Intelligent assessment system and method for child development disorder based on multi-modal data

Abstract

The present application belongs to the technical field of child development assessment, and particularly relates to a child development disorder intelligent assessment system and method based on multi-modal data. The present application can distinguish between occasional abnormalities that occur for a short time and abnormal performance that persists by simultaneously considering the duration and deviation of abnormal performance, and can form comprehensive abnormal representations of children in different development ability dimensions by classifying in combination with different development ability dimensions, thereby enhancing the representation ability of the assessment results for persistent development abnormalities and abnormal performance in different ability dimensions. By comprehensively considering the ability dimension, duration and deviation of abnormal characteristics, the present application avoids assessment based on only a single index or single abnormal data, so that the assessment results can reflect the category, degree and persistent characteristics of child development abnormalities, and the accuracy of child development state assessment is improved.

Description

Technical Field

[0001] This invention belongs to the field of child development assessment technology, specifically relating to an intelligent assessment system and method for child developmental disorders based on multimodal data. Background Technology

[0002] Children are in a critical stage of growth and development, and their cognitive, language, motor, social interaction, attention, and emotional regulation abilities change continuously with age and developmental stage. When a child shows persistent deviations in one or more of these developmental ability dimensions, it may manifest as developmental abnormalities such as inattention, insufficient motor control, delayed language expression, poor motor coordination, and abnormal social interaction. Therefore, in the process of child health care and developmental behavior assessment, it is necessary to comprehensively obtain information from children's behavioral performance, physiological responses, and standardized test results in different scenarios to form a comprehensive assessment of the child's developmental status. Current child development assessments typically use developmental scales, behavioral observation, medical records, and some physiological indicators. Developmental scales can evaluate children's developmental levels from different areas such as language, motor, adaptive, and social behavior, while behavioral observation mainly relies on the assessment personnel recording children's behavioral performance in specific tasks or interactive scenarios. Related research also shows that developmental abnormalities in children often have multidimensional manifestations. For example, language developmental abnormalities are not necessarily limited to language ability itself, but may also be accompanied by abnormalities in other ability dimensions such as motor, adaptive, and social behavior, thus requiring a comprehensive developmental assessment.

[0003] Current child development assessments suffer from several problems: inconsistent time benchmarks for multimodal data, difficulty in accurately linking behavior, physiology, and assessment events, and the use of uniform abnormality standards across different developmental stages, which can lead to assessment bias. Reliance on single indicators or single assessments is susceptible to short-term fluctuations in the child's condition. Furthermore, there is a lack of comprehensive representation of the process, duration, and degree of deviation of abnormalities. Coordinated abnormalities between different developmental ability dimensions are difficult to identify effectively, resulting in insufficient completeness and reliability of the final assessment results. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent assessment system and method for children's developmental disorders based on multimodal data. This system can accurately identify and grade developmental abnormalities by performing temporal correlation and multidimensional feature comprehensive characterization of children's behavioral, physiological and assessment data, thereby improving the accuracy of children's developmental status assessment.

[0005] The specific technical solution adopted by this invention is as follows: Intelligent assessment methods for children's developmental disorders based on multimodal data include: Acquire multimodal data of target children in preset scenarios, divide the multimodal data into developmental stages to obtain multimodal time series data, and combine preset rules to obtain dimensional feature data corresponding to each developmental ability dimension; Based on dimensional feature data, the locations of physiological changes and behavioral abnormalities are determined, and the trigger points for developmental abnormalities are determined based on the time overlap between the locations of physiological changes and behavioral abnormalities. Based on the developmental abnormality trigger point, the abnormality observation time window is determined, and the target dimension features corresponding to the abnormality observation time window are extracted from the dimensional feature data to obtain the trigger feature set; Based on the trigger feature set, the dimensional feature data are compared under the same capability dimension to generate the deviation label data of the corresponding capability dimension; Based on preset reference data, the target feature range corresponding to each developmental ability dimension and developmental stage is obtained. The trigger feature set and deviation identification data are compared with the target feature range set to obtain the unachieved features of each developmental ability dimension and form an abnormal feature sequence in chronological order. Based on the abnormal feature sequence, the degree of abnormality and deviation of the unachieved features are determined, and combined with the unachieved features corresponding to different developmental ability dimensions, developmental abnormality characterization data are obtained. Based on developmental abnormality representation data, and combined with the degree of abnormality persistence, degree of deviation and corresponding developmental ability dimensions, the developmental abnormality category and degree of abnormality are determined, and the child's developmental status assessment results are generated.

[0006] In a preferred embodiment, multimodal data of the target child in a preset scenario is acquired, the multimodal data is divided into developmental stages to obtain multimodal time-series data, and combined with preset rules, dimensional feature data corresponding to each developmental ability dimension is obtained, including: Acquire multimodal data of the target children in a preset scenario, including behavioral data, physiological data, and assessment data; The multimodal data were divided into developmental stages based on behavioral and assessment data, and the physiological data were allocated accordingly according to the developmental stage division results to obtain the stage datasets for the corresponding developmental stages. Based on the stage dataset, a unified time reference is constructed, and the behavioral data, physiological data, and evaluation data are time-calibrated to obtain multimodal time series data, which includes behavioral time series data, physiological time series data, and evaluation time series data. Based on multimodal time-series data, behavioral feature sequences, physiological feature sequences, and evaluation feature sequences are extracted from behavioral time-series data, physiological time-series data, and evaluation time-series data respectively to form multi-source feature sequences; Based on the preset developmental ability dimension division rules, behavioral feature sequences, physiological feature sequences and assessment feature sequences are assigned to the corresponding developmental ability dimensions. Among them, the developmental ability dimensions include social interaction ability dimension, attention control ability dimension, language expression ability dimension, motor coordination ability dimension and emotion regulation ability dimension, thus obtaining the dimension feature set corresponding to each developmental ability dimension. Based on the dimensional feature set, the behavioral feature sequence, physiological feature sequence, and assessment feature sequence under the same developmental ability dimension are integrated and processed to generate dimensional feature data corresponding to each developmental ability dimension. The dimensional feature data includes behavioral representation information, physiological response information, and assessment scale information under the corresponding developmental ability dimension.

[0007] In a preferred embodiment, based on dimensional feature data, the locations of physiological changes and behavioral abnormalities are determined, and the trigger points for developmental abnormalities are determined based on the temporal overlap between the locations of physiological changes and behavioral abnormalities, including: Extract physiological feature sequences corresponding to each developmental ability dimension from dimensional feature data, detect the change amplitude of physiological feature sequences according to adjacent temporal positions, and determine the temporal positions where the change amplitude exceeds the preset change threshold as physiological change positions. Extract the behavioral feature sequence corresponding to each developmental ability dimension from the dimensional feature data, and determine the temporal position where the behavioral feature exceeds the corresponding reference range of change as the behavioral abnormal position; The locations of physiological changes and abnormal behaviors are matched in chronological order to determine the time overlap between the two within a preset time interval. The temporal position that simultaneously corresponds to the location of physiological changes and the location of behavioral abnormalities within the time overlap interval is determined as the developmental abnormality trigger point.

[0008] In a preferred embodiment, an abnormality observation time window is determined based on the developmental abnormality trigger point, and the target dimensional features corresponding to the abnormality observation time window are extracted from the dimensional feature data to obtain a trigger feature set, including: Obtain the trigger timing position corresponding to the developmental abnormality trigger point, and extend it by a preset time in the direction before and after the trigger timing position to obtain the abnormality observation time window; The dimensional feature data are filtered by time according to the anomaly observation time window to extract behavioral features, physiological features and evaluation features located within the anomaly observation time window. Based on the extracted behavioral features, physiological features, and assessment features corresponding to the developmental ability dimensions, the target dimension features corresponding to each developmental ability dimension are categorized to obtain the trigger feature set.

[0009] In a preferred embodiment, based on a trigger feature set, dimensional feature data are compared within the same capability dimension to generate deviation identification data for the corresponding capability dimension, including: The developmental ability dimension is obtained based on the trigger feature set, and the corresponding dimension feature data is obtained from the dimension feature data. Based on preset reference data, obtain reference feature data corresponding to the developmental ability dimension and developmental stage. The feature data in the trigger feature set is compared with the reference feature data within the corresponding time range to obtain feature difference information, which includes behavioral feature differences, physiological feature differences and evaluation feature differences. Based on the feature difference information, determine whether each feature in the feature difference information exceeds the corresponding preset difference range, and obtain the deviation determination result; Based on the deviation determination results, corresponding deviation label data for the developmental ability dimension is generated.

[0010] In a preferred embodiment, the target feature range corresponding to each developmental ability dimension and developmental stage is obtained based on preset reference data. The trigger feature set and deviation identification data are compared with the target feature range set to obtain the unachieved features for each developmental ability dimension. An abnormal feature sequence is then formed in chronological order, including: Based on preset reference data, the target feature range corresponding to each developmental ability dimension is obtained, and a target feature range set is constructed; Based on the trigger feature set, behavioral features, physiological features and assessment features under each developmental ability dimension are extracted, and the corresponding abnormal feature items are screened out in combination with the deviation identification data to obtain the feature data to be compared for each developmental ability dimension. The feature data to be compared for each developmental ability dimension are matched with the target feature range set according to time position and feature type to obtain the feature alignment result; Based on the feature alignment results, obtain the difference information between each feature to be compared and the corresponding target feature range; Based on the difference information and the preset judgment threshold, each feature to be compared is judged to see whether it exceeds the target feature range, and the judgment result is corrected by combining the deviation identification data to obtain the feature deviation result. Based on the feature deviation results, feature items that do not fall within the corresponding target feature range are selected as unachieved features and formed into an abnormal feature sequence in chronological order.

[0011] In a preferred embodiment, the degree of abnormality and deviation of unachieved features are determined based on the abnormal feature sequence, and developmental abnormality characterization data are obtained by combining the unachieved features corresponding to different developmental capacity dimensions, including: Obtain the consecutive occurrence time sequence positions corresponding to each unachieved feature from the abnormal feature sequence, and determine the abnormal persistence degree of each unachieved feature based on the consecutive occurrence time sequence positions; Obtain the deviation range between the unachieved feature and the corresponding reference feature, and determine the degree of deviation for each unachieved feature; The degree of abnormality persistence and deviation are classified according to the developmental ability dimension to which the unachieved feature belongs, and the abnormal feature representation corresponding to each developmental ability dimension is obtained. The abnormal feature representations corresponding to each developmental ability dimension are aggregated to obtain developmental abnormality representation data.

[0012] In a preferred embodiment, based on developmental abnormality characterization data, and combined with the degree of abnormality persistence, deviation, and corresponding developmental ability dimensions, the category and severity of developmental abnormality are determined, generating a child's developmental status assessment result, including: The abnormal feature representations corresponding to each developmental ability dimension are obtained from the developmental abnormality representation data, and the abnormal feature level is determined according to the degree of abnormality persistence and deviation of the abnormal feature representations. The abnormality category of the target child is determined by classifying the abnormality feature level corresponding to each developmental ability dimension. The degree of abnormality corresponding to the developmental abnormality category is determined based on the level of abnormality characteristics and the degree of abnormality persistence. The developmental abnormality categories, abnormality severity, and corresponding developmental ability dimensions are compiled to generate a child development status assessment result.

