An individualized social skill training system for autism

Abstract

This invention relates to the field of autism spectrum disorder technology, specifically to a personalized social skills training system for autism, comprising a multimodal data acquisition module, a personalized cognitive state assessment module, a parameterized social scenario generation module, an adaptive decision control module, and a closed-loop neural modulation module. The system synchronously acquires the user's electroencephalogram (EEG), eye movements, and other physiological signals, calculates quantitative indicators such as the cognitive load index in real time, and dynamically generates adjustment instructions for the difficulty parameters of virtual social scenarios using methods such as model predictive control, so as to match the training task with the user's real-time cognitive load. Simultaneously, the system can trigger non-invasive brain stimulation synchronized with the user's own EEG phase when specific neural oscillation characteristics are detected. This invention achieves personalized, adaptive social skills training based on real-time feedback of intrinsic physiological state, and also possesses the ability to target and modulate abnormal neural activity patterns.

Description

Technical Field

[0001] This invention relates to the field of autism spectrum disorder technology, specifically to a personalized social skills training system for autism. Background Technology

[0002] Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder, one of the core symptoms of which is persistent impairment in social communication and interaction. These impairments are widespread across multiple areas, including social-emotional interaction, nonverbal communication behaviors, and the development, maintenance, and understanding of interpersonal relationships. Traditional social skills interventions, such as behavior analysis, structured instruction, and social stories, while achieving some success in improving specific behaviors, generally suffer from limitations such as rigid training patterns, difficulty in generalization, and inability to adapt to significant individual differences. These methods largely rely on external behavioral observation and subjective reports, lacking real-time, objective quantification of the user's internal cognition and neural state, thus hindering truly personalized and adaptive training.

[0003] With the development of technology, computer-assisted intervention and virtual reality technology have been introduced into the field of autism rehabilitation, providing a controllable, safe, and repeatable simulated environment for social skills training. These systems can present rich social scenarios, allowing users to practice without real-world pressure. However, most existing technology-driven solutions still have significant shortcomings. Their training content and difficulty are usually preset and fixed, or can only be adjusted in simple, step-by-step manner based on the user's behavior, failing to respond in real-time to the user's fluctuating internal cognitive load and emotional state during training. This one-size-fits-all or semi-static adjustment mode may result in training tasks that are either too simple to promote learning or too difficult to cause anxiety, frustration, or even behavioral problems, thus weakening the intervention effect. A deeper bottleneck lies in the lack of an endosensory channel in existing systems that can accurately perceive and interpret the user's neurophysiological state. Behind the social difficulties of individuals with autism, there are often specific neural activity patterns, such as abnormal neural oscillations, low efficiency in sensory information processing, or an imbalance in the allocation of cognitive resources when faced with social information. Traditional systems cannot capture these latent, real-time neurophysiological indicators, thus their interactions are open-loop, meaning the training content output by the system cannot be finely controlled in a closed-loop manner based on the user's real-time brain and body state. This limits the system's ability to deeply understand the root causes of user challenges, prevent overload during training, and intervene in the most favorable neural states. Summary of the Invention

[0004] The purpose of this invention is to provide a personalized social skills training system for autism. By collecting and analyzing users' physiological signals such as EEG in real time to quantify their cognitive state, the system dynamically adjusts the difficulty parameters of virtual social situations based on this data. At the same time, it can provide closed-loop synchronous stimulation of specific neural oscillations to achieve intervention that combines personalized adaptive training with neuromodulation.

[0005] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution:

[0006] A personalized social skills training system for autism includes:

[0007] The multimodal data acquisition module is used to simultaneously acquire the user's multi-channel EEG signals, eye-tracking data, and peripheral physiological signals;

[0008] A personalized cognitive state assessment module is connected to the data acquisition module and is used to calculate at least one quantitative cognitive state index in real time based on the acquired physiological signals.

[0009] The parameterized social context generation module is used to present virtual social interaction scenarios with adjustable difficulty parameters;

[0010] An adaptive decision control module is connected to the personalized cognitive state assessment module and the parameterized social context generation module, respectively, and is used to generate real-time adjustment instructions for the difficulty parameter by solving a constrained optimization problem based on the quantified cognitive state index.

