Children autism spectrum disorder prediction method of brain enlightenment causal large model

Abstract

The invention discloses a children autism spectrum disorder prediction method based on a brain enlightenment causal large model. The method comprises the following steps: S1, collecting and processing behavior data and brain data; forming a multi-modal data set by the behavior data and the brain data; s2, forming a unified multi-modal data set, and constructing a causal graph by adopting a structural equation model; s3, integrating data processing on the core dimension; s4, coding the knowledge graph text by adopting a mask language model; s5, Bayesian network learning and causal rule mining; s6, adopting an anti-fact strategy and a depolarization strategy; s7, proposing a training and optimizing method; and S8, predicting and evaluating the children autism spectrum disorder based on the brain enlightenment causal large model. According to the children autism spectrum disorder prediction method based on the brain enlightenment causal large model, multi-modal data are integrated, the causal large model is constructed by using the knowledge graph, and the potential causal relationship among the factors of the autism spectrum disorder ASD is revealed.

Description

Technical Field

[0001] The present invention relates to the technical field of autism spectrum disorder (ASD) models, and in particular to a method for predicting autism spectrum disorder in children based on a brain-inspired causal large model. Background Art

[0002] Autism spectrum disorder (ASD) is a complex neurodevelopmental disorder whose core symptoms include social communication disorders, narrow interests, and repetitive stereotyped behaviors, which may be accompanied by symptoms such as sensory abnormalities. Its early identification and intervention are crucial for children's growth and rehabilitation. However, the clinical manifestations of autism spectrum disorder (ASD) are diverse, and there are significant differences among different ages, genders, and individuals, which makes the accurate diagnosis of ASD a huge challenge. Traditional diagnostic methods rely on the doctor's clinical experience and scale assessment, but these methods are often time-consuming, subjective, and easily limited by the doctor's experience and knowledge level.

[0003] In recent years, with the rapid development of artificial intelligence technology, especially breakthroughs in big data, machine learning and deep learning, new opportunities have been provided for the early prediction of autism spectrum disorder (ASD). Among them, multimodal data fusion technology can make full use of information from different channels, such as behavioral observation and brain data, to provide more comprehensive and in-depth data support for the prediction of autism spectrum disorder (ASD).

[0004] However, relying solely on multimodal data fusion is not enough to solve all problems in the prediction of autism spectrum disorder (ASD). The pathogenesis of autism spectrum disorder (ASD) involves multiple complex biological, psychological, and social factors, and the interactions and causal relationships between these factors are often difficult to directly observe and quantify. Therefore, a method that can reveal these potential causal relationships is needed to more accurately predict the occurrence and development of autism spectrum disorder (ASD).

[0005] Therefore, we use knowledge graph technology to build a brain-inspired causal model for the prediction of autism spectrum disorder in children. This model integrates multi-dimensional information such as behavioral data, brain data, psychological and emotional state, external behavioral patterns, and achievement development trajectories of patients with autism spectrum disorder (ASD). Through sophisticated algorithms and logic, it constructs an intricate causal network. In this causal model, we deeply explore the internal connections and potential laws between the various factors of autism spectrum disorder (ASD), accurately determine the causal relationship between them, and conduct in-depth causal inference, so as to make more accurate and reliable predictions of autism spectrum disorder in children.

[0006] On this basis, the use of counterfactual strategies and debiasing strategies can significantly weaken the bias effects in large causal models, providing more solid support for the accurate prediction of autism spectrum disorders (ASD) in children. These two strategies enhance the model's ability to identify and predict complex causal relationships by deeply analyzing and correcting potential sources of bias in the model, thereby helping to improve the accuracy and reliability of autism spectrum disorders (ASD) predictions. Summary of the invention

[0007] The purpose of the present invention is to provide a method for predicting autism spectrum disorder in children based on a brain-inspired causal macromodel, integrate multimodal data, use knowledge graphs to construct a causal macromodel, reveal the potential causal relationship between various factors of autism spectrum disorder ASD, and provide more accurate and effective guidance for the early prediction of autism spectrum disorder ASD.

[0008] The present invention provides a method for predicting autism spectrum disorders in children based on a brain-inspired causal macromodel, comprising the following steps: S1. Collect and process behavioral data and brain data; combine the processed behavioral data and brain data into a multimodal data set; S2. Clean, normalize and feature select the multimodal data set. When processing unstructured text data, introduce causal embedding technology to fuse the behavioral data and brain data output after processing by causal embedding technology to form a unified multimodal data set. Based on the unified multimodal data set, use the structural equation model to construct a causal diagram and display the causal relationship between variables.