[0013] The present invention also provides an intelligent assessment system for children's developmental disorders based on multimodal data, used in the aforementioned intelligent assessment method for children's developmental disorders based on multimodal data, comprising: The dimensional feature module is used to acquire multimodal data of the target child in a preset scenario, divide the multimodal data into developmental stages to obtain multimodal time series data, and combine preset rules to acquire dimensional feature data corresponding to each developmental ability dimension. The abnormality triggering module determines the location of physiological changes and abnormal behavior based on dimensional feature data, and determines the developmental abnormality trigger point based on the time overlap interval between the locations of physiological changes and abnormal behavior. The trigger feature module determines the abnormal observation time window based on the developmental abnormality trigger point, and extracts the target dimension features corresponding to the abnormal observation time window from the dimension feature data to obtain the trigger feature set; The deviation identification module, based on the trigger feature set, compares the dimensional feature data under the same capability dimension and generates deviation identification data for the corresponding capability dimension. The module for unachieved features obtains the target feature range corresponding to each developmental ability dimension and developmental stage based on preset reference data, compares the trigger feature set and deviation identification data with the target feature range set, obtains the unachieved features for each developmental ability dimension, and forms an abnormal feature sequence in chronological order. The anomaly characterization module determines the degree of persistence and deviation of unachieved features based on the anomaly feature sequence, and obtains developmental anomaly characterization data by combining the unachieved features corresponding to different developmental ability dimensions. The developmental assessment module determines the type and severity of developmental abnormalities based on data representing developmental abnormalities, combined with the degree of abnormality persistence, deviation, and corresponding developmental ability dimensions, and generates assessment results of the child's developmental status.

[0014] And, a smart assessment terminal for children's developmental disorders based on multimodal data, including: One or more processors; A storage device on which one or more programs are stored; When one or more programs are executed by one or more processors, the one or more processors implement an intelligent assessment method for children's developmental disorders based on multimodal data.

[0015] The technical effects achieved by this invention are as follows: This invention addresses the challenge of matching behavioral, physiological, and assessment data due to varying collection times and granularities by dividing developmental data from different sources into developmental stages and standardizing the time signature. It aggregates multimodal features according to developmental capacity dimensions, enabling joint representation of different modalities around the same developmental capacity dimension. This improves the accuracy and completeness of subsequent abnormal feature localization and developmental status assessment. By simultaneously utilizing the temporal synchronization features of physiological changes and behavioral abnormalities to determine developmental abnormality trigger points, it reduces the possibility of misjudgment caused by instantaneous fluctuations in single-modal data compared to relying solely on single behavioral or physiological indicators. Furthermore, by establishing an abnormality observation time window around the developmental abnormality trigger point, it eliminates the need for indiscriminate processing of the entire child's developmental data. Instead, it extracts multimodal features from the time periods preceding and following the abnormality trigger point, preserving the feature changes before and after the abnormality while reducing interference from data unrelated to the abnormal event, thus improving the targeting of abnormal feature extraction. This invention avoids misjudgments caused by normal developmental differences between children at different developmental stages by comparing features within the same developmental ability dimension, the same developmental stage, and the corresponding time range. By separately confirming deviations in three categories of features—behavioral, physiological, and assessment—abnormal manifestations can be cross-represented from different data sources, thereby improving the reliability of abnormal feature identification. By using the target feature range corresponding to the developmental stage to perform secondary screening of the identified abnormal features, it avoids directly using features that fluctuate briefly or fall within the normal developmental range as evidence of developmental abnormalities. The unachieved features obtained from the screening are arranged in chronological order to form an abnormal feature sequence, giving the originally discrete abnormal features a continuous temporal representation. This invention, by simultaneously considering the persistence and deviation of abnormal manifestations, can distinguish between short-term, occasional abnormalities and persistent abnormal manifestations. By categorizing these abnormalities according to different developmental ability dimensions, it can form a comprehensive abnormal representation of children across these dimensions. This enhances the assessment results' ability to represent persistent developmental abnormalities and abnormal manifestations across different ability dimensions. By comprehensively considering the ability dimension to which the abnormal characteristics belong, the persistence, and the deviation, it avoids assessment based solely on a single indicator or a single instance of abnormal data. This allows the assessment results to reflect the category, degree, and persistence characteristics of children's developmental abnormalities, thereby improving the accuracy of children's developmental status assessment. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method provided by the present invention; Figure 2 This is a system module diagram provided by the present invention. Detailed Implementation

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0018] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0019] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in a preferred embodiment" appearing in different places throughout this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0020] Furthermore, the present invention will be described in detail with reference to the schematic diagrams. When describing the embodiments of the present invention in detail, the schematic diagrams are merely examples for ease of explanation and should not limit the scope of protection of the present invention.

[0021] Please see the appendix Figure 1 As shown, an intelligent assessment method for children's developmental disorders based on multimodal data is provided, including: S1. Obtain multimodal data of the target child in a preset scenario, divide the multimodal data into developmental stages to obtain multimodal time series data, and combine with preset rules to obtain dimensional feature data corresponding to each developmental ability dimension. S2. Based on dimensional feature data, determine the location of physiological changes and the location of behavioral abnormalities, and determine the trigger point of developmental abnormalities based on the time overlap interval between the locations of physiological changes and behavioral abnormalities. S3. Determine the abnormal observation time window based on the developmental abnormality trigger point, and extract the target dimension features corresponding to the abnormal observation time window from the dimension feature data to obtain the trigger feature set; S4. Based on the trigger feature set, compare the dimensional feature data under the same capability dimension to generate the deviation label data of the corresponding capability dimension; S5. Based on preset reference data, obtain the target feature range corresponding to each developmental ability dimension and developmental stage, compare the trigger feature set and deviation identification data with the target feature range set, obtain the unachieved features of each developmental ability dimension, and form an abnormal feature sequence in chronological order. S6. Based on the abnormal feature sequence, determine the degree of abnormality and deviation of the unachieved features, and combine the unachieved features corresponding to different developmental ability dimensions to obtain developmental abnormality characterization data. S7. Based on the developmental abnormality representation data, and combined with the degree of abnormality persistence, degree of deviation and corresponding developmental ability dimensions, determine the type and degree of developmental abnormality, and generate the child's developmental status assessment results.

[0022] As described in steps S1 to S7 above, behavioral data, physiological data, and assessment data of the target child in a preset scenario are acquired. Based on the behavioral and assessment data, the developmental stage of the target child is determined, and the physiological data is assigned to the corresponding developmental stage. A unified time reference is established, and data from different sources are time-calibrated so that the behavioral, physiological, and assessment data form multimodal time-series data with a unified time position. According to the preset developmental ability dimension division rules, the feature sequences from different sources are assigned to the corresponding developmental ability dimensions. The behavioral, physiological, and assessment features under the same developmental ability dimension are integrated to form corresponding dimensional feature data. By dividing the developmental stage and unifying the time calibration of the child's developmental data from different sources, the problem of difficulty in corresponding behavioral, physiological, and assessment data due to different collection times and data granularities can be solved. Multimodal features are collected according to the developmental ability dimension so that different modal data can be jointly represented around the same developmental ability dimension, thereby improving the accuracy and completeness of subsequent abnormal feature localization and developmental status assessment. Physiological and behavioral feature sequences are extracted from the data. For the physiological feature sequences, the location of physiological changes is determined based on the changes between adjacent time-series positions. For the behavioral feature sequences, the location of behavioral anomalies is determined based on whether the behavioral features exceed the corresponding reference change range. Then, the physiological change locations and behavioral anomaly locations are matched according to a unified time benchmark. When the physiological change location and the behavioral anomaly location overlap within a preset time interval, the corresponding time-series position is determined as the developmental abnormality trigger point. By simultaneously utilizing the time synchronization features of physiological changes and behavioral anomalies to determine the developmental abnormality trigger point, compared to judging anomalies based on only a single behavioral indicator or a single physiological indicator, the possibility of misjudgment caused by instantaneous fluctuations in single modality data can be reduced. Based on the triggering time position corresponding to the developmental abnormality trigger point, a preset time period is extended in the direction before and after it to form an abnormality observation time window. Then, the dimensional feature data is filtered by time using the abnormality observation time window, and only the behavioral features, physiological features, and assessment features located within the time window are extracted. These features are then classified and aggregated according to the corresponding developmental ability dimensions to form a trigger feature set. By establishing an abnormality observation time window around the developmental abnormality trigger point, it is no longer necessary to process the entire child's developmental data indiscriminately. Instead, it is possible to extract multimodal features of the time period before and after the abnormality trigger position. This can preserve the feature change process before and after the abnormality occurs and reduce data interference unrelated to the abnormal event, thereby improving the targeting of abnormal feature extraction. Based on the developmental ability dimension to which the trigger feature belongs, feature data corresponding to the same developmental ability dimension is obtained from the dimension feature data, and reference feature data corresponding to the developmental ability dimension and developmental stage is obtained. The behavioral features, physiological features, and assessment features in the trigger feature set are compared with the reference feature data to determine the feature differences of different modal features. Based on the preset difference range, it is judged whether there is a significant deviation in each feature. Finally, deviation label data corresponding to the developmental ability dimension is formed. By comparing features within the same developmental ability dimension, the same developmental stage, and the corresponding time range, erroneous judgments caused by normal developmental differences between children at different developmental stages can be avoided. By confirming the deviations of the three types of features (behavioral, physiological, and assessment), abnormal performance can be cross-characterized from different data sources, thereby improving the reliability of abnormal feature identification. The target feature ranges corresponding to different developmental ability dimensions and different developmental stages are obtained and a target feature range set is formed. Behavioral, physiological and assessment features under different developmental ability dimensions are extracted from the trigger feature set, and abnormal feature items are screened out using deviation identification data. The screened abnormal features are matched with the target feature range according to time position and feature type to determine whether each abnormal feature enters the corresponding target feature range. Feature items that do not enter the target feature range are screened as unachieved features. Finally, the unachieved features are arranged in chronological order to form an abnormal feature sequence. By using the target feature range corresponding to the developmental stage to screen the already identified abnormal features, it is possible to avoid directly using features with short-term fluctuations or those within the normal developmental range as the basis for developmental abnormalities. The unachieved features are arranged in chronological order to form an abnormal feature sequence, so that the originally discrete abnormal features have a continuous temporal representation. The system obtains the consecutive occurrence positions of each unachieved feature in the time series and determines the degree of abnormality based on the consecutive occurrence. It also determines the degree of deviation based on the deviation range of the unachieved feature relative to the corresponding reference feature. According to the developmental ability dimension to which the unachieved feature belongs, the system classifies the degree of abnormality and the degree of deviation corresponding to different unachieved features to form abnormal feature representations corresponding to different developmental ability dimensions. The abnormal feature representations of each developmental ability dimension are then aggregated to obtain developmental abnormality representation data. By considering both the degree of persistence and the degree of deviation of abnormal manifestations, it is possible to distinguish between short-term occasional abnormalities and persistent abnormal manifestations. Combining different developmental ability dimensions for classification can form a comprehensive abnormal representation of children in different developmental ability dimensions, which can enhance the ability of the assessment results to represent persistent developmental abnormalities and abnormal manifestations in different ability dimensions. Based on developmental abnormality data, the duration and deviation of abnormalities are comprehensively graded according to the developmental ability dimension to which the abnormal features belong. Unachieved features with the same ability dimension and common abnormal characteristics are categorized to determine the corresponding developmental abnormality category. The degree of abnormality is determined by combining the abnormal feature level, duration, and deviation. Finally, the results are compiled to form the child's developmental status assessment. By comprehensively considering the ability dimension to which the abnormal features belong, the duration, and the deviation, the assessment avoids relying solely on a single indicator or a single abnormal data point. This ensures that the assessment results can reflect the category, degree, and duration of the child's developmental abnormalities, thereby improving the accuracy of the child's developmental status assessment.

