Method and device for early screening of neurodegenerative diseases

Abstract

This invention discloses an early screening method and device for neurodegenerative diseases, relating to the fields of medical artificial intelligence and tabular data deep learning. This method replaces manual heuristic grouping with a data-driven grouping module, incorporating the feature segmentation process into end-to-end training. This allows feature grouping to adaptively update with the task objective, reducing reliance on domain-specific manual rules. By using redundant decoupling loss to simultaneously constrain mutual information and correlation coefficients, it helps reduce intra-group redundancy and information overlap, making each feature group more compact and possessing relatively independent discriminative capabilities. Parallel modeling through full feature branching, intra-group branching, and inter-group branching captures global dependencies, local defects, and cross-group association rules, respectively. Furthermore, intra-group and inter-group branches are jointly constrained by the data-driven grouping results, balancing the commonalities and diversity of disease signals and avoiding information omissions caused by single-path interactions.

Description

Technical Field

[0001] This invention relates to the fields of medical artificial intelligence and deep learning of tabular data, and in particular to a method and apparatus for early screening of neurodegenerative diseases. Background Technology

[0002] Neurodegenerative diseases include Alzheimer's disease, Parkinson's disease, and amyotrophic lateral sclerosis (ALS). Early screening for these diseases is crucial for expanding the treatment window, improving long-term patient outcomes, and reducing the burden on the healthcare system. While neuroimaging techniques such as magnetic resonance imaging (MRI) and positron emission tomography (PET) have high diagnostic value, they face challenges in community, home, and large-scale continuous screening scenarios, including high equipment costs, limited accessibility, and restricted testing frequency.

[0003] With the development of natural human-computer interaction technology, behavioral or physiological signals such as writing, speech, and gait can be collected at a lower cost and further converted into structured tabular features. For example, writing tests can generate handwriting kinematic indicators, speech tests can generate acoustic statistical indicators, and gait tests can generate indicators such as step height, gait frequency, swing phase, and variability. Building automated screening models based on this tabular data is beneficial for developing non-invasive, low-cost, and scalable early screening programs for neurodegenerative diseases.

[0004] However, screening data for neurodegenerative diseases typically present the following technical challenges. First, the features are heterogeneous, originating from writing, speech, gait, posture estimation, or other sensory channels, lacking the natural spatial sequence structure found in images or text. Second, features exhibit redundancy and overlap; different descriptors may reflect similar physiological or behavioral changes, and direct involvement in interaction modeling can amplify redundant noise and mask subtle pathological clues. Third, disease-related signals encompass fine-grained defects within local feature groups, cross-feature group association rules, and global dependencies across the entire feature space; a single interaction path cannot simultaneously accommodate all of this information.

[0005] Existing methods for learning tabular data mainly include tree-based ensemble models, kernel methods, deep tabular models, and feature cross-models. Tree models such as Random Forest, XGBoost, CatBoost, and LightGBM can handle irregular distributions and are robust to noise to a certain extent, but their split-driven mechanisms are difficult to optimize end-to-end and are not conducive to fully representing complex high-order feature dependencies. Deep tabular models such as DANets, TabPFN, and T2G-Former can learn dense representations, but they often handle interactions through implicit aggregation or self-attention mechanisms, making it difficult to clearly enumerate and maintain explicit high-order feature combinations that are meaningful for medical screening. Feature cross-models such as DeepFM, FiBiNET, and EulerNet enhance feature interaction capabilities, but they usually use single-branch or double-branch pipelines, or perform interactions within a predefined feature space, failing to fully address the dynamic partitioning and redundancy decoupling issues of heterogeneous feature domains.

[0006] Compared to classic feature interaction models, DeepFM mainly achieves explicit and implicit interactions by combining factorization machines with deep networks, but it does not perform data-driven grouping of heterogeneous features; FiBiNET uses bilinear interactions to enhance feature combinations, but it lacks intra-group and inter-group separation modeling driven by redundancy constraints; EulerNet can model high-order interactions in complex spaces, but it is difficult to isolate redundant disease signals in tabular medical data; multi-branch models such as DLF can enhance explicit and implicit interactions, but they usually still interact in the preset or original feature space, and do not use mutual information and correlation coefficients together for end-to-end grouping constraints. Summary of the Invention

[0007] In view of the above problems, the present invention provides an early screening method and apparatus for neurodegenerative diseases to overcome or at least partially solve the above problems, and solves the technical problems in the prior art such as data dependence on heuristic feature grouping in early screening of neurodegenerative diseases, difficulty in suppressing intra-group redundancy, difficulty in explicitly mining inter-group associations, and difficulty in balancing global dependence and local defects in modeling.

[0008] This invention provides the following solution: An early screening method for neurodegenerative diseases includes: Obtain the tabular feature data generated by the subject's natural human-computer interaction test, and preprocess the tabular feature data to obtain the input feature matrix; Based on the statistics of the input feature matrix in the batch dimension and the learnable mapping parameters, grouping logical values ​​are generated, and differentiable grouping matrices are generated through the Gumbel-Softmax function. A redundancy decoupling loss is constructed based on the differentiable grouping matrix, the feature mutual information matrix, and the feature correlation coefficient matrix. The redundancy decoupling loss is used to constrain information overlap and correlation within the same feature group. Multiple feature groups are obtained based on the differentiable grouping matrix. The multiple feature groups are respectively input into the intra-group branch and the inter-group branch. At the same time, the complete input feature representation output by the preprocessing module is input into the full feature branch to form a three-branch network architecture composed of parallel grouping paths and full feature paths. Implicit interaction modeling is performed on each feature group through the intra-group branch; multiplicative cross-interaction, attention weighting, dimension-preserving projection and residual output are performed on the group representations of different feature groups through the bilinear attention block in the inter-group branch; and global interaction modeling is performed on the complete feature representation through the full feature branch. The outputs of the intra-group branch, the inter-group branch, and the full-feature branch are gated and weighted to obtain a risk score for early screening of neurodegenerative diseases.

