Chronic overuse risk prediction system and method based on dual-mode progressive assessment

Abstract

The application discloses a chronic strain risk prediction system and method based on a dual-mode progressive evaluation, comprising: a data acquisition module for collecting open-eye resting-state electroencephalogram signals and closed-eye resting-state electroencephalogram signals of a subject, and collecting task-state multi-modal signals when a first risk level or a first risk score reaches a preset transfer-in condition; a quality control and alignment module for pre-processing the open-eye resting-state electroencephalogram signals and the closed-eye resting-state electroencephalogram signals to obtain first evaluation mode features; and pre-processing the task-state multi-modal signals to obtain second evaluation mode features; a first evaluation mode screening module for generating a first risk score and a first risk level based on the first evaluation mode features; a second evaluation mode evaluation module for generating second evaluation mode features based on the second evaluation mode features; and a cross-mode fusion module for projecting the first evaluation mode features and the second evaluation mode features to a common space to generate a cross-mode consistency score and a comprehensive risk result.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of artificial intelligence, biomedical engineering, rehabilitation engineering, wearable health monitoring, and chronic musculoskeletal injury risk assessment, and more specifically to a chronic strain risk prediction system and method based on dual-mode progressive assessment. Background Technology

[0002] Chronic musculoskeletal strain is widely seen in the neck, shoulders, lower back, and knees, and is usually associated with repetitive movements, poor posture, cumulative local load, pain avoidance behaviors, and sensorimotor modulation abnormalities. In addition to subjective pain and functional impairment, these individuals often experience objective physiological changes such as alterations in central nervous system activity, abnormal muscle coordination, abnormal autonomic nervous system responses, and decreased quality of motor execution. Therefore, objectively and comprehensively assessing these physiological changes is crucial for early screening and precise intervention.

[0003] Currently, assessment techniques for chronic strain injuries and pain mainly focus on the following areas: scales and questionnaires, monomodal physiological signal recognition schemes, and kinematic or wearable kinesiological assessment schemes. Although these technologies have made some progress in their respective fields, significant technical bottlenecks remain when applied to large-scale screening and accurate diagnosis of chronic strain injuries:

[0004] 1) The lack of a "screening-refinement" tiered evaluation mechanism makes it difficult to balance efficiency and accuracy: Current technologies typically employ a "one-size-fits-all" assessment approach. They either rely on simple scales or single physiological indicators for rapid screening, leading to high false negative rates and poor objectivity; or they directly employ complex multimodal laboratory-grade tests, which are expensive, cumbersome, and time-consuming, making them difficult to implement in primary healthcare institutions or large-scale population screenings. Current technologies lack a progressive assessment framework that can dynamically adjust the depth of assessment based on initial screening results.

[0005] 2) Insufficient mining of resting-state central features and lack of state comparison mechanisms: The brains of patients with chronic strain often exhibit specific functional connectivity abnormalities at rest. However, current technologies mostly focus on signal changes during task states or only collect resting EEG data in a single state. The lack of a comparative analysis mechanism based on the differences between "eyes-open resting" and "eyes-closed resting" makes it difficult to effectively characterize the differences in central regulatory capacity between states of blocked and open sensory input, thus failing to accurately capture the central sensitization characteristics of chronic pain.

[0006] 3) Spatiotemporal alignment of multimodal data is difficult, and there is a lack of a unified trial re-slicing mechanism: In assessments involving motor tasks, signals such as electroencephalograms (EEG), electromyograms (EMG), and accelerometers (ACC) differ in sampling rates, data structures, and physical meanings. Existing technologies struggle to handle temporal synchronization between different modalities, particularly lacking a mechanism for unified reslicing and trial alignment of multimodal data around "action initiation and cessation events." This makes it difficult to establish precise causal relationships between central commands (EEG) and peripheral execution (EMG / ACC) during fusion analysis.

[0007] 4) Lack of comprehensive consistency analysis across patterns, resulting in a single basis for decision-making: Existing multimodal fusion methods often involve simple feature stitching, lacking in-depth analysis of consistency between different modalities. When a modal signal is disturbed (such as by motion artifacts) or is partially missing, existing systems struggle to ensure the robustness of the evaluation results through cross-modal consistency checks. Summary of the Invention

[0008] In view of the above problems, the present invention is proposed to provide a chronic strain risk prediction system and method based on dual-mode progressive assessment to overcome or at least partially solve the above problems. The system completes rapid screening in a low-burden resting state through a first assessment mode, completes detailed assessment in a task state through a second assessment mode, and outputs comprehensive risk results through cross-mode consistency analysis, thereby balancing the efficiency of early screening and the accuracy of detailed assessment.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, embodiments of the present invention provide a chronic strain injury risk prediction system based on dual-mode progressive assessment, comprising: The data acquisition module is used to collect the subject's open-eye resting-state EEG signals and closed-eye resting-state EEG signals, as well as task-state multimodal signals when the first risk level or first risk score reaches the preset transition conditions. The quality control and alignment module is used to preprocess the open-eye resting-state EEG signals and the closed-eye resting-state EEG signals to obtain the first evaluation mode characterization; and to preprocess the task-state multimodal signals to obtain the second evaluation mode features. The first assessment mode screening module is used to generate a first risk score and a first risk level based on the first assessment mode representation. The second evaluation mode evaluation module is used to generate a second evaluation mode representation based on the characteristics of the second evaluation mode. The cross-modal fusion module is used to project the representations of the first and second evaluation modalities into a common space to generate a cross-modal consistency score and a comprehensive risk result.

[0010] Preferably, the task-mode multimodal signal includes: Acquire task-state EEG signals and corresponding stimulation intensity calibration information under electrical stimulation tasks; Acquire task-state EEG signals, task-state photoplethysmography / skin conductance signals, task-state electromyography signals, and motion acceleration signals under motion-evoked tasks.

[0011] Preferably, the quality control and alignment module is specifically used for: Filtering, artifact removal, rereference, resampling, and normalization are performed on open-eye resting-state EEG signals, closed-eye resting-state EEG signals, and / or task-mode multimodal signals; Task-state EEG signals under electrical stimulation tasks and corresponding stimulation intensity calibration information are extracted according to the stimulation event and the corresponding time period. Task-state EEG signals, task-state photoplethysmography / skin conductance signals, task-state electromyography signals, and motor acceleration signals under motion-evoked tasks were determined according to the action initiation event and the action termination event, and then sliced ​​according to the uniform trial boundary. The trial structures under different part protocols are uniformly aligned using valid trial markers; Missing trials, trials of abnormal length, or partially missing modes are padded or masked.