[0009] S3. Process behavioral data, brain data and causal relationships into structured data and store them in the form of triples; integrate the data on the three core dimensions of psychological and emotional state, external behavioral patterns and achievement development trajectory into structured data, also store them in the form of triples, and initially construct a knowledge graph; and construct a complete knowledge graph together with the data of the causal graph; S4. Convert all acquired data into text form and use masked language model to encode the knowledge graph text; verify the effectiveness of the masked entity model by predicting the masked entities in the triples under the masked entity model classification mechanism; enhance the integrity and accuracy of the knowledge graph through semantic embedding and structural embedding technology; S5. Introduce Bayesian network learning and differential causal rule mining. First, introduce the Bayesian network learning mechanism to build a probabilistic relationship model, and then use differential causal rule mining technology; perform causal query and causal path identification through metadata semantics; finally generate a causal large model for the prediction of autism spectrum disorders in children; S6. A causal macromodel for predicting autism spectrum disorders in children, using counterfactual and debiasing strategies; S7. Propose a method for training and optimizing a large brain-inspired causal model for predicting autism spectrum disorders in children under the guidance of counterfactuals. S8. Prediction and assessment of autism spectrum disorders in children based on brain-inspired causal big models.

[0010] Preferably, in step S1, a pinhole imaging model is used to collect and process the behavioral data of the child by observing the head rotation of the observed person; brain data is obtained by magnetic resonance imaging MRI technology and electroencephalogram EEG technology in graph theory analysis, and the obtained brain data is regarded as a variable using a functional connection method to construct a brain network and measure the global characteristics of the observed person.

[0011] Preferably, in step S2, the specific method of constructing the causal diagram is as follows: according to the research hypothesis and data characteristics, the latent variables, the observed variables and the relationship between them are set, and then the factor loading matrix, the structural coefficient matrix and the variance of the error term are estimated by the maximum likelihood estimation method, and the model fitting is compared by the chi-square test to verify the validity and accuracy of the model. Finally, based on the estimated parameters and model structure, a causal diagram is drawn to intuitively display the causal relationship between the variables, as shown in the following formula: ; Where A represents the processed behavioral data vector; M represents the processed brain data vector; Represents a latent variable vector that cannot be directly observed in the data; Represents the factor loading matrix corresponding to the behavioral data; represents the factor loading matrix corresponding to the brain data; represents the random error term introduced when observing behavioral data; represents the random error term introduced when observing brain data; Q The structural coefficient matrix representing the causal relationship between latent variables; Represents the influence matrix of exogenous variables on latent variables; represents the random error term introduced when estimating the latent variable.

[0012] Preferably, in step S3, in terms of psychological and emotional state, the psychological and emotional data of the observed person is obtained and quantitatively constructed into a set of psychological and emotional instances. ; In terms of behavioral patterns, obtain the observed behavior performance data and quantify it, and construct a set of behavioral performance instances ; In terms of achievement development trajectory, obtain the achievement level data of the observed and quantify and construct a set of achievement level instances .

[0013] Among them, I represents the psychological emotion of the observed person, The set represents the value of psychological emotions, H represents the behavior of the observed person, The set represents the value of the behavior performance, L represents the achievement level of the observed person, A collection of values ​​representing achievement levels.

[0014] Preferably, in step S4, two strategies, entity description prediction and incomplete triple prediction, are adopted. In entity description prediction, the triple is first encoded and a final hidden state is generated; then, it is passed as the embedding of the head entity to the incomplete triple prediction to predict the target entity; Specifically, for a given head entity and relationship , the model predicts the tail entity The probability of is shown in the following formula: ; In the formula, Indicates that the model predicts the tail entity The probability of Represents the header entity ,relation and candidate tail entity The score between Represents the tail entity set; Preferably, in step S5, the difference causal rule is expressed as: if the control variable W changes, the effect variable G changes accordingly with probability P. In this framework, W plays the role of the control variable, G acts as the effect variable, and P quantitatively reflects the magnitude of the probability change of the effect variable G after the control variable W changes; as shown in the following formula; ; In the formula, and They represent the conditional probability of the control variable W under the new and old states of the effect variable G; This is the change in the probability of the effect variable.

[0015] Preferably, in step S6, firstly, a potential set of relevant features is identified by deeply analyzing the nodes and edges in the causal graph, and counterfactual reasoning is used.

[0016] Preferably, in step S7, the brain-inspired causal big model of autism spectrum disorders in children is placed in a corpus of text data for pre-training. After the pre-training stage, the brain-inspired causal big model of autism spectrum disorders in children is causally fine-tuned using the annotated dataset, and the annotated dataset is further used to perform supervised fine-tuning on the brain-inspired causal big model of autism spectrum disorders in children.

[0017] Preferably, in step S8, after the training is completed, the brain-inspired causal big model for autism spectrum disorders in children is deployed to practical applications, and sophisticated prompt engineering technology is used to enable the brain-inspired causal big model for autism spectrum disorders in children to receive multi-dimensional and highly complex data information as input, and the brain-inspired causal big model for autism spectrum disorders in children accurately outputs relevant behavioral data, brain data, and prediction results of autism spectrum disorders in children; finally, the performance of the brain-inspired causal big model for autism spectrum disorders in children is comprehensively and rigorously evaluated, and the evaluation dimensions cover task-specific key indicators.

