptive enhancement of key information channels and feature dimensions through a dual mechanism of lightweight channel attention and feature attention. The model uses a 5-3-3 progressive convolutional kernel design and combines dynamic learning rate scheduling, class-adaptive weighting, and multi-level regularization strategies. While maintaining the model's lightweight and efficient nature, it achieves a balance between high accuracy and strong generalization ability, providing a reliable analytical model for assistive identification of dementia.

[0096] Therefore, this invention solves the problems in the prior art, such as how to construct a standardized preprocessing process to reduce manual intervention, how to systematically integrate the multidimensional features of EEG signals to improve the model's ability to distinguish dementia, and how to design a neural network structure that combines local feature extraction and global nonlinear modeling capabilities to improve the model's accuracy in identifying dementia. Ultimately, it realizes end-to-end intelligent analysis from raw EEG data to dementia identification results output, and constructs a fully automated solution from data preprocessing to intelligent classification, which significantly improves the accuracy and practicality of early dementia identification.

[0097] Based on the same concept, this invention also discloses a dementia-like disease identification system based on spatiotemporal frequency multidimensional feature fusion, such as... Figure 7 As shown, the system includes: The preprocessing module is used to preprocess the EEG signals of the target user; The feature extraction module is used to extract the spatiotemporal frequency multidimensional features of the preprocessed EEG signal; The disease identification module is used to classify based on the spatiotemporal frequency multidimensional features to obtain the identification results of dementia-related diseases.

[0098] Based on the same concept, this invention also discloses a computer device, which may be a terminal, a laptop computer, a desktop computer, a server, a computer cluster, or other types of computer devices. For example... Figure 8 As shown, the computer device may include at least one processor and memory. The processor can execute instructions stored in the memory. The processor is communicatively connected to the memory via a data bus. In addition to the memory, the processor can also be communicatively connected to input devices, output devices, and communication devices via the data bus.

[0099] The processor can be any conventional processor. Processors may include central processing units (CPUs), graphics processing units (GPUs), field-programmable gate arrays (FPGAs), systems on chips (SoCs), application-specific integrated circuits (ASICs), or combinations thereof.

[0100] Memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0101] In this embodiment of the invention, an executable instruction is stored in a memory. The processor can read the executable instruction from the memory and execute the instruction to implement all or part of the steps of the method of the invention.

[0102] Based on the same concept, the present invention also discloses a computer-readable storage medium including a computer program product or storing the computer program product. The computer product includes computer program instructions that can be executed by a processor to perform all or part of the steps described in the exemplary embodiments above.

[0103] Computer program products can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. These programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages, and scripting languages ​​(e.g., Python). The program code can be executed entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0104] Computer-readable storage media can take the form of any combination of one or more readable media. A readable medium can be a readable signal medium or a readable storage medium. A readable storage medium can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media include: static random access memory (SRAM) having one or more electrically connected wires; electrically erasable programmable read-only memory (EEPROM); erasable programmable read-only memory (EPROM); programmable read-only memory (PROM); read-only memory (ROM); magnetic storage; flash memory; magnetic disk or optical disk; or any suitable combination thereof.

[0105] Although specific embodiments of the invention have been disclosed for illustrative purposes to aid in understanding and implementing the invention, those skilled in the art will understand that various substitutions, variations, and modifications are possible without departing from the spirit and scope of the invention and the appended claims. Therefore, the invention should not be limited to the content disclosed in the preferred embodiments, and the scope of protection claimed by the invention is defined by the claims.

Claims

1. A method for identifying dementia-like diseases based on spatiotemporal-frequency multidimensional feature fusion, characterized in that, The method includes: Preprocessing the target user's EEG signals; Extracting spatiotemporal frequency multidimensional features from preprocessed EEG signals; Based on this spatiotemporal frequency multidimensional feature, classification was performed to obtain the identification results of dementia-related diseases; This includes preprocessing the target user's EEG signals, including: Use a sliding window to segment the EEG signal and construct a Raw object based on the segmentation results; Electrode location is performed on the Raw object by combining electrode location information to obtain a Raw object with electrode location information; among them, for the Raw object with electrode location information, bad conductors are marked by interactive visualization and the bad conductors are repaired by interpolation using surrounding normal electrode data. The rereferenced Raw object is obtained based on the average electrode value of all Raw objects with electrode position information. Filter the rereferenced Raw object; Artifact removal is performed on the filtered Raw object to obtain the preprocessed EEG signal; Among them, the spatiotemporal frequency multidimensional features of the preprocessed EEG signals are extracted, including: Based on the brain regions where the lead channels are located, the EEG signals are divided into several brain region datasets; wherein, the brain region datasets include the prefrontal region dataset, frontal region dataset, central region dataset, left temporal region dataset, right temporal region dataset, parietal region dataset, and occipital region dataset; For each brain region dataset, EEG data in the delta, theta, alpha, beta, and gamma bands are extracted respectively. The temporal and frequency domain features of the EEG data for each frequency band are extracted separately; wherein, the frequency domain features include spectral energy and differential entropy, and the temporal features include zero-crossing rate, peak-to-peak value, and variance; Based on the brain location of the lead channels, the EEG signals are divided into left hemisphere datasets and right hemisphere datasets. Based on the EEG signals in the left and right hemisphere datasets, hemisphere functional asymmetry features are extracted. The hemisphere functional asymmetry features include power asymmetry index and complexity asymmetry index. Based on the brain location of the lead channels, the EEG signals are divided into anterior brain region datasets and posterior brain region datasets. Based on the EEG signals in the anterior and posterior brain region datasets, functional gradient features of the anterior and posterior brain regions are extracted. The functional gradient features of the anterior and posterior brain regions include the fronto-occipital power gradient ratio and the fronto-occipital complexity gradient ratio. Based on the temporal and frequency domain features of EEG data in each frequency band of the brain region dataset, as well as hemispheric functional asymmetry features and anterior and posterior brain region functional gradient features, the spatiotemporal-frequency multidimensional features of the preprocessed EEG signal are obtained. Among them, classification based on this spatiotemporal frequency multidimensional feature yields dementia disease identification results, including: A hybrid classification model is constructed, comprising a CNN branch, an MLP branch, a concatenation layer, and an output layer. The CNN branch is used to extract local spatiotemporal features of EEG signals, the MLP branch is used to extract global abstract features of EEG signals, the concatenation layer is used to fuse local spatiotemporal features and global abstract features, and the output layer is used to classify the fused results. By inputting spatiotemporal frequency multidimensional features into a hybrid classification model, the identification results of dementia-related diseases are obtained.

