Method for constructing multi-dimensional intervention scheme for intrinsic ability of old people based on neural network

Abstract

The invention discloses a neural network-based old people internal ability multi-dimensional intervention scheme construction method. The method comprises the steps of obtaining preprocessed multi-modal sensory data; generating an edge exception marking result and uploading the edge exception marking result; outputting a second multi-scale feature representation by using a channel sensing scale selection mechanism; outputting a third self-supervised representation vector by adopting a mask rate dynamic adjustment mechanism driven by sensory characteristics; generating a fourth fusion feature tensor; outputting a fifth time sequence feature sequence, and obtaining a sensory decline trend regression result, a sensory decline risk grading result and a sensory pathway priority protection list; and generating an intervention matrix including an intervention scene, an intervention action, an intervention dose and an intervention frequency, and defining the intervention matrix as an old people internal ability multi-dimensional intervention scheme. According to the method, the high-risk sensory pathway and the critical decline moment can be recognized earlier and more accurately, and support is provided for continuous health management and health care service of the old.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intrinsic ability of the elderly, and in particular to a neural network-based multi-dimensional intervention scheme construction method for intrinsic ability of the elderly. BACKGROUND

[0002] With the acceleration of population aging in China, the problem of sensory ability decline of the elderly is increasingly prominent, which seriously affects their quality of life and social participation ability. Comprehensive assessment and early intervention of multiple sensory functions such as vision, hearing, touch and balance of the elderly have become an important research direction in the field of intelligent health and digital rehabilitation. Existing sensory ability decline assessment and intervention technologies mainly rely on traditional quantitative examination, questionnaire scale or single modality sensory detection equipment, but they generally have strong dependence on manual annotation, strong subjectivity in assessment, limited scope of application, and difficulty in capturing small changes and interaction in the multi-modal sensory decline process.

[0003] In the fusion modeling and dynamic intervention scheme construction of multi-modal sensory data, the existing technology usually adopts a single task, static template or one-size-fits-all rehabilitation training mode, which lacks dynamic tracking and graded response to individual sensory decline trends. The intervention scheme is difficult to balance personalized efficacy, compliance and resource utilization efficiency. For low-annotation sample or cross-institutional collaborative training sensory decline assessment tasks, the existing method lacks generalization ability and data security protection of the model, and is prone to problems such as model overfitting, inaccurate assessment and user privacy risks. SUMMARY

[0004] One object of the present application is to provide a neural network-based multi-dimensional intervention scheme construction method for intrinsic ability of the elderly, which can identify high-risk sensory pathways and critical decline moments earlier and more accurately, providing support for continuous health management and health care services for the elderly.

[0005] According to the neural network-based multi-dimensional intervention scheme construction method for intrinsic ability of the elderly according to an embodiment of the present application, the method comprises:

[0006] Collecting multi-modal sensory data of the elderly and pre-processing to obtain pre-processed multi-modal sensory data;

[0007] Performing feature preliminary screening on the pre-processed multi-modal sensory data based on a lightweight Res2Net-SimMIM sub-model, generating a first edge feature vector and uploading it to the cloud together with the synchronous metadata, performing decline anomaly detection based on the first edge feature vector, and generating and uploading an edge anomaly label result;

[0008] performing multi-scale feature extraction on the preprocessed multi-modal sensory data based on a Res2Net multi-scale convolutional network in the cloud, and outputting a second multi-scale feature representation by using a channel perception scale selection mechanism;

[0009] performing self-supervised pre-training on the preprocessed multi-modal sensory data based on a SimMIM self-supervised mask image modeling framework in the cloud, and outputting a third self-supervised representation vector by using a sensory feature driven mask rate dynamic adjustment mechanism;

[0010] inputting the edge anomaly labeling result, the second multi-scale feature representation, and the third self-supervised representation vector into a cross-modal attention fusion module to generate a fourth fusion feature tensor;

[0011] inputting the fourth fusion feature tensor into a multi-scale time series modeling module to complete short-term fluctuation modeling and long-term trend modeling, and outputting a fifth time series feature sequence, inputting the fifth time series feature sequence into a multi-task evaluation module to obtain a sensory decline trend regression result, a sensory decline risk classification result, and a sensory pathway priority protection list;

[0012] Based on the sensory decline trend regression result, the sensory decline risk classification result, and the sensory pathway priority protection list, calling an intervention strategy generation network and combining a multi-objective optimization module to generate an intervention matrix including an intervention scenario, an intervention action, an intervention dose, and an intervention frequency, and defining the intervention matrix as an old person's internal ability multi-dimensional intervention scheme.

[0013] Optionally, the construction of the preprocessed multi-modal sensory data comprises:

[0014] Collecting multi-modal sensory data of the elderly, the multi-modal sensory data including visual sensory data, auditory sensory data, tactile sensory data, and balance sensory data, and completing time synchronization, format unification, and noise suppression on the edge side to obtain preprocessed multi-modal sensory data.

[0015] Optionally, the generation of the edge anomaly labeling result comprises:

[0016] defining the preprocessed multi-modal sensory data as a group of data samples;

[0017] Each data sample includes a spatial height, a spatial width, and a number of channels, and all data samples form a batch input.

[0018] dividing each data sample into a plurality of patches according to a fixed patch size by using a deep separable convolution Patch-Embed module, and obtaining an initial patch feature sequence for each patch by linear mapping;

[0019] input each initial patch feature sequence into a lightweight Res2Net feature extraction unit, divide the initial patch feature sequence into multiple scale sub-branches, directly output the first scale sub-branch as the output feature of the first scale branch, for each scale sub-branch greater than the first, first perform a depth separable convolution operation on the output feature of the previous scale branch, then add it to the current scale sub-branch to obtain the output feature of the current scale branch, concatenate the output features of all scale branches and perform a one-time one-multiplication convolution operation to obtain a multi-scale fusion feature tensor;

[0020] apply SimMIM self-supervised masks to each multi-scale fusion feature tensor to generate a mask matrix, multiply the mask matrix and the multi-scale fusion feature tensor element by element to obtain a mask feature tensor, and input the mask feature tensor into a reconstruction decoder to output a reconstruction feature;

[0021] obtain a first edge feature vector for each reconstruction feature through a global average pooling operation, and calculate an anomaly score based on the first edge feature vector;

[0022] determine the size relationship between the anomaly score and the anomaly determination threshold, when the anomaly score is greater than or equal to the anomaly determination threshold, generate an edge anomaly label result as abnormal, otherwise generate an edge anomaly label result as normal.

[0023] Optionally, the generation of the second multi-scale feature representation comprises:

[0024] send the data sample into a cloud Res2Net multi-scale convolution network to perform first-stage convolution feature extraction to obtain a convolution feature tensor;

[0025] In a cloud environment, according to the statistical differences of different sensory modalities, the convolution feature tensor is divided into multiple scale feature groups, and for each scale branch in each feature group, convolution feature transmission and branch-by-branch fusion are performed in sequence to obtain a scale feature output set adapted to channel perception;

[0026] concatenate the multi-scale feature output sets of each sensory modality in the channel dimension respectively, and perform one-multiplication convolution compression to generate a second multi-scale fusion feature tensor;

[0027] perform sensory channel feature grouping and weighting processing on the second multi-scale fusion feature tensor, divide the sensory channels into a visual group, an auditory group, a tactile group and a balance group according to the source of the sensory modality, and calculate the global sensory statistics of each sensory modality group;

[0028] During the channel perception weight calculation process, the global sensory statistics of each modality group are combined with the decay trend prior vector calculated from the preceding time window to form a fused sensory prior. The fused sensory prior is then weighted and synthesized with the current channel response value to obtain the time-adaptive fused channel response value.

