A method and system for modeling and intervention simulation of children with Chinese reading and writing difficulties based on digital twin brain technology

Abstract

The application discloses a Chinese reading and writing difficulty child modeling and intervention simulation method and system based on digital twin brain technology, a data acquisition module acquires behavior data and brain imaging data of Chinese reading and writing difficulty children and normal control children; a model construction module is used to construct a Chinese reading and writing digital twin base model, simulate the function shunting and consciousness formation dynamics in the brain reading and writing process; an individualized modeling module adopts a double constraint optimization method, and individualized fine-tuning is performed on the Chinese reading and writing digital twin base model to establish an individualized digital twin model; an intervention simulation module verifies and mechanism analyzes the individualized digital twin model, simulates different intervention strategies, predicts the improvement effect of the individual reading and writing behavior and brain representation, recommends a personalized intervention scheme and outputs; the application is aimed at the individualized digital twin modeling of Chinese reading and writing difficulty, constructs an exclusive individual digital twin model for each tester, and analyzes the individual heterogeneity of reading and writing difficulty.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of cognitive neuroscience and artificial intelligence, specifically relating to a modeling and intervention simulation method and system for Chinese dyslexia children based on digital twin brain technology, used to reveal the neural mechanisms of dyslexia and to simulate and predict the intervention effects. Background Technology

[0002] Dyslexia is a common learning disability affecting 5%–17.5% of school-aged children. Existing research is largely based on population-level brain imaging data analysis, which struggles to characterize individual differences and cannot provide mechanistic deduction or intervention prediction. While traditional functional magnetic resonance imaging (fMRI) and behavioral association studies can reveal abnormalities in certain brain regions, they cannot dynamically simulate inter-brain interactions and their causal relationships with behavior. Furthermore, current methods primarily focus on conscious processing, neglecting the role of unconscious processing in automatization of reading and writing, thus limiting the completeness of mechanistic explanations.

[0003] Over the past three decades, various non-invasive brain imaging techniques, such as functional magnetic resonance imaging (fMRI), magnetoencephalography (MEG), and electroencephalography (EEG), have been widely applied in this field, gradually revealing functional abnormalities in multiple brain regions, from the cortex to the subcortex, in dyslexia. Studies have shown that dyslexia is primarily related to insufficient activation of the left-sided language network, involving brain regions including the superior temporal gyrus, middle temporal gyrus, temporoparietal junction, and visual word form area. Abnormal activity in these areas affects character recognition and speech decoding, leading to decreased reading and writing efficiency. However, how these abnormal activities specifically cause dyslexia, and the core pathogenic mechanisms, remain unclear. Limited by the inherent limitations of traditional research methods, the synergistic mechanisms between different brain regions have not yet been systematically elucidated.

[0004] Cross-linguistic research has shown that the neurophenotypes of dyslexia possess both universality and language specificity. Taking Chinese as an example, as an ideographic writing system, it differs significantly from alphabetic writing systems in its visual structure and the correspondence between form, sound, and meaning. Studies have found that the core brain region for Chinese dyslexia is located in the left middle frontal gyrus, unlike the temporoparietal lobe abnormalities commonly found in alphabetic writing systems. Therefore, research on the mechanisms and intervention systems for Chinese dyslexia must be built upon its linguistic characteristics and cannot simply follow the research framework of alphabetic writing systems. Moreover, due to technological and ethical limitations, the mechanisms of the brains of children with dyslexia can only be measured indirectly, making it difficult to fundamentally understand the working principles of the brain and establish causal relationships. Due to the lack of guidance from research on brain mechanisms, the design of interventions for reading disorders is often based on phenomenological levels, failing to reach the core mechanisms and resulting in inconsistent intervention effectiveness.

[0005] In recent years, the combination of digital twin technology and deep learning has provided a new approach to constructing computable "brain-behavior" models at the individual level. Digital twin technology can establish interpretable brain function models at the individual level, achieving a leap from data description to mechanism simulation. Combined with deep neural networks, personalized deep neural network models can be further constructed. By adjusting parameters such as neural excitability and connection strength, the cognitive behavior and brain activity characteristics of an individual can be reproduced in a virtual model, realizing "neuro-digital twins." Personalized deep neural network models constructed using this method have successfully revealed the neurophysiological mechanisms of learning disabilities in alphabetic languages. The models can highly fit the behavioral and brain activity characteristics of children with learning disabilities, simulating phenomena such as reduced learning accuracy, slower learning rate, neural over-excitation, and decreased differentiation of neural representations of numerical problems. However, most existing neuro-digital twin models are designed for phonetic writing systems or general visual recognition tasks, and there is still a lack of specific modeling paradigms and frameworks for difficulties in reading and writing Chinese. Summary of the Invention

[0006] To address the aforementioned technical problems in existing technologies, this invention provides a method and system for modeling and intervening in Chinese-speaking children with dyslexia based on digital twin brain technology. Through multimodal data fusion, bio-inspired network architecture, and individualized parameter optimization, a computational model that accurately reflects the individual's dyslexia neural mechanisms is constructed, and virtual simulation and effect prediction of intervention strategies are supported.

[0007] The specific plan adopted is as follows:

[0008] On the one hand, this invention provides a method for modeling and intervening in Chinese-speaking children with dyslexia based on digital twin brain technology, comprising the following steps:

[0009] Collect behavioral and brain imaging data of children with Chinese reading and writing difficulties and normal control children;

[0010] A digital twin base model for Chinese reading and writing was constructed, including a Siamese dual coding network, a multi-task discriminative branch, a Chinese component sensitive module, and a loop feedback mechanism, to simulate the functional diversion and dynamic formation of consciousness in the brain during the reading and writing process.

[0011] A dual-constraint optimization method, combining behavioral data alignment and brain imaging data alignment, was used to fine-tune the Chinese reading and writing digital twin base model, thereby establishing an individualized digital twin model.