[0023] In a preferred embodiment, multimodal data of the target child in a preset scenario is acquired, the multimodal data is divided into developmental stages to obtain multimodal time-series data, and combined with preset rules, dimensional feature data corresponding to each developmental ability dimension is obtained, including: S101. Obtain multimodal data of the target child in a preset scenario, including behavioral data, physiological data and assessment data; S102. Divide the multimodal data into developmental stages based on behavioral data and assessment data, and allocate the physiological data accordingly according to the developmental stage division results to obtain the stage dataset for the corresponding developmental stage. S103. Based on the stage dataset, construct a unified time reference and perform time calibration processing on behavioral data, physiological data and evaluation data to obtain multimodal time series data, which includes behavioral time series data, physiological time series data and evaluation time series data. S104. Based on multimodal time series data, extract behavioral feature sequences, physiological feature sequences, and evaluation feature sequences from behavioral time series data, physiological time series data, and evaluation time series data respectively to form multi-source feature sequences; S105. Based on the preset developmental ability dimension division rules, the behavioral feature sequence, physiological feature sequence and assessment feature sequence are assigned to the corresponding developmental ability dimension. The developmental ability dimension includes the social interaction ability dimension, attention control ability dimension, language expression ability dimension, motor coordination ability dimension and emotion regulation ability dimension, and the dimension feature set corresponding to each developmental ability dimension is obtained. S106. Based on the dimensional feature set, integrate and process the behavioral feature sequence, physiological feature sequence and assessment feature sequence under the same developmental ability dimension to generate dimensional feature data corresponding to each developmental ability dimension. The dimensional feature data includes behavioral representation information, physiological response information and assessment scale information under the corresponding developmental ability dimension.

[0024] As described in steps S101 to S106 above, behavioral data, physiological data, and assessment data of the target child are collected in a preset scenario. Behavioral data describes the child's external behavior during the actual interaction process. This data can be continuously collected using a camera device, capturing information such as the child's body movements, facial expressions, gaze direction, voice, and activity trajectory. For example, in a 10-minute parent-child interaction scenario, video data is acquired at a sampling frequency of 25 frames per second, resulting in 25 x 60 = 1500 frames per minute, and 1500 x 10 = 15000 frames per minute. Behavioral data such as the child's gaze duration, limb movement duration, number of movements, and number of verbal responses are also collected from the video. Physiological data reflects the child's physiological response in the same scenario and can be acquired using a heart rate sensor, skin conductance sensor, or other wearable data collection devices. For example, heart rate data is acquired at a frequency of once per second for 10 minutes. It can obtain 1 multiplied by 600, which equals 600 heart rate sampling values; the assessment data is used to reflect the child's existing developmental assessment results and can be obtained through child health assessment scales, language ability assessment items, attention assessment items, and motor ability assessment items. For example, it records the child's score, completion time, and number of errors for each task according to the preset assessment items; after the collection is completed, the collection time, data type, and identification information corresponding to the target child are added to each data point. For example, if the collection time of a certain behavioral data is 125 seconds, the collection time of physiological data is 125 seconds, and the assessment item is completed from 124 seconds to 128 seconds, then the above data are recorded at the corresponding time positions, thus forming multimodal raw data with source information and time information; by acquiring behavioral, physiological, and assessment data simultaneously, the problem of missing information between the child's actual behavioral performance, physiological response, and scale performance caused by relying on only a single data source can be avoided; Behavioral indicators reflecting a child's developmental stage are extracted from behavioral data, such as the number of language responses, continuous attention duration, action completion time, and number of social interactions. Simultaneously, assessment results for the corresponding age group are obtained from assessment data. The developmental stage corresponding to the current data is then determined based on the target child's actual age, assessment completion status, and the range of the behavioral indicators. For example, the child's data is divided into three consecutive developmental stages: Stage A, Stage B, and Stage C, with corresponding ranges set for indicators such as continuous attention duration and the number of language responses. Assuming the continuous attention duration range for Stage B is 5 to 10 minutes, and the target child's continuous attention duration in a single interaction scenario is 7 minutes, then 7 minutes falls between 5 and 10 minutes, therefore this behavioral data is classified as Stage B. If the corresponding language assessment score is 82 points, and the assessment reference range for Stage B is 75 to 90 points, then 82 points... Since it also falls within this range, the data is confirmed to belong to stage B. After completing the stage division of behavioral and assessment data, the physiological data is then assigned to the developmental stage corresponding to the same time range based on the collection time of each physiological data point. For example, if the data from the 1st to the 300th second is determined to be stage B, then all 600 physiological samples collected between the 1st and 300th seconds are assigned to the stage dataset corresponding to stage B. For data spanning multiple stages, it can be split according to its actual position in each stage based on its collection time. For example, if the first 6 seconds of a 10-second physiological data segment belong to stage B and the last 4 seconds belong to stage C, then the first 6 seconds of data are assigned to the stage B dataset, and the last 4 seconds are assigned to the stage C dataset. This ensures that behavioral, assessment, and physiological data remain consistent across developmental stages, avoiding the misinterpretation of normal stage differences as abnormal performance due to varying normal developmental levels in children at different stages. A preset scene start time is selected as the zero point of a unified time reference. For example, if the start time of the interactive scene is recorded as 0 seconds, then the first frame of the video corresponds to a time position near second 0. The data collected by the heart rate sensor 3 seconds after the scene starts corresponds to the time position of second 3. If the evaluator starts performing the evaluation item at 125 seconds, then the evaluation data corresponds to the evaluation time interval starting from 125 seconds. For data with different sampling frequencies, each data point is mapped to a unified time axis according to its collection time. For example, behavioral data generates a video frame every 0.04 seconds, physiological data generates a heart rate value every 1 second, and evaluation data generates an evaluation record according to the time of event occurrence. Then, the position of each data point on the unified time axis is determined according to the timestamp. For example, if behavioral characteristics are generated at 10.00 seconds, 10.04 seconds, and 10.08 seconds, and physiological characteristics are generated at 10 seconds, the evaluation data recording time is 10.05 seconds. If the data is collected in seconds, the data will be labeled to the corresponding time position around the 10th second according to the order of time. For scenarios that require a unified time granularity, 100 milliseconds can be set as a unified time slice, that is, every 100 consecutive milliseconds form a time slice; for example, the 100th time slice is formed from the 10.00 second to the 10.10 second. If two behavioral feature values, one physiological feature value, and one evaluation event are collected in this time slice, then the above data will be labeled to this time slice. If there is no corresponding data in a certain time slice, the missing state of the time slice will be retained, and the data of other time slices will not be directly copied to this time slice. By using a unified time reference, data from different modalities can be correlated at the same time position. For example, when a child shows attention shift behavior at the 35th second, the heart rate change and evaluation event around the 35th second can be obtained simultaneously, reducing abnormal position mismatches caused by inconsistent data collection times. For behavioral time-series data, behavioral features are extracted according to preset behavioral feature extraction items. For example, from continuous video data, gaze duration, number of verbal responses, number of actions, duration of actions, and interaction response time between people are extracted in chronological order. For instance, in a continuous 60-second interaction, if a child makes a verbal response at the 5th, 15th, 32nd, and 48th seconds, the number of verbal responses is accumulated as 1, 2, 3, and 4, respectively, thus forming a sequence of verbal response features arranged in chronological order. For physiological time-series data, heart rate changes, skin conductance changes, or other physiological response features are extracted from continuous physiological data. For example, if the heart rate is 82 beats / minute at the 20th second, 83 beats / minute at the 21st second, and 90 beats / minute at the 22nd second, then a sequence is formed according to the acquisition time. The system extracts physiological feature sequences at times 82, 83, and 90, and can record changes between adjacent time-series positions. For example, subtracting 82 from 83 yields 1, and subtracting 83 from 90 yields 7. For evaluation time-series data, features such as evaluation scores, task completion time, number of errors, and project completion status are extracted. For example, if the scores for the three evaluation items are 8, 6, and 5, an evaluation feature sequence of 8, 6, and 5 is formed. After extracting the three types of features, the behavioral feature sequence, physiological feature sequence, and evaluation feature sequence are mapped according to a unified time reference to form a multi-source feature sequence. By converting raw videos, continuous physiological signals, and discrete evaluation records into feature sequences with clear meanings, the amount of data generated by directly processing raw data can be reduced, while retaining information on changes in behavior, physiology, and evaluation performance over time. A pre-established rule for classifying developmental ability dimensions was developed. The dimensions to which each characteristic belonged were determined through child health assessment projects, commonly used clinical developmental evaluation projects, and the actual representation of each data feature. For example, developmental ability was divided into five dimensions: social interaction ability, attention control ability, language expression ability, motor coordination ability, and emotion regulation ability. Correspondence rules between features and dimensions were established: eye fixation duration, number of interactive responses, and number of proactive interactions were assigned to the social interaction ability dimension; continuous attention time, number of attention shifts, and task retention time were assigned to the attention control ability dimension; the number of language responses, number of words, and language response time were assigned to the language expression ability dimension; action completion time, number of actions, and deviation of action trajectory were assigned to the motor coordination ability dimension; and the number of emotional changes, emotional recovery time, and duration of emotional expression were assigned to the emotion regulation ability dimension. For example, if a child's eye fixation duration is 4 seconds... The heart rate was 86 beats / minute, and the number of language responses was 3. The fixation duration and corresponding physiological response can be classified into the social interaction ability dimension, while the number of language responses can be classified into the language expression ability dimension. If there is also a social interaction score in the assessment scale at the same time location, it can also be classified into the social interaction ability dimension. When a feature can reflect multiple ability dimensions, its classification can be determined according to the assessment item corresponding to the feature or the pre-defined main representation dimension. For example, task retention time is mainly used to reflect attention control ability, so it is classified into the attention control ability dimension. The behavioral, physiological and assessment features that were originally scattered according to the data source are reorganized according to the child's developmental ability to form a dimensional feature set corresponding to each developmental ability dimension. This allows for the simultaneous acquisition of behavioral performance, physiological response and assessment performance for a specific developmental ability dimension, thereby improving the correspondence between multimodal information. Behavioral, physiological, and assessment features at the same time point are extracted from the feature set corresponding to each developmental ability dimension. For example, in the social interaction ability dimension corresponding to 40 seconds, the behavioral feature is recorded as a child's gaze duration of 3 seconds, the physiological feature is recorded as a heart rate of 88 beats / minute, and the assessment feature is recorded as a score of 7 points in the social interaction item. These three features are then mapped to the unified time point of 40 seconds, forming behavioral representation information, physiological response information, and assessment scale information for that developmental ability dimension. For continuous time data, they are integrated sequentially according to a unified time slice. For example, if 100 milliseconds is set as a time slice, and the behavioral feature corresponding to 20 seconds to 20.1 seconds is a gaze duration of 0.08 seconds, the physiological feature is a heart rate of 85 beats / minute, and the assessment feature is task retention state 1, then all three are recorded together in the dimensional feature data corresponding to 20 seconds to 20.1 seconds. For different features with different numerical units... In this case, instead of directly adding the values, the original units and characteristic meanings of each data point are preserved. For example, heart rate is still recorded in beats per minute, time in seconds, and scale scores in points. For assessment data in the form of events, the events are mapped to corresponding behaviors and physiological characteristics according to their occurrence time. For example, if an assessment task error occurs at the 60th second, the number of assessment errors is increased by 1 in the dimension feature data corresponding to the 60th second. Ultimately, each developmental ability dimension forms dimension feature data containing behavioral representation information, physiological response information, and assessment scale information. Without destroying the original physical meaning of different data types, multimodal information is organized according to the same developmental ability dimension, the same time location, and different data sources. This allows for the direct extraction of physiological and behavioral feature sequences from the corresponding dimension feature data, and further determination of the location of physiological changes and behavioral abnormalities, improving the data continuity of the entire assessment process.

[0025] In a preferred embodiment, based on dimensional feature data, the locations of physiological changes and behavioral abnormalities are determined, and the trigger points for developmental abnormalities are determined based on the temporal overlap between the locations of physiological changes and behavioral abnormalities, including: S201. Extract the physiological feature sequence corresponding to each developmental ability dimension from the dimensional feature data, detect the change amplitude of the physiological feature sequence according to the adjacent temporal positions, and determine the temporal position where the change amplitude exceeds the preset change threshold as the physiological change position. S202. Extract the behavioral feature sequence corresponding to each developmental ability dimension from the dimensional feature data, and determine the temporal position where the behavioral feature exceeds the corresponding reference range of change as the behavioral abnormal position. S203. Match the locations of physiological changes with the locations of behavioral abnormalities in chronological order to determine the time overlap interval between the two within a preset time interval; S204. The temporal position of the location that simultaneously corresponds to the location of physiological change and the location of behavioral abnormality within the time overlap interval is determined as the developmental abnormality trigger point.