[0029] For the fusion channel response value of each scale branch, cross-scale dynamic normalization is performed, and the sharpness of the normalization is adjusted to generate a set of channel-aware scale selection weight coefficients.

[0030] The output features of each scale branch are weighted and fused according to the set of weight coefficients selected based on the channel perception scale to obtain the second multi-scale feature representation.

[0031] Optionally, the generation of the third self-supervised representation vector includes:

[0032] Each multimodal sensory data sample is input into the patch embedding module of the cloud-based SimMIM self-supervised mask image modeling framework. Each multimodal sensory data sample is divided into several patches according to a fixed patch size. Each patch is mapped to a patch embedding vector with a fixed embedding dimension through a linear transformation. All patch embedding vectors are sequentially concatenated to form a patch embedding sequence.

[0033] Based on the sensory feature-driven mask ratio dynamic adjustment mechanism, the patch importance score of each patch embedding vector in the current sensory modality decay assessment task is calculated.

[0034] A dynamic mask matrix is ​​generated based on the importance scores of all patches;

[0035] The dynamic mask matrix is ​​multiplied element-wise with the patch embedding sequence. The patch embedding vector values ​​at the masked positions are set to zero, while the original values ​​of the patch embedding vectors at the unmasked positions are retained to obtain the mask feature sequence.

[0036] The mask feature sequence is input into the reconstruction decoder, and the patch embedding vectors of all masked positions are recovered using the patch embedding vectors of all unmasked positions. The reconstructed feature sequence is then output.

[0037] A global average pooling operation is performed on the reconstructed feature sequence, and the reconstructed embedding vectors at all patch positions are averaged within the batch to obtain the third self-supervised representation vector.

[0038] Optionally, the generation of the fourth fusion feature tensor includes:

[0039] Based on edge anomaly labeling results, second multi-scale feature representation, and third self-supervised representation vector, a cross-modal feature heterogeneity measurement mechanism is established in the cross-modal attention fusion module. The cross-modal feature heterogeneity measurement mechanism is used to quantify the information complementarity and difference of different sensory modal features in the current scenario of sensory ability decline in the elderly, and obtain a set of cross-modal heterogeneity quantification factors.

[0040] A dynamic adaptive gating mechanism for modal importance is constructed based on the set of cross-modal heterogeneity quantization factors, and dynamically adjusted modal fusion weight coefficients are assigned to the second multi-scale feature representation and the third self-supervised representation vector, respectively.

[0041] A cross-modal multi-head collaborative attention mechanism is established based on dynamically adjusted modal fusion weight coefficients. Attention features are collaboratively optimized across different sensory modalities, and a cross-modal collaborative attention feature vector is output.

[0042] Based on cross-modal collaborative attention feature vectors, a decaying sensory modality compensation and enhancement mechanism is constructed. The feature representation strength of the sensory modality with a high degree of decay is explicitly enhanced through the sensory modality compensation matrix, and the compensated and enhanced cross-modal feature representation is output.

[0043] The compensated and enhanced cross-modal feature representation is input into the residual feedforward fusion network. Nonlinear mapping is performed on the compensated and enhanced cross-modal feature representation, and a residual connection structure is introduced to stabilize the fusion training, and the fourth fusion feature tensor is calculated.

[0044] Optionally, obtaining the sensory decline trend regression results, sensory decline risk classification results, and sensory pathway priority protection list includes:

[0045] In the multi-scale temporal modeling module, the fourth fusion feature tensor is divided into multi-scale time windows in the time dimension. The multi-scale time windows include short-term modeling windows and long-term modeling windows.

[0046] For each short-term modeling window, the feature sequence within the window is extracted and input into the temporal convolutional network. The temporal convolutional network uses a fixed convolutional kernel width to extract local temporal patterns from the sequence to obtain a set of short-term temporal feature vectors.

[0047] For each long-term modeling window, the feature sequence within the window is extracted and input into the bidirectional long short-term memory network. The bidirectional long short-term memory network models the global time dependency in both the forward and reverse directions and outputs a set of long-term time-series feature vectors.

[0048] The short-term time series feature vector set and the long-term time series feature vector set are concatenated along the feature dimension to form a multi-scale time series feature set, and then input into the cross-scale time series attention mechanism module. The cross-scale time series attention mechanism assigns attention weights to features in different time windows to obtain a multi-scale fused time series feature set.

[0049] The multi-scale fused temporal feature set is input into the feedforward fusion layer and mapped to the fifth temporal feature sequence through a fully connected layer and a normalization operation.

[0050] The fifth time-series feature sequence is input into the multi-task evaluation module, which includes shared low-level features and three task prediction heads:

[0051] The sensory decline trend regression task predicts the sensory ability change curve for future time intervals based on the fifth time series feature sequence, and outputs the sensory decline trend prediction results.

[0052] Optionally, the generation of the intervention matrix includes:

[0053] The intervention strategy generation network receives and integrates edge anomaly labeling results, second multi-scale feature representation, third self-supervised representation vector, fourth fusion feature tensor, and fifth temporal feature sequence as input. It combines sensory decline trend regression results, sensory decline risk classification results, and sensory pathway priority protection list with personal health records and historical intervention response records to generate a set of intervention candidate programs.

[0054] The set of intervention candidate protocols is input into the multi-objective optimization module. The multi-objective optimization module takes maximizing efficacy, optimizing compliance, and minimizing resource utilization as objective functions. For each candidate protocol, efficacy, compliance, and resource consumption are scored based on full-process characteristics and indicators. The multi-objective optimization module uses a weighted summation method or Pareto optimality method to comprehensively score and screen to obtain the optimal intervention parameter set. The optimal intervention parameter set includes a combination of four dimensions: intervention scenario, intervention action, intervention dose, and intervention frequency.

[0055] An intervention matrix is ​​generated based on the optimal set of intervention parameters. The intervention matrix is ​​defined as a multi-dimensional intervention plan for the intrinsic abilities of the elderly. Each row of the intervention matrix corresponds to a sensory pathway intervention unit, and each column represents four types of parameters: intervention scenario, intervention action, intervention dose, and intervention frequency.

[0056] Optionally, the output rules of the intervention matrix include:

[0057] When the sensory decline risk classification result is high risk and the sensory ability change curve shows a downward trend greater than the threshold, for the sensory modality that is given priority protection, the intervention scenario parameters in the intervention matrix are set to an enhanced training environment based on the edge anomaly labeling results. The intervention action parameters are selected as high-intensity multi-sensory stimulation with reference to the second multi-scale feature representation and the priority protection list. The intervention dose parameters are increased to twice a day, each lasting twenty minutes, based on the third self-supervised representation vector and the decline trend. The intervention frequency is set to be implemented continuously for four weeks.