[0018] Therefore, the present invention adopts the above-mentioned brain-inspired causal big model prediction method for children with autism spectrum disorder, integrates multimodal data, and uses knowledge graphs to construct a causal big model to reveal the potential causal relationship between various factors of autism spectrum disorder ASD, providing more accurate and effective guidance for the early prediction of autism spectrum disorder ASD.

[0019] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is an overall flow chart of a method for predicting autism spectrum disorders in children based on a brain-inspired causal model of the present invention; Figure 2 A schematic diagram of constructing a causal graph of multimodal data of a method for predicting autism spectrum disorders in children based on a brain-inspired causal macromodel of the present invention; Figure 3 This is a schematic diagram of a causal macromodel for predicting autism spectrum disorders in children according to a method for predicting autism spectrum disorders in children based on a brain-inspired causal macromodel of the present invention. DETAILED DESCRIPTION

[0021] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0022] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0023] Embodiment 1 like Figure 1-Figure 3 As shown, the present invention provides a method for predicting autism spectrum disorders in children based on a brain-inspired causal model, comprising the following steps: S1. Collection and processing of behavioral data and brain data; combining behavioral data and brain data into a multimodal data set; The pinhole imaging model and the head rotation of the observed person are used to collect and process the children's behavior data; the center point of the observed person's eyes is defined as the observation point B, and the optical center O of the camera corresponds to the position of the pinhole in the pinhole imaging model. When the observation point B in the real space is mapped to the physical imaging plane O-xyz through the pinhole imaging model, its landing point is recorded as , let the coordinates of B be (X, Y, Z), then The coordinates of The focal length of the camera, that is, the distance from the camera plane to the physical imaging plane, is denoted by f. Based on the principle of similarity of triangle focal lengths, we can deduce that B and The proportional relationship between the coordinates is shown in formula (1).

[0024] (1); Where f represents the focal length of the camera; X, Y, Z represent the X, Y, and Z coordinates of B respectively; , represent The X and Y coordinates of the

[0025] In the camera system, the image is finally presented as pixels on the electronic screen. Therefore, by quantizing the imaging process, the conversion from the image coordinate system to the pixel coordinate system is realized. This conversion process is composed of a scaling transformation and a translation transformation, which is specifically expressed as scaling a times on the m-axis and b times on the n-axis. Based on this, the point on the physical imaging plane is The coordinates on the pixel plane can be obtained by formula (2); (2); Among them, m represents the horizontal axis of the pixel point in the pixel coordinate system; n represents the vertical axis of the pixel point in the pixel coordinate system; a represents the scaling factor on the m-axis; b represents the scaling factor on the n-axis; on the pixel plane, the coordinates of a pixel point can be expressed as (m, n).

[0026] In order to express this conversion process more concisely, af is merged into , merge bf into , K is the intrinsic parameter matrix of the camera. Formula (2) is rewritten into a matrix form represented by homogeneous coordinates, see formula (3).

[0027] (3); Among them, m represents the coordinate of the pixel point on the m-axis; n represents the coordinate of the pixel point on the n-axis; represents the focal length on the m-axis; represents the focal length on the n-axis; (X, Y, Z) represent the coordinates of B respectively; Z represents the distance from the camera lens to the pixel, that is, the depth value of the pixel in the camera coordinate system. The 3×3 matrix in the middle of the equation is the intrinsic parameter matrix K of the camera.

[0028] The head rotation of the observed person needs to be determined based on the movement and rotation of the observation point B. The rotation is represented by the rotation matrix R and the translation vector t, as shown in formula (4).

[0029] (4); The head rotation is divided into three angles: pitch angle, yaw angle and roll angle. The pitch angle is used to describe the changes in the posture of the head in flexion and extension in the sagittal plane, which reflects the degree of inclination of the head in the vertical direction; the yaw angle reflects the posture distribution of the head rotating left and right in the horizontal plane, revealing the rotation of the head in the horizontal direction; and the roll angle characterizes the posture distribution of the head in the coronal plane, that is, the changes in the left and right inclination of the head.

[0030] Brain data is obtained through magnetic resonance imaging (MRI) and electroencephalogram (EEG) technology in graph theory analysis. The functional connection method is used to treat the obtained brain data as variables, construct a brain network, and measure the global characteristics of the observed person.

[0031] Graph analysis: In the study of brain data for children with autism spectrum disorder, magnetic resonance imaging (MRI) technology was used to obtain high-resolution structural and functional images of the brain of the observed person in a non-invasive manner. These images can clearly show the anatomical structure inside the brain and the functional activities of different brain regions. At the same time, the electrical signals on the scalp of the observed child are recorded, and the electroencephalogram (EEG) technology is used to reflect the electrical activity of the brain. The electroencephalogram (EEG) can capture the changes in electrical signals of the brain during specific tasks or in a resting state, and these changes are closely related to the functional state of the brain. By analyzing the EEG data, we can further understand the brain function characteristics of ASD patients in terms of cognition, emotion, and social interaction.