[0012] To validate and analyze the mechanism of personalized digital twin models, revealing the causal mechanism of reading and writing difficulties;

[0013] Based on the individualized digital twin model, different intervention strategies are simulated to predict their effects on improving individual reading and writing behavior and brain representation, and personalized intervention plans are recommended for individuals.

[0014] The behavioral data consists of reaction time and accuracy obtained through successive trials of speech discrimination tasks, character shape discrimination tasks, component character formation discrimination tasks, and shape similarity discrimination tasks. The brain imaging data consists of task-state and resting-state whole-brain BOLD signals acquired using functional magnetic resonance imaging. The key regions of interest include the left fusiform gyrus visual word shape region, the left temporoparietal junction region, the inferior frontal gyrus, the middle frontal gyrus, and the primary visual cortex. These key regions of interest are used for alignment constraints within the internal representation of the Chinese reading and writing digital twin base model.

[0015] More preferably, in the process of establishing the Chinese reading and writing digital twin base model, four key neurophysiological parameters are introduced: global neural gain G, internal noise σ, cyclic feedback gain R, and neuron activation threshold T. A Bayesian optimization or genetic algorithm global search method is adopted, with the total individual loss as the evaluation function, to find the optimal combination of (G, σ, R, T) in the parameter space.

[0016] In the Chinese reading and writing digital twin base model, the global neural gain G is used to characterize the balance between cortical excitability and inhibition, and is achieved by multiplying by a coefficient after each layer of convolution output;

[0017] The internal noise σ is simulated by adding Gaussian noise after each layer of convolution calculation to simulate the inherent random fluctuations of neural activity.

[0018] The recurrent feedback gain R is determined by multiplying the weights of each layer of recurrent connections by a coefficient to determine the strength of the feedback loop.

[0019] The neuron activation threshold T controls the minimum input intensity required for a neuron to transition from a resting state to an activated state by adjusting the threshold parameter of the activation function.

[0020] Preferably, the specific method for constructing a Chinese reading and writing digital twin base model is as follows:

[0021] For the pairwise stimulus discrimination paradigm, a Siamese dual coding network is constructed;

[0022] A parallel convolutional branch structure with two shared parameters is adopted to simulate the parallel processing mechanism of paired stimuli in the human brain and to share the visual coding backbone.

[0023] Each branch uses a four-layer recurrent convolutional structure (V1→V2→V4→IT) to extract hierarchical visual features.

[0024] A fusion layer is introduced after the IT layer, and a splicing and differential strategy is used to integrate the dual-branch features.

[0025] After sharing the visual encoding backbone, four parallel task discrimination branches emerge, corresponding to the neural mechanisms by which the brain processes different information pathways in parallel during reading and writing.

[0026] The four-layer recurrent convolutional structure V1→V2→V4→IT corresponds to the neural anatomy of the ventral visual pathway in primates, specifically:

[0027] The V1 layer corresponds to the primary visual cortex and extracts basic visual features such as edges and orientation.

[0028] The V2 layer corresponds to the secondary visual cortex, integrating simple features to form more complex shape representations;

[0029] The V4 layer corresponds to the high-level visual area, extracting mid-level features such as shape and texture;

[0030] The IT layer corresponds to the inferior temporal cortex, forming a high-level representation of objects and text.

[0031] Furthermore, a fusion layer is introduced after the IT layer, which integrates the dual-branch features using a concatenation and difference strategy. Specifically, the feature vectors of the two branches are concatenated end to end, and the element-wise difference is calculated and fed into the subsequent decision module. The training adopts supervised multi-task classification, with each discrimination task having a clear binary classification label, and the output is supervised by cross-entropy loss.

[0032] Furthermore, a radical attention branch is added after the V4 layer. The pre-trained component detector selectively responds to Chinese character radicals and high-frequency components, generates an attention heatmap, and applies it to the original feature map through element-wise product, thereby enhancing the representation ability of key component regions in subsequent IT layer processing.

[0033] Furthermore, the four parallel task discrimination branches employed include:

[0034] The phonetic discrimination branch corresponds to the dorsal phonological pathway, extending from the left temporoparietal junction to the inferior frontal gyrus, to determine whether two words are pronounced the same.

[0035] The character shape discrimination branch corresponds to the ventral visual word shape area and determines whether two stimuli are the same Chinese character.

[0036] The component-based character formation branches correspond to the processing path of the left-side middle-angle return shape, determining whether the radical can form part of the whole character;

[0037] Shape similarity branches are used for purely visual graphic judgment and serve as a visual baseline for children with dyslexia.

[0038] Furthermore, a recurrent feedback mechanism with local recurrent connections is introduced into the four-layer recurrent convolutional structure V1→V2→V4→IT, which performs multiple recurrent cycles of information updates for each input in each layer, simulating the feedback processing in the formation of visual consciousness.

[0039] Furthermore, the behavioral data alignment uses the individual's actual reaction as a supervision signal to fit its error patterns and reaction time characteristics; the brain imaging data alignment matches the model's internal representation with the individual's functional magnetic resonance imaging activation patterns; the correspondence between each layer of the Chinese reading and writing digital twin base model and brain regions is as follows: the IT layer corresponds to the left fusiform gyrus visual word shape area, the speech branch hidden layer corresponds to the left temporoparietal junction area and inferior frontal gyrus, the component module corresponds to the left middle frontal gyrus, and the early convolutional layer corresponds to the primary visual cortex.

[0040] Preferably, the personalized digital twin model is validated and its mechanism analyzed, specifically including:

[0041] The accuracy of the personalized digital twin model in predicting individual behavior was verified on untested trials.

[0042] By comparing the similarity between the internal representations of individualized digital twin models and the multivoxel activation patterns of brain regions, and through module ablation and parameter intervention experiments, we simulated specific pathway damage or parameter adjustments in individualized digital twin models, inferred their impact on behavior, and revealed the causal mechanism.