[0058] When the sensory decline risk classification result is medium risk and the sensory ability change curve shows a slow decline less than the threshold, for the sensory modality that is prioritized for protection, the intervention scenario parameters in the intervention matrix are set to the standard training environment with reference to the edge anomaly labeling results, the intervention action parameters are selected as a medium-intensity sensory activation task with reference to the second multi-scale feature representation and the priority protection list, the intervention dose parameter is once a day, each lasting fifteen minutes, and the intervention frequency is set to be performed every other day for four weeks.

[0059] When the risk classification of sensory decline is low and the sensory ability change curve shows a stable or fluctuating trend, for the sensory modality that is given priority protection, the intervention scenario parameter in the intervention matrix is ​​set to the daily health care environment, the intervention action parameter is selected as a low-intensity sensory activation task, the intervention dosage parameter is twice a week, each lasting ten minutes, and the intervention frequency is set to be implemented for eight weeks.

[0060] The beneficial effects of this invention are:

[0061] (1) This invention utilizes the deep fusion of Res2Net multi-scale convolution and SimMIM self-supervised learning to effectively improve the accuracy and generalization ability of low-label multimodal sensory decline assessment. By combining the Res2Net multi-scale residual grouping convolution mechanism with the SimMIM mask self-supervised learning framework, fine-grained feature mining and cross-modal information reconstruction are achieved on visual, auditory, tactile and balance multimodal sensory data of the elderly. In scenarios with extremely low labeling or even unlabeled samples, Res2Net can fully capture local minor pathological changes in different sensory modalities, while SimMIM mask reconstruction can mine potential sensory decline precursor features, significantly reducing the dependence on manual labels and suppressing the risk of overfitting.

[0062] (2) In the cross-modal attention fusion module, the present invention introduces a heterogeneity quantification factor for the sensory ability decline scenario of the elderly, automatically measures the information complementarity and synergy between the visual, auditory, tactile and balance sensory pathways. The system can adjust the intervention scenario, action, dosage and frequency parameters in a targeted manner according to the real-time sensory assessment results, and realize the closed loop of the assessment-intervention decision-making process and iterative optimization.

[0063] (3) This invention uses multi-scale temporal window joint convolution and bidirectional long short-term memory network, combined with cross-scale temporal attention mechanism to jointly model the short-term fluctuations and long-term changes in the sensory abilities of the elderly; at the same time, it uses a multi-task evaluation module to synchronously output the sensory decline trend regression results, risk classification results and priority protection pathway list, providing high-dimensional decision support for the generation of intervention strategies, and can identify high-risk sensory pathways and critical decline moments earlier and more accurately, providing support for the continuous health management and health care services of the elderly. Attached Figure Description

[0064] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0065] Figure 1 This is a flowchart illustrating a method for constructing a multi-dimensional intervention program for the intrinsic abilities of the elderly based on neural networks, as proposed in this invention. Detailed Implementation

[0066] Example 1:

[0067] refer to Figure 1 A method for constructing a multi-dimensional intervention program for the intrinsic abilities of the elderly based on neural networks, comprising:

[0068] Multimodal sensory data of elderly people were collected and preprocessed to obtain preprocessed multimodal sensory data;

[0069] In this embodiment, the construction of preprocessed multimodal sensory data includes:

[0070] Multimodal sensory data of elderly people were collected, including visual, auditory, tactile, and balance sensory data. Time synchronization, format uniformity, and noise suppression were performed at the edge to obtain preprocessed multimodal sensory data.

[0071] The lightweight Res2Net-SimMIM sub-model is used to perform initial feature screening on the preprocessed multimodal sensory data, generate the first edge feature vector and upload it to the cloud along with synchronous metadata, and perform decay anomaly detection based on the first edge feature vector to generate edge anomaly labeling results and upload them.

[0072] In this embodiment, the generation of edge anomaly marking results includes:

[0073] The preprocessed multimodal sensory data is defined as a set of data samples;

[0074] Each data sample includes spatial height, spatial width, and number of channels, and all data samples constitute a batch input.

[0075] The depthwise separable convolutional Patch-Embed module is used to divide each data sample into several patches according to a fixed patch size, and the initial patch feature sequence is obtained for each patch through linear mapping;

[0076] In Example 1, the depthwise separable convolution Patch-Embed module is used to perform feature extraction on each preprocessed multimodal sensory data sample. Each multimodal sensory data sample is slidably segmented according to a fixed spatial size. Each image patch obtained from the slidwise segmentation is used as an independent input. For each image patch, a depthwise separable convolution operation is applied to extract spatial features. The depthwise separable convolution operation includes first performing channel-wise convolution in the spatial dimension, and then performing point-wise convolution in the channel dimension to obtain the embedding features of each image patch. The embedding features of all patches are sequentially concatenated to form an initial patch feature sequence. The number of patches in the initial patch feature sequence is equal to the product of the spatial height and spatial width of each multimodal sensory data sample divided by the patch size. The channel dimension of the embedding features is equal to the set embedding dimension.

[0077] Each initial patch feature sequence is input into a lightweight Res2Net feature extraction unit, which divides the initial patch feature sequence into multiple scale sub-branches. For the first scale sub-branch, the output is directly used as the output feature of the first scale branch. For each scale sub-branch larger than the first one, a depthwise separable convolution operation is first performed on the output feature of the previous scale branch, and then added to the current scale sub-branch to obtain the output feature of the current scale branch. The output features of all scale branches are concatenated and subjected to a one-to-one convolution operation to obtain a multi-scale fused feature tensor.

[0078] SimMIM self-supervised mask is applied to each multi-scale fusion feature tensor to generate a mask matrix. The mask matrix is ​​then multiplied element-wise with the multi-scale fusion feature tensor to obtain the mask feature tensor. The mask feature tensor is then input into the reconstruction decoder to output the reconstructed features.

[0079] In Example 1, a SimMIM self-supervised mask is applied to each multi-scale fusion feature tensor. Specifically, this involves: randomly selecting some patch positions from all patch features according to a set mask ratio; marking the mask corresponding to the selected patch position as masked and the mask corresponding to the unselected patch position as unmasked, thus obtaining a mask matrix. Each element of the mask matrix corresponds to the position of each patch feature in the multi-scale fusion feature tensor. The mask matrix is ​​then multiplied element-wise with the multi-scale fusion feature tensor. For patch features at masked positions, their values ​​are set to zero; for patch features at unmasked positions, their original values ​​are retained, resulting in a mask feature tensor. This mask feature tensor is then fed as input to the reconstruction decoder. The reconstruction decoder uses all unmasked patch features and employs an encoder-decoder structure to recover the patch features at all masked positions, outputting reconstructed features. Each position of the reconstructed features corresponds one-to-one with the original multi-scale fusion feature tensor. The mask ratio is equal to the sum of all mask positions in the mask matrix divided by the total number of patches.

[0080] For each reconstructed feature, a first edge feature vector is obtained through global average pooling. Based on the first edge feature vector, an anomaly score is calculated.

[0081] The dimension of the first edge feature vector is equal to the embedding dimension.