2. The method according to claim 1, characterized in that, Filtering is performed on the rereferenced Raw object, including: Set the bandpass filter range; Capture within the bandpass filter range frequency band frequency band frequency band frequency band and The EEG data in the frequency band was filtered using a notch filter.

3. The method according to claim 1, characterized in that, Artifact removal is performed on the filtered Raw object to obtain the preprocessed EEG signal, including: Perform independent component analysis on the filtered Raw object to obtain the separation matrix; The separation matrix and the filtered Raw object are multiplied together to obtain the preprocessed EEG signal.

4. The method according to claim 1, characterized in that, The CNN branch includes several convolutional blocks and a global average pooling layer, and the convolutional blocks form a pyramid structure with increasing feature depth. Each convolutional block is composed of a convolutional layer, a batch normalization layer, an activation function layer, a pooling layer, and a Dropout layer connected in sequence. A lightweight channel attention module is introduced after the first convolutional block. This lightweight channel attention module generates channel weights based on global average pooling and shared MLP.

5. The method according to claim 1, characterized in that, The MLP branch includes a feature attention module and several fully connected layers. The feature attention module is used to generate a feature weight matrix based on the fully connected layers and Sigmoid activation to weight the spatiotemporal frequency multidimensional features based on their importance. Several fully connected layers form a refined structure for feature compression, and each fully connected layer is followed by a batch normalization layer and a Dropout layer.

6. The method according to claim 1, characterized in that, The hybrid classification model is trained based on a multi-level comprehensive regularization strategy; wherein, the multi-level comprehensive regularization strategy includes structural regularization, training process regularization, and data-level regularization.

7. A dementia-like disease identification system based on spatiotemporal frequency multidimensional feature fusion, characterized in that, The system includes: The preprocessing module is used to preprocess the EEG signals of the target user; The feature extraction module is used to extract the spatiotemporal frequency multidimensional features of the preprocessed EEG signal; The disease identification module is used to classify based on the spatiotemporal frequency multidimensional features to obtain the identification results of dementia-related diseases; This includes preprocessing the target user's EEG signals, including: Use a sliding window to segment the EEG signal and construct a Raw object based on the segmentation results; Electrode location is performed on the Raw object by combining electrode location information to obtain a Raw object with electrode location information; among them, for the Raw object with electrode location information, bad conductors are marked by interactive visualization and the bad conductors are repaired by interpolation using surrounding normal electrode data. The rereferenced Raw object is obtained based on the average electrode value of all Raw objects with electrode position information. Filter the rereferenced Raw object; Artifact removal is performed on the filtered Raw object to obtain the preprocessed EEG signal; Among them, the spatiotemporal frequency multidimensional features of the preprocessed EEG signals are extracted, including: Based on the brain regions where the lead channels are located, the EEG signals are divided into several brain region datasets; wherein, the brain region datasets include the prefrontal region dataset, frontal region dataset, central region dataset, left temporal region dataset, right temporal region dataset, parietal region dataset, and occipital region dataset; For each brain region dataset, EEG data in the delta, theta, alpha, beta, and gamma bands are extracted respectively. The temporal and frequency domain features of the EEG data for each frequency band are extracted separately; wherein, the frequency domain features include spectral energy and differential entropy, and the temporal features include zero-crossing rate, peak-to-peak value, and variance; Based on the brain location of the lead channels, the EEG signals are divided into left hemisphere datasets and right hemisphere datasets. Based on the EEG signals in the left and right hemisphere datasets, hemisphere functional asymmetry features are extracted. The hemisphere functional asymmetry features include power asymmetry index and complexity asymmetry index. Based on the brain location of the lead channels, the EEG signals are divided into anterior brain region datasets and posterior brain region datasets. Based on the EEG signals in the anterior and posterior brain region datasets, functional gradient features of the anterior and posterior brain regions are extracted. The functional gradient features of the anterior and posterior brain regions include the fronto-occipital power gradient ratio and the fronto-occipital complexity gradient ratio. Based on the temporal and frequency domain features of EEG data in each frequency band of the brain region dataset, as well as hemispheric functional asymmetry features and anterior and posterior brain region functional gradient features, the spatiotemporal-frequency multidimensional features of the preprocessed EEG signal are obtained. Among them, classification based on this spatiotemporal frequency multidimensional feature yields dementia disease identification results, including: A hybrid classification model is constructed, comprising a CNN branch, an MLP branch, a concatenation layer, and an output layer. The CNN branch is used to extract local spatiotemporal features of EEG signals, the MLP branch is used to extract global abstract features of EEG signals, the concatenation layer is used to fuse local spatiotemporal features and global abstract features, and the output layer is used to classify the fused results. By inputting spatiotemporal frequency multidimensional features into a hybrid classification model, the identification results of dementia-related diseases are obtained.

Images