[0009] Preferably, the preprocessing of the table feature data includes: Select a preset number of features based on the mutual information between the input features and the training labels; The normalization parameters are calculated on the training set and applied to both the training and test sets.

[0010] Preferably, the table feature data generated by the natural human-computer interaction test includes at least one of the following: handwriting speed, pressure, pause time, trajectory geometric statistics, pitch period entropy, MFCC, formants, jitter, step height, step frequency, swing phase, and gait variability.

[0011] Preferably, the number of feature groups in the differentiable grouping matrix is ​​determined based on the singular value energy threshold of the input feature matrix, and the group representation dimension is determined based on the cumulative explained variance threshold of the principal components.

[0012] Preferably, the step of constructing the redundant decoupling loss specifically includes: Calculate the mutual information matrix and correlation coefficient matrix between each pair of features, configure the correlation coefficient matrix as an absolute correlation coefficient matrix, and set the diagonal elements of the mutual information matrix and the correlation coefficient matrix to zero; For the Let there be _n_ feature groups, and denote their allocation vector in the grouping matrix as _n_i. The intra-group information overlap penalty and intra-group correlation penalty are obtained by performing quadratic weighting using the following formulas.

[0013] In the formula: For balance coefficient, For the total number of features, For mutual information matrix, The correlation coefficient matrix, This is the transpose symbol.

[0014] Preferably: The total training loss is expressed by the following formula:

[0015] In the formula: To classify task losses, The weights are used for redundancy decoupling loss.

[0016] Preferably, the bilinear attention block in the inter-group branch performs the following steps: The interaction terms of the second-order factorization machine are calculated by multiplicative cross-units to obtain the pairwise interaction signals between groups; Group-level attention scores are generated using the SENET structure through attention weight units. And reweight the group representation V to obtain ; Enumerate all that satisfy of Unordered feature pairs, through a learnable bilinear weight matrix ; Calculate the bilinear cross tensor as well as ; Will The signal is copied to the matching dimension and concatenated with the bilinear attention signal to obtain the concatenated features. ; The dimension-preserving unit compresses the spliced ​​features back to the original group representation dimension through one-dimensional convolution, batch normalization, and ReLU function, and then passes them to the next layer through the residual output unit.

[0017] Preferably, the number of stacked layers of the bilinear attention block is set to 2 to form a third-order inter-group feature crossover, balancing expressive power and sample efficiency under the premise of approximately linear growth in computational complexity.

[0018] Preferably, the gated weighted fusion step specifically includes: Dynamic branch weights are generated based on the full feature output, within-group output, and between-group output through a gated weight calculation unit based on the Sigmoid function. When local feature defects dominate, the weight of intra-group branches increases adaptively; when cross-group associations dominate, the weight of inter-group branches increases adaptively. The outputs of the three branches are weighted and summed to obtain the early screening risk score.

[0019] An early screening device for neurodegenerative diseases, used to perform the above-described early screening method for neurodegenerative diseases, the device comprising: The preprocessing unit is used to acquire tabular feature data generated by the subject's natural human-computer interaction test, and to preprocess the tabular feature data to obtain the input feature matrix; A differentiable grouping matrix generation unit is used to generate grouping logical values ​​based on the statistics of the input feature matrix in the batch dimension and the learnable mapping parameters, and to generate a differentiable grouping matrix through the Gumbel-Softmax function. A redundant decoupling loss construction unit is used to construct a redundant decoupling loss based on the differentiable grouping matrix, the feature mutual information matrix, and the feature correlation coefficient matrix. The redundant decoupling loss is used to constrain information overlap and correlation within the same feature group. The three-branch processing unit is used to obtain multiple feature groups based on the differentiable grouping matrix, input the multiple feature groups into the intra-group branch and the inter-group branch respectively, and input the complete input feature representation output by the preprocessing module into the full feature branch to form a three-branch network architecture composed of parallel grouping paths and full feature paths. The three-way feature interaction modeling unit is used to perform implicit interaction modeling on each feature group through the intra-group branch, perform multiplicative cross, attention weighting, dimension-preserving projection and residual output on the group representation of different feature groups through the bilinear attention block in the inter-group branch, and perform global interaction modeling on the complete feature representation through the full feature branch. The output unit is used to perform gated weighted fusion of the outputs of the intra-group branch, the inter-group branch, and the full-feature branch to obtain a risk score for early screening of neurodegenerative diseases.

[0020] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This application provides an early screening method and apparatus for neurodegenerative diseases. The method replaces manual heuristic grouping with a data-driven grouping module, incorporates the feature segmentation process into end-to-end training, and enables feature grouping to be adaptively updated according to task objectives, reducing the reliance on domain-specific manual rules.

[0021] By constraining mutual information and correlation coefficients simultaneously through redundancy decoupling loss, it helps to reduce intra-group redundancy and information overlap, making each feature group more compact and having relatively independent discriminative capabilities.

[0022] By modeling in parallel using full-feature branches, intra-group branches, and inter-group branches, global dependencies, local defects, and cross-group association rules are captured respectively. Furthermore, intra-group branches and inter-group branches are both constrained by data-driven grouping results, which can balance the commonality and diversity of disease signals and avoid information omissions caused by single-path interactions.

[0023] Interpretable explicit inter-group feature crossings are achieved through bilinear attention blocks, and multi-order stacking is supported through dimension preservation and residual connections, making the computational complexity grow approximately linearly rather than exponentially as the inter-group crossing depth increases.