[0012] Preferably, the first evaluation mode screening module is specifically used for: Feature extraction was performed on the resting-state EEG signals with eyes open and resting-state EEG signals with eyes closed to obtain resting-state EEG features with eyes open and resting-state EEG features with eyes closed. Construct state difference features between open-eye resting EEG features and closed-eye resting EEG features, and calculate resting-state functional connectivity features; The state difference features and resting state functional connectivity features are fused to generate the first risk score and the first risk level.

[0013] Preferably, the second evaluation mode evaluation module is specifically used for: Task-state EEG features were obtained by extracting features from the preprocessed electrical stimulation task; stimulation intensity calibration information was used as calibration features and cross-fused with the task-state EEG features to obtain a pain task representation and output an electrical stimulation pain score. Multimodal trial-secondary features were constructed from preprocessed task-state EEG, task-state photoplethysmography / conductivity-of-skin (TES), task-state EMG, and motor acceleration signals under a motion-evoked task. Adaptive modal alignment was performed on these multimodal trial-secondary features to obtain multimodal weights. These weights were then weighted, concatenated, and fused to obtain a trial-secondary fusion representation. The motion-evoked pain load assessment result was obtained through this trial-secondary fusion representation. The second assessment modality was derived from the results of exercise-induced pain load assessment and pain task representation.

[0014] Preferably, the cross-mode fusion module is specifically used for: Project the first evaluation model representation and the second evaluation model representation onto a common feature space; Calculate the Euclidean distance between the projected bimodal representations and generate a cross-modal consistency score based on the negative exponential form of the Euclidean distance. Based on the cross-modal consistency score and the interactive features of the bimodal representation, a comprehensive risk result is generated.

[0015] Preferably, it also includes an output module for outputting cross-modal consistency scores and overall risk results.

[0016] Preferably, the first evaluation mode screening module is further used for: The central feature is obtained by concatenating the mean pooling features of the open-eye resting EEG token sequence, the mean pooling features of the closed-eye resting EEG token sequence, the resting-state functional connectivity features, and the state difference features, and then performing linear mapping, GELU activation, and layer normalization. After concatenating the central features, state difference features, resting state functional connectivity features, strain location features and demographic features, the data are input into the shared feature extraction head. The shared representation is obtained after two layers of linear mapping, GELU activation, random deactivation and layer normalization. The shared representation outputs a pain sensitivity score through a linear layer and a sigmoid activation function.

[0017] Secondly, embodiments of the present invention provide a method for predicting the risk of chronic strain injury based on a dual-mode progressive assessment, including: The subjects' resting-state EEG signals with eyes open and resting-state EEG signals with eyes closed were collected; Preprocessing of open-eye resting-state EEG signals and closed-eye resting-state EEG signals yielded the first assessment mode representation; The first risk score and the first risk level are generated based on the first assessment mode; when the first risk level or the first risk score reaches the preset transition condition, the acquisition task state multimodal signal is triggered. The task-state multimodal signal is preprocessed to obtain the second evaluation mode features; Generate a representation of the second evaluation model based on its characteristics; The first and second assessment model representations are projected into a common space to generate cross-modal consistency scores and comprehensive risk results.

[0018] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a chronic strain injury risk prediction system and method based on dual-mode progressive assessment, which has the following advantages: A dual-mode progressive assessment approach is adopted, which allows for low-burden initial screening via resting-state EEG in primary care settings and further task-based refined assessment in hospital settings.

[0019] Explicitly utilizing the state difference characteristics between open-eye rest and closed-eye rest can enhance the identification of abnormal baseline regulatory capacity in individuals with chronic strain.

[0020] Incorporating EEG information from electrical stimulation tasks and multimodal information from motor evoked tasks into a unified assessment framework is beneficial for simultaneously characterizing sensory and motor functions.

[0021] Performing multimodal trial re-cutting and alignment based on action events is beneficial for improving the fusion stability under different part protocols.

[0022] Adding cross-modal consistency analysis helps to generate more structured and comprehensive risk outputs and supports referral or additional testing decisions. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of a chronic strain injury risk prediction system based on dual-mode progressive assessment provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the processing flow of the first evaluation mode screening module provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the processing flow of the evaluation module for the second evaluation mode provided in this embodiment of the invention; Figure 4 This is a schematic diagram of the cross-mode fusion module processing flow provided in an embodiment of the present invention; Figure 5 This is a detail diagram of action event-driven alignment provided in an embodiment of the present invention; Figure 6 This is a flowchart of the chronic strain injury risk prediction method based on dual-mode progressive assessment provided in this embodiment of the invention. Figure 7 This is a diagram of the multi-scale temporal coding and state comparison structure provided in this embodiment of the invention; Figure 8 An overall structural diagram is generated to characterize the second evaluation mode provided in the embodiments of the present invention. Detailed Implementation

[0025] 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 some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] This invention discloses a chronic strain injury risk prediction system based on a dual-mode progressive assessment, such as... Figure 1 As shown, it includes: The data acquisition module is used to collect the subject's open-eye resting-state EEG signals and closed-eye resting-state EEG signals, as well as task-state multimodal signals when the first risk level or first risk score reaches the preset transition conditions. The quality control and alignment module is used to preprocess the open-eye resting-state EEG signals and the closed-eye resting-state EEG signals to obtain the first evaluation mode characterization; and to preprocess the task-state multimodal signals to obtain the second evaluation mode features. The first assessment mode screening module is used to generate a first risk score and a first risk level based on the first assessment mode representation. The second evaluation mode evaluation module is used to generate a second evaluation mode representation based on the characteristics of the second evaluation mode. The cross-modal fusion module is used to project the representations of the first and second evaluation modalities into a common space to generate a cross-modal consistency score and a comprehensive risk result.

[0027] The two assessment modes of this invention are conditionally triggered progressive relationships—the first assessment mode outputs the risk level, and the second assessment mode is only activated when the level reaches a preset threshold. This can reduce the data collection burden and quickly provide the results of chronic strain risk, while improving the assessment accuracy. The system of this invention can be used as a control and judgment strategy for chronic strain risk assessment and can be applied to scenarios such as community screening, physical examinations in elderly care institutions, initial screening in primary clinics, detailed assessments in rehabilitation centers, occupational strain risk monitoring, and pain management assistance.