[0043] Preferably, the specific method for recommending the optimal intervention plan is as follows: based on the parameters and representation features of the individualized digital twin model, children with dyslexia are classified into subtypes, and the subtype classifications used include: pure speech mapping disorder, visual attention deficit, and mixed deficit.

[0044] A virtual intervention simulation platform for "silicon-based brain" was established. Through individualized digital twin models, the impact of various hypothetical intervention methods on reading and writing was simulated in a computer. The intervention strategies adopted included speech reinforcement training simulation, visual word form training simulation, and attention regulation training simulation.

[0045] By combining simulations of different intervention strategies, the effectiveness of interventions can be predicted, and the optimal intervention plan can be recommended for individual children with dyslexia.

[0046] On the other hand, the present invention also provides a modeling and intervention simulation system for Chinese dyslexia children based on digital twin brain technology, comprising:

[0047] The data acquisition module is used to collect multimodal data from children with Chinese reading and writing difficulties and normal control children;

[0048] The model building module is used to build a Chinese reading and writing digital twin base model, which includes a visual encoding unit, a task discrimination unit, a component sensitivity unit, and a loop feedback unit.

[0049] The individualized modeling module employs a dual-constraint optimization method combining behavioral data alignment and brain imaging data alignment to fine-tune the Chinese reading and writing digital twin base model, establishing an individualized digital twin model; it includes: a behavioral alignment unit, a brain imaging alignment unit, and a parameter optimization unit.

[0050] The intervention simulation module is used to simulate different intervention strategies and predict intervention effects, including an intervention strategy simulation unit, an effect prediction unit, a subtype classification unit, and a program recommendation unit.

[0051] The output module is used to output individualized diagnostic reports and personalized intervention recommendations.

[0052] The technical solution of the present invention has the following advantages:

[0053] A. This invention is the first to achieve individualized digital twin modeling for Chinese reading and writing difficulties. Unlike previous analyses based on group averages, this invention constructs a unique individual digital twin model for each test subject, which can analyze the individual heterogeneity of reading and writing difficulties.

[0054] B. This invention integrates the dual processing perspectives of consciousness and unconsciousness, revealing the neural mechanisms of dyslexia more comprehensively. By simulating the temporal dynamics of consciousness formation through a cyclical feedback mechanism, it expands the research scope to the long-neglected unconscious level.

[0055] C. This invention establishes a transparent mapping of "brain-behavior-computation" through interpretable network architecture and parameter design. It constructs an interpretable deep network based on neuroscience theory, so that each part of the personalized digital twin model has a clear biological meaning, which makes it easy to map the results back to the brain mechanism.

[0056] D. This invention summarizes the differences in complex brain states through four key neural parameters, which is more explanatory and operable than previous methods that required a large number of parameters to be freely adjusted. By simply adjusting parameters such as global gain, an individualized digital twin model may change from a normal reading and writing mode to a reading and writing difficulty mode.

[0057] E. This invention supports virtual simulation and effect prediction of intervention strategies, providing a scientific basis for personalized educational interventions. It establishes an experimental platform that can simulate interventions in a computer, testing the impact of various hypothetical interventions on reading and writing in a virtual space without any risk and greatly accelerating the development of personalized treatment plans.

[0058] F. This invention has good scalability and can be extended to modeling and intervention research of other learning disabilities, such as writing disorders, reading disorders, and attention disorders. Attached Figure Description

[0059] To more clearly illustrate the specific embodiments of the present invention, the accompanying drawings used in the specific embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0060] Figure 1 This is a flowchart of the modeling and intervention simulation method for children with Chinese reading and writing difficulties provided by this invention;

[0061] Figure 2 This is a diagram of the digital twin base model architecture provided by the present invention;

[0062] Figure 3 This is a flowchart of the individualized modeling and parameter optimization process provided by the present invention;

[0063] Figure 4 This is a schematic diagram of the intervention effect simulation and recommendation system provided by the present invention;

[0064] Figure 5 This is a diagram of the modeling and intervention simulation system provided by the present invention. Detailed Implementation

[0065] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0066] like Figure 1 and Figure 2 As shown, this invention provides a method for modeling and intervening in Chinese dyslexia children based on digital twin brain technology, comprising the following steps:

[0067]

S01

[0068] Behavioral data collection:

[0069] This invention designs four core cognitive tasks to obtain trial-by-trial response data. These include a speech discrimination task, a character shape discrimination task, a component-based character discrimination task, and a shape similarity discrimination task.

[0070] The speech discrimination task presents two Chinese characters, and participants judge whether the two characters are pronounced the same (homophone / heterophone judgment). This task examines the function of the speech processing pathway.

[0071] The character shape discrimination task presents two Chinese characters, and the participants judge whether they are the same Chinese characters (same character / different character judgment). This task examines the visual word shape recognition function.

[0072] The component-based character discrimination task presents a component and a whole character. Participants judge whether the component is a part of the whole character. This task examines the ability to process the morphological structure of Chinese characters and is a cognitive link unique to Chinese reading and writing.

[0073] The shape similarity discrimination task presents two non-textual graphics, and participants judge whether the two graphics are similar. This task serves as a non-verbal visual baseline to separate domain-specific defects.

[0074] The task design covers key cognitive aspects of Chinese reading and writing, from basic visual processing to speech mapping and morphological analysis, and includes a non-verbal visual baseline task to isolate domain-specific deficiencies, and records the accuracy and reaction time for each trial.

[0075] Brain imaging data acquisition:

[0076] This invention preferably uses a 3T MRI scanner to acquire whole-brain BOLD signals. Task-based data acquisition includes the four cognitive tasks mentioned above, as well as text processing tasks under conscious and unconscious conditions. Resting-state data is also acquired to obtain spontaneous brain functional activity.