[0024] It is applicable to medical tabular data from different natural human-computer interaction sources such as writing, voice, and gait, and can be used as a computer model component in early screening or risk assessment to improve the robustness of screening in heterogeneous, small-sample, and redundant feature scenarios. Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0026] Figure 1 This is a flowchart of an early screening method for neurodegenerative diseases provided in an embodiment of the present invention; Figure 2 A flowchart of the three-branch feature cross network system provided in this embodiment of the invention; Figure 3 A schematic diagram of a data-driven grouping module provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a bilinear attention block structure provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the dynamic gating fusion module provided in an embodiment of the present invention; Figure 6 This is a diagram of a three-branch feature cross-network architecture provided in an embodiment of the present invention; Figure 7 A schematic diagram of an early screening device for neurodegenerative diseases provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of an early screening device for neurodegenerative diseases provided in an embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0028] See Figure 1 This invention provides an early screening method for neurodegenerative diseases, such as... Figure 1 As shown, the method may include: S101: Obtain the table feature data generated by the subject's natural human-computer interaction test, and preprocess the table feature data to obtain the input feature matrix; in specific implementation, the table feature data generated by the natural human-computer interaction test may include at least one of the following: handwriting speed, pressure, pause time, trajectory geometric statistics, pitch period entropy, MFCC, formants, jitter, step height, step frequency, swing phase, and gait variability.

[0029] Furthermore, the preprocessing of the table feature data includes: Select a preset number of features based on the mutual information between the input features and the training labels; The normalization parameters are calculated on the training set and applied to both the training and test sets.

[0030] S102: Generate grouped logical values ​​based on the statistics and learnable mapping parameters of the input feature matrix in the batch dimension, and generate a differentiable grouped matrix through the Gumbel-Softmax function; in specific implementation, the embodiments of the present invention can provide that the number of feature groups in the differentiable grouped matrix is ​​determined according to the singular value energy threshold of the input feature matrix, and the group representation dimension is determined according to the principal component cumulative explained variance threshold.

[0031] S103: Construct a redundant decoupling loss based on the differentiable grouping matrix, the feature mutual information matrix, and the feature correlation coefficient matrix. This redundant decoupling loss is used to constrain information overlap and correlation within the same feature group. Specifically, in this embodiment of the invention, the step of constructing the redundant decoupling loss may include: Calculate the mutual information matrix and correlation coefficient matrix between each pair of features, configure the correlation coefficient matrix as an absolute correlation coefficient matrix, and set the diagonal elements of the mutual information matrix and the correlation coefficient matrix to zero; For the Let there be _n_ feature groups, and denote their allocation vector in the grouping matrix as _n_i. The intra-group information overlap penalty and intra-group correlation penalty are obtained by performing quadratic weighting using the following formulas.

[0032] In the formula: For balance coefficient, For the total number of features, For mutual information matrix, The correlation coefficient matrix, This is the transpose symbol.

[0033] The total training loss is expressed by the following formula:

[0034] In the formula: To classify task losses, The weights are used for redundancy decoupling loss.

[0035] S104: Based on the differentiable grouping matrix, multiple feature groups are obtained. These feature groups are then input into the intra-group branch and the inter-group branch, respectively. Simultaneously, the complete input feature representation output by the preprocessing module is input into the full feature branch to form a three-branch network architecture composed of parallel grouping paths and full feature paths. In specific implementation, this embodiment of the invention can provide the bilinear attention block in the inter-group branch to perform the following steps: The interaction terms of the second-order factorization machine are calculated by multiplicative cross-units to obtain the pairwise interaction signals between groups; Group-level attention scores are generated using the SENET structure through attention weight units. And reweight the group representation V to obtain ; Enumerate all that satisfy The disordered feature pairs are learned through a bilinear weight matrix. ; Calculate the bilinear cross tensor as well as ; Will The signal is copied to the matching dimension and concatenated with the bilinear attention signal to obtain the concatenated features. ; The dimension-preserving unit compresses the spliced ​​features back to the original group representation dimension through one-dimensional convolution, batch normalization, and ReLU function, and then passes them to the next layer through the residual output unit.

[0036] S105: Implicit interaction modeling is performed on each feature group through the intra-group branch, and multiplicative cross, attention weighting, dimension-preserving projection and residual output are performed on the group representations of different feature groups through the bilinear attention block in the inter-group branch. Global interaction modeling is performed on the complete feature representation through the full feature branch. In specific implementation, the embodiment of the present invention can provide that the stacking layer of the bilinear attention block is set to 2 layers to form a third-order inter-group feature cross, balancing expressive power and sample efficiency under the premise of approximately linear growth in computational complexity.

[0037] S106: Gated weighted fusion of the outputs of the intra-group branch, the inter-group branch, and the full-feature branch to obtain a risk score for early screening of neurodegenerative diseases.

[0038] In specific implementation, embodiments of the present invention may provide the following specific steps for the gated weighted fusion: Dynamic branch weights are generated based on the full feature output, within-group output, and between-group output through a gated weight calculation unit based on the Sigmoid function. When local feature defects dominate, the weight of intra-group branches increases adaptively; when cross-group associations dominate, the weight of inter-group branches increases adaptively. The outputs of the three branches are weighted and summed to obtain the early screening risk score.

[0039] The early screening method for neurodegenerative diseases provided by this invention simultaneously supplies data-driven grouping results guided by redundancy decoupling loss to both intra-group and inter-group branches, and performs gated fusion with the full feature branch, thus forming a collaborative technical solution that differs from existing technologies. It can preprocess tabular feature data generated from natural human-computer interaction tests, perform data-driven grouping, redundancy decoupling constraints, three-branch interaction modeling, and gated fusion output. During end-to-end training, it automatically learns non-redundant feature groups and captures intra-group local dependencies, inter-group association rules, and global higher-order dependencies, thereby improving the accuracy, robustness, and interpretability of early screening in scenarios with small sample sizes, heterogeneous, and redundant medical tabular data.

[0040] The following is a detailed description of the early screening method for neurodegenerative diseases provided in this application.