[0028] In a preferred embodiment, such as Figure 5 As shown, the quality control and alignment module is specifically used for: Filtering, artifact removal, rereference, resampling, and normalization are performed on open-eye resting-state EEG signals, closed-eye resting-state EEG signals, and / or task-mode multimodal signals; Task-state EEG signals under electrical stimulation tasks and corresponding stimulation intensity calibration information are extracted according to the stimulation event and the corresponding time period. Task-state EEG signals, task-state photoplethysmography / skin conductance signals, task-state electromyography signals, and motor acceleration signals under motion-evoked tasks were determined according to the action initiation event and the action termination event, and then sliced ​​according to the uniform trial boundary. The trial structures under different part protocols are uniformly aligned using valid trial markers; Missing trials, trials of abnormal length, or partially missing modes are padded or masked.

[0029] In a preferred embodiment, such as Figure 2 As shown, the first evaluation mode screening module is specifically used for: Feature extraction was performed on the preprocessed open-eye resting-state EEG signal and closed-eye resting-state EEG signal to obtain open-eye resting-state EEG features and closed-eye resting-state EEG features. Specifically, the time-domain features, frequency-domain features or time-frequency combined features of the open-eye resting-state EEG signal and the time-domain features, frequency-domain features or time-frequency combined features of the closed-eye resting-state EEG signal were extracted. State difference features were constructed between open-eye resting EEG features and closed-eye resting EEG features, and resting-state functional connectivity features were calculated to enhance the characterization of resting regulation capacity and central features related to chronic pain. The state difference features and resting-state functional connectivity features are fused to generate a first risk score and a first risk level. Furthermore, the first assessment mode screening module also outputs a pain sensitivity score based on the fused shared representation (which is obtained by concatenating central features (mean pooling features of open-eye resting EEG token sequences, mean pooling features of closed-eye resting EEG token sequences, resting-state functional connectivity features, and state difference features, followed by linear mapping, GELU activation, and layer normalization from the central feature projection module), state difference features, resting-state functional connectivity features, strain site features (mapped from strain site categories registered before screening to category identifiers, followed by a trainable embedding layer), and demographic features (normalized by dividing the registered age by 100 and combining it with gender codes to form a two-dimensional vector, followed by linear mapping, GELU activation, and layer normalization), inputted into the shared feature extraction head, and processed through two layers of linear mapping, GELU activation, random inactivation, and layer normalization to obtain a 96-dimensional feature). This is achieved through a linear layer and a Sigmoid activation function. The supervision label for this score is derived from the reciprocal of the mean of the low, medium, and high stimulation perception thresholds measured by the subjects during the electrical stimulation calibration phase. This value is then normalized to the [0, 1] interval using min-max normalization, with higher values ​​indicating stronger pain sensitivity. The pain sensitivity score can serve as a reference baseline for the electrical stimulation task in the subsequent second assessment mode. Specifically, the central feature projection module maps the 384-dimensional input formed by concatenating four 96-dimensional features to a 96-dimensional central feature; the strain site category is embedded to obtain 16-dimensional site features, and the 2-dimensional demographic vector composed of age and gender is mapped to 16-dimensional demographic features. The input dimension of the shared feature extraction head is 320, which is the dimension obtained by concatenating three 96-dimensional features and two 16-dimensional conditional features, and is subsequently mapped to 192-dimensional intermediate features and 96-dimensional shared representations, with a random inactivation probability of 0.2. Both the first risk score output head and the pain sensitivity output head map the 96-dimensional shared representations to a continuous score, while the first risk level output head maps the 96-dimensional shared representations to classification scores for three risk levels. The 100 in the age divided by 100 is the age normalization scale, used to scale the age value to a range similar to other demographic inputs.

[0030] In terms of network implementation, the open-eye resting-state branch and the closed-eye resting-state branch can respectively employ convolutional neural networks, temporal convolutional networks, recurrent neural networks, Transformer encoders, or combinations thereof for feature extraction. More specifically, a multi-scale temporal encoder is used for feature extraction of both open-eye resting-state EEG signals and closed-eye resting-state EEG signals. The multi-scale temporal encoder contains three parallel branches with different convolutional kernel scales, capturing short-term local patterns, medium-time-scale waveform envelopes, and long-term slow oscillation trends, respectively. The outputs of each branch are concatenated and then adaptively weighted through a channel compression-excited attention module, followed by a temporal self-attention layer to establish global temporal dependencies. The first evaluation mode representation further includes: resting-state functional connectivity difference features calculated based on the difference in the upper triangular region of the correlation matrix between open-eye and closed-eye EEG channels, and state difference features constructed based on an attention-weighted state comparison module. The state comparison module applies multi-head self-attention to the difference temporal sequences to learn key difference time points, introduces cross-state cross-attention to capture nonlinear relationships between open and closed eyes, and fuses it with a simple mean difference three-way method. The complete network structure of the multi-scale temporal encoder is as follows: the stem projection layer uses a one-dimensional convolution with a kernel size of 5 to map the input channels to the hidden dimension; the three parallel convolutional kernels of the EEG encoding branch have sizes of 9, 25, and 51, the EMG encoding branch has sizes of 7, 15, and 31, and the PPG / GSR encoding branch has sizes of 5, 11, and 21; each branch also includes downsampling convolutions with a stride of 2 and adaptive average pooling to a uniform token length; the channel compression excitation module generates channel gating coefficients through global average pooling and two fully connected layers; the temporal self-attention layer uses a 4-head TransformerEncoderLayer, with a feedforward dimension twice the hidden dimension. The attention-weighted state comparison module includes a 4-head differential sequence self-attention layer, a 4-head cross-state attention layer, and a fusion network consisting of three concatenated paths that undergo two linear transformations, as shown in the specific structure below. Figure 7 As shown. Figure 7 In the multi-scale temporal encoder, the encoding branches for EEG, EMG, PPG / GSR, and motion acceleration (ACC) are defined. Their common structure includes trunk projection, multi-scale convolutional branches, channel compression excitation attention, temporal self-attention, and token sequence output. The main difference between the modalities lies in the convolutional kernel group settings: 9, 25, and 51 for EEG; 7, 15, and 31 for EMG and ACC; and 5, 11, and 21 for PPG / GSR. Functional connectivity differential features (FHDs) are not part of the internal convolutional branches of the multi-scale encoder but are resting-state FHDs computed in parallel after encoding open-eye and closed-eye EEG. The attention-weighted state comparison module corresponds to... Figure 7 The state comparison branch in the model includes a three-way fusion component: differential sequence self-attention, cross-state cross attention, and mean difference.