[0077] This invention focuses on the following regions of interest when acquiring whole-brain BOLD signals:

[0078] Left fusiform visual word form region: responsible for character shape recognition and visual word form representation;

[0079] Left temporoparietal junction: supports speech decoding and phonetic-speech conversion;

[0080] Left inferior frontal gyrus: involved in speech rehearsal, semantic processing, and phono-semantic integration;

[0081] Left midfrontal gyrus: a morphological processing area unique to Chinese reading and writing;

[0082] Primary visual cortex V1 / V2: responsible for basic visual feature extraction.

[0083] Subject information collection:

[0084] The study recorded covariate information including cognitive ability assessment results, literacy test scores, educational background, and family literacy environment. The inclusion criteria for the dyslexia group were based on the DSM-5 diagnostic criteria, requiring significantly lower than age-appropriate literacy skills but normal intelligence, excluding sensory impairments and other neurodevelopmental disorders.

[0085]

S02

[0086] like Figure 2 As shown, this invention designs a bio-inspired deep neural network architecture, the core of which is based on neuroscience theory to construct an interpretable deep network that explicitly encodes known brain functional organization principles into the network structure.

[0087] (a) Siamese dual coding network

[0088] For the paired stimulus discrimination paradigm, a Siamese dual encoding network is constructed. This network consists of two parallel convolutional branches with shared parameters, simulating the mechanism by which the human brain processes two stimuli in parallel and extracts relational features in comparison tasks.

[0089] Each branch employs a four-layer recurrent convolutional structure (V1→V2→V4→IT) to extract hierarchical visual features. This hierarchical structure corresponds to the neural anatomy of the ventral visual pathway in primates. Specifically:

[0090] Layer V1 corresponds to the primary visual cortex, extracting basic visual features such as edges and orientations;

[0091] The V2 layer corresponds to the secondary visual cortex, integrating simple features to form more complex shape representations;

[0092] Layer V4 corresponds to the high-level visual area, extracting mid-level features such as shape and texture;

[0093] The IT layer corresponds to the inferior temporal cortex, forming a high-level representation of objects and text.

[0094] This invention further introduces a fusion layer after the IT layer to integrate the dual-branch features, employing a "concatenation + difference" strategy: the feature vectors of the two branches are concatenated end-to-end, and the element-wise difference is calculated before being fed into the subsequent decision module. Training employs supervised multi-task classification, with each discrimination task having a clear binary classification label, and cross-entropy loss is used to supervise the output. Simultaneously, drawing on metric learning ideas, contrastive loss is added to assist in constraint, encouraging inputs of the same class to be closer in the representation space, thereby improving generalization ability.

[0095] (ii) Multi-task discrimination branch

[0096] After sharing the visual encoding backbone, four parallel task discrimination branches emerge, corresponding to the neural mechanisms by which the brain processes different information pathways in parallel during reading and writing.

[0097] Phonetic discrimination branch: Corresponding to the dorsal speech pathway, from the left temporoparietal junction to the inferior frontal gyrus, it judges whether the pronunciations of two characters are the same. This branch learns the mapping relationship from glyph to rhyme through the hidden layer. During training, a large number of Chinese character pairs with phonetic notations are used as samples for supervision, enabling the Chinese reading and writing digital twin base model to form encodings similar to pinyin tones internally.

[0098] Glyph discrimination branch: Corresponding to the ventral visual word form area, it judges whether two stimuli are the same Chinese character. A trainable fully connected layer is adopted. The input fusion feature contains differential information, and the output is a binary classification probability. The network self-learns which feature differences are most crucial for glyph recognition.

[0099] Component formation branch: Corresponding to the morphological processing pathway in the left middle frontal gyrus, it judges whether a radical can form part of a whole character. A convolutional matching mechanism is adopted: The component features are used as templates to perform cross-correlation operations on the whole character feature map, and the maximum response value is taken as the matching degree. This simulates the brain's ability to deconstruct and analyze the internal structure of Chinese characters.

[0100] Shape similarity branch: Used for pure visual graphic judgment, serving as the visual baseline for reading and writing. This branch uses mid-early features (V4 layer) for discrimination, corresponding to the processing mainly relying on the primary visual cortex and parietal attention areas when the human brain quickly and roughly distinguishes shapes, without delving into semantic details.

[0101] This multi-task design has multiple advantages: Parameter sharing reduces the risk of overfitting; The gradients of different tasks jointly shape the visual representation, making it more general and abstract; It promotes functional differentiation, and the hidden node groups behind different branches perform their respective functions, corresponding to the functional differentiation pattern of the reading and writing brain regions.

[0102] (III) Chinese component-sensitive module

[0103] In view of the unique component structure of Chinese characters, the present invention further adds a radical attention branch after the V4 layer.

[0104] First, a component detector is pre-trained on an image library of 214 radicals and high-frequency components, enabling the convolutional kernel to obtain a selective response to the shape of the target component. For example, a certain filter specifically responds to "氵", and another filter specifically responds to "口", etc.

[0105] Then these filters are applied to the whole character feature map output by the V4 layer, generating high responses at the matching positions. The response mappings of all component filters are combined into an attention heat map, which acts on the original feature map through element-wise multiplication. When the attention weight is high, it indicates that there is a component pattern at that place, and the original feature is amplified, making the key component area more prominent in the subsequent IT layer processing.

[0106] The biological rationale for this design lies in the fact that when humans learn Chinese characters, they consciously master radicals and other components to aid in character recognition, and the brain accordingly forms selective activation pathways for these components. Compared to simply deepening the network, this architectural modification based on cognitive priors more closely resembles the actual neural mechanisms of Chinese reading and writing, and only slightly increases the number of parameters and computational load.

[0107] (iv) Circular Feedback Mechanism

[0108] This invention introduces local cyclic connections at each layer, allowing information to be processed iteratively within the same layer, simulating the temporal dynamics of visual awareness formation. For each input, the Chinese reading and writing digital twin base model performs multiple cyclic information updates at each layer, rather than a single-path feedforward. For example, after receiving input, layer V1 generates an initial activation, then corrects the activation by combining it with its previous state through feedback. After several iterations, the steady-state output is passed to layer V2; layer V2 also iterates, and so on.