[0041] like Figure 2 and Figure 6 As shown, this invention provides an early screening method for neurodegenerative diseases. The reference numerals in the accompanying drawings are as follows: 100 is a three-branch feature crossover network system; 101 is an input module; 102 is a preprocessing module; 103 is a data-driven grouping module; 104 is an intra-group branch; 105 is an inter-group branch; 106 is a full-feature branch; 107 is a gated fusion module; 108 is an output module; 201 is an input feature matrix; 202 is a batch statistics unit; 203 is a grouping mapping unit; 204 is a differentiable sampling unit; 205 is a first feature group; 206 is other feature groups; 207 is a redundancy decoupling unit; 301 is a group representation; 302 is a multiplicative crossover unit; 303 is an attention weight unit; 304 is a dimension-preserving unit; 305 is a residual output unit; 401 is a full-feature output; 402 is an intra-group output; 403 is an inter-group output; 404 is a weight calculation unit; 405 is a fusion classification unit.

[0042] Figure 2 The overall steps of the method are shown. Figure 6 The modular network architecture and data flow of a three-branch feature cross-network system 100 are demonstrated. The method includes the following steps: The table feature data generated by the subject's natural human-computer interaction test is obtained, and the table feature data is preprocessed to obtain the input feature matrix.

[0043] Based on the statistics of the input feature matrix in the batch dimension and the learnable mapping parameters, grouping logical values ​​are generated, and differentiable grouping matrices are generated through the Gumbel-Softmax function.

[0044] A redundancy decoupling loss is constructed based on the differentiable grouping matrix, the feature mutual information matrix, and the feature correlation coefficient matrix. The redundancy decoupling loss is used to constrain information overlap and correlation within the same feature group.

[0045] like Figure 2 As shown, multiple feature groups are obtained according to the differentiable grouping matrix. The multiple feature groups are respectively input into the intra-group branch 104 and the inter-group branch 105. At the same time, the complete input feature representation output by the preprocessing module 102 is directly input into the full feature branch 106 to form a three-branch network architecture composed of parallel grouping paths and full feature paths.

[0046] Implicit interaction modeling is performed on each feature group through the intra-group branch, and multiplicative cross-interaction, attention weighting, dimension-preserving projection and residual output are performed on the group representations of different feature groups through the bilinear attention block in the inter-group branch. Global interaction modeling is performed on the complete feature representation through the full feature branch.

[0047] The outputs of the intra-group branch, the inter-group branch, and the full-feature branch are gated and weighted to obtain a risk score for early screening of neurodegenerative diseases.

[0048] In one alternative implementation, feature selection selects a preset number of features based on the mutual information between the input features and the training labels; normalization parameters are calculated on the training set and applied to both the training and test sets.

[0049] In one alternative implementation, the number of feature groups The group representation dimension is determined based on the singular value energy threshold of the input feature matrix. Determined based on the cumulative explained variance threshold of the principal components.

[0050] In one alternative implementation, the redundancy decoupling loss The mutual information matrix (MI) and the correlation coefficient matrix (CC) are jointly determined, and the balance coefficients can be learned or preset. Adjust the weights of both. The total training loss is... ,in To classify task losses, The weights are used for redundancy decoupling loss.

[0051] In one alternative implementation, the number of stacked layers of bilinear attention blocks is used to control the order of inter-group feature crossing; one bilinear attention block corresponds to a second-order inter-group crossing, and the crossing order increases by one level for each additional bilinear attention block. In particular, the crossing order is set to three to achieve a balance between expressive power, sample efficiency, and computational complexity.

[0052] In one alternative implementation, the dynamic gating fusion module includes a gating weight calculation unit based on the Sigmoid function. The gating weight calculation unit generates branch weights based on the outputs of different branches and performs weighted fusion on the outputs of the three branches.

[0053] In practice: Three-branch feature cross-network method.

[0054] For early screening of neurodegenerative diseases, the input consists of tabular data generated by subjects in a natural human-computer interaction test. Each row corresponds to a subject or a test instance, and each column corresponds to a behavioral, physiological, or statistical characteristic. Training labels can be binary training labels for healthy controls and disease-related risk states, or they can be expanded to multi-category training labels.

[0055] The output is an early screening risk score or probability value obtained by a computer model based on tabular feature data. This output is used for medical personnel, researchers, or screening platforms to make subsequent assessments. It does not directly determine whether the subject has a specific disease, nor does it replace the diagnosis made by a doctor based on clinical examination.

[0056] First, the original table data is preprocessed. Combined with... Figure 2 and Figure 6 As shown, the input module 101 receives the subject's natural human-computer interaction test data, and the preprocessing module 102 performs standardization, encoding, missing value processing, and feature selection on the data. For continuous features, standardization can be performed based on training set statistics; for categorical features, one-hot encoding or embedding encoding can be used; for missing values, the mean, median, model estimation, or missing indicator variables can be used for processing. In one implementation, features are first selected based on the mutual information between each feature and the training labels. One characteristic, It can be set to 50, or adjusted according to the device's computing power, sample size, and task requirements. The input feature matrix is ​​obtained after preprocessing. ,in Indicates the number of samples. This represents the number of features; the input feature matrix serves as... Figure 6 The common data foundation for subsequent grouping paths and full feature paths.

[0057] Secondly, such as Figure 3 and Figure 6As shown, the input feature matrix 201 is input into the data-driven grouping module 103. This module calculates a grouping matrix. ,in Indicates the number of feature groups. The Middle Line number Column elements represent the first The first feature was assigned to the first... The probability of each feature group. Specifically, the batch statistics unit 202 first calculates the mean or other statistics of the input features along the batch dimension, the grouping mapping unit 203 then obtains the grouping logic value through a learnable linear mapping, and the differentiable sampling unit 204 uses the Gumbel-Softmax function to generate an approximately discrete but backpropagable grouping matrix. Thus, each input feature is dynamically assigned to the first feature group 205 or other feature groups 206; Figure 6 In the network architecture shown, the grouping results are used by the intra-group branch 104 to extract local dependencies within the group and by the inter-group branch 105 to perform explicit cross-group crossing. During the inference phase, the feature group to which it belongs can be determined based on the highest probability.