[0032] Based on the graph theory analysis method, the functional connection method is used to treat the acquired brain data as variables to construct a brain network. The brain network is constructed to measure the global characteristics of the observed person. First, different brain regions are represented as nodes, their functional connections are represented as edges, and 90 fMRI time series using the AAL human brain neuroanatomical atlas are selected as nodes. The functional connection method is used to obtain the dependency between neuronal activity patterns in brain regions. In this type of connection, although brain regions are spatially separated, they are functionally interconnected and constantly exchange and share information with each other. The mutual dependence between two brain variables is measured by mutual information, and the shared information between two time series is evaluated based on Shannon entropy. The specific details are shown in formula (5).

[0033] (5); Where MI(u,v) represents a suitable indicator of functional connectivity, u and v represent two brain variables, p(u) represents the probability density function of u, p(v) represents the probability density function of v, and p(u,v) represents the joint probability density function of u and v. represents the probability density function of u under the condition of known brain variable v, represents the probability density function of v under the condition of known brain variables u.

[0034] The brain network was constructed based on the correlation between two brain variables. Then, features were extracted from the brain network and the degree of separability and coherence of network functions was determined. Six global features were calculated and measured using the binary brain network, including clustering coefficient C, average strength , eccentricity E, transmissibility , Global Efficiency and local efficiency .

[0035] S2. Clean, normalize and feature select the multimodal data set. When processing unstructured text data, introduce causal embedding technology to fuse the behavioral data and brain data output after processing by causal embedding technology to form a unified multimodal data set. Based on the unified multimodal data set, use the structural equation model to construct a causal diagram and display the causal relationship between variables.

[0036] Through behavioral data, the subjects' reaction time T after hearing the sound stimulus and the posture distribution of the head rotation at multiple angles (pitch angle, yaw angle, roll angle) were collected. At the same time, the global features of 6 key brain data were measured using binary brain network calculations, including clustering coefficient C, average intensity , eccentricity E, transmissibility , Global Efficiency and local efficiency Together, these data constitute the multimodal dataset for analysis.

[0037] When constructing a multimodal data causal graph for children with autism spectrum disorder, the collected multimodal datasets are first cleaned, normalized, and feature selected to ensure the quality and consistency of the data. However, when processing unstructured text data, although large language models are good at learning statistical correlation patterns from large-scale corpora, they have difficulty directly inferring the causal dynamics between entities, events, and behaviors. To overcome this limitation, causal embedding technology is introduced to convert the original text into a numerical representation that the model can handle, and to convert unstructured text data into numerical vectors to provide input for subsequent causal reasoning. The behavioral data and text data output by brain data after causal embedding processing are then fused to form a unified multimodal dataset representation.

[0038] In step S2, the specific method of constructing the causal diagram is as follows: according to the research hypothesis and data characteristics, the latent variables, observed variables and the relationship between them are set, and then the maximum likelihood estimation method is used to estimate the factor loading matrix, structural coefficient matrix and the variance of the error term. The chi-square test is used to compare the quality of model fitting and verify the validity and accuracy of the model. Finally, based on the estimated parameters and model structure, a causal diagram is drawn to intuitively display the causal relationship between the variables, as shown in formula (6).

[0039] (6); Where A represents the processed behavioral data vector; M represents the processed brain data vector; Represents a latent variable vector that cannot be directly observed in the data; Represents the factor loading matrix corresponding to the behavioral data; represents the factor loading matrix corresponding to the brain data; represents the random error term introduced when observing behavioral data; represents the random error term introduced when observing brain data; Q represents the structural coefficient matrix of the causal relationship between latent variables; represents the influence matrix of exogenous variables on latent variables, represents the random error term introduced when estimating the latent variable.

[0040] S3. Process behavioral data, brain data and causal relationships into structured data and store them in the form of triples; integrate the data on the three core dimensions of psychological and emotional state, external behavioral patterns and achievement development trajectory into structured data, also store them in the form of triples, and initially construct a knowledge graph; and construct a complete knowledge graph together with the data of the causal graph; In terms of psychological and emotional state, ASD individuals show characteristics such as flat emotional response and slow emotional fluctuations, which reflects the inherent obstacles in emotional processing and expression. Obtain the psychological and emotional data of the observed person and quantify and construct a set of psychological and emotional instances .

[0041] Secondly, in terms of behavioral patterns, the frequency of social interactions of ASD patients is significantly reduced, and they often show repetitive and stereotyped behavioral patterns. Obtain the behavioral performance data of the observed person and quantify it and construct a collection of behavioral performance examples .

[0042] Finally, in terms of achievement development trajectory, ASD patients usually lack due attention and pursuit of academic performance, social achievement, etc., which reflects the limitations of their interest range and achievement motivation. .