[0031] and Figure 7 Correspondingly, in this embodiment, the hidden dimension hidden_dim is 96, the temporal token length token_length is 16, and the number of attention heads num_heads is 4. When the EEG is 32-dimensional, the functional connection difference module extracts the upper triangular difference vector of the channel correlation matrix, which has a dimension of 32×31 / 2=496, and then projects it onto the 96-dimensional hidden space. Figure 7 The numbers in the table are all model parameters or tensor dimensions mentioned above, used to explain the correspondence between multi-scale coding, functional connectivity difference, and state contrast modules in the first evaluation mode.

[0032] In another alternative embodiment, the first evaluation mode may further receive resting-state EMG and / or wristband PPG / GSR signals as auxiliary inputs, and supplement and enhance the resting-state EEG representation through central-peripheral fusion; however, the above-mentioned auxiliary inputs are not a necessary condition for implementing the main technical route of the present invention.

[0033] In a preferred embodiment, when the first risk level or the first risk score reaches the preset transition condition, the second assessment mode is activated.

[0034] In the second assessment mode, an electrical stimulation task was performed, task-state EEG signals were collected, and an electrical stimulation pain score was generated by combining the stimulation intensity calibration information.

[0035] In the second assessment mode, a motor evoked task was performed, and task-state EEG signals, task-state EMG signals, task-state photoplethysmography (PPG) / skin conductance response (GSR) signals, and motor acceleration signals were collected simultaneously.

[0036] Based on the start and end events of the action, the multimodal signals in the second evaluation mode are re-cut, uniformly aligned and fused in multiple trials to generate the evaluation results of the motion-induced task.

[0037] More specifically, there is a conditionally triggered, gated progressive relationship between the first and second assessment modes. Specifically, after the first assessment mode outputs a logits vector containing the risk level classification results, the system takes its argmax (i.e., the operation corresponding to the index of the maximum value in the vector) as the first risk level. Simultaneously, it inputs a shared representation into an independent first risk score output header, passes it through a linear layer and a sigmoid function to obtain a continuous first risk score. The first risk level or first risk score is compared with a preset gating threshold. The system activates the second assessment mode for forward computation only when the first risk level or first risk score of at least one subject meets the preset transition condition. If neither the first risk level nor the first risk score meets the preset transition condition, the system directly outputs the result of the first assessment mode without activating the second assessment mode. During the training phase, the gating threshold is preferably set to negative one to disable gating logic and ensure all samples participate in the two-stage joint training. During the inference deployment phase, the gating threshold is preferably set to zero or one to actually enable the progressive assessment route. In addition, the second assessment mode simultaneously includes four modalities (EEG, EMG, PPG / GSR and motion acceleration) fused through adaptive attention weights, as well as an independent electrical stimulation task branch.

[0038] In a preferred embodiment, such as Figure 3 and Figure 8 As shown, the second assessment mode assessment module includes an electrical stimulation task branch and a motor evoked task branch, which together form the second assessment mode representation.

[0039] The electrical stimulation task branch is used to construct a pain task representation and output an electrical stimulation pain score. Specifically, the task-state EEG is first input into the electrical stimulation task EEG encoding module; this module employs a multi-scale temporal encoder, including a stem convolution with a kernel length of 5, three multi-scale convolutional branches with kernel lengths of 9, 25, and 51, channel compression excitation, a Transformer encoder layer, and layer normalization, outputting an EEG token sequence with a length of 16 and a hidden dimension of 96; this sequence is then subjected to mean pooling to obtain task-state EEG feature A. The stimulation intensity calibration information fusion module uses low, medium, and high stimulation intensity calibration information as a 3D calibration vector, which is processed by Linear(3,96), GELU, and layer normalization to obtain calibration feature B. Using task-state EEG feature A as the query and B as the key and value, attention-enhanced EEG feature C is obtained through 4-head multi-head attention, residual connections, and layer normalization. Attention-enhanced EEG feature C is then concatenated with calibration feature B, and the pain task representation is obtained through Linear(192,96), GELU, and layer normalization. The electrical stimulation pain score is output by the electrical stimulation task output head Linear(96,1). In Linear(3,96), 3 represents the number of input features, corresponding to the low, medium, and high stimulation intensity calibration values, and 96 represents the number of output features, corresponding to the unified hidden feature dimension. In Linear(192,96), 192 represents the input dimension after concatenating the attention-enhanced EEG feature and the calibration feature, and 96 represents the output dimension of the pain task representation. In Linear(96,1), 96 represents the dimension of the pain task representation, and 1 represents the output dimension of the electrical stimulation pain score.

[0040] The exercise-evoked task branch is used to generate exercise-evoked pain load assessment results. Specifically, EEG, EMG, PPG / GSR, and ACC signals are encoded by their respective encoding modules to obtain 96-dimensional four-modal trial-level features. These four-modal trial-level features are a collective term for the four types of trial features: EEG, EMG, PPG / GSR, and ACC. The ACC trial-level features are the ACC modal components within these features. The ACC encoding module outputs two types of results simultaneously: one is the ACC trial-level features, which are used as ACC modal features in the subsequent adaptive modality alignment attention module; the other is the ACC trial-level action quality score, which is obtained by sequentially executing Linear(96,48), GELU, Linear(48,1), and Sigmoid on the ACC trial-level features using the action quality assessment head. The ACC trial-level action quality score is not input for missing modality processing, nor is it used as input for the four-modality fusion weight calculation. Instead, it is aggregated by trial weights after the trial attention weights are determined to form a subject-level action quality assessment. The ACC trial-level features are then aggregated by mean based on the valid trial labels to obtain the ACC action features. In Linear(96,48), 96 is the input dimension of the ACC trial features, and 48 is the intermediate feature dimension of the motion quality assessment head; in Linear(48,1), 48 is the input dimension, and 1 indicates that each trial outputs one motion quality score.