[0026] As described in steps S201 to S204 above, based on the dimensional feature data, corresponding physiological response information is extracted according to the developmental ability dimension, and the physiological response information is sorted by time according to a unified time benchmark to form a physiological feature sequence. For example, under the attention control ability dimension, heart rate data from the 1st to the 10th second are obtained sequentially at sampling time positions of 1 second each, resulting in a physiological feature sequence of 82, 83, 84, 85, 86, 87, 95, 96, 94, and 93 beats per minute. Then, the change amplitude of two adjacent physiological feature values ​​is compared sequentially according to adjacent time positions, for example, at the 2nd second... Subtracting 82 from the first second (83) gives 1 beat / minute; subtracting 83 from the second second (84) gives 1 beat / minute; subtracting 87 from the sixth second (95) gives 8 beats / minute; subtracting 95 from the seventh second (96) gives 1 beat / minute. A preset change threshold is used to determine if a significant change in physiological characteristics has occurred. For example, if the change threshold is set to 5 beats / minute, meaning the change in physiological characteristics at adjacent time points reaches or exceeds 5 beats / minute, the corresponding next time point is determined as the location of the physiological change. In the above sequence, the seventh second increases by 8 beats / minute compared to the sixth second. Since 8 is greater than 5, therefore... The 7th second is identified as the location of physiological change, while the changes at the 2nd and 3rd seconds, etc., are all less than 5 (1 change / minute), and therefore are not considered physiological change locations. In other embodiments, the change threshold can be set according to the sampling frequency of physiological data and the child's developmental stage. For example, it can be set in the range of 3 to 10 changes / minute. For physiological data with a high sampling frequency, the change trend of 3 to 5 consecutive sampling locations can be used for confirmation to avoid misjudgment caused by noise from a single sampling point. The change threshold can be obtained from historical normal children's samples. Specifically, physiological data of multiple children at the same developmental stage under the same preset scenario can be collected, time-calibrated in the above manner, and then the physiological change amplitude can be obtained according to adjacent time sequence positions. The range of change amplitude under normal conditions can be statistically analyzed. For example, if 1000 sets of adjacent physiological data are obtained, of which 950 sets of change amplitude are between 0 and 3 changes / minute, 40 sets are between 3 and 5 changes / minute, and 10 sets exceed 5 changes / minute, then 5 changes / minute can be used as the change threshold in the embodiment to reduce the situation where normal physiological fluctuations are identified as significant changes. This can filter out significant changes in continuous physiological data from a large number of normal fluctuations and retain their specific occurrence time. Behavioral representation information is extracted from dimensional feature data and formed into a sequence of behavioral features corresponding to the developmental ability dimension in chronological order. For example, in the dimension of attention control ability, continuous attention duration, number of attention shifts, and task retention time are extracted. Continuous attention duration can be selected as a feature for judging behavioral abnormalities. Assuming that behavioral feature values ​​are acquired every second during a continuous 10-second observation period, resulting in values ​​of 8, 8, 7, 7, 6, 2, 3, 3, 7, and 8 seconds, a reference range of variation corresponding to the same developmental stage as the target child is further obtained. For example, this can be achieved by collecting data from normal children at the same developmental stage in the same environment. The behavioral data in the context, after time calibration and feature extraction, yields a normal behavioral feature sequence. The normal range of continuous attention time is then statistically analyzed at the same time position, and the interval covering 95% of the normal samples is determined as the reference range. For example, if the statistical analysis shows that the normal continuous attention time at a corresponding time position is typically between 5 and 10 seconds, then 5 to 10 seconds is set as the reference range for this embodiment. When the target child's continuous attention time at the 6th second is 2 seconds, 2 is less than 5, therefore the 6th second exceeds the reference range, and the 6th second is determined as the abnormal behavioral position. The 7th second is 3 seconds, also less than 5, therefore... The 7th second was also identified as a location of behavioral abnormality, while the 1st second (8 seconds), 2nd second (8 seconds), etc., all fall within the range of 5-10 seconds, and therefore are not considered locations of behavioral abnormality. For different developmental ability dimensions, behavioral characteristics and reference ranges of variation corresponding to that dimension can be used. For example, the social interaction ability dimension can use the duration of eye contact and the number of interaction responses; the language expression ability dimension can use the number of language responses and the language response time; the motor coordination ability dimension can use the action completion time and the deviation of the action trajectory; and the emotion regulation ability dimension can use the emotion recovery time and the duration of emotion changes. Each reference range of variation can be obtained from historical normal children at the same developmental stage. For example, collecting data from 50 children at the same developmental stage, with each child interacting in the same scene for 10 minutes, yields 50 multiplied by 600, which equals 30,000 behavioral characteristic records with a time granularity of 1 second. Then, the normal value range is statistically analyzed according to the characteristic type, and the extreme top 5% and bottom 5% of the data are removed. The maximum and minimum values ​​of the remaining 90% of the data are used as the reference range for the example. The child's behavioral characteristics are compared with the normal behavioral range of variation for their corresponding developmental stage to avoid judging abnormalities solely based on the magnitude of behavioral values. The process involves acquiring the location of each physiological change and its corresponding time, for example, identifying the 37th, 82nd, and 125th seconds as physiological change locations. Then, it identifies the locations of behavioral abnormalities, such as the 35th, 37th, 38th, 80th, 81st, and 100th seconds. Next, the two types of locations are sorted by time according to a unified time benchmark, and a preset time interval is set. For example, setting the time interval to 2 seconds means that if the time difference between a physiological change location and a behavioral abnormality location does not exceed 2 seconds, they are considered to be within the same abnormality observation time range. For example, if the physiological change occurs at 37 seconds, and the abnormal behavior occurs at 35 seconds, 37 seconds, and 38 seconds, then the difference between 35 seconds and 37 seconds is 2 seconds, the difference between 37 seconds and 38 seconds is 0 seconds, and the difference between 38 seconds and 37 seconds is 1 second. All three abnormal behavior locations satisfy the condition of not exceeding 2 seconds, therefore, the period from 35 seconds to 38 seconds is determined as the corresponding time overlap interval. Similarly, if the physiological change occurs at 82 seconds, and the abnormal behavior occurs at 80 seconds and 81 seconds, then the difference between 80 seconds and 82 seconds is 2 seconds, and the difference between 81 seconds and 82 seconds is 1 second, thus forming the 8th... The time overlaps from 0 seconds to 82 seconds; however, the location of the behavioral abnormality at 100 seconds differs from the location of the physiological change at 82 seconds by 18 seconds, which is greater than 2 seconds. Therefore, it is not matched with the location of the physiological change at 82 seconds. The preset time interval can be set according to the data sampling frequency and behavioral response delay. For example, for data sampled once per second, it can be set to 1 to 5 seconds. This embodiment uses 2 seconds. If there is a significant response time difference between behavioral data and physiological data, it can be set to 3 seconds or 5 seconds. The time interval can be obtained through historical scene data. Specifically, it can record the physiological response time when the child has a clear behavioral change. For example, if 100 events with obvious behavioral changes are collected, and 80 physiological changes occur within 2 seconds before and after the behavioral change, 15 occur within 3 seconds, and 5 exceed 3 seconds, then 2 seconds can be selected as a stricter time matching interval, or 3 seconds can be selected as an embodiment parameter to cover more effective events. This can exclude situations where physiological changes and behavioral abnormalities exist but their actual occurrence times are far apart, so that subsequent developmental abnormality trigger points have a clear time synchronization basis, thereby improving the targeting of abnormality trigger locations. Each overlapping time interval is obtained, for example, two overlapping time intervals are obtained: 35 seconds to 38 seconds and 80 seconds to 82 seconds. Then, it is checked whether there are physiological change locations and behavioral abnormality locations within each interval. For example, within the 35-38 second interval, there is a physiological change location at 37 seconds, and behavioral abnormality locations exist at 35, 37, and 38 seconds. Therefore, this interval is determined to have the characteristic of both physiological change and behavioral abnormality. Corresponding to the time positions, 37 seconds can be taken as the temporal position corresponding to both physiological change and behavioral abnormality locations, and 37 seconds is determined as the developmental abnormality trigger point. For the 80-82 second interval, if there is a physiological change location at 82 seconds, and behavioral abnormality locations at 80 and 81 seconds, then since both are within a preset 2-second time interval, 82 seconds can be determined as the developmental abnormality trigger point for this interval. If a certain overlapping time interval has a behavioral abnormality location but no actual physiological change location, then the developmental abnormality trigger point is uncertain. For example, near 100 seconds, only... If there is no corresponding physiological change location at a location of behavioral abnormality, that location will not be identified as a developmental abnormality trigger point. For situations where multiple time-series locations meet the criteria within a time-overlapping interval, the trigger point can be determined based on the location that appears first simultaneously. For example, if the physiological change and behavioral abnormality conditions are met simultaneously for the first time at second 51, and the conditions are also met at second 52 and second 53, then second 51 is taken as the developmental abnormality trigger point for that time-overlapping interval, and the interval from second 51 to second 53 is retained as the trigger duration interval. Physiological change locations and behavioral abnormality locations can be further integrated into a developmental abnormality trigger point with a clear temporal location. This allows for direct expansion of the abnormality observation time window forward and backward from this trigger point, and extraction of behavioral, physiological, and assessment features within the corresponding time range. This forms a continuous processing chain of physiological change locations, behavioral abnormality locations, time-overlapping intervals, and developmental abnormality trigger points. Compared to judging a single physiological abnormality or a single behavioral abnormality separately, this further highlights the collaborative abnormal manifestations of multimodal data at the same time location.

[0027] In a preferred embodiment, an abnormal observation time window is determined based on the developmental abnormality trigger point, and the target dimensional features corresponding to the abnormal observation time window are extracted from the dimensional feature data to obtain a trigger feature set, including: S301. Obtain the trigger timing position corresponding to the developmental abnormality trigger point, and extend it by a preset duration in the direction before and after the trigger timing position to obtain the abnormality observation time window; S302. Filter the dimensional feature data according to the time window of abnormal observation, and extract the behavioral features, physiological features and evaluation features located within the time window of abnormal observation. S303. Classify the developmental ability dimensions corresponding to the extracted behavioral features, physiological features and assessment features to obtain the target dimension features corresponding to each developmental ability dimension. Collect the target dimension features corresponding to each developmental ability dimension to obtain the trigger feature set.