[0058] Furthermore, redundant decoupling loss is introduced during training through redundant decoupling unit 207. For the input feature set, the mutual information matrix MI and correlation coefficient matrix CC are calculated between each pair of features. The mutual information matrix measures the strength of information that can be shared between two features, and the correlation coefficient matrix measures the linear or statistical similarity between two features. For the g-th feature group, its allocation vector in the grouping matrix M is denoted as... Then it can be based on Applying a quadratic weighting to MI and CC yields the intra-group information overlap penalty and the intra-group correlation penalty. One possible implementation is:

[0059] in, These are the balance coefficients. Furthermore, MI and CC can be normalized to the [0,1] interval, the correlation coefficient matrix CC can use the absolute correlation coefficient, and the diagonal elements of MI and CC can be set to zero, or... Terms sharing the same feature are excluded to avoid autocorrelation affecting intra-group redundancy penalties. During model training, the total loss is:

[0060] in, To classify task losses, To avoid redundant decoupling of loss weights, the preferred loss for classification tasks is cross-entropy loss.

[0061] Through the aforementioned grouping and constraints, features within the same group tend to retain independent discriminative power and reduce redundant overlap, while high-value cross-group relationships are guided to inter-group branches for explicit modeling. This process does not require manually pre-defined feature domains, nor does it rely on fixed medical expert grouping rules, thus enabling it to adapt to screening datasets of different sources and sizes.

[0062] Subsequently, build Figure 6 The diagram shows three parallel branches. The full-feature branch 106 receives the complete feature representation output by the preprocessing module 102 and learns global dependencies through a multilayer perceptron, attention layer, or other high-order implicit interaction structures. The intra-group branch 104 receives the feature representations of each group obtained by the data-driven grouping module 103 and performs implicit interaction modeling on each group to highlight local defects and fine-grained consistency within the group. The inter-group branch 105 receives the encoded representations of multiple feature groups and performs explicit cross-group cross-interaction through a bilinear attention block to discover association rules between different physiological or behavioral feature groups. These three branches are parallel in the network architecture, preserving information at different granularities, and are adaptively merged in the subsequent gated fusion module 107.

[0063] In intergroup branch 105, such as Figure 4 and Figure 6 As shown, the bilinear attention block receives the group representation output from the previous layer and integrates factorization machine interaction and bilinear attention interaction to perform explicit inter-group feature crossing. First, the multiplicative crossing unit 302 calculates the second-order FM interaction term based on the previous layer group representation to obtain the inter-group pairwise interaction signal. Second, the attention weighting unit 303 uses the SENET structure to generate group-level attention scores:

[0064] And received a reweighted group statement Subsequently, for all those that satisfy... We enumerate the unordered feature pairs; for the p-th pair, its index is... and And set a learnable bilinear weight matrix. Thus, the original group represents the bilinear cross tensor involved:

[0065] And the bilinear cross tensor involved in the output of the previous layer:

[0066] Two bilinear cross tensors are input into a single-head self-attention module to obtain the attention representation:

[0067] Then, the feature aggregation unit copies the FM signal to the matching dimension and concatenates it with the three bilinear attention signals mentioned above. The dimension-preserving unit 304 compresses the 4E-dimensional cross features back to E-dimensionality through continuous one-dimensional convolution, batch normalization, and the ReLU function, and reshapes them into group representation dimensions, thus obtaining the first E-dimensional cross feature. A bilinear attention block output The residual output unit 305 passes the output to the next layer of bilinear attention block, so that the first... The output of the nth block is used as the nth The input to each block is processed by a Transformer layer with shared parameters, followed by concatenation and linear mapping to obtain the final output of the inter-group branches. The data is then fed into the dynamic gating fusion module. Through this structure, BAB can form explicit inter-group crosses of a controllable order according to the number of stacked layers, while using a dimension-preserving layer to prevent the feature dimension from expanding with the cross order.

[0068] In one specific implementation, the number of bilinear attention block stacking layers, L, determines the inter-group crossover order. When L is 1, it corresponds to second-order inter-group crossover; as L increases, the crossover order increases accordingly. Experiments show that third-order crossover is stable on multiple neurodegenerative disease screening datasets, therefore it can be used as the default configuration. In other tasks, L can be determined through a validation set or automatic hyperparameter search.

[0069] After the three branches yield the full feature output 401, the intra-group output 402, and the inter-group output 403 respectively, as follows: Figure 5 and Figure 6 As shown, the input is given to the dynamic gating fusion module 107. The weight calculation unit 404 generates gating weights based on the outputs of the three branches and adaptively adjusts the contributions of different branches. When the local feature defects of a dataset are more obvious, the weight of the intra-group branch 104 can be increased; when the cross-group association is more discriminative, the weight of the inter-group branch 105 can be increased; when the global distribution pattern is more stable, the weight of the full feature branch 106 can be increased. The fusion classification unit 405 performs a weighted summation of the predicted values ​​or latent representations of the three branches based on the gating weights, and obtains the early screening risk score through the output module 108.

[0070] The system implementation, as described in this embodiment, provides a software system that can be deployed on servers, workstations, edge computing devices, or early medical screening platforms. The system includes the following modules.

[0071] The data acquisition and preprocessing module receives handwriting, speech, gait, or other natural human-computer interaction test data from subjects and converts the raw signals into structured tabular features. This module can also perform feature selection, missing value handling, normalization, and batch construction.