[0082] The anomaly score is equal to the Euclidean distance between the first edge feature vector and the mean vector of the edge baseline features divided by the standard deviation vector of the edge baseline features. The dimensions of the mean vector of the edge baseline features and the standard deviation vector of the edge baseline features are both the same as those of the first edge feature vector.

[0083] The relationship between the anomaly score and the anomaly determination threshold is determined. When the anomaly score is greater than or equal to the anomaly determination threshold, the edge anomaly labeling result is generated as an anomaly; otherwise, the edge anomaly labeling result is generated as normal.

[0084] Multi-scale feature extraction is performed on preprocessed multimodal sensory data in the cloud based on the Res2Net multi-scale convolutional network, and a second multi-scale feature representation is output using the channel-aware scale selection mechanism.

[0085] In this embodiment, the generation of the second multi-scale feature representation includes:

[0086] The data samples are sent to the Res2Net multi-scale convolutional network in the cloud to perform the first stage of convolutional feature extraction and obtain the convolutional feature tensor.

[0087] In a cloud environment, based on the statistical differences of different sensory modal features, the convolutional feature tensor is divided into multi-scale feature groups. For the multi-scale branches in each feature group, convolutional feature transfer and branch-by-branch fusion are performed in sequence to obtain a set of scale features that are adapted to channel perception.

[0088] The multi-scale feature output sets of each sensory modality are concatenated in the channel dimension and compressed by one-to-one convolution to generate a second multi-scale fusion feature tensor.

[0089] The second multi-scale fusion feature tensor is subjected to sensory channel feature grouping weighting processing. The sensory channels are divided into visual, auditory, tactile and balance groups according to the source of sensory modality. The global sensory statistics of each sensory modality group are calculated respectively.

[0090] In Example 1, the second multi-scale fusion feature tensor is subjected to sensory channel feature grouping weighting processing. According to the source of sensory modality, the sensory channels are divided into visual group, auditory group, tactile group and balance group. For each sensory modality group, the feature values ​​of all channels in the sensory modality group at all positions in the spatial dimension are summed and divided by the total number of spatial positions to obtain the global average value of the sensory modality group. The feature values ​​of all channels in the sensory modality group at all positions in the spatial dimension are subtracted from the global average value and squared respectively. The sum of all squared values ​​is divided by the total number of spatial positions and the square root is taken to obtain the global standard deviation of the sensory modality group. The global average value and the global standard deviation are used together as the global sensory statistics of the sensory modality group.

[0091] During the channel perception weight calculation process, the global sensory statistics of each modality group are combined with the decay trend prior vector calculated from the preceding time window to form a fused sensory prior. The fused sensory prior is then weighted and synthesized with the current channel response value to obtain the time-adaptive fused channel response value.

[0092] For the fusion channel response value of each scale branch, cross-scale dynamic normalization is performed, and the sharpness of the normalization is adjusted to generate a set of channel-aware scale selection weight coefficients.

[0093] The output features of each scale branch are weighted and fused according to the set of weight coefficients selected based on the channel perception scale to obtain the second multi-scale feature representation.

[0094] In the cloud, the preprocessed multimodal sensory data is self-supervised pre-trained based on the SimMIM self-supervised mask image modeling framework, and the third self-supervised representation vector is output by adopting a sensory feature-driven mask rate dynamic adjustment mechanism.

[0095] In this embodiment, the generation of the third self-supervised representation vector includes:

[0096] Each multimodal sensory data sample is input into the patch embedding module of the cloud-based SimMIM self-supervised mask image modeling framework. Each multimodal sensory data sample is divided into several patches according to a fixed patch size. Each patch is mapped to a patch embedding vector with a fixed embedding dimension through a linear transformation. All patch embedding vectors are sequentially concatenated to form a patch embedding sequence.

[0097] Based on the sensory feature-driven mask ratio dynamic adjustment mechanism, the patch importance score of each patch embedding vector in the current sensory modality decay assessment task is calculated.

[0098] In Example 1, for each patch embedding vector, based on its sensory modality, the patch embedding vector is fused with prior feature parameters related to the sensory modality. The feature fusion operation includes concatenating or weighted summing the patch embedding vector and the prior feature parameters of the sensory modality on the same dimension. The fused features are input into a fully connected layer and the patch importance score is calculated through an activation function. The patch importance score reflects the discriminative value of the patch in this round of sensory decay assessment. Based on the patch importance scores of all patches, a ranking method is used to obtain a set of patch importance scores. The set of patch importance scores guides the dynamic generation of the mask matrix. Patches with high importance scores are preferentially retained in the masking operation, and patches with low importance scores are preferentially masked, thereby realizing the dynamic adjustment of the masking rate as the sensory modality features and assessment task change. The patch importance score reflects the discriminative value of each patch in the sensory decay assessment.

[0099] A dynamic mask matrix is ​​generated based on the importance scores of all patches;

[0100] Each element in the dynamic mask matrix corresponds one-to-one with the position of each patch embedding vector in the patch embedding sequence. For selected patch positions, the corresponding dynamic mask matrix element is set to be masked, and for unselected patch positions, the corresponding dynamic mask matrix element is set to be unmasked. The mask ratio of the dynamic mask matrix is ​​equal to the number of elements at all masked positions in the dynamic mask matrix divided by the total number of patches.

[0101] The dynamic mask matrix is ​​multiplied element-wise with the patch embedding sequence. The patch embedding vector values ​​at the masked positions are set to zero, while the original values ​​of the patch embedding vectors at the unmasked positions are retained to obtain the mask feature sequence.

[0102] The mask feature sequence is input into the reconstruction decoder, and the patch embedding vectors of all masked positions are recovered using all unmasked patch embedding vectors. The reconstructed feature sequence is then output.

[0103] A global average pooling operation is performed on the reconstructed feature sequence, and the reconstructed embedding vectors at all patch positions are averaged within the batch to obtain the third self-supervised representation vector.

[0104] The edge anomaly labeling results, the second multi-scale feature representation, and the third self-supervised representation vector are input into the cross-modal attention fusion module to generate the fourth fusion feature tensor;

[0105] In this embodiment, the generation of the fourth fusion feature tensor includes:

[0106] Based on edge anomaly labeling results, second multi-scale feature representation, and third self-supervised representation vector, a cross-modal feature heterogeneity measurement mechanism is established in the cross-modal attention fusion module. The cross-modal feature heterogeneity measurement mechanism is used to quantify the information complementarity and difference of different sensory modal features in the current scenario of sensory ability decline in the elderly, and obtain a set of cross-modal heterogeneity quantification factors.

[0107] In Example 1, the cross-modal feature heterogeneity measurement mechanism specifically includes: concatenating the second multi-scale feature representation and the third self-supervised representation vector with the anomaly label embedding vector in the channel dimension to obtain two concatenated feature vectors; calculating the cosine similarity between the feature vectors by concatenating the feature vectors; the cosine similarity measures the complementarity of visual, auditory, tactile and balance sensory modalities in the current sensory decline state of the elderly; and the cosine similarity normalization result constitutes a set of cross-modal heterogeneity quantification factors.