[0011] The closed-loop neuromodulation module, connected to the multimodal data acquisition module, is used to trigger non-invasive brain stimulation that is phase-synchronized with the user's own brainwave oscillations when a neural oscillation feature that conforms to a preset pattern is detected.

[0012] Furthermore, the personalized cognitive state assessment module calculates a quantitative cognitive state index called the cognitive load index CL(t), which is calculated using the following formula:

[0013]

[0014] in, Indicates within the time window Internally, from the prefrontal cortex region of the brain Extracted from EEG signals acquired by individual electrode channels Frequency band (center frequency) Average power, calculated using the following formula:

[0015]

[0016] This indicates that within the same time window, brain regions located in the parietal or occipital lobes... Extracted from EEG signals acquired by individual electrode channels Frequency band (center frequency) Average power, calculated using the following formula:

[0017]

[0018] The length of the integration time window, with values ​​ranging from... Instant In a matter of seconds; and These are the time-frequency representations of the EEG signals for the corresponding channels and frequency bands.

[0019] Furthermore, the constrained optimization problem solved by the adaptive decision control module is a finite-time model predictive control problem, mathematically expressed as follows:

[0020]

[0021] Subject to

[0022]

[0023]

[0024]

[0025]

[0026]

[0027] in, For continuous time, For discrete-time indexing; The predicted time domain length is a positive integer; For at any time right The predicted value of the system state at any given time, whose state vector at least includes the cognitive load index. ; The control input sequence to be optimized, whose elements correspond to the difficulty parameters; This is a reference value for the target state. and It is a symmetric positive definite weight matrix; and A discrete-time system matrix describing the dynamic state; These are the upper and lower bounds of the state constraints; To control the upper and lower bounds of input constraints; The upper and lower bounds are used to control the rate of change of the input.

[0028] Furthermore, the neural oscillation characteristics of the preset mode detected by the closed-loop neural modulation module are as follows: -γ trans-band phase amplitude coupling events, their coupling strength Calculated using the following formula:

[0029]

[0030] in, Represents at discrete time points Extracted from EEG signals of specific brain regions (such as the prefrontal cortex or the temporoparietal junction). Frequency band (center frequency) The instantaneous amplitude envelope (Hz); Indicates at the same point in time Extracted from EEG signals of relevant brain regions Frequency band (center frequency) The instantaneous phase (Hz); Time window used for calculation Number of sampling points within, The length of the time window; when The value exceeds the preset threshold (in When a valid event is detected, a stimulus is triggered.

[0031] Furthermore, the non-invasive brain stimulation applied by the closed-loop neuromodulation module is transcranial alternating current stimulation, and its output current waveform... for:

[0032]

[0033] in, The peak amplitude of the stimulation current ranges from 0.5 mA to 2.0 mA. The stimulation frequency is selected from values ​​related to the target nerve oscillation frequency band. Hz; It is an adjustable fixed phase offset; The dynamic phase adjustment amount, based on real-time EEG phase feedback, is used to achieve phase synchronization between stimulation and EEG. Its adjustment strategy is as follows:

[0034]

[0035] in This is the proportional gain coefficient. For the target phase to be locked, This refers to the instantaneous phase of the target frequency band as measured in real time.

[0036] Furthermore, the target state reference value Cognitive load target component It is dynamically updated based on the user's individual historical training data. The update rule is based on a performance-load fitting model, specifically:

[0037]

[0038] in, Forgetting factor, satisfying The number of historical task types; For the first Performance rating for each type of task and The performance scores for this type of task vary with cognitive load. The mean and variance of the fitted Gaussian distribution; For the first Weight coefficients for task categories; This represents a feasible search interval for cognitive load. The mean is variance is The Gaussian probability density function.