[0043] It is also stored in the form of triples, and together with the data of the causal graph, it constructs a complete knowledge graph; First, the information of behavioral data and brain data and their causal relationships are processed into structured data and stored in the form of triples ZS=(M,R,O), where M represents the subject (such as a specific behavior or brain area activity), O represents the object (such as the object of the behavior or the response of the brain), and R represents the relationship between the subject and the object (such as causal relationship, correlation, etc.). This step aims to convert the raw data into a form that is easy to analyze and reason about.

[0044] At the same time, the data information instance sets obtained from the three dimensions of psychological and emotional state, external behavior pattern and achievement development trajectory are processed into structured data, which are also stored in the form of triples ZS=(M,R,O). The above triple data are then initially constructed to form a knowledge graph.

[0045] S4. Convert all acquired data into text form and use masked language model to encode the knowledge graph text; verify the effectiveness of the masked entity model by predicting the masked entities in the triples under the masked entity model classification mechanism; enhance the integrity and accuracy of the knowledge graph through semantic embedding and structural embedding technology; Under the classification mechanism of the masked entity model, the effectiveness of the model is verified by predicting the masked entity in the triple. In this process, two strategies, entity description prediction and incomplete triple prediction, are adopted. Entity description prediction first encodes the triple and generates the final hidden state. Then, it is passed as the embedding of the head entity to the incomplete triple prediction task to predict the target entity. Specifically, for a given head entity and relationship , the model predicts the tail entity The specific probability is shown in formula (7).

[0046] (7); In the formula, Represents the header entity ,relation and candidate tail entity The score between Represents the set of all possible tail entities.

[0047] Finally, the tail entities in the knowledge graph are completed through task description, few-shot examples, and test input.

[0048] In order to transform the triple information in the knowledge graph into a representation that can be used in deep learning models, semantic embedding and structural embedding techniques are used. First, the complete text information of the triple is input into the large language model (LLMs) to calculate the head entity ,relation and tail entity The corresponding LLMs outputs are averaged and pooled to obtain their respective embedding representations. Specifically, for each triple (M, R, O), its embedding representation is as shown in formula (8).

[0049] (8); In the formula, LLMs_output(x) represents the output representation of the large language model for the input variable x (i.e., head entity, relation, or tail entity). In this process, a graph-based method is used to further process and analyze the embedded representation to extract deeper knowledge and relations, and to infer missing facts in the knowledge graph, thereby enhancing the integrity of the knowledge graph.

[0050] Emb(M,R,O) represents the embedding representation of a triple (M,R,O). LLMs_output(M) represents the output representation of the large language model for the head entity M; LLMs_output(R) represents the output representation of the large language model for relation R; LLMs_output(O) represents the output representation of the tail entity O by the large language model.

[0051] S5. Introduce Bayesian network learning and differential causal rule mining. First, introduce the Bayesian network learning mechanism to build a probabilistic relationship model, and then use differential causal rule mining technology; perform causal query and causal path identification through metadata semantics; finally generate a causal large model for the prediction of autism spectrum disorders in children; In order to more accurately reveal the intricate association patterns between variables, the Bayesian network learning mechanism is introduced to construct a probabilistic relationship model, and based on this, a joint probability distribution P(X) covering all variable sets X is proposed. This joint probability distribution P(X) deeply depicts the probability map of all variables occurring simultaneously, however, it does not directly reveal the causal chain between variables.

[0052] In view of this, a differential causal rule mining technique is adopted, which cleverly embeds the idea of ​​control variables in the knowledge graph to carefully examine the potential causal relationship.

[0053] Specifically, the differential causal rule can be expressed as follows: if the control variable W changes, the effect variable G changes accordingly with probability P. In this framework, W plays the role of a control variable and is the initiator of the potential causal chain; G, as an effect variable, is the manifestation of the terminal of the causal chain; and P quantitatively reflects the magnitude of the probability change of the effect variable G after the control variable W changes. This rule is specifically shown in formula (9).

[0054] (9); In the formula, and They represent the conditional probability of the control variable W under the new and old states of the effect variable G; That is, the change in the probability of the effect variable, which intuitively reflects the strength of the causal relationship.

[0055] Then, through the powerful function of metadata semantics, causal query in the knowledge graph is realized, so that the subject and object in the knowledge graph can clearly indicate the causal relationship and allow causal inference. Through the in-depth analysis of the relationship between the subject and the object in the knowledge graph, the causal path can be accurately identified and the causal effect can be inferred based on it. This process is specifically shown in formula (10).

[0056] (10); In the formula, Represents the collection of subjects To the effect set The set of causal paths of and They represent the adjacent nodes on the path, i.e. the previous and next variables in the causal chain. The causal relationship between nodes is clearly pointed out, and finally a large causal model for predicting autism spectrum disorders in children is generated.