[0108] A dynamic adaptive gating mechanism for modal importance is constructed based on the set of cross-modal heterogeneity quantization factors, and dynamically adjusted modal fusion weight coefficients are assigned to the second multi-scale feature representation and the third self-supervised representation vector, respectively.

[0109] In Example 1, the modal importance dynamic adaptive gating mechanism specifically includes: combining the set of cross-modal heterogeneity quantification factors with the current sensory decline risk level of the elderly, obtaining the initial modal fusion weight coefficient by dynamically predicting the contribution trend of sensory modalities to the current state through a bidirectional long short-term memory network, adaptively adjusting the initial modal fusion weight coefficient through a sigmoid gating function with learnable parameters, and outputting the dynamically adjusted modal fusion weight coefficient. The dynamically adjusted modal fusion weight coefficient reflects the dynamic weight contribution of different sensory modalities in the scenario of sensory ability decline in the elderly.

[0110] A cross-modal multi-head collaborative attention mechanism is established based on dynamically adjusted modal fusion weight coefficients. Attention features are collaboratively optimized across different sensory modalities, and a cross-modal collaborative attention feature vector is output.

[0111] In Example 1, the cross-modal multi-head collaborative attention mechanism specifically includes: using the second multi-scale feature representation as the query object and the third self-supervised representation vector as the key and value object, introducing dynamically adjusted modal fusion weight coefficients in the multi-head attention calculation, and coordinating the information flow direction of different sensory modalities among the multi-head attention heads by adjusting the calculation method of attention weights, thereby calculating the cross-modal collaborative attention feature vector.

[0112] Based on cross-modal collaborative attention feature vectors, a decaying sensory modality compensation and enhancement mechanism is constructed. The feature representation strength of the sensory modality with a high degree of decay is explicitly enhanced through the sensory modality compensation matrix, and the compensated and enhanced cross-modal feature representation is output.

[0113] In Example 1, the sensory modality degradation compensation enhancement mechanism specifically includes: generating a sensory modality compensation matrix using the sensory degradation risk level results output by the pre-assessment module; sensory modalities with high sensory degradation risk levels have larger compensation coefficients in the compensation matrix, while sensory modalities with low degradation risk levels have smaller compensation coefficients; multiplying the cross-modal collaborative attention feature vector element by element with the sensory modality compensation matrix to output the compensated and enhanced cross-modal feature representation.

[0114] The compensated and enhanced cross-modal feature representation is input into the residual feedforward fusion network. Nonlinear mapping is performed on the compensated and enhanced cross-modal feature representation, and a residual connection structure is introduced to stabilize the fusion training, and the fourth fusion feature tensor is calculated.

[0115] In Example 1, the residual connection structure stable fusion training specifically includes: passing the compensated and enhanced cross-modal feature representation through at least two nonlinear fully connected layers, introducing batch normalization operations and activation functions between each nonlinear fully connected layer, performing nonlinear mapping of the fused features, and implementing residual connections with the initial compensated and enhanced cross-modal feature representation after each nonlinear mapping to calculate the fourth fused feature tensor.

[0116] The fourth fusion feature tensor is input into the multi-scale time series modeling module to complete short-term fluctuation modeling and long-term trend modeling, and outputs the fifth time series feature sequence. The fifth time series feature sequence is input into the multi-task evaluation module to obtain the sensory decay trend regression results, sensory decay risk classification results, and sensory pathway priority protection list.

[0117] In this embodiment, the acquisition of sensory decline trend regression results, sensory decline risk classification results, and sensory pathway priority protection list includes:

[0118] In the multi-scale temporal modeling module, the fourth fusion feature tensor is divided into multi-scale time windows in the time dimension. The multi-scale time windows include short-term modeling windows and long-term modeling windows.

[0119] Short-term modeling windows capture the rapid fluctuations in sensory decline in older adults, while long-term modeling windows capture the gradual trend of sensory decline in older adults.

[0120] For each short-term modeling window, the feature sequence within the window is extracted and input into the temporal convolutional network. The temporal convolutional network uses a fixed convolutional kernel width to extract local temporal patterns from the sequence to obtain a set of short-term temporal feature vectors.

[0121] For each long-term modeling window, the feature sequence within the window is extracted and input into the bidirectional long short-term memory network. The bidirectional long short-term memory network models the global time dependency in both the forward and reverse directions and outputs a set of long-term time-series feature vectors.

[0122] The short-term time series feature vector set and the long-term time series feature vector set are concatenated along the feature dimension to form a multi-scale time series feature set, and then input into the cross-scale time series attention mechanism module. The cross-scale time series attention mechanism assigns attention weights to features in different time windows to obtain a multi-scale fused time series feature set.

[0123] The total feature dimension of the multi-scale temporal feature set is equal to the sum of the short-term temporal feature vector set and the long-term temporal feature vector set.

[0124] The multi-scale fused temporal feature set is input into the feedforward fusion layer and mapped to the fifth temporal feature sequence through a fully connected layer and a normalization operation.

[0125] The fifth time-series feature sequence is input into the multi-task evaluation module, which includes shared low-level features and three task prediction heads:

[0126] The sensory decline trend regression task predicts the sensory ability change curve for future time intervals based on the fifth time series feature sequence, and outputs the sensory decline trend prediction results.

[0127] The sensory decline risk classification task outputs sensory decline risk classification results in three levels: high risk, medium risk, and low risk, based on historical and current time series characteristics.

[0128] The task of prioritizing sensory pathway protection outputs a list of sensory pathways for priority protection based on the degradation rate and importance of each sensory modality in the time series.

[0129] In Example 1, the sensory decay trend regression task specifically includes: using the fifth temporal feature sequence as input, further extracting global temporal features through a fully connected network or a temporal recursive network, using regression analysis to predict the sensory ability change values ​​for each sensory modality at several subsequent time steps, forming a sensory ability change curve for the future time interval, with each sampling point of the sensory ability change curve corresponding one-to-one with the predicted time step and sensory modality, and outputting the sensory ability change curve as the sensory decay trend prediction result.

[0130] The sensory decay risk classification task specifically includes: using the fifth time-series feature sequence and historical time-series statistical features as input, aggregating the comprehensive representation of each sensory modality and each time step through a fully connected layer, and using a multi-class classifier to calculate the probability distribution of each sample under three risk categories: high risk, medium risk, and low risk. The risk category corresponding to the maximum value of the probability distribution is selected as the sensory decay risk classification result, and the high-risk, medium-risk, and low-risk results correspond one-to-one with the sample number and sensory modality.

[0131] The sensory pathway priority protection task specifically includes: using the fifth temporal feature sequence to calculate the degradation rate and importance score of each sensory modality in the time dimension. The degradation rate is obtained by calculating the rate of change of sensory modality-related features in the time dimension, and the importance score is obtained by the weighted contribution of sensory modality features in multi-task loss. After weighted sorting of degradation rate and importance score, the sensory modalities with the highest ranking are selected to form a sensory pathway priority protection list. The sensory pathway priority protection list is arranged in order of protection priority, and each protected object corresponds one-to-one with sample number and sensory modality type.