[0028] As described in steps S301 to S303 above, the trigger timing position corresponding to the developmental abnormality trigger point is obtained. For example, the 60th second is determined as the developmental abnormality trigger point. Then, the developmental ability dimension corresponding to the trigger point and the multimodal time-series data before and after the trigger point are obtained. The forward and backward extension time lengths are determined according to the preset time window setting rules. For example, the forward extension time is set to 3 seconds and the backward extension time is set to 5 seconds. Then, taking the 60th second as the center, the data from the 57th second to the 60th second is obtained forward, and the data from the 60th second to the 65th second is obtained backward, thus forming the abnormality observation time window from the 57th second to the 65th second. Among them, the forward extension time is used to retain the change process of the child's behavior and physiological state before the abnormality trigger, and the backward extension time is used to retain the continuous performance after the abnormality trigger. Therefore, the forward and backward times can be set to different values. For example, the forward time can be set in the range of 2 to 10 seconds, and the backward time can be set in the range of 3 to 15 seconds. This embodiment uses 3 seconds and 5 seconds. The preset time length can be determined according to the historical time overlap interval. Specifically, multiple sets of abnormal events can be collected under the same preset scenario. The time from the first appearance of physiological changes or behavioral abnormalities to the trigger point and the duration of abnormal behavior after the trigger point can be recorded for each set of events. For example, after collecting 100 sets of events, it was found that 80 sets of abnormal behaviors occurred within 3 seconds before the trigger point, and 85 sets of abnormal behaviors lasted until 5 seconds after the trigger point. Therefore, the forward 3 seconds and backward 5 seconds can be used as the time window parameters in the embodiment. If a trigger point is located at the edge of the entire data collection interval, for example, the trigger point is the 2nd second, then the data is extracted from the 0th second according to the actual available data range, without generating non-existent data. This can avoid losing the change process before and after the abnormality by only extracting the developmental abnormality trigger point itself, so that the state before the trigger, the state at the trigger, and the continuous state after the trigger can be obtained simultaneously. This provides a complete time range for judging whether the abnormality is persistent and whether different modal features change synchronously. At the same time, by limiting the observation time window, the interference of irrelevant time period data on the subsequent abnormal feature extraction can also be reduced. Obtain the start and end times of the anomaly observation time window, for example, the time window from 57 seconds to 65 seconds. Then, read the time markers corresponding to each behavioral feature, physiological feature, and evaluation feature in the dimensional feature data, and determine whether the time markers of each feature fall within the range of 57 seconds to 65 seconds. For example, if behavioral data has feature records at 56 seconds, 58 seconds, 60 seconds, and 64 seconds, then retain the behavioral features at 58 seconds, 60 seconds, and 64 seconds, and delete the time position corresponding to 56 seconds; if physiological data has features at 57 seconds... If sampling values ​​exist at 58 seconds, 59 seconds, and 60 to 65 seconds, then all physiological data within these time windows will be extracted. For assessment data, if an assessment item starts at 61 seconds and continues until 63 seconds, then that assessment event will be recorded as an assessment feature within the range of 61 to 63 seconds. For features with different sampling frequencies, instead of simply concatenating the data, time correspondence will be performed according to an established unified time benchmark. For example, behavioral data will be recorded every 0.04 seconds, and physiological data will be recorded every 1 second. Evaluation data is recorded according to the events generated, and all data is mapped to the anomaly observation time window through time markers. For example, the behavioral feature corresponding to the 60th second is a gaze duration of 3 seconds, the physiological feature is a heart rate of 92 beats / minute, and the evaluation feature is 1 task error. All three features are retained and recorded at the time position corresponding to the 60th second. For missing data, the corresponding missing marker can be retained. For example, if there is no evaluation data at the 62nd second, it is recorded that there is no evaluation event at the 62nd second, instead of using the evaluation results at the 61st or 63rd second. For continuous behavioral features, they can be divided according to a preset time granularity. For example, if 100 milliseconds is set as a time slice, then the 57th to 65th second contains a total of 8 seconds. Calculated according to each time slice of 100 milliseconds, 8 seconds contains 8 multiplied by 10 equals 80 time slices. Then, the behavioral, physiological, and evaluation features located in these 80 time slices are extracted separately. Data sets from different sources within the local time period corresponding to the developmental abnormality trigger point can be extracted so that the trigger features are no longer interfered with by a large amount of normal data in the entire collection cycle, and the multimodal change information before and after the abnormality occurs is retained. The developmental ability dimension corresponding to each extracted behavioral, physiological, and assessment feature is obtained. This dimension is derived from the established developmental ability dimension classification rules, so there is no need to re-establish the dimension categories. For example, eye fixation duration, number of interaction responses, and corresponding physiological responses are classified into the social interaction ability dimension; continuous attention time and number of attention shifts are classified into the attention control ability dimension; language response count and language response time are classified into the language expression ability dimension; action completion time and number of actions are classified into the action coordination ability dimension; and the number of emotional changes and emotional recovery time are classified into the emotion regulation ability dimension. For example, if 8 behavioral features, 9 physiological features, and 2 assessment features are extracted within the abnormal observation time window from the 57th to the 65th second, and the behavioral features corresponding to the social interaction ability dimension include 3 eye fixation durations and 2 interaction responses, the corresponding physiological features include 3 heart rate changes, and the corresponding assessment feature includes 1 social interaction score, then these features are all classified into the social interaction ability dimension. If 4 continuous attention time and 2 physiological responses are also extracted from the attention control ability dimension, then they are classified into the attention control ability dimension. After completing the classification, in each Within each developmental ability dimension, features from different sources are mapped according to a unified time position. For example, the social interaction ability dimension at the 60th second includes a behavioral gaze duration of 3 seconds, a physiological heart rate of 92 beats / minute, and an assessment score of 7 points. These three are then used as the target dimension features corresponding to the 60th second. The target dimension features corresponding to each developmental ability dimension are then aggregated to form a trigger feature set. For example, in this embodiment, feature sets are formed for three dimensions: social interaction ability dimension trigger features, attention control ability dimension trigger features, and language expression ability dimension trigger features. Information such as feature type, collection time, original value, developmental stage, and data source for each feature is retained. Local multimodal features can be further converted into structured trigger features organized according to developmental ability dimensions. This allows for direct comparison of the trigger features with the complete dimension feature data and reference features corresponding to the same developmental ability dimension. Since the trigger feature set retains the time position, it is possible not only to determine whether a feature is abnormal, but also to further determine the specific time position of the abnormality and whether different modal features appear within the same time range, thereby improving the precision of abnormal feature localization.

[0029] In a preferred implementation, based on a trigger feature set, dimensional feature data are compared within the same capability dimension to generate deviation identification data for the corresponding capability dimension, including: S401. Obtain the developmental ability dimension according to the trigger feature set, and obtain the dimension feature data corresponding to the developmental ability dimension from the dimension feature data. S402. Based on preset reference data, obtain reference feature data corresponding to the developmental ability dimension and developmental stage. S403. Compare the feature data in the trigger feature set with the reference feature data within the corresponding time range to obtain feature difference information, wherein the feature difference information includes behavioral feature differences, physiological feature differences and evaluation feature differences. S404. Based on the feature difference information, determine whether each feature in the feature difference information exceeds the corresponding preset difference range, and obtain the deviation determination result; S405. Based on the deviation determination results, generate the corresponding deviation label data for the developmental ability dimension.