[0041] The EEG, EMG, PPG / GSR, and ACC trial features from the four modal trial features are included. Figure 8The adaptive modality alignment attention module shown applies missing modality handling to modal features and their weights: first, the features are concatenated into a 4×96-dimensional vector, and then processed by Linear(384,96), GELU, Linear(96,4), and Softmax to obtain four modal weights; when the ACC modality is missing, the corresponding ACC weight is reduced to 10% of its original value and then renormalized; then, the weighted four modal features are concatenated and processed by Linear(384,96), GELU, and LayerNorm to obtain a trial-level fusion representation. The trial-level fusion representation, along with location and demographic conditions, forms the trial context, which is then aggregated by trial attention to obtain the subject-level task embedding; the subject-level task embedding, ACC action features, location conditions, and demographic conditions are fed into the output head to obtain the motion-evoked pain load assessment result. Thus, the second assessment modality representation is formed by the pain task representation and electrical stimulation pain score output from the electrical stimulation task branch, and the subject-level task embedding, ACC action features, and motion-evoked pain load assessment result output from the motion-evoked task branch. The four 96-dimensional modal features are concatenated to form a 384-dimensional input. Linear(384,96) maps this input to a 96-dimensional hidden space, where the 4 in Linear(96,4) represents the weights of the four modalities: EEG, EMG, PPG / GSR, and ACC. Location features and demographic features are each 16-dimensional, and these are concatenated with the 96-dimensional trial-level fusion representation to form a 128-dimensional trial context, which is then mapped to 96-dimensional trial features via Linear(128,96). The trial attention aggregation module sequentially uses Linear(96,96) and Linear(96,1) to generate an attention score for each trial, and then weights and sums the 96-dimensional features of each trial to obtain a 96-dimensional subject-level task embedding; the random deactivation probability in this module is 0.2. The task embedding, 96-dimensional ACC action features, 16-dimensional location features, and 16-dimensional demographic features are concatenated to form a 224-dimensional input. This input is then mapped sequentially by the subject-level output head to 192-dimensional intermediate features and 96-dimensional shared representations of the second evaluation mode, with a random inactivation probability of 0.2. The motion-induced pain load output head then maps this 96-dimensional representation to a single score. When masking invalid trials, their attention score is set to -1e4, where -1e4 represents a minimum value approximating negative infinity, making the weights of invalid trials after Softmax normalization close to zero.

[0042] This invention first extracts resting-state risk characteristics based on the differences between open-eye and closed-eye EEG, and performs a low-burden initial screening of subjects. Then, for subjects who meet the preset conditions, it further combines EEG information from the electrical stimulation task and EEG, EMG, PPG / GSR, and motor acceleration signals from the motor evoked task to further assess the subjects' pain sensitivity, motor-evoked pain load, and overall risk. This reduces reliance on single subjective scales, single manual observations, and experience-based judgments, and improves the objectivity, consistency, and repeatability of the assessment process.

[0043] In a preferred embodiment, such as Figure 4 As shown, the specific implementation of the cross-mode fusion module is as follows: after the first evaluation mode representation and the second evaluation mode representation are respectively linearly projected and layer normalized to obtain the aligned representation, the element-wise absolute difference term and the element-wise product term of the two are constructed. The aligned representation, the difference term and the product term are input into the cross-mode fusion network to obtain the fused representation; the consistency score is calculated in the form of a negative exponential of the Euclidean distance between the two-stage aligned representations.

[0044] Based on the interaction features of cross-modal consistency scores and bimodal representations, a comprehensive risk result is generated. The comprehensive risk result may include a fusion risk score, a fusion risk level, a fusion pain score, and the difference between potential ability and actual performance. Specifically, a 96-dimensional cross-modal fusion representation is concatenated with 16-dimensional strain site features to form a 112-dimensional input, which is then mapped sequentially to 192-dimensional intermediate features and a 96-dimensional comprehensive representation via a shared comprehensive output head, with a random inactivation probability of 0.2. The fusion risk score output head and the fusion pain score output head each map this 96-dimensional comprehensive representation to a continuous value, and the fusion risk level output head maps it to a classification score of three risk levels.

[0045] In a preferred embodiment, during training and deployment, the first evaluation mode screening module, the second evaluation mode evaluation module, and the cross-mode fusion module can be trained independently, sequentially, or jointly. In a specific embodiment, the first evaluation mode training can be completed first, followed by the second evaluation mode training, and finally, the outputs of the first two can be used for cross-mode fusion training. In actual deployment, the first evaluation mode screening module can be used for rapid screening at the front end, while the second evaluation mode evaluation module and the cross-mode fusion module can be used for fine-grained evaluation and comprehensive judgment at the back end.

[0046] In a preferred embodiment, an output module is also included for outputting risk score, risk level, task status score, consistency score, and comprehensive evaluation result.

[0047] In this embodiment, the specific implementation of the core algorithm and network structure of the present invention will be described in detail with reference to the accompanying drawings. In this embodiment, the hidden dimension `hidden_dim` is set to 96, the temporal token length `token_length` is set to 16, and the number of attention heads `num_heads` is set to 4. The main symbols and English parameters involved in this embodiment have the following meanings: B is the batch size, C... in T represents the number of input channels and the number of time sampling points; Conv1d is a one-dimensional convolution, kernel_size is the kernel length, padding is the padding length, and stride is the convolution stride; BatchNorm1d (BN) is a one-dimensional batch normalization; GELU is a Gaussian error linear unit activation function; AdaptiveAvgPool1d is an adaptive one-dimensional average pooling; FC is a fully connected layer; Sigmoid is a sigmoid activation function; Tanh is a hyperbolic tangent activation function; Softmax is a normalized exponential function; TransformerEncoderLayer is a Transformer encoding layer; SE (Squeeze-and-Excitation) is a channel compression-excited attention module. Additional notes: LayerNorm is layer normalization, MultiheadAttention is multi-head attention, Softmax is a normalized exponential function, and the token sequence represents a fixed-length temporal feature sequence output by the encoder.

[0048] (1) Specific structure of multi-scale timing encoder The multi-scale temporal encoder performs end-to-end feature learning on the input multi-channel temporal signal and outputs a fixed-length token sequence. Its complete processing flow is as follows: Step A1: Stem projection. For the input signal x(B, C) in (, T), where B represents the batch size, C in The input channel number is represented by T, and the time sampling points are represented by T. A one-dimensional convolution Conv1d(C) is applied. in , 96, kernel_size=5, padding=2), followed by BatchNorm1d(96) and GELU activation, mapping the number of input channels to a unified 96-dimensional hidden space. Wherein, Conv1d(C inIn BatchNorm1d(96), 96 represents the number of output channels of the convolution, which is used as the unified hidden feature dimension of the subsequent encoder; kernel_size=5 indicates that the convolution kernel covers 5 time sampling points, and padding=2 indicates that the time dimension padding length is 2. In BatchNorm1d(96), 96 represents the number of channels for batch normalization.