[0132] Based on the regression results of sensory decline trends, the risk classification results of sensory decline, and the priority protection list of sensory pathways, the intervention strategy generation network is invoked and combined with the multi-objective optimization module to generate an intervention matrix that includes intervention scenarios, intervention actions, intervention dosages, and intervention frequencies. The intervention matrix is ​​defined as a multi-dimensional intervention program for the intrinsic abilities of the elderly.

[0133] In this embodiment, the generation of the intervention matrix includes:

[0134] The intervention strategy generation network receives and integrates edge anomaly labeling results, second multi-scale feature representation, third self-supervised representation vector, fourth fusion feature tensor, and fifth temporal feature sequence as input. It combines sensory decline trend regression results, sensory decline risk classification results, and sensory pathway priority protection list with personal health records and historical intervention response records to generate a set of intervention candidate programs.

[0135] When generating a set of intervention candidate schemes, the intervention strategy generation network dynamically adjusts the intervention scenario level based on the edge anomaly labeling results, recommends intervention actions in collaboration with the second multi-scale feature representation and the sensory pathway priority protection list, dynamically optimizes the intervention dosage and frequency based on the third self-supervised representation vector, sensory decline trend regression results, and sensory decline risk classification results, and adaptively screens candidate schemes based on the fourth fusion feature tensor and the fifth temporal feature sequence.

[0136] The set of intervention candidate protocols is input into the multi-objective optimization module. The multi-objective optimization module takes maximizing efficacy, optimizing compliance, and minimizing resource utilization as objective functions. For each candidate protocol, efficacy, compliance, and resource consumption are scored based on full-process characteristics and indicators. The multi-objective optimization module uses a weighted summation method or Pareto optimality method to comprehensively score and screen to obtain the optimal intervention parameter set. The optimal intervention parameter set includes a combination of four dimensions: intervention scenario, intervention action, intervention dose, and intervention frequency.

[0137] An intervention matrix is ​​generated based on the optimal set of intervention parameters. The intervention matrix is ​​defined as a multi-dimensional intervention plan for the intrinsic abilities of the elderly. Each row of the intervention matrix corresponds to a sensory pathway intervention unit, and each column represents four types of parameters: intervention scenario, intervention action, intervention dose, and intervention frequency.

[0138] In this embodiment, the output rules of the intervention matrix include:

[0139] When the sensory decline risk classification result is high risk and the sensory ability change curve shows a downward trend greater than the threshold, for the sensory modality that is given priority protection, the intervention scenario parameters in the intervention matrix are set to an enhanced training environment based on the edge anomaly labeling results. The intervention action parameters are selected as high-intensity multi-sensory stimulation with reference to the second multi-scale feature representation and the priority protection list. The intervention dose parameters are increased to twice a day, each lasting twenty minutes, based on the third self-supervised representation vector and the decline trend. The intervention frequency is set to be implemented continuously for four weeks.

[0140] When the sensory decline risk classification result is medium risk and the sensory ability change curve shows a slow decline less than the threshold, for the sensory modality that is prioritized for protection, the intervention scenario parameters in the intervention matrix are set to the standard training environment with reference to the edge anomaly labeling results, the intervention action parameters are selected as a medium-intensity sensory activation task with reference to the second multi-scale feature representation and the priority protection list, the intervention dose parameter is once a day, each lasting fifteen minutes, and the intervention frequency is set to be performed every other day for four weeks.

[0141] When the risk classification of sensory decline is low and the sensory ability change curve shows a stable or fluctuating trend, for the sensory modality that is given priority protection, the intervention scenario parameter in the intervention matrix is ​​set to the daily health care environment, the intervention action parameter is selected as a low-intensity sensory activation task, the intervention dosage parameter is twice a week, each lasting ten minutes, and the intervention frequency is set to be implemented for eight weeks.

[0142] Example 2:

[0143] During a certain implementation period, the system continuously collects multimodal sensory data from a group of elderly users. The collected raw data includes retinal structure images, high-frequency speech signals, finger vibration response sequences, and gait balance pressure distribution. A total of 12,000 raw data samples were collected. The system first performs time synchronization, format normalization, and noise suppression on each piece of raw data to obtain a multimodal sensory dataset that can be used for analysis.

[0144] The preprocessed multimodal sensory data is input into the lightweight Res2Net-SimMIM sub-model at the edge. Taking user number "U0012" as an example, the average edge anomaly score of this user in the past two weeks is 2.7, which exceeds the anomaly judgment threshold of 2.2 for three consecutive days. The system generates an edge anomaly marker and automatically uploads the multimodal synchronous metadata and anomaly marker results of user number U0012 to the cloud platform.

[0145] After receiving the uploaded content from user U0012, the cloud platform uses a Res2Net multi-scale convolutional network to perform feature grouping and channel-aware scale selection processing on the user's multimodal data, including retinal images, auditory spectra, and tactile signals, to generate a second multi-scale feature representation. Taking the tensor "F_multi2_0012" as an example, this tensor shows fine-grained anomalous distributions in the retinal microvascular region, the high-frequency band of the auditory spectrum, and the fingertip tactile peak region.

[0146] The system uses the SimMIM mask self-supervised algorithm on the same batch of data. Through the dynamic adjustment of the mask rate driven by sensory features, the importance score of each patch feature is calculated, which masks 38% of the low importance area. The third self-supervised representation vector "V_sim_0012" is output by the reconstruction decoder, which reflects that the user's auditory sensory mask reconstruction error is 2.3 times higher than the group mean, indicating the risk of early high-frequency hearing loss.

[0147] The platform inputs the edge anomaly labeling results, second multi-scale feature representation, and third self-supervised representation vector of user U0012 into the cross-modal attention fusion module. The system detects high complementarity between the user's auditory and tactile modalities through a cross-modal heterogeneity measurement mechanism, adjusting the modal importance gating coefficients to 0.43 for auditory, 0.37 for tactile, 0.15 for visual, and 0.05 for balance. The multi-head collaborative attention mechanism outputs the cross-modal feature tensor "F_fuse4_0012".

[0148] In the temporal modeling phase, the system uses F_fuse4_0012 as input, setting the time window to 7 days for the short term and 30 days for the long term. For user U0012, an increase in the fluctuation range of auditory feature values ​​was detected in the short term, while the long-term trend was a continuous decrease. The short-term convolutional feature D_short_0012 and the long-term bidirectional LSTM feature D_long_0012 are concatenated and then output as a fused feature V_multiT_0012 through a cross-scale temporal attention mechanism. This is further mapped to the fifth temporal feature sequence V_seq5_0012 through a feedforward fusion layer.

[0149] The multi-task assessment module performs a multi-dimensional sensory decline trend regression analysis on user U0012 based on V_seq5_0012, outputting that the maximum rate of decline in hearing ability over the next two months is -3.8 dB / month, classifying it as high-risk; the tactile change trend is stable, classifying it as low-risk. The pathway priority protection list automatically recommends hearing, followed by vision and touch.