[0072] The data-driven grouping module generates a differentiable grouping matrix based on the input feature matrix. This module includes a batch statistics unit, a grouping logic mapping unit, and a Gumbel-Softmax sampling unit. The batch statistics unit outputs the feature mean or other statistics; the grouping logic mapping unit obtains the allocation logic of each feature to each group through learnable parameters; and the Gumbel-Softmax sampling unit outputs a differentiable approximate discrete grouping result.

[0073] The redundancy decoupling constraint module is used to calculate the mutual information matrix and correlation coefficient matrix during the training phase, and generate the redundancy decoupling loss accordingly. Mutual information can be obtained using maximum information coefficient, discretization estimation, kernel estimation, or neural estimation; correlation coefficients can be obtained using Pearson correlation coefficient, Spearman correlation coefficient, or other similarity measures.

[0074] The full-feature interaction module is used for implicit high-order interaction modeling of the complete feature space. This module can employ multilayer perceptrons, residual networks, attention networks, or combinations thereof.

[0075] The intra-group interaction module is used to process each feature group separately and extract local pathological clues. This module can share parameters for each feature group, or use independent parameters for different groups.

[0076] The inter-group interaction module is used to perform explicit cross-group cross-interaction on the group representation through one or more bilinear attention blocks. This module includes a multiplicative cross-interaction unit, a bilinear attention unit, a dimension-preserving unit, and a residual output unit.

[0077] The dynamic gating fusion module is used to calculate the fusion weights based on the outputs of the three branches. This module can use sigmoid gating, softmax gating, or conditional gating networks.

[0078] The prediction output module is used to output early screening risk scores or probability values, and can return risk alerts to doctors, researchers, or screening platforms for subsequent assessment.

[0079] Training and inference process.

[0080] The training phase includes the following process.

[0081] First, construct training, validation, and test sets, and compute normalized parameters on the training set. Second, perform preliminary feature selection based on mutual information or other criteria. Third, according to... Figure 6The network architecture shown initializes the parameters of the input module 101, preprocessing module 102, data-driven grouping module 103, intra-group branch 104, inter-group branch 105, full-feature branch 106, gated fusion module 107, and output module 108, including grouping matrix generation parameters, branch parameters, bilinear attention block parameters, and gated fusion parameters. Then, training batches are input into the network to obtain prediction results, and the redundancy decoupling loss is calculated based on the current grouping matrix. Finally, the model parameters are updated using AdamW or another optimizer, with the weighted sum of the task loss and the redundancy decoupling loss as the optimization objective.

[0082] The reasoning phase includes the following process.

[0083] First, the tabular characteristics of the subjects to be screened in the early stage were processed using the preprocessing parameters determined in the training phase, and then... Figure 6 The input module 101 feeds into the preprocessing module 102. Next, the data-driven grouping module 103 generates group representations based on the input features and sends them to the intra-group branch 104 and the inter-group branch 105, respectively. Simultaneously, the complete feature representation is sent to the full feature branch 106. Then, the three branches calculate the global, intra-group, and inter-group interaction results, respectively. Finally, the dynamic gating fusion module 107 fuses the outputs from the three branches, and the output module 108 outputs an early screening risk score or probability value. This risk score or probability value can be used for subsequent evaluation.

[0084] Applicable data and parameter configuration.

[0085] This invention is applicable to tabular data such as handwriting characteristics of Alzheimer's disease, speech characteristics of Parkinson's disease, speech characteristics of amyotrophic lateral sclerosis (ALS), and gait characteristics of Alzheimer's disease. Typical input features include, but are not limited to, handwriting speed, pressure, pause time, trajectory geometric statistics, pitch periodic entropy, MFCC, formants, jitter, dim light, harmonic structure, step height, step frequency, swing phase, and gait variability.

[0086] In one specific configuration, the number of input features C is set to 50 after mutual information selection; the group representation dimension E can explain variance exceeding a threshold based on the cumulative variance of the principal components. The minimum dimension is determined; the number of feature groups G can be determined based on the singular value energy exceeding a threshold. The minimum number of singular values ​​is determined; the preferred crossover order of the bilinear attention block is 3; the optimizer can be AdamW, with a learning rate of 0.001, weight decay of 0.001, batch size of 256, and maximum training epochs of 100, and early stopping based on validation loss. These parameters respectively act on... Figure 6 The data-driven grouping module 103, intra-group branch 104, inter-group branch 105, and full-feature branch 106 are included.

[0087] The parameters described above are merely examples. Those skilled in the art can adjust the parameters based on sample size, number of features, disease type, and computing power of the deployed equipment without affecting the core concept of this invention: early screening through data-driven grouping, redundancy decoupling, three-branch interaction, and bilinear attention blocks.

[0088] Computational complexity.

[0089] The computational overhead of this invention mainly comes from Figure 6 The data-driven grouping module 103 and the bilinear attention block in the inter-group branch 105 are examples. The data-driven grouping module needs to calculate mutual information and correlation coefficients, and its complexity is related to the number of samples N and the number of features C, which can be approximated as follows: The bilinear attention block performs cross-group interactions across G feature groups and computes interactions for each group's representation dimension E. Its single-layer complexity can be approximated as... Since this invention maps the cross-results of each layer back to a fixed dimension through a dimension-preserving module, the complexity of stacking L layers of bilinear attention blocks can be approximated as follows: This avoids the exponential growth of higher-order explicit crossovers.

[0090] Therefore, while improving the ability to express cross features, the present invention can maintain acceptable computational complexity, making it suitable for deployment in medical screening platforms, workstations, or edge devices.

[0091] In summary, the early screening method for neurodegenerative diseases provided by this invention replaces manual heuristic grouping with a data-driven grouping module, incorporates the feature segmentation process into end-to-end training, and enables feature grouping to be adaptively updated according to task objectives, thereby reducing reliance on domain-specific manual rules.

[0092] By constraining mutual information and correlation coefficients simultaneously through redundancy decoupling loss, it helps to reduce intra-group redundancy and information overlap, making each feature group more compact and having relatively independent discriminative capabilities.