[0039] Furthermore, it also includes a skills transfer graph management module, used to model the transfer relationships between social skills, with the graph represented as a directed graph. , where vertex set represent A different social sub-skill, edge set The weight matrix represents the potential transfer paths between skills. Each element Indicates from skills Training to skills Positive migration strength; weight Perform Bayesian updates based on user training data:

[0040]

[0041] in, For the observed user training data, For a given migration strength Data observed at the time The likelihood function, the likelihood function is based on skill Performance improvement rate and The linear or nonlinear relationship model between them is defined.

[0042] Furthermore, the parameterized social context generation module provides fine-grained control over the facial expressions of virtual characters by adjusting the parameter vectors based on the facial motion coding system. Implementation, where each element Corresponding to the The activation intensity of each facial movement unit; for expressing a specific target emotion. The parameter vector is generated by the following formula:

[0043]

[0044] in, A parameter vector for neutral facial expressions; For target emotions The template parameter vector; [0, 1] represents the global facial expression intensity adjustment parameter, which is determined by the output of the adaptive decision control module; The coefficient of natural variability in facial expressions; Each component is independent and identically distributed and follows a standard normal distribution. Random vectors are used to introduce natural, subtle changes in facial expressions.

[0045] The beneficial effects of this invention are:

[0046] Traditional methods rely on post-training behavioral assessment, resulting in significant delays in intervention and adjustment. This invention utilizes a personalized cognitive state assessment module to calculate the cognitive load index CL(t) in real time. The numerator of the index... Power derived from the theta band in the prefrontal cortex is closely related to executive control and cognitive effort; denominator Originating from the alpha band power of the parietal-occipital lobe and associated with sensory gating function, this study specifically characterizes the imbalance between excessive cognitive control resources and insufficient sensory input inhibition commonly observed in autistic individuals when processing social information. The adaptive decision control module uses this continuously quantified physiological state as the system state variable and embeds it within the framework of a finite-time model predictive control problem. By solving this constrained optimization problem online, the system can proactively generate the optimal adjustment sequence for the difficulty parameters of virtual social situations, aiming to smoothly and stably approach the individualized target reference value for the predicted cognitive load state in the future time domain. This achieves dynamic matching between training challenge and the user's neurocognitive load level, ensuring the training process remains within the user's zone of proximal development, effectively avoiding anxiety and avoidance caused by overly difficult tasks, or low learning efficiency caused by overly easy tasks.

[0047] This invention uses a closed-loop neural modulation module to continuously calculate the θ-γ cross-band phase amplitude coupling strength extracted from signals in brain regions such as the prefrontal cortex. When the coupling strength exceeds a threshold set based on the individual's resting baseline, it indicates a neural event that may be associated with local network overexcitation and difficulty in cross-brain region integration. At this point, the system triggers transcranial alternating current stimulation, and the stimulation current waveform... The phase of the signal is dynamically synchronized with the instantaneous phase of the user's target frequency band EEG oscillations through a proportional control strategy. This gently modulates abnormal neural oscillation rhythms and cross-frequency coupling relationships in a manner consistent with physiological rhythms, potentially optimizing the collaborative efficiency of neural networks in processing social information. This signifies a shift in intervention from simply adapting to the external environment to providing real-time, targeted guidance of the internal neural computational state, offering a parallel means of promoting neuroplasticity for behavioral correction.

[0048] Existing virtual reality training systems often employ finite, discrete difficulty levels, resulting in stiff and unnatural adjustments. This invention's parameterized social context generation module maps key social cues into a continuously adjustable parameter space. The virtual character's facial expressions are generated by parameter vectors based on a facial motion coding system. Driven, and through formula The generation process is achieved by directly deriving the expression intensity adjustment parameter λ from the optimized output of the adaptive decision control module, enabling a seamless transition between neutral and target emotional expressions. Simultaneously, a random perturbation term is introduced. This imbues facial expressions with natural, subtle changes. Based on real-time user status assessments, complex and intense facial expressions can be softened into more easily interpretable simplified versions, or conversely, the challenge can be increased, achieving millisecond-level continuous adjustment of the intensity of social emotional cues. Similarly, parameters such as the semantic complexity of the dialogue can also be continuously adjusted. This refined generation capability based on a continuous parameter space enables the system to create highly dynamic, natural, and precisely matched personalized social scenarios that match the user's cognitive load, ensuring the safety and tolerability of the intervention process while maintaining the validity of the training ecosystem.