[0109] The cognitive significance of cyclic processing lies in the fact that the first feedforward scan only yields a rough representation, requiring a feedback loop to focus on details and eliminate ambiguity. This corresponds to the classic theory that "without feedback, there is no consciousness," and the generation of visual consciousness depends on the repeated interaction of bottom-up and top-down signals. This invention employs a cyclic mechanism to enable the Chinese reading and writing digital twin base model to have a stronger ability to recognize complex Chinese characters, and can record the number of iterations required to achieve stable output under different tasks, analogous to the speed at which recognition enters consciousness.

[0110]

S03

[0111] like Figure 3 As shown, individualized modeling is performed based on the constructed Chinese reading and writing digital twin foundation model, making the individualized digital twin model a digital twin of each child's brain. Specifically, this includes dual-constraint optimization and the introduction of key neural parameters.

[0112] (I) Dual Constraint Optimization

[0113] This invention takes into account the dual constraints of behavioral and brain imaging data, so that every change in the internal state of the individualized digital twin model corresponds to a synergistic change in behavioral output and brain activation, thus achieving a unified interpretation of brain-behavior.

[0114] Behavioral data alignment: Subject responses are used as supervisory signals. For each stimulus pair experienced by the child, the individualized digital twin model calculates the output probability of the corresponding task branch with the same input. This invention uses the child's actual response, rather than the objective correct answer, as the supervisory signal, enabling the individualized digital twin model to reproduce the child's error patterns. The loss function is the standard binary classification cross-entropy.

[0115] Furthermore, this invention utilizes reaction time data to calibrate the model's decision-making process: an auxiliary loss is introduced to encourage a negative correlation between the model's output confidence and reaction time; a low confidence indicates a slow reaction. This fits both right / wrong binary behavior and partially fits the continuous characteristics of the behavior.

[0116] Brain imaging data alignment: activation patterns of regions of interest related to reading and writing tasks were extracted from the functional magnetic resonance imaging data of each subject to constrain the internal representation of individualized digital twin models.

[0117] A correspondence was established between model layers and brain regions: the IT layer corresponds to the left fusiform gyrus visual word-form area, the phonological branch hidden layer corresponds to the left temporoparietal junction and inferior frontal gyrus, the component module corresponds to the left middle frontal gyrus, and the early convolutional layer corresponds to the primary visual cortex. For each region of interest, the average BOLD change relative to the baseline under each task condition was calculated, and the brain alignment loss was constructed as a weighted sum of terms for each region of interest. By minimizing the brain alignment loss, the internal representation of the individualized digital twin model was guided to converge with the subject's brain representation.

[0118] The overall objective function is the weighted sum of behavioral loss and brain alignment loss. A phased training approach is adopted: first, behavior is given higher weight to ensure the individualized digital twin model "learns the task," and then the weight of brain constraints is gradually increased to fine-tune the internal representation, avoiding forced fitting of brain signals at the beginning that could disrupt task learning.

[0119] (II) Key Neural Parameters

[0120] This invention further introduces four global neurophysiological parameters to summarize differences in brain states:

[0121] Global neural gain G: Characterizes the balance between excitability and inhibition in the cortex. It is achieved by multiplying the output of each convolutional layer by a coefficient G. A G greater than 1 indicates amplification of the output of all layers, corresponding to over-excitation; a G less than 1 indicates overall suppression, corresponding to enhanced inhibition. Studies have found that children with dyslexia have higher glutamate concentrations in the visual cortex, suggesting that over-excitation may impair signal accuracy. The level of G is determined based on data: if a child's brain imaging shows abnormally high amplitude BOLD fluctuations, the individualized digital twin model simulates this by increasing G; conversely, G is decreased.

[0122] Internal noise σ simulates the inherent random fluctuations of neural activity. Gaussian noise with a mean of 0 and a variance of σ² is added after each convolutional layer calculation. The larger the σ, the lower the discrimination stability, corresponding to the "intermittent good and bad" behavioral variations in children with dyslexia. Physiologically, it may originate from increased background discharge due to insufficient inhibitory function, or from visual signal temporal disorder caused by weak macrocellular pathway function.

[0123] The cyclic feedback gain R determines the strength of the feedback loop and corresponds to the weighting of the top-down signal's influence on perceptual processing. It is achieved by multiplying the weights of each layer's cyclic connection by a coefficient R. R less than 1 indicates weakened feedback, tending towards single-feedforward; R greater than 1 indicates amplified feedback, requiring repeated adjustments to stabilize. Changes in R affect performance: excessively low feedback increases errors under difficult conditions (lack of higher-level guidance), while excessively high feedback introduces spurious signals and confusion.

[0124] Neuronal activation threshold T: This is the minimum input intensity required for a neuron to transition from a resting state to an activated state, and is a key determinant of the excitability of the nervous system. It is achieved by adjusting the threshold parameter of the activation function, for example, using a thresholded ReLU variant. When T is high, neurons require stronger input to produce a response, corresponding to a "sluggish response" or "difficult activation" state; when T is low, neurons are easily activated, corresponding to an "excited" or "oversensitive" state.

[0125] Compared to thousands of network weights, the above four parameters are extremely simple and easy to assign clear physiological meanings. Optimization employs global search methods such as Bayesian optimization or genetic algorithms, using the total loss of each individual as the evaluation function to search the parameter space for the optimal combination of (G, σ, R, T).

[0126] (III) Layered Training Strategy

[0127] This invention employs a three-stage training approach to fully utilize available information and reduce the risk of overfitting:

[0128] Phase 1 – Basic Visual Pre-training: A Chinese reading and writing digital twin base model was pre-trained on a large-scale dataset of 3000 commonly used Chinese characters, enabling it to perceive general Chinese characters. A character recognition task was employed, with the IT layer followed by Softmax classification. After dozens of training rounds, an accuracy rate exceeding 90% was achieved, indicating that the Chinese reading and writing digital twin base model has learned the hierarchical features from low-level strokes and mid-level components to high-level whole characters.