[0057] S6. A causal macromodel for predicting autism spectrum disorders in children, using counterfactual and debiasing strategies; In the application of causal large models for knowledge graphs, the causal nodes and edges in the causal graph may be masked, resulting in incorrect prediction counterfactual subgraphs, which greatly affects the accuracy of model predictions. To this end, the counterfactual strategy is used to significantly enhance the distinction between the most relevant causal features and irrelevant features, and to facilitate the construction of a causal subgraph that only contains the former and accurately excludes the latter.

[0058] Specifically, the strategy first identifies a set of potential relevant features by deeply analyzing the nodes and edges in the causal graph. Then, counterfactual reasoning is used to strengthen the ability to identify the true causal features in these feature sets, making the boundary between them and irrelevant features clearer. This process is specifically shown in formula (11).

[0059] (11); In the formula, represents the set of relevant features identified; Represents a single feature in a feature set; Features The causal score of , which measures its likelihood of being a true causal feature; is the threshold set to filter features with high causal scores.

[0060] On this basis, a resampling-based feature selection strategy is further applied to effectively exclude spurious correlation features from the constructed causal subgraph. This strategy selects highly discriminative instances from different sampling distributions as the basis for constructing causal instance selection. These instances, due to their unique properties, can more accurately reflect the true relationship between causal features, thus helping to identify and eliminate spurious correlation features.

[0061] Furthermore, in order to eliminate the influence of spurious correlation features on model prediction, a causal invariance constraint is introduced. This constraint requires that each causal subgraph can still maintain a largely correct prediction when combined with various spurious correlation subgraphs, as shown in formula (12).

[0062] (12); In the formula, represents any combination of causal subgraphs and spurious correlation subgraphs, For the prediction accuracy. represents any combination of a causal subgraph and a spurious correlation subgraph, represents a causal subgraph, represents a spurious correlation subgraph.

[0063] After introducing the spurious correlation subgraph, the overall prediction accuracy should remain basically unchanged, thus verifying the effectiveness of the causal subgraph in extracting true causal features.

[0064] The debiasing strategy is applied to the causal large model for predicting autism spectrum disorder in children, aiming to effectively prevent the model from using data set bias for generation, while avoiding the inherent defects of fine-tuning methods. This strategy deeply analyzes the essential difference between biased information and semantic information, that is, there is a lack of stable causal relationship between biased information and subsequent text, which makes it possible to distinguish the two in the model.

[0065] Specifically, we first identify instances that cause the model to fail to make consistent predictions. These instances contain bias information in their representations and are therefore considered potential bias instances. For example, if there is an instance a that satisfies (in are model parameters, a and are similar inputs but differ in key characteristics), then a or May contain biased information.

[0066] Once the deviation instance is identified, its deviation representation vector is further extracted. In order to reduce the computational complexity and improve the clustering effect, the principal component analysis (PCA) technique is first applied to reduce the dimension of the deviation representation vector to two dimensions. The specific expression of PCA is shown in formula (13).

[0067] U = VD (13); Where V represents the original deviation representation matrix; D represents the principal component weight matrix; and U represents the representation matrix after dimensionality reduction.

[0068] However, as the data dimension increases, the effectiveness of traditional distance metrics (such as Euclidean distance) will significantly decrease, which in turn affects the performance of clustering algorithms. Therefore, a density-based clustering method DBSCAN is used to perform clustering analysis based on the deviation representation vector after dimensionality reduction. The core idea of ​​DBSCAN is to divide data points in high-density areas into the same cluster by calculating the local density of data points, while data points in low-density areas are regarded as noise or boundary points.

[0069] After clustering, the deviation pairs are extracted from each cluster and fed into a graph convolutional neural network (GCN) to summarize the deviation patterns. GCN can effectively learn and represent the characteristics of deviation patterns by capturing the local graph structure information between nodes.

[0070] Finally, we exploit automatically induced bias patterns to explicitly guide LLMs to avoid using specific types of information during reasoning in the form of additional text.

[0071] S7. Propose a method for training and optimizing a large brain-inspired causal model for predicting autism spectrum disorders in children under the guidance of counterfactuals. After successfully building a large causal model for predicting autism spectrum disorders in children guided by counterfactuals, it is necessary to start training and optimizing the model to verify whether it can accurately output relevant behavioral data, brain data, and prediction results of autism spectrum disorders in children by inputting multi-dimensional data information such as psychological and emotional states, external behavioral patterns, and achievement development trajectories. First, the model is placed in a corpus containing massive and diverse text data for extensive pre-training. This process uses self-supervised learning technology to enable the model to autonomously learn language patterns, structural features, and general understanding of contextual information from the data. Specifically, as shown in formula (14).

[0072] (14); In the formula, represents the model parameters; F represents the data distribution in the corpus; y represents a sample randomly drawn from the data; Indicates that the model has given parameters The probability of generating sample y under the condition of ; by maximizing the log-likelihood function, the optimal model parameters are found .