[0030] As described in steps S401 to S405 above, the trigger feature set is read. Each feature data in the trigger feature set retains its corresponding developmental ability dimension identifier, time position, feature type, feature value, and data source. The developmental ability dimension identifier is derived from the developmental ability dimension classification rules. For example, features related to sustained fixation, task retention, and attention target switching are classified into the attention control ability dimension; features related to word comprehension, word expression, and sentence organization are classified into the language expression ability dimension; features related to eye contact, interactive response, and social movement are classified into the social interaction ability dimension; features related to limb movement continuity, hand-eye coordination, and posture transition are classified into the motor coordination ability dimension; and features related to emotional fluctuations, emotional recovery time, and emotional stimulus response are classified into the emotion regulation ability dimension. The above classification rules can be manually configured during the system establishment phase through child health care scale items, child development screening items, and clinically collected indicators. For example, a developmental ability dimension mapping table can be established, where sustained fixation duration, task retention duration, and attention target switching frequency correspond to attention... The mapping table is structured as follows: The number of active words, sentence length, and language response time correspond to the language expression ability dimension; the number of active gazes, interaction responses, and duration of joint attention correspond to the social interaction ability dimension; gait stabilization time, action completion time, and hand-eye coordination frequency correspond to the motor coordination ability dimension; and emotional recovery time, crying duration, and post-stimulation emotional fluctuation amplitude correspond to the emotion regulation ability dimension. This mapping table can be created by pediatric healthcare professionals based on actual assessment items. It is then verified by comparing historical assessment records of, for example, 100 children of different ages. The dimensions to which the same indicator belongs in different records are manually reviewed. When an indicator belongs to the same dimension in at least 95% of the reviewed records, that classification is used as the default mapping relationship. For indicators with cross-dimensional attributes, the original data is retained, and corresponding dimension indexes are created for each, rather than directly deleting the indicator. Subsequently, based on the developmental ability dimension identifiers in the trigger feature set, a search is performed in the formed dimension feature data. For example, the trigger feature set may include the attention control ability dimension, 60 to 65 seconds, and sustained gaze duration of 2.Features such as 5 seconds, task retention time of 3 seconds, and heart rate of 92 beats / minute determine the developmental ability dimension to be attention control ability. The corresponding attention control ability dimension data for the child is then extracted from all dimension feature data. For example, data such as sustained gaze duration from 0 to 120 seconds, number of attention target transitions, task retention time, and heart rate changes collected synchronously with attentional behavior are extracted, rather than just extracting partial data from 60 to 65 seconds. For instance, if a child forms a time position every second from 0 to 120 seconds, this dimension can yield 120 consecutive temporal positions, each containing at least... It includes one or more of behavioral characteristics, physiological characteristics, and assessment characteristics, thus forming complete dimensional feature data for the child in the attention control ability dimension; it can remap local trigger features to the complete ability dimension data space, enabling subsequent steps to determine whether the abnormality during the trigger period has a significant deviation relative to other time positions of the same ability dimension of the child, while avoiding direct mixing and comparison of data from different developmental ability dimensions, so that subsequent reference feature acquisition and feature difference comparison are limited to the same developmental ability dimension, thereby reducing misjudgments caused by different feature scales and collection methods of different ability dimensions; Pre-set reference data can come from historical assessment data generated by child health institutions during routine developmental assessments, or from desensitized child developmental screening databases, age-stratified data corresponding to standardized developmental scales, and normal developmental sample data generated within the institution. When establishing pre-set reference data, historical data is first grouped according to the child's actual age or developmental stage. For example, 2-3 years, 3-4 years, and 4-5 years can be considered different developmental stages. Then, within each developmental stage, secondary classification is performed based on social interaction ability, attention control ability, language expression ability, motor coordination ability, and emotional regulation ability. For example, when collecting attention control ability data from 3-4 year old children, it can be collected from 50 children... Data on sustained gaze duration, task retention time, number of attentional target transitions, and corresponding physiological responses were obtained from historical assessment records of children aged 3 to 4. Each child was recorded at one time point per second, with each assessment lasting 600 seconds. Therefore, 50 children yielded 30,000 time point data points (50 x 600 = 30,000). The data was then subjected to quality screening, removing data with obvious missing values, device disconnections, incomplete tasks, and mismatched timestamps. For example, if 1,200 time points out of the original 30,000 time points had consecutive physiological data gaps exceeding 5 seconds, these time segments were marked as invalid. The remaining data was then filtered for outliers; for example, for heart rate data, values ​​below 40 beats per minute were removed. Alternatively, records exceeding 180 times / minute are marked as candidate data for device anomalies and confirmed in conjunction with the original device status records. For behavioral characteristics, reasonable ranges are set according to specific indicators; for example, sustained gaze time is set to 0 to 30 seconds, and data exceeding 30 seconds is further examined to determine whether it belongs to a task-continuous state rather than a true gaze state. After data cleaning, the normal distribution range of each characteristic is statistically analyzed within the same developmental stage and the same ability dimension. For example, for sustained gaze time in the attention control ability dimension of 3- to 4-year-old children, the normal range for most children is 4 to 12 seconds in the valid reference data, so 4 to 12 seconds is taken as the target reference range for this characteristic; for task retention time, it is 6 to 15 seconds. For the number of attentional target switching times, a reference range of 2 to 8 times per minute is obtained, and for the corresponding physiological response, a reference range of 70 to 110 times per minute is obtained. It should be noted that different features do not use a uniform numerical range, but rather establish reference ranges separately based on their historical data and specific physical meaning. The determination of the reference range can be achieved by sorting and removing extreme data at both ends. For example, after sorting 1000 valid samples in ascending order, the lowest and highest 50 samples are removed, and the lowest and highest values ​​are determined from the remaining 900 samples as the initial reference range. If the 51st data point after sorting the continuous fixation duration is 4 seconds and the 950th data point is 12 seconds, then a reference range of 4 to 12 seconds is formed.In addition, the number of samples, developmental stage, ability dimension, type of collection device, collection scenario, and data update time corresponding to the reference data can be recorded. This allows the same indicator to be used to establish reference data in different scenarios. For example, different reference data can be established for static image gaze scenarios and interactive game scenarios, instead of directly mixing them. The reference feature data obtained in the above way includes not only feature values, but also the corresponding developmental stage, developmental ability dimension, feature type, time granularity, and reference range source. It can compare features of the same child, the same ability dimension, the same developmental stage, and the same or corresponding time range. This is to incorporate age stage differences and ability dimension differences into the reference process, avoiding the misjudgment of young children as abnormal or incorrect comparisons between different ability dimensions due to the use of a uniform threshold. At the same time, a traceable reference benchmark is formed by historical children's data, which improves the stability of bias judgment. The time position in the trigger feature set is read. For example, if the abnormal observation time window is determined to be 57 to 65 seconds, then behavioral, physiological, and assessment characteristics during the period of 57 to 65 seconds are obtained from the trigger feature set. At the same time, reference features corresponding to the 3- to 4-year-old and attention control ability dimensions are extracted from the obtained reference feature data. Since the sampling frequency of the target child may be different from that of the reference data, time alignment is first performed according to the established unified time benchmark. For example, the target child's behavioral data is sampled once every 100 milliseconds, while the reference data is statistically analyzed once every 1 second. Then, 10 consecutive 100-millisecond data points of the target child are merged into the same 1-second time position according to their time positions, and different summarization methods are used according to specific characteristics. For example, the sustained gaze time is accumulated using the effective gaze duration within the 1 second, resulting in a gaze time of 0.7 seconds within the 1 second; the heart rate is the median value of the effective heart rate data within the 1 second, for example, 91, 92, 93, 92, and 92 times were collected in the 1 second. If the value is 92 times / minute, then 92 times / minute is taken as the physiological characteristic value for that second. The assessment scale data is then mapped to the actual recording time of each assessment item. For example, if a task is completed at 60 seconds and one error is recorded, it is mapped to the 60-second time position. Subsequently, difference information is generated item by item according to the characteristic type. For example, the reference range for sustained gaze time in the attention control ability dimension of 3- to 4-year-old children is 4 to 12 seconds, while the target child's sustained gaze time is between 60 and 65 seconds for six consecutive periods. The time intervals are 2 seconds, 2.5 seconds, 3 seconds, 2.5 seconds, 3 seconds, and 3.5 seconds. For the 60-second position, the target value of 2 seconds differs from the lower limit of the reference range of 4 seconds by 2 seconds; the 61-second position is 2.5 seconds, differing from 4 seconds by 1.5 seconds; the 62-second position is 3 seconds, differing from 4 seconds by 1 second; the 63-second position is 2.5 seconds, differing from 4 seconds by 1.5 seconds; the 64-second position is 3 seconds, differing from 4 seconds by 1 second; and the 65-second position is 3.5 seconds, differing from 4 seconds by 0 seconds.Five seconds are used to establish behavioral differences. For physiological characteristics, for example, the reference heart rate range for the corresponding time period is 70 to 110 beats per minute, while the target child's heart rate at the corresponding time points is 92, 93, 95, 98, 99, and 101 beats per minute. These data are all within the reference range, but the difference from the center of the reference range can be further preserved. For example, if the midpoint of the reference range is 90 beats per minute, this is obtained by summing 70 and 110 to get 180, then dividing 180 by 2 to get 90. The target child's heart rate at 62 seconds is 95 beats per minute, therefore it is 5 beats per minute higher than the reference center value. For assessment characteristics, for example, the reference error count for the corresponding assessment item is 0 to 1, while the target child's error counts between 60 and 65 seconds are 0, 1, 2, 2, 3, and 2 respectively. Therefore, after 62 seconds, the difference gradually increases beyond the reference center value. The evaluation characteristics of the reference range are differentiated; ultimately, various differences are saved in the form of time location, ability dimension, feature type, target value, reference value or reference range, and difference value. For example, behavioral difference information is formed for the attention control ability dimension, 62 seconds, continuous fixation for 3 seconds, reference lower limit for 4 seconds, and below the reference lower limit for 1 second; physiological difference information is formed for the attention control ability dimension, 62 seconds, heart rate for 95 beats / minute, reference center for 90 beats / minute, and above 5 beats / minute; and evaluation difference information is formed for the attention control ability dimension, 62 seconds, number of errors for 2, reference upper limit for 1, and more than 1 error. The original multimodal data is transformed into feature difference information with clear time location and ability dimension attribution, so that the three data sources of behavior, physiology, and evaluation can be examined together under the same time benchmark, while retaining the physical meaning of each type of data and avoiding the simple superposition of data of different dimensions. The preset difference range is derived from historical samples of the reference data, and is set in conjunction with the measurement accuracy of specific features and the normal fluctuation range of children. For example, for the behavioral feature of sustained gaze duration, statistical analysis of historical normal samples shows that its short-term fluctuations usually do not exceed 1 second, so the allowable difference from the reference range boundary is set to 1 second. For the physiological feature of heart rate, since children's normal fluctuations during activities are relatively large, the allowable difference can be set to 10 times / minute. For the number of assessment errors, since it is a discrete counting feature, the allowable difference can be set to 1 time. These difference ranges can be specifically obtained from historical data. For example, collecting data from 50 children aged 3 to 4 years, with each child collecting 600 seconds of attention task data, yields 30,000 time positions. Statistical analysis of the sustained gaze duration difference between adjacent time positions shows that if 27,000 differences do not exceed 0.8 seconds and 28,500 differences do not exceed 1 second, then 1 second can be used as the allowable short-term variation range. For heart rate, statistical analysis is conducted within the same scenario. For changes in position between adjacent time periods, if 95% of normal samples show changes of no more than 10 times / minute, then 10 times / minute is set as the range for physiological characteristic differences. For the number of assessment errors, statistics are compiled based on the task records of 50 children. If, under normal conditions, the number of errors between adjacent task stages typically changes by no more than one time, then one error is set as the allowable difference. In practical applications, different difference ranges can be set for different developmental stages. For example, the allowable difference for sustained gaze in children aged 2-3 years is set to 1.5 seconds, in children aged 3-4 years to 1 second, and in children aged 4-5 years to 0.8 seconds, thus avoiding using the same difference range to cover all age stages. After configuring the difference ranges, each characteristic difference is judged item by item. For example, if the target child's sustained gaze time in 62 seconds is 3 seconds, and the reference lower limit is 4 seconds, the difference is 1 second. Since the allowable difference range is 1 second, this characteristic can be judged as a critical deviation or within the deviation range. If the sustained gaze time in 63 seconds is 2.5 seconds, the difference is 1.5 seconds from the reference lower limit of 4 seconds, 1.If a difference of 5 seconds exceeds the allowable difference of 1 second, it is considered to be outside the deviation range. For example, a heart rate of 95 beats / minute over 62 seconds differs from the reference center of 90 beats / minute by 5 beats / minute, while the allowable difference for physiological characteristics is 10 beats / minute; therefore, this physiological characteristic is not considered an abnormal deviation. Another example is an assessment error count of 2 times over 62 seconds, while the reference upper limit is 1 time, a difference of 1 time. If the allowable difference for discrete assessment characteristics is set to 1 time, it can be marked as a critical state. When the number of errors reaches 3 times, subtracting 1 from 3 yields 2 times, which is greater than the allowable difference of 1 time, thus indicating that the deviation range is exceeded. This ultimately forms the deviation judgment result for each feature, such as three levels: normal deviation, critical deviation, and exceeding the deviation range. The judgment of exceeding the deviation range also records the direction of the deviation, such as sustained gaze below the lower reference limit or the number of errors above the upper reference limit, thus preserving abnormal direction information. This method can filter short-term physiological and behavioral fluctuations in children under normal conditions, and sets different difference ranges for different types of features such as behavior, physiology, and assessment, improving the specificity and stability of deviation identification. The deviation judgment results corresponding to each feature are read and categorized according to the developmental ability dimension. For example, in the attention control ability dimension, there are 6 time points in the sustained gaze time between 60 and 65 seconds. Among them, 60 seconds is 2 seconds below the reference lower limit, 61 seconds is 1.5 seconds below the reference lower limit, 62 seconds is 1 second below the reference lower limit, 63 seconds is 1.5 seconds below the reference lower limit, 64 seconds is 1 second below the reference lower limit, and 65 seconds is 0.5 seconds below the reference lower limit. At the same time, the number of assessment errors between 62 and 65 seconds are 2, 2, 3, and 2, respectively. The heart rate of 92 to 101 beats per minute is within the physiological range. Therefore, the above results are merged according to time position and feature type. To avoid being influenced by a single feature... A deviation in the ability dimension is generated when a feature momentarily exceeds its range. Further dimension-level labeling rules can be set. For example, within an abnormal observation window, if at least two different time points show out-of-range behavioral features for the same ability dimension, or if at least one behavioral feature and one assessment feature simultaneously exceed their corresponding difference range, a deviation label for that ability dimension is generated. This rule can be verified using historical data. For instance, normal assessment data and data with confirmed developmental abnormalities can be selected from the historical records of 100 children, and the stability of deviation under different combinations of conditions can be statistically analyzed. If it is found that the occurrence of out-of-range features at two or more consecutive time points can reliably distinguish between short-term fluctuations and persistent abnormalities, then the two consecutive time points are considered as... Conditions for dimensional deviation formation: For example, if the sustained gaze time of 60 to 65 seconds is below the reference range for multiple consecutive time points, and the number of evaluation errors exceeds the reference range from 62 to 65 seconds, then the conditions of consecutive behavioral feature deviations and synchronous evaluation feature deviations are met, and deviation labeling data corresponding to the attention control ability dimension can be generated. The deviation labeling data can specifically include the ability dimension label, deviation start time, deviation end time, abnormal feature type, deviation direction, number of deviation duration points, and deviation degree. For example, it can form the attention control ability dimension, 60 to 65 seconds, behavioral feature deviation, no physiological feature deviation, evaluation feature deviation, behavioral feature below the reference range, and deviation at 6 consecutive time points. The labeling results are as follows: For the dimensions of social interaction ability, language expression ability, motor coordination ability, and emotion regulation ability, corresponding labels are generated according to the same data organization method. For example, if the language expression ability dimension shows a continuous language response time exceeding the reference upper limit between 80 and 90 seconds, a deviation label for the language expression ability dimension is generated; if the motor coordination ability dimension shows a continuous action completion time exceeding the reference range between 100 and 108 seconds, a deviation label for the motor coordination ability dimension is generated. Multiple ability dimensions can have deviation labels simultaneously, but each deviation label retains its own ability dimension, time range, and feature type. Multiple dimensions cannot be simply merged into a dimensionless outlier.Finally, deviation markers for each ability dimension are compiled chronologically to form deviation marker data, which is then indexed with trigger feature sets and dimensional feature data. For example, it can be linked by the child's assessment record number, developmental stage, ability dimension marker, and time location. This allows for further screening of unmet characteristics and the formation of abnormal feature sequences. Individual deviations scattered across behavioral, physiological, and assessment data can be transformed into structured markers with ability dimensions, time ranges, deviation directions, and persistence locations. This preserves the specific source of the original abnormal characteristics while avoiding repeated large-scale data screening in subsequent steps. The deviation markers not only indicate whether a characteristic has deviated, but also clearly identify which developmental ability dimension the deviation occurred in, when it occurred, how long it lasted, and which types of characteristics contributed to it.