[0129] Phase 2 – Multi-Task Joint Training (Population Baseline Model): The Chinese reading and writing digital twin baseline model is fine-tuned based on behavioral and functional magnetic resonance imaging (fMRI) data from normal children, enabling them to initially master four discrimination tasks. The training objective is to minimize total loss. At this stage, brain alignment loss is based on the average brain activation of the normal group, and behavioral loss is based on the average behavioral patterns of the normal group. Task-specific batch normalization or dynamic weight adjustment is used to balance the losses of different tasks. The result is a "population average model," whose behavioral performance and internal activation are consistent with the average of typical readers and writers, and can be regarded as a computer mapping of the brain of a normal reader and writer.

[0130] Phase 3 – Individual Fine-tuning (Digital Twin Model): Fine-tuning is performed on data from each child with dyslexia. A small step size and few iterations are used to prevent overfitting. Most early layers are frozen (low-level visual features should be similar for all children), and only the weights of higher layers and task branches are adjusted. Parameters (G, σ, R, T) are updated simultaneously. Multiple fine-tuning sessions are conducted for each child to find the optimal result: starting with the group model, training is performed separately with different initial parameters, and the individualized digital twin model with the lowest validation set loss is selected.

[0131] [S04] Verify and analyze the mechanism of the personalized digital twin model to reveal the causal mechanism of reading and writing difficulties.

[0132] This invention designs a multi-level verification and analysis system to evaluate the effectiveness of digital twin models.

[0133] (a) Validation of model prediction capability

[0134] Cross-validation predictive ability: A portion of trials are removed from each child's data and not used for fine-tuning. The fine-tuned model is then used to predict the results of these unseen trials, and the trial-by-trial matching rate and Pearson correlation are calculated. If the validation performance significantly exceeds the random level, it indicates that the model has learned stable behavioral patterns in children rather than simply memorizing them.

[0135] Similarity of brain activation patterns: Compare the correlation between the model and the brain's activation distribution on unseen stimuli. For repeated stimuli or probe stimulation during the scan, record the activation of the child's brain regions, use the model to predict the activation of the corresponding layers, and calculate Spearman's rank correlation or Pearson's correlation. If the correlation remains significantly positive, it indicates that the model has successfully captured the child's brain's response patterns to a wide range of stimuli. Also check the reproduction of differences at the group level: For example, if real data shows that the activation of the left visual word shape area in the dyslexia group is 20% lower than that in the normal group, does the average activity of the IT layer in the dyslexia model also decrease by approximately 20%?

[0136] Representation similarity analysis: A set of stimuli is selected, and the activation vectors of these stimuli in a hidden layer of the model are collected. The pairwise distance matrix between the stimuli is then calculated. Similarly, the distance matrix is ​​calculated for the multi-voxel responses of the corresponding brain regions of the subjects. If the model is an effective simulation of real brain representations, the two distance matrices should be statistically correlated in key brain regions. Representation similarity analysis can go beyond simple activation intensity analysis and test the consistency of high-dimensional representation spaces.

[0137] (II) Mechanism Analysis and Causal Inference

[0138] Module ablation experiments: Intervention experiments were conducted on well-fitting models to explore the causal effects of each pathway on reading and writing behavior. For example, the output of the speech discrimination branch was forcibly set to random (simulating Broca's area loss of function), and the accuracy of homophony tasks was tested to see if it significantly decreased. If the decrease was comparable to the impact on real patients with impaired speech pathways, it would support the importance of this module. Similarly, the enhancement effect of the lateral attention module was temporarily set to 0 to observe its impact on component tasks.

[0139] Parametric intervention simulation: This simulation tests whether reducing the internal noise σ of the model improves performance, providing a "noise reduction" intervention effect for children with dyslexia. Alternatively, it might ask, "Does appropriately reducing the global gain G to avoid overactivation improve the accuracy of distinguishing different characters?" Systematically adjusting parameters to evaluate the impact on performance helps identify which factors have the greatest influence on dyslexia. For example, if increasing the feedback gain R helps a certain type of model correct errors, it suggests that these children can compensate by improving their attention regulation abilities through training.

[0140] Comparison of representation distributions: In-depth analysis of the differences in representation patterns between the dyslexia group and the normal group models. Examination of the distribution variance and signal-to-noise ratio of hidden layer activations: If the hidden layer distance of the dyslexia model is smaller (more overlapping representations) in distinguishing similar-looking characters and homophones, it indicates more chaotic information encoding and decreased discriminability, consistent with the hypothesis that "excessive noise / excitability leads to representation confusion." Comparison of learning rates: If the number of training iterations required for the dyslexia model to reach a certain accuracy is several times that of the normal model, it will simulate the "more effort, less reward" situation in their real learning.

[0141]

S05

[0142] like Figure 4 As shown, intervention simulation and personalized recommendations are based on an individualized digital twin model.

[0143] (a) Subtype classification

[0144] Based on in-depth analysis of model parameters and internal representations, fine-grained subtype classification of children with dyslexia is achieved. For example:

[0145] Pure speech mapping disorder type: mainly manifested by abnormal speech branch parameters, significant impairment in the word-sound discrimination task, while other tasks are relatively preserved;

[0146] Visual attention deficit type: mainly manifested as insufficient function of component-sensitive modules, and significant impairment in component character formation and character shape recognition tasks;

[0147] Mixed defect type: Multiple pathways are abnormal at the same time, resulting in comprehensive impairment of multiple tasks.

[0148] This classification system provides a direct theoretical basis for subsequent precision teaching and rehabilitation intervention.

[0149] (ii) Virtual intervention simulation

[0150] Establish a "silicon-based brain" virtual intervention simulation platform to simulate the impact of various hypothetical intervention methods on reading and writing in a computer through a digital twin model.