[0049] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Detailed Implementation

[0050] The technical solutions in the embodiments of the present invention will be clearly and completely described below. 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.

[0051] Example 1

[0052] The personalized social skills training system for autism described in this embodiment includes: a multimodal data acquisition module specifically implemented by using a multi-channel physiological signal synchronous acquisition instrument with an integrated high-precision clock source to provide a unified time reference for all sensing sub-modules.

[0053] In this embodiment, as an example, EEG signal acquisition can be performed using a 32-lead electrode cap conforming to the 10-20 international standard. This is used to calculate P. θ The prefrontal cortex electrodes for calculating Pα(t) exemplarily include Fp1, Fpz, Fp2, etc.; the parieto-occipital cortex electrodes for calculating Pα(t) exemplarily include P3, Pz, P4, etc. This invention is not limited to the specific number or location of the electrodes described above; any electrode arrangement that covers the relevant brain regions falls within the scope of protection of this invention.

[0054] Eye tracking is achieved using an infrared eye tracker integrated into the head-mounted display, with a sampling rate of 120Hz. Peripheral physiological signals are acquired through accessory modules of the same device: fingertip photoplethysmography (PPG) sensors acquire heart rate variability signals, and hand electrodes acquire skin conductance signals; both have a sampling rate of 256Hz. The analog-to-digital converters of all submodules are triggered by the same frame synchronization signal from the main control unit. This design ensures data acquisition from the source. , The timestamps of other modal signals are aligned with microsecond precision. The acquired data is then transmitted to the processing unit in real time via a high-speed data bus.

[0055] It should be understood that the personalized cognitive state assessment module calculates the cognitive load index. The specific process includes: first, setting the integration time window length. s. For each prefrontal channel The original signal needs to be processed by a 4-8Hz zero-phase digital filter to obtain the filtered signal. At discrete time points Calculate the channel within the time window The average power within is calculated using the following formula: ,in This represents the number of sampling points within the window. Subsequently, for all... The calculation results of each channel are taken as an arithmetic mean to obtain the global prefrontal cortex. Band average power . occipital lobe Band average power The calculation process is exactly the same, the difference being that for each apical leaf channel... The original signal was filtered using an 8-12Hz bandpass filter to obtain... And finally calculate Ultimately, the real-time cognitive load index is calculated using the formula... Provided.

[0056] As an example, the system state vector can be constructed as follows: ,in The real-time cognitive load index is normalized. This is a performance score calculated based on task accuracy and reaction time. The control input vector is defined as follows: ,in This represents the parameter for adjusting the intensity of facial expressions in a virtual social context. This represents the level of semantic ambiguity of the dialogue statements. The target state reference value is set to... It should be understood that the composition of the state vector and control input is not limited to this, and the dimensions or meanings can be added or removed or adjusted according to the actual training scenario.

[0057] For example, matrices A and B of the state-space model can be obtained through system identification, such as Parameter matrix for The prediction time domain is set to The values ​​above are for illustrative purposes only and can be adjusted according to the user's model in actual applications.

[0058] The weight matrix in the optimization problem is set as the state deviation weight matrix. Control input weight matrix The constraints are specifically set as follows: lower bound of the state variables. and upper limit Lower limit of control input and upper limit ; Lower limit of control input increment and upper limit In each control cycle, the module updates the current measurement status. Using the initial conditions, the interior-point method is used to solve this quadratic programming problem online, and the first set of optimized control commands is then applied. Output to the parameterized social context generation module.