[0049] Step A2: Multi-scale parallel convolution. Set up 3 parallel branches with a branch width of branch_width = max(96 / / 3,8) = 32, where branch_width represents the channel width of each convolutional branch. The structure of each branch is: Conv1d(96, 32, kernel_size=k i padding=k i / / 2, where k i The kernel length of the i-th branch is represented by (where padding represents the padding length) + BN + GELU + Conv1d(32, 32, kernel=3, stride=2, padding=1) + BN + GELU + AdaptiveAvgPool1d(16). The three branches of the EEG mode have convolution kernels of k1=9, k2=25, and k3=51, which correspond to time receptive fields of approximately 72ms, 200ms, and 408ms at a sampling rate of 125Hz. The convolution kernel of the EMG mode is (7, 15, 31), the kernel of the PPG / GSR mode is (5, 11, 21), and the kernel of the ACC motion acceleration mode is (7, 15, 31). In max(96 / / 3,8), 96 is the hidden feature dimension, 3 is the number of parallel convolution branches, the symbol " / / " represents integer division down, 8 is the lower limit of the channel width of a single branch, and the calculated result 32 is the channel width of each branch in this embodiment. In Conv1d(96,32,...), 96 and 32 are the number of input and output channels, respectively; in Conv1d(32,32,kernel=3,stride=2,padding=1), the two 32s are the number of input and output channels, 3 is the kernel length, 2 is the convolution stride, and 1 is the padding length; in AdaptiveAvgPool1d(16), 16 is the uniform output token length. Each number in the above convolution kernel group represents the number of time sampling points covered by the corresponding one-dimensional convolution kernel, and the three numbers correspond to short, medium, and long time scales, respectively.

[0050] Step A3: Branch Fusion. The outputs of the three branches are concatenated along the channel dimension to obtain a feature tensor with batch size B, 96 channels, and a token length of 16. This tensor is then fused using a 1x1 convolution Conv1d(96, 96, kernel_size=1), followed by Batch Normalization (BN) and GELU. In Conv1d(96, 96, kernel_size=1), the first and last 96 represent the number of input and output channels, respectively, and kernel_size=1 indicates that a convolutional kernel of length 1 is used for channel fusion.

[0051] Step A4: Channel compression stimulates the attention module, i.e., the SE module. The fused features are globally averaged and pooled to (B, 96, 1). Channel weights are generated through two fully connected layers and a sigmoid activation function, and then adaptively reweighted by multiplying them element-wise with the original features. In (B, 96, 1), B represents the batch size, 96 represents the number of channels, and 1 indicates that the time dimension is compressed into a single value after global average pooling. Specifically, the two fully connected layers sequentially compress the 96-dimensional channel description vector into a 24-dimensional intermediate vector, and then restore it to 96-dimensional channel weights. The 24 is obtained by dividing 96 by a compression ratio of 4, which reduces the number of parameters in the channel weight generation process.

[0052] Step A5: Temporal Self-Attention. Transpose (B, 96, 16) to (B, 16, 96) and input it into `TransformerEncoderLayer(d_model=96,nhead=4,dim_feedforward=192,dropout=0.1,activation=GELU, batch_first=True)`, which is the Transformer encoder layer. The parameters represent the hidden dimension `d_model` as 96, the number of attention heads `nhead` as 4, the feedforward dimension `dim_feedforward` as 192, and the dropout as 0.1. GELU activation is performed using Gaussian error linear units. Global temporal interaction modeling is then performed. Finally, `LayerNorm` outputs a token sequence of shape (B, 16, 96). Here, `LayerNorm` is layer normalization, B represents the batch dimension, 16 represents the uniform token length, and 96 represents the hidden dimension.

[0053] (2) Specific calculation of functional connectivity difference features For a 32-lead EEG signal x(32, T), channel-normalized: Z(c,t) = (x(c,t) - mean c ) / (std c + 1e-6). Calculate the Pearson correlation matrix: R = Z·Z T / T, resulting in a 32x32 symmetric matrix. Extracting the upper triangular elements (excluding the diagonal), the vector dimension is 32. 31 / 2 = 496. Calculate the difference between the EEG results for open and closed eyes: FC diff =Upper(R open ) - Upper(R closed Where T represents the number of time sampling points, c represents the EEG channel index, t represents the time index, and mean c and std c Let Z represent the mean and standard deviation of the c-th channel over time, respectively, with 1e-6 being a very small positive number introduced to prevent the standard deviation from being zero; T R represents the transpose of the normalized matrix Z; open and R closed These represent the Pearson correlation matrices obtained after standardization of open-eye resting EEG and closed-eye resting EEG, respectively; Upper(·) represents extracting the off-diagonal elements of the upper triangular part of the matrix and flattening them into a vector in a fixed order; FC diff This represents the difference feature between the functional connectivity vectors for open and closed eyes. Finally, it is projected into the hidden space using Linear(496, 96) + GELU + LayerNorm. Here, 496 in Linear(496, 96) represents the number of off-diagonal elements in the upper triangle of the 32-lead correlation matrix, and 96 represents the dimension of the projected hidden feature.

[0054] (3) Specific structure of the three-way fusion status comparison module First path: Calculate diff tokens = open tokens - closed tokens , where open tokens Represents the open-eyes token sequence, closed tokens Represents the closed-eye token sequence, diff tokens The token sequence (B, 16, 96) represents the state difference. MultiheadAttention (embed_dim=96, num_heads=4, batch_first=True) self-attention is performed, followed by LayerNorm, and time-dimension mean pooling yields (B, 96). Second path: using open... tokens For Query, closed tokensFor the Key and Value, perform MultiheadAttention(96, 4) cross-attention, followed by LayerNorm and mean pooling. Third path: Calculate the mean of the open-eye token sequence and the closed-eye token sequence along the time dimension dim=1, then calculate the difference. The three paths are concatenated to obtain (B, 288), which is then processed by Linear(288, 192) + GELU + Linear(192, 96) + LayerNorm to output (B, 96). Here, in MultiheadAttention(embed_dim=96,num_heads=4), 96 represents the feature dimension of each token, and 4 represents the number of attention heads; 288 in (B, 288) is obtained by concatenating three 96-dimensional features; Linear(288, 192) and Linear(192, 96) respectively represent mapping the 288-dimensional three-path concatenation features to 192-dimensional intermediate features, and then to 96-dimensional state contrast features. The random inactivation probability of the differential sequence in the self-attention layer and the cross-state attention layer is 0.1. This value is used to suppress overfitting of the attention weights.