[0150] Subsequently, the platform inputs the evaluation results of user U0012 into the intervention strategy generation network. Targeting the auditory pathway, the system generates candidate intervention actions based on anomaly markers, feature trends, and priorities, including "high-frequency speech training," "musical sound recognition training," and "scenario-simulated dialogue training." The multi-objective optimization module calculates efficacy scores (e.g., expected hearing threshold improvement), compliance scores (based on past participation rates and user behavior models), and resource consumption scores (training duration and system load) for each plan. The optimal intervention parameter set is determined as follows: "Training scenario: home smart speaker," "Training twice daily, 20 minutes each time," and "Course content: high-frequency phrase repetition and high / low frequency switching auditory recognition." The generated intervention matrix is ​​shown in Table 1 below.

[0151] Table 1 Intervention Matrix for User U0012

[0152]

[0153]

[0154] During the intervention period of the method of this invention, the compliance rate of user U0012 was 93%, and the hearing ability improved by 6.3dB after two months. No high-risk events occurred. In contrast, user U0059 in the same batch of traditional template group, who used hearing exercises once a week for 10 minutes each time, had a compliance rate of 71%, and the hearing ability improved by only 2.1dB after two months. One high-risk event (sudden interruption of communication) occurred.

[0155] Further analysis of the training data from 120 user samples yielded the following comparative data (average) in Table 2:

[0156] Table 2 Comparison data between traditional methods and the method of this invention

[0157]

[0158] Meanwhile, in long-term trend prediction, the system's mean squared error in predicting sensory ability decline is 0.021 lower than that of traditional methods. Intervention matrix allocation results show that 90% of users receive a one-size-fits-all training scheme using traditional methods, while 96% of users receive a personalized parameter matrix using the method of this invention, with the prioritized protection pathways highly matched to actual risks.

[0159] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for constructing a multi-dimensional intervention program for the intrinsic abilities of the elderly based on neural networks, characterized in that, include: Multimodal sensory data of elderly people were collected and preprocessed to obtain preprocessed multimodal sensory data; Based on the lightweight Res2Net-SimMIM sub-model, the preprocessed multimodal sensory data is subjected to initial feature screening to generate the first edge feature vector, and decay anomaly detection is performed to generate edge anomaly labeling results and upload them. Multi-scale feature extraction is performed on preprocessed multimodal sensory data in the cloud based on the Res2Net multi-scale convolutional network, and a second multi-scale feature representation is output using the channel-aware scale selection mechanism. In the cloud, the preprocessed multimodal sensory data is self-supervised pre-trained based on the SimMIM self-supervised mask image modeling framework, and the third self-supervised representation vector is output. The edge anomaly labeling results, the second multi-scale feature representation, and the third self-supervised representation vector are input into the cross-modal attention fusion module to generate the fourth fusion feature tensor; The fourth fusion feature tensor is input into the multi-scale time series modeling module to complete short-term fluctuation modeling and long-term trend modeling, and outputs the fifth time series feature sequence. The fifth time series feature sequence is input into the multi-task evaluation module to obtain the sensory decay trend regression results, sensory decay risk classification results, and sensory pathway priority protection list. Based on the regression results of sensory decline trends, the risk classification results of sensory decline, and the priority protection list of sensory pathways, an intervention strategy generation network is invoked and combined with a multi-objective optimization module to generate an intervention matrix. The intervention matrix is ​​defined as a multi-dimensional intervention program for the intrinsic abilities of the elderly.

2. The method for constructing a multi-dimensional intervention program for the intrinsic abilities of the elderly based on neural networks according to claim 1, characterized in that, The construction of the preprocessed multimodal sensory data includes: Multimodal sensory data of elderly people were collected, including visual, auditory, tactile, and balance sensory data. Time synchronization, format uniformity, and noise suppression were performed at the edge to obtain preprocessed multimodal sensory data.

3. The method for constructing a multi-dimensional intervention program for the intrinsic abilities of the elderly based on neural networks according to claim 1, characterized in that, The generation of the edge anomaly marking results includes: S21. Define the preprocessed multimodal sensory data as a set of data samples; Each data sample includes spatial height, spatial width, and number of channels, and all data samples constitute a batch input. The depthwise separable convolutional Patch-Embed module is used to divide each data sample into several patches according to a fixed patch size, and the initial patch feature sequence is obtained for each patch through linear mapping; Each initial patch feature sequence is input into a lightweight Res2Net feature extraction unit, which divides the initial patch feature sequence into multiple scale sub-branches. For the first scale sub-branch, the output is directly used as the output feature of the first scale branch. For each scale sub-branch larger than the first one, a depthwise separable convolution operation is first performed on the output feature of the previous scale branch, and then added to the current scale sub-branch to obtain the output feature of the current scale branch. The output features of all scale branches are concatenated and subjected to a one-to-one convolution operation to obtain a multi-scale fused feature tensor. SimMIM self-supervised mask is applied to each multi-scale fusion feature tensor to generate a mask matrix. The mask matrix is ​​then multiplied element-wise with the multi-scale fusion feature tensor to obtain the mask feature tensor. The mask feature tensor is then input into the reconstruction decoder to output the reconstructed features. For each reconstructed feature, a first edge feature vector is obtained through global average pooling. Based on the first edge feature vector, an anomaly score is calculated. The relationship between the anomaly score and the anomaly determination threshold is determined. When the anomaly score is greater than or equal to the anomaly determination threshold, the edge anomaly labeling result is generated as an anomaly; otherwise, the edge anomaly labeling result is generated as normal.

4. The method for constructing a multi-dimensional intervention program for the intrinsic abilities of the elderly based on neural networks according to claim 1, characterized in that, The generation of the second multi-scale feature representation includes: The data samples are sent to the Res2Net multi-scale convolutional network in the cloud to perform the first stage of convolutional feature extraction and obtain the convolutional feature tensor. In a cloud environment, based on the statistical differences of different sensory modal features, the convolutional feature tensor is divided into multi-scale feature groups. For the multi-scale branches in each feature group, convolutional feature transfer and branch-by-branch fusion are performed in sequence to obtain a set of scale features that are adapted to channel perception. The multi-scale feature output sets of each sensory modality are concatenated in the channel dimension and compressed by one-to-one convolution to generate a second multi-scale fusion feature tensor. The second multi-scale fusion feature tensor is subjected to sensory channel feature grouping weighting processing. The sensory channels are divided into visual, auditory, tactile and balance groups according to the source of sensory modality. The global sensory statistics of each sensory modality group are calculated respectively. During the channel perception weight calculation process, the global sensory statistics of each modality group are combined with the decay trend prior vector calculated from the preceding time window to form a fused sensory prior. The fused sensory prior is then weighted and synthesized with the current channel response value to obtain the time-adaptive fused channel response value. For the fusion channel response value of each scale branch, cross-scale dynamic normalization is performed, and the sharpness of the normalization is adjusted to generate a set of channel-aware scale selection weight coefficients. The output features of each scale branch are weighted and fused according to the set of weight coefficients selected based on the channel perception scale to obtain the second multi-scale feature representation.