[0059] It should be understood that the specific process by which the closed-loop neuromodulation module detects the theta-γ cross-band phase-amplitude coupling event includes: applying zero-phase FIR bandpass filters of 4-8Hz and 30-50Hz to the EEG signals collected from preset brain regions, such as the signal from the prefrontal cortex electrode Fz, to extract the theta and γ oscillation components. Subsequently, the instantaneous phase of the theta band is obtained through Hilbert transform. and the instantaneous amplitude envelope of the γ band Coupling strength In length The calculation is performed within a sliding time window, and the calculation formula is as follows: ,in The normalized envelope of the γ amplitude is represented by L, where L is the number of sampling points within the window. Preset threshold. This is determined by analyzing the user's baseline resting-state data, typically taking this state as the reference. The 90th percentile of the distribution. Calculated in real time. When this occurs, the system determines it as a valid coupling event and triggers the corresponding neural stimulation.

[0060] In some embodiments, the transcranial alternating current stimulation applied by the closed-loop neuromodulation module is implemented as follows: the stimulation waveform is given by the formula... Definition. Wherein, stimulus amplitude 1.0mA. Stimulation frequency. The stimulation frequency is dynamically selected based on the current task context. When the system determines that the user is performing a task requiring deep working memory involvement, a stimulation frequency of 6Hz is selected; when performing a task requiring rapid social perception, a stimulation frequency of 40Hz is selected. (Basic phase shift) Set as Radius, or 180 degrees. Dynamic phase adjustment amount. Updates are performed via a proportional controller, with the following update rules: Among them, proportional gain Target phase The purpose of this is to align the peak value of the stimulation current with a specific phase of the user's own brainwave oscillations. For real-time measurement, and the current The corresponding instantaneous phase of the target EEG frequency band. This closed-loop control mechanism enables the phase of the stimulation current to dynamically track and lock onto the user's own EEG oscillation phase.

[0061] It should be understood that the target state reference value Cognitive load components The dynamic update rules are implemented according to the following steps: The system continuously records user activity... Historical performance data across different social subtasks. For the first... For similar tasks, the system fits the performance score using Gaussian process regression. With cognitive load The nonlinear relationship between them yields a continuous function. and variance At the time of parameter update, the system is within the feasible cognitive load range. The system searches for the optimal cognitive load that maximizes the weighted expected performance. Its mathematical expression is ,in As a preset ideal performance reference point, These are the weighting coefficients assigned to different task types. Ultimately, the new cognitive load target value is calculated using the formula... The calculation yielded the forgetting factor. .

[0062] In some embodiments, the skill transfer graph management module is implemented as follows: Define a directed graph model. , where vertex set Include Different social sub-skills. Initial edge weight matrix. Pre-set based on cognitive neuroscience theory, for example, starting from the "facial expression recognition" peak. To the peak of "empathic accuracy" The migration path, whose initial weight is preset to . When users complete the skill-related task... After specialized training, in subsequent skills-related matters If the performance improvement observed in the task is Then, the transition weights are adjusted according to the Bayesian update rule. The likelihood function used for the update is defined as follows: ,in As a training efficiency factor, To observe the noise variance, according to Bayes' theorem, the posterior weight of this path is proportional to the product of its prior weights and the likelihood function, i.e. After the weights are updated, the system will process all data from the vertices. The weights of the starting migration paths are normalized to ensure that they meet the requirements. This yields the updated global weight matrix. .

[0063] It should be understood that the parameterized social context generation module achieves fine-grained control over the facial expressions of virtual characters through the following process: The system has a built-in facial motion coding system parameter database, which stores 46-dimensional motion unit template vectors corresponding to basic emotions and neutral expressions, denoted as... and When the module receives the facial expression intensity adjustment parameters from the control module... First, the basic parameter vector is generated. Subsequently, to increase the naturalness of the expressions, the system introduced a natural mutation vector: generating a 46-dimensional random vector. Each of its components independently follows a standard normal distribution. The final facial motion unit parameter vector used to drive the animation is given by the formula. clip The calculation shows that, among which The coefficient of natural variability. The function is responsible for restricting each component of the parameter vector to a valid interval of [0, 1]. The calculated... Vectors are fed into the 3D character animation engine in real time, driving the deformation of the facial mesh to render the target expression with a specified intensity and natural, subtle fluctuations.