[0055] (4) Specific implementation of the electrical stimulation task branch The task-state EEG features f are obtained by encoding the electrical stimulation EEG using a multi-scale temporal encoder (kernels=9,25,51) and then performing mean pooling. eeg Its dimension is B×96. The calibration triplets composed of low, medium, and high stimulus intensity calibration information are used to obtain the calibration feature f through Linear(3, 96) + GELU + LayerNorm. cal Its dimension is also B×96. Using f eeg For Query, f cal MultiheadAttention (96, 4) is applied to the Key / Value pairs, along with residuals and LayerNorm. The concatenated strings are then processed by Linear(192, 96) + GELU + LayerNorm to output a pain task representation, and an electrical stimulation pain score is output via Linear(96, 1). Specifically, the task-state EEG first enters the electrical stimulation task EEG encoding module, where a multi-scale temporal encoder obtains the EEG token sequence, which is then subjected to temporal mean pooling to form f. eeg The low, medium, and high stimulus intensity calibration triplets enter the stimulus intensity calibration information fusion module, and are then processed by the linear layer, Gaussian error linear unit (GELU), and layer normalization (LayerNorm) to form f. cal ; then f eeg As a query, with f calCross-attention is performed on the key and value, and residual normalization is applied to integrate the individualized calibration context into the task-state EEG features. Finally, the attention-enhanced EEG features are concatenated with the calibration features, and the resulting fusion projection yields the pain task representation. An electrical stimulation pain score is then generated from the output head of the electrical stimulation task. In kernels=9,25,51, 9, 25, and 51 represent the kernel lengths at short, medium, and long time scales, respectively. In Linear(3,96), 3 and 96 represent the input and output dimensions of the calibration triplet, respectively. In MultiheadAttention(96,4), 96 and 4 represent the attention feature dimension and the number of attention heads, respectively. In Linear(192,96), 192 is obtained by concatenating two 96-dimensional features, outputting a 96-dimensional pain task representation. Linear(96,1) maps the 96-dimensional pain task representation to a single electrical stimulation pain score. The random deactivation probability of this cross-attention layer is 0.1.

[0056] (5) Specific calculation of adaptive modal alignment attention The four modal features of EEG, EMG, PPG / GSR, and ACC are concatenated into a (B, 384) vector, and then weights w are generated by applying Linear(384, 96) + GELU + Linear(96, 4) + Softmax. eeg , w emg , w ppg / gsr , w acc ]. Among them, w eeg w emg w ppg / gsr and w acc represents the normalized weights for the EEG mode, EMG mode, PPG / GSR peripheral physiological mode, and ACC motor acceleration mode, respectively. If ACC does not exist, then w acc Multiply by 0.1 and renormalize. Concatenate the weighted features (B, 384) and then perform Linear(384, 96) + GELU + LayerNorm to obtain the secondary fusion representation. Here, 384 in (B, 384) is obtained by concatenating four 96-dimensional modal features; 384 and 96 in Linear(384, 96) are the input dimension and hidden dimension, respectively; 4 in Linear(96, 4) corresponds to the weight output of the four modalities EEG, EMG, PPG / GSR and ACC; 0.1 is the weight reduction coefficient used when ACC is missing.

[0057] (6) Specific calculation of cross-stage integration and consistency score Projection respectively: f r (First projection feature) = LayerNorm(Linear(rest)emb (Characteristics of the first evaluation model), 96)), f t (Second projection feature) = LayerNorm(Linear(task) emb (Second evaluation model representation), 96). Constructing interactive features: [f r , f t , |f r -f t |, f r f t The concatenation is (B, 384). The fused representation is obtained by applying Linear(384, 192) + GELU + Dropout(0.2) + Linear(192, 96) + GELU + LayerNorm. Consistency score: consistency = exp(-||f r -f t ||2). Where ||·||2 represents the Euclidean norm; alignment penalty during training: L align =mean((f r -f t ) 2 ), with a weight of 0.2. Where, f r and f t Both are 96-dimensional projection features, where 384 in (B,384) is derived from f. r f t The four 96-dimensional features are concatenated: the absolute difference between the two and the element-wise product. Linear(384,192) and Linear(192,96) map the 384-dimensional interactive features to 192-dimensional intermediate features and 96-dimensional fused representations, respectively. In Dropout(0.2), 0.2 is the random inactivation probability, and another 0.2 is the weight of the alignment penalty term in the total loss.

[0058] (7) Training strategies A three-stage training method is used, with the optimizer AdamW (lr=3e-4, weight_decay=1e-4) and gradient clipping of 1.0. The loss for the first stage is L1 = 1.0. SmoothL1(risk) + 0.4 SmoothL1 (sensitivity) + 0.5 CE (level). Second-stage loss: L2 = SmoothL1(pain) + SmoothL1(motor) pain ) + 0.25 SmoothL1 (trial) painIn the third stage, the parameters of the first two stages are frozen, and the loss is: L3 = 1.0. SmoothL1(fused risk ) + 1.0 SmoothL1(fused pain ) + 0.5 CE(level) + 0.2 L align Fifty participants were randomly assigned at a ratio of 0.7:0.15:0.15, with a seed of 42. Here, risk represents the primary risk score supervision target, sensitivity represents the pain sensitivity supervision target, level represents the risk level classification supervision target, pain represents the electrical stimulation pain score supervision target, and motor... pain This indicates the monitoring target for exercise-induced pain load assessment, trial pain This represents the secondary pain rating supervision objective. Among the parameters above, 3e-4 is the learning rate of the AdamW optimizer, 1e-4 is the weight decay coefficient, and 1.0 is the gradient clipping threshold; the values ​​before each loss term represent the loss weights of the corresponding supervision objective; 0.7, 0.15, and 0.15 are the split ratios of the training, validation, and test sets, respectively, and 42 is the random seed to ensure reproducible data splitting. In the second-stage loss, pain and motor... pain All coefficients not explicitly stated before are 1.0; trial pain Specifically, it represents the monitoring target for pain scores in each valid trial during the exercise-induced task.