[0073] The model may have a "causal blind spot". This "causal blind spot" phenomenon seriously restricts the potential of LLMs to perform meaningful causal reasoning without fine-tuning or prompt engineering, which in turn limits its practicality in practical application scenarios that require deep causal understanding. To overcome this limitation, a causal fine-tuning strategy is introduced to optimize the model in a targeted manner. After the pre-training stage, the model is further supervised and fine-tuned using annotated datasets to adapt to specific downstream tasks such as machine translation, text summarization, or question answering. This process is achieved by minimizing the task-specific loss function, as shown in formula (15).

[0074] (15); In the formula, represents the labeled dataset; j and h represent the input and the corresponding label respectively; Indicates that the model has given parameters The predicted output for input j is: Represents the loss function between the predicted output and the true label; by minimizing the loss function, the model parameters that best suit the downstream task are found. Based on supervised fine-tuning, causal knowledge is further introduced to fine-tune the model to enhance its causal reasoning ability.

[0075] S8. Prediction and assessment of autism spectrum disorders in children based on brain-inspired causal big models.

[0076] After the training is completed, the causal large model for predicting autism spectrum disorders in children guided by counterfactuals is deployed in practical applications. Its core operating mechanism lies in sophisticated prompt engineering technology. Through input prompt design, the model can use its deep representation learning ability to generate clear and logically coherent text content, efficiently retrieve relevant information, or actively participate in dialogue exchanges in various natural language processing (NLP) tasks. Specifically, the model can receive multi-dimensional and highly complex data information including psychological and emotional states, external behavior patterns, and achievement development trajectories as input, and based on its powerful causal reasoning ability, accurately output relevant behavioral data, brain data, and prediction results of autism spectrum disorders in children.

[0077] In order to ensure that the model can perform as expected in practical applications, its performance is comprehensively and rigorously evaluated, covering key indicators such as task-specific accuracy, generalization ability to unknown data, ethical alignment, and robustness. Among them, the task-specific accuracy is quantified, as shown in formula (16).

[0078] (16); Among them, Accuracy represents the accuracy, Number of Correct Predictions represents the number of correctly predicted samples, and Total Number of Predictions represents the total number of predicted samples.

[0079] This formula is used to measure the accuracy of the model's prediction results on a specific task. For the evaluation of generalization ability, methods such as cross-validation are used to test the performance of the model on unseen data.

[0080] In terms of ethical alignment, we ensure that the output of the model is consistent with human values, ethical principles, and expected behaviors. To this end, we introduce human feedback reinforcement learning (RLHF) technology to optimize the model's response strategy through direct human judgment. The specific formula is shown in formula (17).

[0081] (17); in, represents the parameters of the model, represents a dataset containing human feedback, e and r represent the input and the corresponding output respectively, represents the probability that the model will generate an output given its parameters and input, represents the human reward function for the output, Represents a weight coefficient for balancing generation probability and ethical rewards. By maximizing this objective function, we can find model parameters that are more in line with social and ethical standards while maintaining high generation quality.

[0082] Therefore, the present invention adopts the above-mentioned brain-inspired causal big model prediction method for children with autism spectrum disorder, integrates multimodal data, and uses knowledge graphs to construct a causal big model to reveal the potential causal relationship between various factors of autism spectrum disorder ASD, providing more accurate and effective guidance for the early prediction of autism spectrum disorder ASD.

[0083] The above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for predicting autism spectrum disorders in children based on a brain-inspired causal model, characterized in that: The following steps are involved: S1. Collection and processing of behavioral and brain data; Combining the processed behavioral data and brain data into a multimodal dataset; S2. Clean, normalize and feature select the multimodal data set. When processing unstructured text data, introduce causal embedding technology to fuse the behavioral data and brain data output after causal embedding technology to form a unified multimodal data set. Based on the unified multimodal data set, use the structural equation model to construct a causal diagram and display the causal relationship between variables. S3. Process behavioral data, brain data, and causal relationships into structured data and store them in the form of triples; integrate the data on the three core dimensions of psychological and emotional state, external behavioral patterns, and achievement development trajectory into structured data, also stored in the form of triples, and initially construct a knowledge graph; Together with the data of the causal graph, a complete knowledge graph is constructed; S4. Convert all acquired data into text form and use the masked language model to encode the knowledge graph text; Under the classification mechanism of masked entity model, the effectiveness of masked entity model is verified by predicting masked entities in triples; the integrity and accuracy of knowledge graph are enhanced through semantic embedding and structural embedding techniques; S5. Introduce Bayesian network learning and differential causal rule mining. First, introduce the Bayesian network learning mechanism to build a probabilistic relationship model, and then use differential causal rule mining technology; perform causal query and causal path identification through metadata semantics; finally generate a causal large model for the prediction of autism spectrum disorders in children; S6. A causal macromodel for predicting autism spectrum disorders in children, using counterfactual and debiasing strategies; S7. Propose a method for training and optimizing a large brain-inspired causal model for predicting autism spectrum disorders in children under the guidance of counterfactuals. S8. Prediction and assessment of autism spectrum disorders in children based on brain-inspired causal big models.