[0064] Example 2

[0065] The personalized social cognitive function assessment and enhancement closed-loop regulation method for autism spectrum disorders described in this embodiment includes the following implementation process: equipment preparation and individualized baseline establishment for perceptual sensitivity, real-time multi-dimensional state assessment for social information processing characteristics, adaptive social situation regulation based on prediction models, closed-loop intervention for neural integration abnormalities, and long-term system optimization for individual differences and generalization difficulties.

[0066] The user-worn integrated multimodal sensing device includes a 32-lead EEG electrode cap made of ultra-flexible materials and an adjustable low-pressure structure. Based on research into the social neural mechanisms of autism, multiple prefrontal and parieto-occipital electrodes are pre-installed for subsequent specialized analyses. A head-mounted display device with an integrated infrared eye tracker is used concurrently. The virtual social context presented has adjustable initial brightness, contrast, and animation speed to accommodate individual differences in visual sensitivity. Finger-tip photoplethysmography (PPG) sensors and hand skin conduction electrodes also employ a low-invasive design. All sensing units are connected to a master acquisition unit to ensure synchronized multimodal signal sources. Upon device startup, the system first acquires the user's resting-state physiological data in a low-sensory-stimulation environment. This individualized baseline is used not only to calculate the trigger threshold for subsequent neural modulation but also to conduct a preliminary assessment of the user's perceptual-cognitive characteristics through analysis of the baseline EEG patterns, providing a basis for personalized initial settings of subsequent training parameters.

[0067] This is followed by a multi-dimensional real-time state assessment phase targeting social information processing characteristics. While users perform interactive tasks in virtual social scenarios, the system calculates in real-time a key indicator of autism's social cognitive characteristics: the cognitive load index CL(t). Its numerator... The average power of the theta band (4-8 Hz) in the prefrontal cortex is closely related to cognitive control, task engagement, and effort maintenance under anxiety. Individuals with autism often exhibit prefrontal cortex overactivation in social tasks. The denominator is... The average power of the α-band (8-12 Hz) in the parietal-occipital lobe region is associated with sensory gating and internal attention; insufficient inhibition of this band may be related to social information overload. Therefore, CL(t) can specifically characterize the imbalance between cognitive input and sensory inhibition in autistic individuals when processing social stimuli, which is an important neurophysiological indicator of their social fatigue and avoidance behavior. Simultaneously, the system calculates a real-time task performance score based on the user's accuracy and speed in recognizing and responding to social cues such as facial expressions, tone of voice, and semantics. Together, these two factors constitute a state vector reflecting the current social information processing efficiency of autistic individuals.

[0068] Based on the above assessment, the system enters the adaptive social situational regulation decision-making stage based on a predictive model. The system incorporates a discrete state-space model corrected for autism data to predict the user's response to changes in external situational parameters. In each control cycle, the adaptive decision control module solves a constrained optimization problem, ensuring that the user's cognitive load index remains within an individualized tolerance window, while seeking optimal situational parameters to guide their performance towards a gradual convergence towards the target. The optimization output is transformed into real-time regulation instructions in two dimensions: first, parameters for adjusting the intensity and directness of the virtual character's facial expressions. For example, when an increase in load is detected, the system automatically adjusts complex, intense negative expressions to neutral or simple positive expressions to reduce the difficulty of interpreting emotions and the resulting anxiety; second, parameters for adjusting the semantic clarity, metaphor content, and contextual dependence of dialogue statements. The system can automatically convert vague or sarcastic statements into direct and specific expressions to reduce the cognitive burden.

[0069] The parameterized social context generation module drives the virtual environment to adapt and change based on control commands. For expression generation, the system interpolates between pre-stored basic emotional facial action codes and neutral expression codes based on expression intensity parameters, and superimposes subtle random perturbations that conform to natural laws to generate a sequence of facial actions that drive the virtual character. For dialogue generation, the system selects or generates sentences with matching complexity from a hierarchical corpus based on semantic ambiguity levels. Thus, the system provides users with personalized social training scenarios that are predictable, manageable, and dynamically adaptable in terms of emotional load and cognitive challenge.