[0031] In a preferred embodiment, the target feature range corresponding to each developmental ability dimension and developmental stage is obtained based on preset reference data. The trigger feature set and deviation identification data are compared with the target feature range set to obtain the unachieved features for each developmental ability dimension. An abnormal feature sequence is then formed in chronological order, including: S501. Based on preset reference data, obtain the target feature range corresponding to each developmental ability dimension, and construct a target feature range set; S502. Based on the trigger feature set, extract the behavioral features, physiological features and assessment features under each developmental ability dimension, and combine them with the deviation identification data to filter out the corresponding abnormal feature items, so as to obtain the feature data to be compared for each developmental ability dimension. S503. Match the feature data to be compared for each developmental ability dimension with the target feature range set according to time position and feature type to obtain the feature alignment result; S504. Based on the feature alignment results, obtain the difference information between each feature to be compared and the corresponding target feature range; S505. Based on the difference information and the preset judgment threshold, determine whether each feature to be compared exceeds the target feature range, and correct the judgment result by combining the deviation identification data to obtain the feature deviation result. S506. Based on the feature deviation results, select the feature items that do not enter the corresponding target feature range as unachieved features, and form an abnormal feature sequence in chronological order.

[0032] As described in steps S501 to S506 above, the preset reference data is read. This preset reference data originates from historical assessment records of child health institutions, standardized child development screening project data, and desensitized assessment data of children of the same age. The process can be as follows: First, collect historical assessment records of children at different developmental stages. For example, collect data from 50 children in each of the three age groups: 2-3 years old, 3-4 years old, and 4-5 years old. Each child completes a 600-second observation task in a unified assessment scenario, and behavioral, physiological, and assessment data are collected at one time point every second. This yields 50 x 600 = 30,000 original time points for each age group. Then, delete the data stored during the assessment process. Data from device disconnections, missing timestamps, children leaving the assessment area for more than 5 seconds, and incomplete assessment tasks (e.g., 1200 invalid timestamps out of 30,000 timestamps for children aged 3-4) resulted in the retention of 28,800 valid timestamps. These valid timestamps were then categorized according to developmental ability dimensions: sustained gaze time, task retention time, and number of attentional target transitions were classified under attention control ability; the number of spontaneous words, language response time, and sentence length were classified under language expression ability; the number of spontaneous gazes, number of interactive responses, and duration of joint attention were classified under social interaction ability; and the time to complete actions, posture retention time, and hand-eye coordination were classified under other dimensions. The number of times a child exhibits strong emotional responses is categorized into the motor coordination ability dimension, while emotional recovery time, stimulus response duration, and frequency of emotional fluctuations are categorized into the emotion regulation ability dimension. Subsequently, within each developmental stage and each ability dimension, feature data groups are further established according to feature type. For example, for the sustained gaze time dimension of attention control ability in 3- to 4-year-old children, all effective sustained gaze records are arranged from low to high values, and the lowest and highest 5% of records are removed to reduce the impact of extreme values ​​on the reference range. For instance, if the lowest and highest 50 samples are removed from 1000 effective sustained gaze samples, the lowest value among the remaining 900 samples is 4 seconds and the highest value is 12 seconds, then the range of 4 to 12 seconds is categorized into... The initial target feature range corresponding to the developmental stage and the characteristic is established; for task retention time, if the effective sample range after processing is 6 to 15 seconds, then a target feature range of 6 to 15 seconds is established; for the physiological characteristic of heart rate, if the effective samples in the same assessment scenario are mainly distributed between 70 beats / minute and 110 beats / minute, then a target feature range of 70 beats / minute to 110 beats / minute is formed; for language response time, if the effective reference data of 3- to 4-year-old children is processed in the same way to obtain 1 to 4 seconds, then a target range of 1 to 4 seconds for language response time is established; for discrete features such as the number of assessment errors, a target range of, for example, 0 to 1 error can be formed based on historical assessment records.It is important to note that the target feature range is not simply a uniform fixed value, but rather established separately for different developmental stages, different ability dimensions, and different feature types. For example, the sustained gaze time for the attention control ability dimension of 3- to 4-year-olds is 4 to 12 seconds, while for 4- to 5-year-olds, a range of 5 to 15 seconds may be formed based on historical data. After determining the ranges for each feature, the developmental stage, ability dimension, feature type, lower limit of the target range, upper limit of the target range, data source, number of samples, collection scenario, and update time are saved as range records. For example, a target range record is formed for 3- to 4-year-olds, attention control ability, sustained gaze time, 4 to 12 seconds, and 1000 valid samples. Multiple target range records are then aggregated to form a target feature range set. This allows subsequent feature comparisons to have clear stage and dimensional benchmarks, avoiding misjudgments caused by using the same range across different age groups and different ability dimensions. At the same time, by retaining the source and sample information of the target range, the target feature range can be traced and updated. Each feature record in the trigger feature set is read. Since the developmental ability dimension identifier, time location, feature type, feature value, and data source have already been configured for the feature records, the trigger features are first classified according to the developmental ability dimension. For example, the trigger feature set covers an abnormal observation time window of 57 to 65 seconds. The attention control ability dimension includes continuous gaze time, task retention time, and number of attention target switching times. The language expression ability dimension includes language response time and number of active words. The social interaction ability dimension includes the number of active gazes and the number of interactive responses. Then, behavioral features, physiological features, and evaluation features are extracted from each ability dimension. For example, in the attention control ability dimension, behavioral features can be obtained from a camera or behavior acquisition device. Specifically, the process involves continuously acquiring videos of the child's face and body movements through a camera device, then correlating the video frame timestamps with other data according to a unified time reference, and obtaining behavioral data such as the child's gaze target, gaze duration, and task retention status from the continuous video frames. Physiological features can come from wearable heart rate sensors or non-contact physiological acquisition devices. For example, an effective heart rate value is acquired every 1 second, and information is deleted based on the device quality identifier. Data of quality below the preset level; assessment features are derived from the results records of children completing standardized assessment items, such as task completion time, number of errors, number of prompts, and assessment item scores; after obtaining the above data, deviation indicator data is read. For example, if it is determined that there is a behavioral characteristic deviation in the attention control ability dimension between 60 and 65 seconds, and an assessment characteristic deviation between 62 and 65 seconds, but no deviation indicator is formed for physiological characteristics, then the sustained gaze time between 60 and 65 seconds and the number of assessment errors between 62 and 65 seconds are first selected as key features to be compared, and those without deviation indicators are excluded. Heart rate data retains its time index but is not considered a primary abnormal feature. Data is filtered according to the type of deviation identified in the deviation marker; for example, if the deviation marker indicates a behavioral characteristic below the target range, the corresponding behavioral characteristic is filtered; if the deviation marker indicates an assessment characteristic above the target range, the corresponding assessment characteristic is filtered. If the same ability dimension exhibits both behavioral and assessment deviations, each type of feature is retained separately to avoid forcibly merging different data types into a single value. For example, a 60-second sustained gaze duration is 2 seconds, 61 seconds is 2.5 seconds, 62 seconds is 3 seconds, 63 seconds is 2.5 seconds, 64 seconds is 3 seconds, and 65 seconds is 3 seconds.If a deviation indicator shows behavioral deviation at these time points, the above six sustained gaze features are used as the behavioral features to be compared. If the number of assessment errors at 62 to 65 seconds are 2, 2, 3, and 2 respectively, and the deviation indicator confirms that they exceed the assessment reference range, then the above four data points are used as the assessment features to be compared. If the heart rate at 60 to 65 seconds is 92, 93, 95, 98, 99, and 101 beats / minute respectively and no deviation indicator is formed, then it is not considered an abnormal feature, but can be retained in the feature data to be compared as an auxiliary feature. Finally, the features are classified according to developmental stage, ability dimension, time location, feature type, and other factors. The comparison feature data is generated using eigenvalues ​​and deviation markers. For example, data records are generated for 3-4 year olds regarding attention control ability, 62 seconds, sustained gaze, 3 seconds, and behavioral deviations, as well as data records for 3-4 year olds regarding attention control ability, 63 seconds, number of assessment errors, 2, and assessment deviations. The deviation markers generated in the previous step are used to narrow down the scope of subsequent comparisons, allowing subsequent processing to focus on features that have shown abnormal trends, while still preserving the original source and time location of behavioral, physiological, and assessment features. This improves the targeting of multimodal feature processing and avoids mixing a large amount of normal data without obvious deviations with abnormal data. Read the feature data to be compared and obtain the developmental stage identifier, ability dimension identifier, time position, and feature type for each data point. Then, search the target feature range set according to the same developmental stage, the same ability dimension, and the same feature type. For example, if the data to be compared is 3 to 4 years old, attention control ability, 62 seconds, continuous gaze duration, and 3 seconds, then first determine the developmental stage as 3 to 4 years old, then determine the ability dimension as attention control ability, and finally determine the feature type as continuous gaze duration. Find the target range from 3 to 4 years old, attention control ability, continuous gaze duration, and 4 to 12 seconds from the target feature range set. If the target child's behavior data is sampled at 100 milliseconds, while the target range is statistically analyzed at 1 second, then it is necessary to merge them according to the time position first, for example, 62.0 seconds to 62.Ten 100-millisecond sampling points within a 9-second interval all correspond to a 62-second time position. If at least seven sampling points are confirmed to be valid, that 1-second time position can be marked as valid. For example, if eight out of ten sampling points are valid, the valid sampling ratio at the 62-second position is 80%, higher than the preset valid ratio of 70%, therefore the 62-second position is retained. If only five sampling points are valid, the ratio is lower than 70%, and this time position is marked as low-quality data and not included in the range determination. For data with different sampling frequencies, they are mapped to a unified time position based on the timestamp; for example, physiological data is sampled at 1-second intervals. Behavioral data is sampled at 100 milliseconds. Evaluation data is recorded discretely at 60, 62, and 65 seconds, and these three are then uniformly mapped to their corresponding second-level time positions. Subsequently, feature-aligned records are established, such as 62 seconds for sustained gaze duration, target value of 3 seconds, target range of 4 to 12 seconds, age 3 to 4 years, and attention control ability; and 63 seconds for the number of evaluation errors, target value of 2, target range of 0 to 1 errors, age 3 to 4 years, and attention control ability. For cases where a completely corresponding target range cannot be found, such as a feature only existing in age 2 to 3 years and age 4 to 5 years... The reference range is determined by the child's age. If the child is currently 3 to 4 years old, data from adjacent stages is not directly selected as the current target range. Instead, the system first checks for existing reference data for 3- to 4-year-olds in the same scenario. If none exists, a common feature range between two adjacent stages is selected according to preset stage supplementation rules, and this range is marked as a stage transition reference. For example, the sustained gaze range for 2- to 3-year-olds is 3 to 10 seconds, and for 4- to 5-year-olds it is 5 to 15 seconds. When a transition reference is needed, the common coverage area of ​​5 to 10 seconds between the two stages can be used as a conservative reference interval. However, the source of this reference should be recorded for future updates. For cases where feature types are inconsistent, such as sustained gaze time not directly matching language response time, no correspondence is established. The final feature alignment result includes at least the feature to be compared, the corresponding target range, developmental stage, ability dimension, feature type, and time position. Multiple matching is performed using developmental stage, ability dimension, feature type, and unified time position to ensure that features from different sources and with different sampling frequencies can accurately find their corresponding target range, avoiding incorrect dimension and age comparisons from a data structure perspective. Read the feature alignment results. For features within the target range, calculate the numerical difference between their distance from the lower and upper limits of the target range. For features below the lower limit, subtract the actual feature value from the lower limit to obtain the difference. For features above the upper limit, subtract the upper limit from the actual feature value to obtain the difference. For example, if the target range for sustained gaze time is 4 to 12 seconds, and the actual sustained gaze time at a certain time point is 3 seconds, then subtracting 3 seconds from 4 seconds gives 1 second, indicating that the feature is 1 second below the lower limit of the target range. If the actual sustained gaze time... If the fixation time is 2.5 seconds, then subtracting 2.5 seconds from 4 seconds gives 1.5 seconds, indicating it's 1.5 seconds below the lower limit of the target range. If the actual sustained fixation time is 14 seconds, then subtracting 12 seconds from 14 seconds gives 2 seconds, indicating it's 2 seconds above the upper limit of the target range. If the actual sustained fixation time is 8 seconds, it falls within the range of 4 to 12 seconds. In this case, subtracting 4 seconds from 8 seconds gives 4 seconds, and subtracting 8 seconds from 12 seconds gives 4 seconds, thus indicating the feature is 4 seconds away from both the lower and upper limits of the target range. For physiological features, such as a heart rate target range of 70 to 110 beats per minute, and an actual heart rate of 95 beats per minute... The clock reading is calculated by subtracting 70 from 95, resulting in 25 times / minute, and by subtracting 95 from 110, resulting in 15 times / minute. This indicates that the feature is within the target range and closer to the upper limit. Regarding the number of evaluation errors, the target range is 0 to 1 times, and the actual number of errors is 3 times. Subtracting 1 from 3 yields 2 times, creating a difference of 2 times above the upper limit of the target range. In addition to the numerical differences mentioned above, directional information can also be generated simultaneously. For example, below the range is recorded as downward deviation, above the range as upward deviation, and within the range as within the range. The actual value, target lower limit, target upper limit, lower limit difference, and upper limit difference are retained. For example, continuous fixation for 3 seconds... The target range is 4 to 12 seconds, below the lower limit by 1 second, and the direction is downward. For the time dimension, the number of time positions where the difference persists can also be recorded. For example, if 6 consecutive time positions from 60 to 65 seconds are all below 4 seconds, the differences corresponding to the 6 time positions are saved as 2 seconds, 1.5 seconds, 1 second, 1.5 seconds, 1 second, and 0.5 seconds, respectively, and the number of consecutive abnormal positions is recorded as 6. This clarifies the direction and degree of deviation of the feature to be compared from the target range, so that subsequent judgments no longer rely solely on whether it exceeds this binary result, but can retain information of different degrees, such as below the lower limit by 0.5 seconds or below the lower limit by 2 seconds. Preset judgment thresholds can be set according to feature types and determined using historical data from normal children. For example, for the behavioral feature of sustained gaze duration, 0.5 seconds can be set as the boundary buffer value, meaning that if the actual value is lower than the reference lower limit but the difference does not exceed 0.5 seconds, it will not be directly judged as not achieved. For the physiological feature of heart rate, 5 beats / minute can be set as the boundary buffer value. For the number of assessment errors, 1 can be set as the discrete feature buffer value. The above thresholds can be obtained from historical normal samples, for example, by collecting attention assessment data from 100 children aged 3 to 4 years, and then setting the threshold for each normal child. The difference between adjacent time positions of the child is statistically analyzed. If 95% of the sustained gaze fluctuations do not exceed 0.5 seconds, then 0.5 seconds is used as the buffer value for judging sustained gaze time; if 95% of the short-term heart rate changes do not exceed 5 beats / minute, then 5 beats / minute is used as the buffer value for judging heart rate; if the number of errors during normal assessment usually does not exceed 1, then 1 error is used as the buffer value for evaluation features. In specific judgment, first determine whether the difference exceeds the corresponding threshold. For example, if the lower limit of the target range is 4 seconds and the actual sustained gaze time is 3.5 seconds, then 0.5 seconds is obtained by subtracting 3.5 seconds from 4 seconds. If the difference exceeds the threshold, then the judgment is made. If the threshold is set at 0.5 seconds, the feature can be marked as a boundary state instead of being directly judged as a significant deviation. If the actual sustained gaze time is 3 seconds, then subtracting 3 seconds from 4 seconds yields 1 second. 1 second is greater than 0.5 seconds, thus forming a deviation result that exceeds the target range. Subsequently, corrections are made using deviation markers. For example, if it has been determined that 60 to 65 seconds falls within the behavioral deviation zone of attention control ability, then if the actual sustained gaze time within this zone is continuously lower than the reference range, the 3.5 seconds that were originally in a boundary state can be further marked as a deviation to be confirmed. Only when the adjacent time position continues to be lower than the reference range can the deviation be further confirmed. A deviation is only confirmed when it occurs. For example, if 60 seconds is 3 seconds, 61 seconds is 3.5 seconds, and 62 seconds is 3 seconds, with a target lower limit of 4 seconds, then the difference for 60 seconds is 1 second, for 61 seconds it is 0.5 seconds, and for 62 seconds it is 1 second. If the preset threshold is 0.5 seconds, then 60 seconds and 62 seconds are out-of-range deviations, while 61 seconds is a boundary state. Since 60 seconds, 61 seconds, and 62 seconds are continuously within a defined deviation range, the boundary state of 61 seconds can be corrected to a continuous deviation state consistent with the preceding and following positions, thus forming a continuous deviation feature. Conversely, if a certain time position only shows 0 once...A 5-second boundary difference, where the preceding and following time positions are within the normal range, is not considered an anomalous feature. Different processing methods are used for different types of deviations. For example, when behavioral features are below the range, the focus is on determining the duration; when physiological features are above the range, the focus is on whether they continuously exceed the physiological threshold; and for evaluation features, the focus is on whether there are consecutive errors exceeding the target range. Finally, for each feature to be compared, a deviation result is generated, including normal, boundary deviation, and confirmed deviation, and the deviation direction, difference, duration, location, and corresponding deviation identification source are recorded. This avoids misjudgment caused by a single anomalous sampling point, while providing further confirmation within the time region where deviation has been identified, thereby improving the stability and continuity of anomalous feature identification. Read all feature deviation results and filter them according to the confirmed deviation result. Feature items within the target range are directly excluded. Feature items that are only at the boundary and have no consecutive deviations are not considered as unachieved features. Feature items that have been confirmed to exceed the target range and are supported by deviation indicators are retained as unachieved features. For example, in the attention control ability dimension, the continuous gaze time from 60 seconds to 65 seconds is 2 seconds, 2.5 seconds, 3 seconds, 2.5 seconds, 3 seconds, and 3.5 seconds, respectively. The target range is 4 seconds to 12 seconds. Among them, the 60-64 seconds are significantly lower than 4 seconds, and the 65-second range is 3 seconds. If the deviation is 5 seconds and below the lower limit by 0.5 seconds, and if the 65-second interval is also determined to be a continuous deviation based on the preceding and following time positions, then the intervals from 60 seconds to 65 seconds are all considered as unachieved features. Meanwhile, the number of evaluation errors from 62 seconds to 65 seconds are 2, 2, 3, and 2 respectively, while the target range is 0 to 1. Subtracting 1 from 2 yields 1 error, subtracting 1 from 2 yields 1 error, subtracting 1 from 3 yields 2 errors, and subtracting 1 from 2 yields 1 error. Therefore, the above time positions all constitute evaluation deviations exceeding the target range and are retained as unachieved features. If the heart rate from 60 seconds to 65 seconds is 92, 93, 95, 98, 99, and 101 beats / min respectively... If the target range is 70 to 110 times per minute, then since these data are all within the target range, no unachieved physiological characteristics are formed. Subsequently, all unachieved characteristics are sorted from earliest to latest according to their time position. For example, first, 60 seconds, attention control, sustained gaze for 2 seconds, below the target range for 2 seconds are formed; then 61 seconds, attention control, sustained gaze for 2.5 seconds, below the target range for 1.5 seconds are formed; then 62 seconds, attention control, sustained gaze for 3 seconds, below the target range for 1 second, 2 evaluation errors, above the target range for 1 time are formed, and so on up to 65 seconds. If multiple features exist at a given location, they are arranged in a preset feature type order, such as first behavioral features, then physiological features, and then assessment features, or further sorted according to the actual collection timestamp; ultimately forming an abnormal feature sequence, which includes at least the time location, developmental stage, developmental ability dimension, feature type, actual feature value, target range, deviation direction, deviation difference, and corresponding deviation identifier. For example, forming 60 seconds, attention control, behavioral features, sustained gaze for 2 seconds, target range for 4 to 12 seconds, downward deviation for 2 seconds; 61 seconds, attention control, behavioral features, sustained gaze for 2.5 seconds, downward deviation for 1 second.Record continuous data such as 5 seconds; 62 seconds, attention control, behavioral characteristics, sustained gaze for 3 seconds, downward deviation for 1 second; 62 seconds, attention control, assessment characteristics, 2 errors, upward deviation for 1 second, etc. If a characteristic is abnormal at 60 and 62 seconds and normal at 61 seconds, the 60 and 62 seconds should not be directly concatenated into a continuous abnormality. Instead, two independent abnormality locations should be retained to ensure accurate identification of the duration of the abnormality later. In addition, for cases where multiple developmental ability dimensions have simultaneously unachieved characteristics, such as an abnormality in the attention control ability dimension between 60 and 65 seconds, or an abnormality in the language expression ability dimension between 80 and 86 seconds, the abnormal feature sequence should be added according to the actual time position, and each should be retained. The ability dimension is identified without direct cross-dimensional splicing. The reference range, feature difference, deviation identifier, and final deviation judgment from the previous steps are uniformly transformed into an abnormal feature sequence arranged chronologically. This allows for direct statistical analysis of the persistence of each unmet feature based on its continuous time position. Furthermore, by combining different developmental ability dimensions, developmental abnormality representation data can be determined. This not only clarifies which features have not reached the target range for the corresponding developmental stage but also retains the time of abnormality occurrence, the location of abnormality persistence, the direction of deviation, and the degree of deviation. This transforms subsequent assessments from single-point abnormality judgments to judgments of abnormal persistence and multi-dimensional combination features, thereby improving the continuity and interpretability of the entire child developmental status assessment process.