[0151] Speech reinforcement training simulation: In the model, it can be simulated to strengthen the weight of speech branches, observe whether it makes up for the defects in speech path, and guide the intensity of practice for pinyin-Chinese character conversion in teaching.

[0152] Visual word form training simulation: In the model, the weights of the visual encoding backbone and word form branches can be simulated to observe the changes in word form recognition accuracy.

[0153] Attention regulation training simulation: In the model, this is reflected in increasing the feedback gain R and reducing the noise σ, and observing whether the recognition error rate decreases and the internal representation improves.

[0154] By systematically scanning different combinations of intervention parameters, the expected effects of various intervention strategies can be quantified.

[0155] (III) Recommendation of Personalized Solutions

[0156] Based on the model parameters before the scan, predict which intervention is more suitable for a child. For example, if the model shows that the main problem is in morphological processing, then speech training may not be very effective, and morphological training should be prioritized. Figure 5 As shown, the system automatically generates personalized auxiliary diagnostic reports that match the subtype diagnosis, and provides prognostic assessment and long-term follow-up functions, forming a complete closed loop from diagnosis to intervention and assessment.

[0157] In addition, the present invention provides a modeling and intervention simulation system for Chinese dyslexia children based on digital twin brain technology, including: a data acquisition module, a model building module, an individualized modeling module, an intervention simulation module, and an output module.

[0158] The data acquisition module is used to collect multimodal data from children with Chinese reading and writing difficulties and normal control children, including behavioral data obtained through four cognitive tasks and brain imaging data obtained through functional magnetic resonance imaging.

[0159] The model building module is used to construct a digital twin base model for Chinese reading and writing, including a visual encoding unit, a task discrimination unit, a component-sensitive unit, and a recurrent feedback unit. The visual encoding unit uses a Siamese dual encoding network structure to achieve parallel processing of paired stimuli; the task discrimination unit contains four parallel discrimination branches to achieve functional splitting; the component-sensitive unit enhances the representation ability of Chinese character structure through a pre-trained component detector; and the recurrent feedback unit introduces local recurrent connections to simulate the consciousness formation process.

[0160] The individualized modeling module is used for individualized modeling and includes a behavior alignment unit, a brain imaging alignment unit, and a parameter optimization unit. The behavior alignment unit fits error patterns using the individual's actual response as a supervisory signal; the brain imaging alignment unit matches the model representation with brain activation patterns; and the parameter optimization unit optimizes four key parameters: global neural gain, internal noise, recurrent feedback gain, and neuron activation threshold.

[0161] The intervention simulation module is used to simulate intervention effects and recommend solutions, including an intervention strategy simulation unit, an effect prediction unit, a subtype classification unit, and a solution recommendation unit.

[0162] The output module is used to generate individualized diagnostic reports and personalized intervention recommendations.

[0163] This invention is the first to achieve individualized digital twin modeling for Chinese reading and writing difficulties. Unlike previous analyses based on group averages, it constructs a unique individual digital twin model for each test subject, which can analyze the individual heterogeneity of reading and writing difficulties.

[0164] Any aspects not covered in this invention are applicable to the prior art.

[0165] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for modeling and intervening in Chinese dyslexia children based on digital twin brain technology, characterized in that, Includes the following steps: Collect behavioral and brain imaging data of children with Chinese reading and writing difficulties and normal control children; A digital twin base model for Chinese reading and writing was constructed, including a Siamese dual coding network, a multi-task discriminative branch, a Chinese component sensitive module, and a loop feedback mechanism, to simulate the functional diversion and dynamic formation of consciousness in the brain during the reading and writing process. A dual-constraint optimization method, combining behavioral data alignment and brain imaging data alignment, was used to fine-tune the Chinese reading and writing digital twin base model, thereby establishing an individualized digital twin model. To verify and analyze the mechanism of personalized digital twin models, revealing the causal mechanism of reading and writing difficulties; Based on the individualized digital twin model, different intervention strategies are simulated to predict their effects on improving individual reading and writing behavior and brain representation, and personalized intervention plans are recommended for individuals. The verification and mechanism analysis of the personalized digital twin model specifically includes: The accuracy of the personalized digital twin model in predicting individual behavior was verified on untested trials. By comparing the similarity between the internal representations of individualized digital twin models and the multivoxel activation patterns of brain regions, and through module ablation and parameter intervention experiments, we simulated specific pathway damage or parameter adjustments in individualized digital twin models, inferred their impact on behavior, and revealed the causal mechanism.

2. The method for modeling and intervening in children with Chinese dyslexia according to claim 1, characterized in that, The behavioral data is obtained through trial-by-trial processing of speech discrimination tasks, character shape discrimination tasks, component character formation discrimination tasks, and shape similarity discrimination tasks, representing reaction time and accuracy. The brain imaging data consists of task-state and resting-state whole-brain BOLD signals acquired using functional magnetic resonance imaging. The key regions of interest include the left fusiform gyrus visual word shape region, the left temporoparietal junction region, the inferior frontal gyrus, the middle frontal gyrus, and the primary visual cortex. These key regions of interest are used for alignment constraints within the internal representation of the Chinese reading and writing digital twin base model.

3. The method for modeling and intervening in children with Chinese dyslexia according to claim 1, characterized in that, In the process of establishing a digital twin base model for Chinese reading and writing, four key neurophysiological parameters are introduced: global neural gain G, internal noise σ, cyclic feedback gain R, and neuron activation threshold T. A global search method using Bayesian optimization or genetic algorithm is adopted, with the total individual loss as the evaluation function, to find the optimal combination of (G, σ, R, T) in the parameter space. In the Chinese reading and writing digital twin base model, the global neural gain G is used to characterize the balance between cortical excitability and inhibition, and is achieved by multiplying by a coefficient after each layer of convolution output; The internal noise σ is simulated by adding Gaussian noise after each layer of convolution calculation to simulate the inherent random fluctuations of neural activity. The recurrent feedback gain R is determined by multiplying the weights of each layer of recurrent connections by a coefficient to determine the strength of the feedback loop. The neuron activation threshold T controls the minimum input intensity required for a neuron to transition from a resting state to an activated state by adjusting the threshold parameter of the activation function.