[0059] This embodiment provides a method for predicting the risk of chronic strain injury based on a dual-mode progressive assessment, such as... Figure 6 As shown, it includes: The subjects were collected with pre-registered information on the type of strain site, age, and gender, as well as resting EEG signals with eyes open and resting EEG signals with eyes closed. Preprocessing of open-eye resting-state EEG signals and closed-eye resting-state EEG signals yielded the first assessment mode representation; The first risk score and the first risk level are generated based on the first assessment mode; when the first risk level or the first risk score reaches the preset transition condition, the acquisition task state multimodal signal is triggered. The task-state multimodal signal is preprocessed to obtain the second evaluation mode features; Generate a representation of the second evaluation model based on its characteristics; The first and second assessment model representations are projected into a common space to generate cross-modal consistency scores and comprehensive risk results.

[0060] Based on the same inventive concept, and since the principle by which the method solves the problem is similar to that of the aforementioned system, the implementation of the aforementioned system can be referred to, and the repeated parts will not be described again.

[0061] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0062] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A chronic strain injury risk prediction system based on dual-mode progressive assessment, characterized in that, include: The data acquisition module is used to collect the subject's open-eye resting-state EEG signals and closed-eye resting-state EEG signals, as well as task-state multimodal signals when the first risk level or first risk score reaches the preset transition conditions. The quality control and alignment module is used to preprocess the open-eye resting-state EEG signals and the closed-eye resting-state EEG signals to obtain the first evaluation mode representation. And preprocessing the task-mode multimodal signal to obtain the second evaluation mode features; The first assessment mode screening module is used to generate a first risk score and a first risk level based on the first assessment mode representation. The second evaluation mode evaluation module is used to generate a second evaluation mode representation based on the characteristics of the second evaluation mode. The cross-modal fusion module is used to project the representations of the first and second evaluation modalities into a common space to generate a cross-modal consistency score and a comprehensive risk result.

2. The system as described in claim 1, characterized in that, Task-mode multimodal signals include: Acquire task-state EEG signals and corresponding stimulation intensity calibration information under electrical stimulation tasks; Acquire task-state EEG signals, task-state photoplethysmography / skin conductance signals, task-state electromyography signals, and motion acceleration signals under motion-evoked tasks.

3. The system as described in claim 2, characterized in that, The quality control and alignment module is specifically used for: Filtering, artifact removal, rereference, resampling, and normalization are performed on open-eye resting-state EEG signals, closed-eye resting-state EEG signals, and / or task-mode multimodal signals; Task-state EEG signals under electrical stimulation tasks and corresponding stimulation intensity calibration information are extracted according to the stimulation event and the corresponding time period. Task-state EEG signals, task-state photoplethysmography / skin conductance signals, task-state electromyography signals, and motor acceleration signals under motion-evoked tasks were determined according to the action initiation event and the action termination event, and then sliced ​​according to the uniform trial boundary. The trial structures under different part protocols are uniformly aligned using valid trial markers; Missing trials, trials of abnormal length, or partially missing modes are padded or masked.

4. The system as described in claim 1, characterized in that, The first assessment mode screening module is specifically used for: Feature extraction was performed on the preprocessed open-eye resting-state EEG signal and closed-eye resting-state EEG signal to obtain open-eye resting-state EEG features and closed-eye resting-state EEG features; Construct state difference features between open-eye resting EEG features and closed-eye resting EEG features, and calculate resting-state functional connectivity features; The state difference features and resting state functional connectivity features are fused to generate the first risk score and the first risk level.

5. The system as described in claim 3, characterized in that, The second evaluation mode evaluation module is specifically used for: Task-state EEG features were obtained by extracting features from the preprocessed electrical stimulation task; stimulation intensity calibration information was used as calibration features and cross-fused with the task-state EEG features to obtain a pain task representation and output an electrical stimulation pain score. Multimodal secondary features were constructed from preprocessed task-state EEG signals, task-state photoplethysmography / skin conductance signals, task-state electromyography signals, and motor acceleration signals under a motion-evoked task. Adaptive modal alignment is performed on the multimodal trial sub-features to obtain multimodal weights. The multimodal weights are then weighted, stitched together, and fused by projection to obtain the trial sub-fusion representation. The exercise-induced pain load assessment result is obtained through the trial sub-fusion representation. The second assessment modality was derived from the results of exercise-induced pain load assessment and pain task representation.

6. The system as described in claim 1, characterized in that, The cross-mode fusion module is specifically used for: Project the first evaluation model representation and the second evaluation model representation onto a common feature space; Calculate the Euclidean distance between the projected bimodal representations and generate a cross-modal consistency score based on the negative exponential form of the Euclidean distance. Based on the cross-modal consistency score and the interactive features of the bimodal representation, a comprehensive risk result is generated.

7. The system as described in claim 1, characterized in that, It also includes an output module for outputting cross-modal consistency scores and overall risk results.

8. The system as described in claim 4, characterized in that, The first assessment mode screening module is also used for: The central feature is obtained by concatenating the mean pooling features of the open-eye resting EEG token sequence, the mean pooling features of the closed-eye resting EEG token sequence, the resting-state functional connectivity features, and the state difference features, and then performing linear mapping, GELU activation, and layer normalization. After concatenating the central features, state difference features, resting state functional connectivity features, strain location features and demographic features, the data are input into the shared feature extraction head. The shared representation is obtained after two layers of linear mapping, GELU activation, random deactivation and layer normalization. The shared representation outputs a pain sensitivity score through a linear layer and a sigmoid activation function.

9. A method for predicting the risk of chronic strain injury based on dual-mode progressive assessment, characterized in that, include: The subjects' resting-state EEG signals with eyes open and resting-state EEG signals with eyes closed were collected; Preprocessing of open-eye resting-state EEG signals and closed-eye resting-state EEG signals yielded the first assessment mode representation; The first risk score and first risk level are generated based on the first assessment model. When the first risk level or the first risk score reaches the preset transition condition, the multimodal signal of the data acquisition task is triggered. The task-state multimodal signal is preprocessed to obtain the second evaluation mode features; Generate a representation of the second evaluation model based on its characteristics; The first and second assessment model representations are projected into a common space to generate cross-modal consistency scores and comprehensive risk results.

Images