2. The method for predicting autism spectrum disorder in children based on a brain-inspired causal model according to claim 1, characterized in that: In step S1, the pinhole imaging model and the head rotation of the observed person are used to collect and process the children's behavioral data; the brain data are obtained through the magnetic resonance imaging MRI technology and electroencephalogram EEG technology in graph theory analysis, and the obtained brain data are regarded as variables using the functional connection method to construct a brain network and measure the global characteristics of the observed person.

3. The method for predicting autism spectrum disorder in children based on a brain-inspired causal model according to claim 1, characterized in that: In step S2, the specific method of constructing the causal diagram is as follows: according to the research hypothesis and data characteristics, the latent variables, observed variables and the relationship between them are set, and then the maximum likelihood estimation method is used to estimate the variance of the factor loading matrix, the structural coefficient matrix and the error term. The chi-square test is used to compare the quality of model fitting and verify the validity and accuracy of the model. Finally, based on the estimated parameters and model structure, a causal diagram is drawn to intuitively display the causal relationship between the variables, as shown in the following formula: ； Where A represents the processed behavioral data vector; M represents the processed brain data vector; Represents a latent variable vector that cannot be directly observed in the data; Represents the factor loading matrix corresponding to the behavioral data; represents the factor loading matrix corresponding to the brain data; represents the random error term introduced when observing behavioral data; represents the random error term introduced when observing brain data; Q The structural coefficient matrix representing the causal relationship between latent variables; Represents the influence matrix of exogenous variables on latent variables; represents the random error term introduced when estimating the latent variable.

4. The method for predicting autism spectrum disorder in children based on a brain-inspired causal model according to claim 1, characterized in that: In step S3, in terms of psychological and emotional state, the psychological and emotional data of the observed person is obtained and quantitatively constructed into a set of psychological and emotional instances. ; In terms of behavioral patterns, obtain the observed behavior performance data and quantify it, and construct a set of behavioral performance instances ; In terms of achievement development trajectory, obtain the achievement level data of the observed and quantify and construct a set of achievement level instances ; Among them, I represents the psychological emotion of the observed person, The set represents the value of psychological emotions, H represents the behavior of the observed person, The set represents the value of the behavior performance, L represents the achievement level of the observed person, A collection of values ​​representing achievement levels.

5. The method for predicting autism spectrum disorder in children based on a brain-inspired causal model according to claim 1, characterized in that: In step S4, two strategies, entity description prediction and incomplete triple prediction, are adopted. In entity description prediction, the triple is first encoded and the final hidden state is generated; then, it is passed as the embedding of the head entity to the incomplete triple prediction to predict the target entity; Specifically, for a given head entity and relationship , the model predicts the tail entity The probability of is shown in the following formula: ； In the formula, Indicates that the model predicts the tail entity The probability of Represents the header entity ,relation and candidate tail entity The score between Represents the tail entity collection.

6. The method for predicting autism spectrum disorder in children based on a brain-inspired causal model according to claim 1, characterized in that: In step S5, the difference causal rule is expressed as follows: if the control variable W changes, the effect variable G changes accordingly with probability P. In this framework, W plays the role of the control variable, G acts as the effect variable, and P quantitatively reflects the magnitude of the probability change of the effect variable G after the control variable W changes; as shown in the following formula; ； In the formula, and They represent the conditional probability of the control variable W under the new and old states of the effect variable G; This is the change in the probability of the effect variable.

7. The method for predicting autism spectrum disorder in children based on a brain-inspired causal model according to claim 1, characterized in that: In step S6, by deeply analyzing the nodes and edges in the causal graph, a potential set of relevant features is identified, using counterfactual reasoning.

8. The method for predicting autism spectrum disorder in children based on a brain-inspired causal model according to claim 1, characterized in that: In step S7, the brain-inspired causal big model of autism spectrum disorders in children is placed in a corpus of text data for pre-training. After the pre-training stage, the brain-inspired causal big model of autism spectrum disorders in children is causally fine-tuned using the annotated dataset, and the annotated dataset is further used to perform supervised fine-tuning on the brain-inspired causal big model of autism spectrum disorders in children.

9. The method for predicting autism spectrum disorder in children based on a brain-inspired causal macromodel according to claim 1, characterized in that: In step S8, after the training is completed, the brain-inspired causal big model of autism spectrum disorder in children is deployed to practical applications. By using sophisticated prompt engineering technology, the brain-inspired causal big model of autism spectrum disorder in children receives multi-dimensional and highly complex data information as input, and the brain-inspired causal big model of autism spectrum disorder in children accurately outputs relevant behavioral data, brain data, and prediction results of autism spectrum disorder in children. Finally, the performance of the brain-inspired causal big model of autism spectrum disorder in children is comprehensively and rigorously evaluated, and the evaluation dimensions cover task-specific key indicators.

Images