[0093] By modeling in parallel using full-feature branches, intra-group branches, and inter-group branches, global dependencies, local defects, and cross-group association rules are captured respectively. Furthermore, intra-group branches and inter-group branches are both constrained by data-driven grouping results, which can balance the commonality and diversity of disease signals and avoid information omissions caused by single-path interactions.

[0094] Interpretable explicit inter-group feature crossings are achieved through bilinear attention blocks, and multi-order stacking is supported through dimension preservation and residual connections, making the computational complexity grow approximately linearly rather than exponentially as the inter-group crossing depth increases.

[0095] This method is applicable to medical tabular data from different natural human-computer interaction sources such as writing, speech, and gait. It can be used as a computer model component in early screening or risk assessment to improve the robustness of screening in heterogeneous, small-sample, and redundant feature scenarios.

[0096] See Figure 7 The present invention can also provide an early screening device for neurodegenerative diseases, such as... Figure 7 As shown, the device may include: The preprocessing unit 701 is used to acquire table feature data generated by the subject's natural human-computer interaction test, and to preprocess the table feature data to obtain an input feature matrix; Differentiable grouped matrix generation unit 702 is used to generate grouped logical values ​​based on the statistics of the input feature matrix in the batch dimension and the learnable mapping parameters, and generate a differentiable grouped matrix through the Gumbel-Softmax function. The redundant decoupling loss construction unit 703 is used to construct a redundant decoupling loss based on the differentiable grouping matrix, the feature mutual information matrix and the feature correlation coefficient matrix. The redundant decoupling loss is used to constrain the information overlap and correlation within the same feature group. The three-branch processing unit 704 is used to obtain multiple feature groups according to the differentiable grouping matrix, input the multiple feature groups into the intra-group branch and the inter-group branch respectively, and input the complete input feature representation output by the preprocessing module into the full feature branch to form a three-branch network architecture composed of parallel grouping paths and full feature paths. The three-way feature interaction modeling unit 705 is used to perform implicit interaction modeling on each feature group through the intra-group branch, perform multiplicative cross, attention weighting, dimension-preserving projection and residual output on the group representation of different feature groups through the bilinear attention block in the inter-group branch, and perform global interaction modeling on the complete feature representation through the full feature branch. Output unit 706 is used to perform gated weighted fusion of the outputs of the intra-group branch, the inter-group branch and the full feature branch to obtain a risk score for early screening of neurodegenerative diseases.

[0097] This invention can also provide an early screening device for neurodegenerative diseases, the device comprising a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the steps of the above-described early screening method for neurodegenerative diseases according to the instructions in the program code.

[0098] like Figure 8As shown in the figure, an early screening device for neurodegenerative diseases provided by an embodiment of the present invention may include: a processor 10, a memory 11, a communication interface 12, and a communication bus 13. The processor 10, the memory 11, and the communication interface 12 all communicate with each other through the communication bus 13.

[0099] In this embodiment of the invention, the processor 10 may be a central processing unit (CPU), an application-specific integrated circuit, a digital signal processor, a field-programmable gate array, or other programmable logic devices.

[0100] The processor 10 can call programs stored in the memory 11. Specifically, the processor 10 can perform operations in embodiments of the early screening method for neurodegenerative diseases.

[0101] The memory 11 is used to store one or more programs. The programs may include program code, which includes computer operation instructions. In this embodiment of the invention, the memory 11 stores at least a program for implementing the following functions: Obtain the tabular feature data generated by the subject's natural human-computer interaction test, and preprocess the tabular feature data to obtain the input feature matrix; Based on the statistics of the input feature matrix in the batch dimension and the learnable mapping parameters, grouping logical values ​​are generated, and differentiable grouping matrices are generated through the Gumbel-Softmax function. A redundancy decoupling loss is constructed based on the differentiable grouping matrix, the feature mutual information matrix, and the feature correlation coefficient matrix. The redundancy decoupling loss is used to constrain information overlap and correlation within the same feature group. Multiple feature groups are obtained based on the differentiable grouping matrix. The multiple feature groups are input into the intra-group branch and inter-group branch respectively. At the same time, the complete input feature representation output by the preprocessing module is input into the full feature branch to form a three-branch network architecture composed of parallel grouping paths and full feature paths. Implicit interaction modeling is performed on each feature group through the intra-group branch; multiplicative cross-interaction, attention weighting, dimension-preserving projection and residual output are performed on the group representations of different feature groups through the bilinear attention block in the inter-group branch; and global interaction modeling is performed on the complete feature representation through the full feature branch. The outputs of the intra-group branch, the inter-group branch, and the full-feature branch are gated and weighted to obtain a risk score for early screening of neurodegenerative diseases.

[0102] In one possible implementation, the memory 11 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function (such as file creation or data read / write). The data storage area may store data created during use, such as initialization data.

[0103] In addition, memory 11 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device or other volatile solid-state storage device.

[0104] Communication interface 12 can be an interface for the communication module, used to connect with other devices or systems.

[0105] Of course, it should be noted that, Figure 8 The structure shown does not constitute a limitation on the early screening device for neurodegenerative diseases in the embodiments of the present invention. In practical applications, the early screening device for neurodegenerative diseases may include more than [other devices]. Figure 8 More or fewer components as shown, or combinations of certain components.

[0106] Embodiments of the present invention may also provide a computer-readable storage medium for storing program code for performing the steps of the above-described method for early screening of neurodegenerative diseases.