5. The method for constructing a multi-dimensional intervention program for the intrinsic abilities of the elderly based on neural networks according to claim 1, characterized in that, The generation of the third self-supervised representation vector includes: Each multimodal sensory data sample is input into the patch embedding module of the cloud-based SimMIM self-supervised mask image modeling framework. Each multimodal sensory data sample is divided into several patches according to a fixed patch size. Each patch is mapped to a patch embedding vector with a fixed embedding dimension through a linear transformation. All patch embedding vectors are sequentially concatenated to form a patch embedding sequence. Based on the sensory feature-driven mask ratio dynamic adjustment mechanism, the patch importance score of each patch embedding vector in the current sensory modality decay assessment task is calculated. A dynamic mask matrix is ​​generated based on the importance scores of all patches; The dynamic mask matrix is ​​multiplied element-wise with the patch embedding sequence. The patch embedding vector values ​​at the masked positions are set to zero, while the original values ​​of the patch embedding vectors at the unmasked positions are retained to obtain the mask feature sequence. The mask feature sequence is input into the reconstruction decoder, and the patch embedding vectors of all masked positions are recovered using all unmasked patch embedding vectors. The reconstructed feature sequence is then output. A global average pooling operation is performed on the reconstructed feature sequence, and the reconstructed embedding vectors at all patch positions are averaged within the batch to obtain the third self-supervised representation vector.

6. The method for constructing a multi-dimensional intervention program for the intrinsic abilities of the elderly based on neural networks according to claim 1, characterized in that, The generation of the fourth fusion feature tensor includes: Based on edge anomaly labeling results, second multi-scale feature representation, and third self-supervised representation vector, a cross-modal feature heterogeneity measurement mechanism is established in the cross-modal attention fusion module. The cross-modal feature heterogeneity measurement mechanism is used to quantify the information complementarity and difference of different sensory modal features in the current scenario of sensory ability decline in the elderly, and obtain a set of cross-modal heterogeneity quantification factors. A dynamic adaptive gating mechanism for modal importance is constructed based on the set of cross-modal heterogeneity quantization factors, and dynamically adjusted modal fusion weight coefficients are assigned to the second multi-scale feature representation and the third self-supervised representation vector, respectively. A cross-modal multi-head collaborative attention mechanism is established based on dynamically adjusted modal fusion weight coefficients. Attention features are collaboratively optimized across different sensory modalities, and a cross-modal collaborative attention feature vector is output. Based on cross-modal collaborative attention feature vectors, a decaying sensory modality compensation and enhancement mechanism is constructed. The feature representation strength of the sensory modality with a high degree of decay is explicitly enhanced through the sensory modality compensation matrix, and the compensated and enhanced cross-modal feature representation is output. The compensated and enhanced cross-modal feature representation is input into the residual feedforward fusion network. Nonlinear mapping is performed on the compensated and enhanced cross-modal feature representation, and a residual connection structure is introduced to stabilize the fusion training, and the fourth fusion feature tensor is calculated.

7. The method for constructing a multi-dimensional intervention program for the intrinsic abilities of the elderly based on neural networks according to claim 1, characterized in that, The acquisition of the sensory decline trend regression results, sensory decline risk classification results, and sensory pathway priority protection list includes: In the multi-scale temporal modeling module, the fourth fusion feature tensor is divided into multi-scale time windows in the time dimension. The multi-scale time windows include short-term modeling windows and long-term modeling windows. For each short-term modeling window, the feature sequence within the window is extracted and input into the temporal convolutional network. The temporal convolutional network uses a fixed convolutional kernel width to extract local temporal patterns from the sequence to obtain a set of short-term temporal feature vectors. For each long-term modeling window, the feature sequence within the window is extracted and input into the bidirectional long short-term memory network. The bidirectional long short-term memory network models the global time dependency in both the forward and reverse directions and outputs a set of long-term time-series feature vectors. The short-term time series feature vector set and the long-term time series feature vector set are concatenated along the feature dimension to form a multi-scale time series feature set, and then input into the cross-scale time series attention mechanism module. The cross-scale time series attention mechanism assigns attention weights to features in different time windows to obtain a multi-scale fused time series feature set. The multi-scale fused temporal feature set is input into the feedforward fusion layer and mapped to the fifth temporal feature sequence through a fully connected layer and normalization operation. The fifth time-series feature sequence is input into the multi-task evaluation module, which includes shared low-level features and three task prediction heads: The sensory decline trend regression task predicts the change curve of sensory ability in future time intervals based on the fifth time series feature sequence, and outputs the sensory decline trend prediction results.

8. The method for constructing a multi-dimensional intervention program for the intrinsic abilities of the elderly based on neural networks according to claim 1, characterized in that, The generation of the intervention matrix includes: The intervention strategy generation network receives and integrates edge anomaly labeling results, second multi-scale feature representation, third self-supervised representation vector, fourth fusion feature tensor, and fifth temporal feature sequence as input. It combines sensory decline trend regression results, sensory decline risk classification results, and sensory pathway priority protection list with personal health records and historical intervention response records to generate a set of intervention candidate programs. The set of intervention candidate protocols is input into the multi-objective optimization module. The multi-objective optimization module takes maximizing efficacy, optimizing compliance, and minimizing resource utilization as objective functions. For each candidate protocol, efficacy, compliance, and resource consumption are scored based on full-process characteristics and indicators. The multi-objective optimization module uses a weighted summation method or Pareto optimality method to comprehensively score and screen to obtain the optimal intervention parameter set. The optimal intervention parameter set includes a combination of four dimensions: intervention scenario, intervention action, intervention dose, and intervention frequency. An intervention matrix is ​​generated based on the optimal set of intervention parameters. The intervention matrix is ​​defined as a multi-dimensional intervention plan for the intrinsic abilities of the elderly. Each row of the intervention matrix corresponds to a sensory pathway intervention unit, and each column represents four types of parameters: intervention scenario, intervention action, intervention dose, and intervention frequency.

9. The method for constructing a multi-dimensional intervention program for the intrinsic abilities of the elderly based on neural networks according to claim 1, characterized in that, The output rules of the intervention matrix include: When the sensory decline risk classification result is high risk and the sensory ability change curve shows a downward trend greater than the threshold, for the sensory modality that is given priority protection, the intervention scenario parameters in the intervention matrix are set to an enhanced training environment based on the edge anomaly labeling results. The intervention action parameters are selected as high-intensity multi-sensory stimulation with reference to the second multi-scale feature representation and the priority protection list. The intervention dose parameters are increased to twice a day, each lasting twenty minutes, based on the third self-supervised representation vector and the decline trend. The intervention frequency is set to be implemented continuously for four weeks. When the sensory decline risk classification result is medium risk and the sensory ability change curve shows a slow decline less than the threshold, for the sensory modality that is prioritized for protection, the intervention scenario parameters in the intervention matrix are set to the standard training environment with reference to the edge anomaly labeling results, the intervention action parameters are selected as a medium-intensity sensory activation task with reference to the second multi-scale feature representation and the priority protection list, the intervention dose parameter is once a day, each lasting fifteen minutes, and the intervention frequency is set to be performed every other day for four weeks. When the risk classification of sensory decline is low and the sensory ability change curve shows a stable or fluctuating trend, for the sensory modality that is given priority protection, the intervention scenario parameter in the intervention matrix is ​​set to the daily health care environment, the intervention action parameter is selected as a low-intensity sensory activation task, the intervention dosage parameter is twice a week, each lasting ten minutes, and the intervention frequency is set to be implemented for eight weeks.

Images