[0070] In addition to external contextual modulation, the system also features closed-loop intervention capabilities for abnormal neural integration. The system continuously monitors the phase-amplitude coupling strength between low-frequency theta oscillations and high-frequency gamma oscillations in the user's prefrontal cortex EEG signals. Abnormally high coupling is thought to be associated with local neural network overexcitation and low long-range information integration efficiency in individuals with autism. When the real-time coupling strength exceeds a threshold set based on the individual's resting baseline, the system identifies a specific neural inefficiency event and triggers transcranial alternating current stimulation. The stimulation frequency is dynamically selected based on the task-driven cognitive needs: low-frequency stimulation is used for tasks requiring understanding social rules and intentions, aiming to modulate brain network activity related to social reasoning; high-frequency stimulation is used for tasks requiring rapid capture of subtle facial expressions, aiming to modulate neural synchronicity related to fine processing of social information. The phase of the stimulation current is dynamically locked to the real-time phase of the user's own target EEG frequency band through a closed-loop controller, creating a neural entrainment effect to gently guide their neural activity patterns.

[0071] It should be understood that the system incorporates a long-term self-optimization mechanism to address the significant individual differences and difficulties in skill generalization. The system continuously records users' historical data across different categories of social sub-tasks and constructs personalized cognitive load-performance relationship models for each task using machine learning methods. The system periodically analyzes these models to identify the individualized cognitive load target range that optimizes overall performance across all tasks, and progressively updates the long-term control target value accordingly, ensuring that the training difficulty remains within the user's zone of proximal development. Furthermore, the system maintains a personalized skill transfer graph, where nodes represent different basic social cognitive skills, and edges represent the potential transfer relationships and strengths between skills. Whenever a user demonstrates improvement in subsequent related tasks after training a specific skill, the system updates the weight estimates of the corresponding transfer path based on Bayesian rules. This continuously evolving graph is used to intelligently plan the most efficient personalized training sequence, prioritizing the training path with the highest efficiency in transferring skills from already mastered skills to target skills, aiming to promote the effective generalization of learned skills in diverse social situations.

[0072] In summary, this invention proposes a personalized social skills training system for autism, comprising a multimodal data acquisition module, a personalized cognitive state assessment module, a parameterized social scenario generation module, an adaptive decision control module, and a closed-loop neural modulation module. The system synchronously acquires users' physiological signals such as EEG and eye movements, calculates quantitative indicators such as the cognitive load index in real time, and dynamically generates adjustment instructions for the difficulty parameters of virtual social scenarios using methods such as model predictive control, so as to match the training task with the user's real-time cognitive load. Simultaneously, the system can trigger non-invasive brain stimulation synchronized with the user's own EEG phase when specific neural oscillation characteristics are detected. This invention achieves personalized, adaptive social skills training based on real-time feedback of intrinsic physiological state, and also possesses the ability to target and modulate abnormal neural activity patterns.

[0073] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A personalized social skills training system for autism, characterized in that, include: The multimodal data acquisition module is used to simultaneously acquire the user's multi-channel EEG signals, eye-tracking data, and peripheral physiological signals; A personalized cognitive state assessment module is connected to the data acquisition module and is used to calculate at least one quantitative cognitive state index in real time based on the acquired physiological signals. The parameterized social context generation module is used to present virtual social interaction scenarios with adjustable difficulty parameters; An adaptive decision control module is connected to the personalized cognitive state assessment module and the parameterized social context generation module, respectively, and is used to generate real-time adjustment instructions for the difficulty parameter by solving a constrained optimization problem based on the quantified cognitive state index. The closed-loop neuromodulation module, connected to the multimodal data acquisition module, is used to trigger non-invasive brain stimulation that is phase-synchronized with the user's own brainwave oscillations when a neural oscillation feature that conforms to a preset pattern is detected.