[0107] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0108] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0109] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0110] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for early screening of neurodegenerative diseases, characterized in that, The method includes: Obtain the tabular feature data generated by the subject's natural human-computer interaction test, and preprocess the tabular feature data to obtain the input feature matrix; Based on the statistics of the input feature matrix in the batch dimension and the learnable mapping parameters, grouping logical values ​​are generated, and differentiable grouping matrices are generated through the Gumbel-Softmax function. A redundancy decoupling loss is constructed based on the differentiable grouping matrix, the feature mutual information matrix, and the feature correlation coefficient matrix. The redundancy decoupling loss is used to constrain information overlap and correlation within the same feature group. Multiple feature groups are obtained based on the differentiable grouping matrix. The multiple feature groups are respectively input into the intra-group branch and the inter-group branch. At the same time, the complete input feature representation output by the preprocessing module is input into the full feature branch to form a three-branch network architecture composed of parallel grouping paths and full feature paths. Implicit interaction modeling is performed on each feature group through the intra-group branch; multiplicative cross-interaction, attention weighting, dimension-preserving projection and residual output are performed on the group representations of different feature groups through the bilinear attention block in the inter-group branch; and global interaction modeling is performed on the complete feature representation through the full feature branch. The outputs of the intra-group branch, the inter-group branch, and the full-feature branch are gated and weighted to obtain a risk score for early screening of neurodegenerative diseases.

2. The method for early screening of neurodegenerative diseases according to claim 1, characterized in that, The preprocessing of the table feature data includes: Select a preset number of features based on the mutual information between the input features and the training labels; The normalization parameters are calculated on the training set and applied to both the training and test sets.

3. The method for early screening of neurodegenerative diseases according to claim 1, characterized in that, The table feature data generated by the natural human-computer interaction test includes at least one of the following: handwriting speed, pressure, pause time, trajectory geometric statistics, pitch period entropy, MFCC, formants, jitter, step height, step frequency, swing phase, and gait variability.

4. The method for early screening of neurodegenerative diseases according to claim 1, characterized in that, The number of feature groups in the differentiable grouping matrix is ​​determined based on the singular value energy threshold of the input feature matrix, and the group representation dimension is determined based on the cumulative explained variance threshold of the principal components.

5. The method for early screening of neurodegenerative diseases according to claim 1, characterized in that, The steps for constructing the redundant decoupling loss specifically include: Calculate the mutual information matrix and correlation coefficient matrix between each pair of features, configure the correlation coefficient matrix as an absolute correlation coefficient matrix, and set the diagonal elements of the mutual information matrix and the correlation coefficient matrix to zero; For the Let there be _n_ feature groups, and denote their allocation vector in the grouping matrix as _n_i. The intra-group information overlap penalty and intra-group correlation penalty are obtained by performing quadratic weighting using the following formulas. In the formula: For balance coefficient, For the total number of features, For mutual information matrix, The correlation coefficient matrix, This is the transpose symbol.

6. The method for early screening of neurodegenerative diseases according to claim 5, characterized in that, The total training loss is expressed by the following formula: In the formula: To classify task losses, The weights are used for redundancy decoupling loss.

7. The method for early screening of neurodegenerative diseases according to claim 1, characterized in that, The bilinear attention block in the intergroup branch performs the following steps: The interaction terms of the second-order factorization machine are calculated by multiplicative cross-units to obtain the pairwise interaction signals between groups; Group-level attention scores are generated using the SENET structure through attention weight units. And reweight the group representation V to obtain ; Enumerate all that satisfy of Unordered feature pairs, through a learnable bilinear weight matrix ; Calculate the bilinear cross tensor as well as ; Will The signal is copied to the matching dimension and concatenated with the bilinear attention signal to obtain the concatenated features. ; The dimension-preserving unit compresses the spliced ​​features back to the original group representation dimension through one-dimensional convolution, batch normalization, and ReLU function, and then passes them to the next layer through the residual output unit.

8. The method for early screening of neurodegenerative diseases according to claim 7, characterized in that, The number of stacked layers of the bilinear attention block is set to 2 to form a third-order inter-group feature crossover, balancing expressive power and sample efficiency under the premise of approximately linear growth in computational complexity.

9. The method for early screening of neurodegenerative diseases according to claim 1, characterized in that, The gated weighted fusion step specifically includes: Dynamic branch weights are generated based on the full feature output, within-group output, and between-group output through a gated weight calculation unit based on the Sigmoid function. When local feature defects dominate, the weight of intra-group branches increases adaptively; when cross-group associations dominate, the weight of inter-group branches increases adaptively. The outputs of the three branches are weighted and summed to obtain the early screening risk score.

10. An early screening device for neurodegenerative diseases, characterized in that, The apparatus is used to perform the early screening method for neurodegenerative diseases according to any one of claims 1-9, the apparatus comprising: The preprocessing unit is used to acquire tabular feature data generated by the subject's natural human-computer interaction test, and to preprocess the tabular feature data to obtain the input feature matrix; A differentiable grouping matrix generation unit is used to generate grouping logical values ​​based on the statistics of the input feature matrix in the batch dimension and the learnable mapping parameters, and to generate a differentiable grouping matrix through the Gumbel-Softmax function. A redundant decoupling loss construction unit is used to construct a redundant decoupling loss based on the differentiable grouping matrix, the feature mutual information matrix, and the feature correlation coefficient matrix. The redundant decoupling loss is used to constrain information overlap and correlation within the same feature group. The three-branch processing unit is used to obtain multiple feature groups based on the differentiable grouping matrix, input the multiple feature groups into the intra-group branch and the inter-group branch respectively, and input the complete input feature representation output by the preprocessing module into the full feature branch to form a three-branch network architecture composed of parallel grouping paths and full feature paths. The three-way feature interaction modeling unit is used to perform implicit interaction modeling on each feature group through the intra-group branch, perform multiplicative cross, attention weighting, dimension-preserving projection and residual output on the group representation of different feature groups through the bilinear attention block in the inter-group branch, and perform global interaction modeling on the complete feature representation through the full feature branch. The output unit is used to perform gated weighted fusion of the outputs of the intra-group branch, the inter-group branch, and the full-feature branch to obtain a risk score for early screening of neurodegenerative diseases.

Images