4. The method for modeling and intervening in children with Chinese dyslexia according to claim 1, characterized in that, The specific method for constructing a digital twin base model for Chinese reading and writing is as follows: For the pairwise stimulus discrimination paradigm, a Siamese dual coding network is constructed; A parallel convolutional branch structure with two shared parameters is adopted to simulate the parallel processing mechanism of paired stimuli in the human brain and to share the visual coding backbone. Each branch uses a four-layer recurrent convolutional structure (V1→V2→V4→IT) to extract hierarchical visual features. A fusion layer is introduced after the IT layer, and a splicing and differential strategy is used to integrate the dual-branch features. After sharing the visual encoding backbone, four parallel task discrimination branches emerge, corresponding to the neural mechanisms by which the brain processes different information pathways in parallel during reading and writing.

5. The method for modeling and intervening in children with Chinese dyslexia according to claim 4, characterized in that, The aforementioned four-layer recurrent convolutional structure V1→V2→V4→IT corresponds to the neural anatomy of the ventral visual pathway in primates, specifically: The V1 layer corresponds to the primary visual cortex and extracts basic visual features such as edges and orientation. The V2 layer corresponds to the secondary visual cortex, integrating simple features to form more complex shape representations; The V4 layer corresponds to the high-level visual area, extracting mid-level features such as shape and texture; The IT layer corresponds to the inferior temporal cortex, forming a high-level representation of objects and text.

6. The method for modeling and intervening in children with Chinese dyslexia according to claim 4, characterized in that, A fusion layer is introduced after the IT layer, which integrates the dual-branch features using a concatenation and difference strategy. Specifically, the feature vectors of the two branches are concatenated end to end, and the element-wise difference is calculated and fed into the subsequent decision module. The training adopts supervised multi-task classification, with each discrimination task having a clear binary classification label, and the output is supervised by cross-entropy loss.

7. The method for modeling and intervening in children with Chinese dyslexia according to claim 6, characterized in that, A radical attention branch is added after the V4 layer. The pre-trained component detector selectively responds to Chinese character radicals and high-frequency components, generates an attention heatmap, and applies it to the original feature map through element-wise product, thereby enhancing the representation ability of key component regions in subsequent IT layer processing.

8. The method for modeling and intervening in children with Chinese dyslexia according to claim 4, characterized in that, The four parallel task discrimination branches employed include: The phonetic discrimination branch corresponds to the dorsal phonological pathway, extending from the left temporoparietal junction to the inferior frontal gyrus, to determine whether two words are pronounced the same. The character shape discrimination branch corresponds to the ventral visual word shape area and determines whether two stimuli are the same Chinese character. The component-based character formation branches correspond to the processing path of the left-side middle-angle return shape, determining whether the radical can form part of the whole character; Shape similarity branches are used for purely visual graphic judgment and serve as a visual baseline for children with dyslexia.

9. The method for modeling and intervening in children with Chinese dyslexia according to claim 4, characterized in that, A recurrent feedback mechanism with local recurrent connections is introduced into the four-layer recurrent convolutional structure V1→V2→V4→IT. For each input, information updates are performed in multiple recurrent cycles in each layer to simulate the feedback processing process in the formation of visual consciousness.

10. The method for modeling and intervening in children with Chinese dyslexia according to claim 1, characterized in that, The behavioral data alignment uses the individual's actual reaction as a supervision signal to fit its error patterns and reaction time characteristics; The brain imaging data alignment matches the model's internal representations with individual functional magnetic resonance imaging activation patterns, establishing the correspondence between each layer of the Chinese reading and writing digital twin base model and brain regions. Specifically, the IT layer corresponds to the left fusiform gyrus visual word shape area, the speech branch hidden layer corresponds to the left temporoparietal junction area and inferior frontal gyrus, the component module corresponds to the left middle frontal gyrus, and the early convolutional layer corresponds to the primary visual cortex.

11. The method for modeling and intervening in children with Chinese dyslexia according to claim 1, characterized in that, The specific methods for recommending personalized intervention programs to individuals are: Based on the parameters and representational features of the individualized digital twin model, children with dyslexia are classified into subtypes. The subtypes used include: pure speech mapping disorder, visual attention deficit, and mixed deficit. A virtual intervention simulation platform for "silicon-based brain" was established. Through individualized digital twin models, the impact of various hypothetical intervention methods on reading and writing was simulated in a computer. The intervention strategies adopted included speech reinforcement training simulation, visual word form training simulation, and attention regulation training simulation. By combining simulations of different intervention strategies, the effectiveness of interventions can be predicted, and the optimal intervention plan can be recommended for individual children with dyslexia.

12. A modeling and intervention simulation system for Chinese dyslexia children based on digital twin brain technology, characterized in that, The system includes: The data acquisition module is used to collect multimodal data from children with Chinese reading and writing difficulties and normal control children; The model building module is used to build a Chinese reading and writing digital twin base model, which includes a visual encoding unit, a task discrimination unit, a component sensitivity unit, and a loop feedback unit. The individualized modeling module employs a dual-constraint optimization method of behavioral data alignment and brain imaging data alignment to fine-tune the Chinese reading and writing digital twin base model, thereby establishing an individualized digital twin model. The intervention simulation module is used to simulate different intervention strategies and predict intervention effects; The output module is used to output individualized diagnostic reports and personalized intervention recommendations; The verification and mechanism analysis of the personalized digital twin model specifically includes: The accuracy of the personalized digital twin model in predicting individual behavior was verified on untested trials. By comparing the similarity between the internal representations of individualized digital twin models and the multivoxel activation patterns of brain regions, and through module ablation and parameter intervention experiments, we simulated specific pathway damage or parameter adjustments in individualized digital twin models, inferred their impact on behavior, and revealed the causal mechanism.

Images