2. The system as described in claim 1, characterized in that, The personalized cognitive state assessment module calculates the quantitative cognitive state index CL(t), which is calculated using the following formula: in, Indicates within the time window Internally, from the prefrontal cortex region of the brain Extracted from EEG signals acquired by individual electrode channels The average power of the frequency band is calculated using the following formula: This indicates that within the same time window, brain regions located in the parietal or occipital lobes... Extracted from EEG signals acquired by individual electrode channels The average power of the frequency band is calculated using the following formula: This is the length of the integration time window; and These are the time-frequency representations of the EEG signals for the corresponding channels and frequency bands.

3. The system as described in claim 2, characterized in that, The constrained optimization problem solved by the adaptive decision control module is a finite-time model predictive control problem, which is mathematically expressed as follows: Subject to in, For continuous time, For discrete-time indexing; The predicted time domain length is a positive integer; For at any time right The predicted value of the system state at any given time, whose state vector at least includes the cognitive load index. ; The control input sequence to be optimized, whose elements correspond to the difficulty parameters; This is a reference value for the target state. and It is a symmetric positive definite weight matrix; and A discrete-time system matrix describing the dynamic state; These are the upper and lower bounds of the state constraints; To control the upper and lower bounds of input constraints; The upper and lower bounds are used to control the rate of change of the input.

4. The system as described in claim 1, characterized in that, The neural oscillation characteristics of the preset pattern detected by the closed-loop neuromodulation module are as follows: -γ trans-band phase amplitude coupling events, their coupling strength Calculated using the following formula: in, Represents at discrete time points Extracted from EEG signals of specific brain regions Instantaneous amplitude envelope of the frequency band; Indicates at the same point in time Extracted from EEG signals of relevant brain regions Instantaneous phase of the frequency band; Time window used for calculation Number of sampling points within, The length of the time window; when The value exceeds the preset threshold When a valid event is detected, a stimulus is triggered.

5. The system as described in claim 4, characterized in that, The non-invasive brain stimulation applied by the closed-loop neuromodulation module is transcranial alternating current stimulation, and its output current waveform... for: in, The peak amplitude of the stimulation current ranges from 0.5 mA to 2.0 mA. The stimulation frequency is selected from values ​​related to the target nerve oscillation frequency band. Hz; It is an adjustable fixed phase offset; The dynamic phase adjustment amount, based on real-time EEG phase feedback, is used to achieve phase synchronization between stimulation and EEG. Its adjustment strategy is as follows: in This is the proportional gain coefficient. For the target phase to be locked, This refers to the instantaneous phase of the target frequency band as measured in real time.

6. The system as described in claim 3, characterized in that, The target state reference value Cognitive load target component It is dynamically updated based on the user's individual historical training data. The update rule is based on the performance-load fitting model, specifically: in, Forgetting factor, satisfying The number of historical task types; For the first Performance rating for each type of task and The performance scores for this type of task vary with cognitive load. The mean and variance of the fitted Gaussian distribution; For the first Weight coefficients for task categories; This represents a feasible search interval for cognitive load. The mean is variance is The Gaussian probability density function.

7. The system according to claim 1, characterized in that, It also includes a skills transfer graph management module, used to model the transfer relationships between social skills, with the graph represented as a directed graph. , where vertex set represent A different social sub-skill, edge set The weight matrix represents the potential transfer paths between skills. Each element Indicates from skills Training to skills Positive migration strength; weight Perform Bayesian updates based on user training data: in, For the observed user training data, For a given migration strength Data observed at the time The likelihood function, the likelihood function is based on skill Performance improvement rate and The linear or nonlinear relationship model between them is defined.

8. The system according to claim 1, characterized in that, The parameterized social context generation module provides fine-grained control over the facial expressions of virtual characters by adjusting the parameter vectors based on the facial motion coding system. Implementation, where each element Corresponding to the The activation intensity of each facial movement unit; for expressing a specific target emotion. The parameter vector is generated by the following formula: in, A parameter vector for neutral facial expressions; For target emotions The template parameter vector; [0, 1] represents the global facial expression intensity adjustment parameter, which is determined by the output of the adaptive decision control module; The coefficient of natural variability in facial expressions; Each component is independent and identically distributed and follows a standard normal distribution. Random vectors are used to introduce natural, subtle changes in facial expressions.

Images