[0033] In a preferred embodiment, the degree of abnormality and deviation of unachieved features are determined based on the abnormal feature sequence, and developmental abnormality characterization data are obtained by combining the unachieved features corresponding to different developmental capacity dimensions, including: S601. Obtain the consecutive occurrence time sequence positions corresponding to each unachieved feature from the abnormal feature sequence, and determine the abnormal persistence degree of each unachieved feature based on the consecutive occurrence time sequence positions; S602. Obtain the deviation range between the unachieved feature and the corresponding reference feature, and determine the degree of deviation for each unachieved feature; S603. Classify the degree of abnormality persistence and deviation according to the developmental ability dimension to which the unachieved feature belongs, and obtain the abnormal feature representation corresponding to each developmental ability dimension. S604. Collect the abnormal feature representations corresponding to each developmental ability dimension to obtain developmental abnormality representation data.

[0034] As described in steps S601 to S604 above, each record in the abnormal feature sequence is read. Each record includes at least the time position, developmental stage, developmental ability dimension, feature type, feature value, target feature range, deviation direction, and deviation difference. Then, the abnormal features are classified according to the developmental ability dimension and feature type. For example, abnormal records of sustained gaze time in the attention control ability dimension are grouped into one group, abnormal records of task retention time in the attention control ability dimension are grouped into another group, and abnormal records of language response time in the language expression ability dimension are grouped into yet another group. Subsequently, the unachieved features in the same group are arranged in ascending order of time position, and it is determined whether two adjacent abnormal time positions are consecutive. For example, if the sustained gaze time in the attention control ability dimension of the abnormal feature sequence is below the target range at 60 seconds, 61 seconds, 62 seconds, 63 seconds, 64 seconds, and 65 seconds, then these six time positions appear consecutively, forming a continuous abnormal interval. If the language response time in another group of language expression ability dimensions exceeds the target only at 80 seconds, 82 seconds, and 85 seconds... If the range is defined, and 81, 83, and 84 seconds fall within the target range, then 80, 82, and 85 seconds cannot be directly identified as the same continuous abnormal interval, but should form three independent abnormal positions. After determining the continuous time sequence positions, the abnormal duration is calculated based on the actual sampling time interval. For example, if the system's unified time granularity is 1 second, and 60 to 65 seconds contains 6 abnormal time positions, then the continuous time length covered from the first abnormal position at 60 seconds to the last abnormal position at 65 seconds is 5 seconds. Furthermore, since each time position corresponds to... The sampling interval is 1 second, so it can be recorded as 6 consecutive abnormal sampling positions, covering a sampling window of about 6 seconds. In order to avoid inconsistent abnormal duration caused by different sampling frequencies, it can also be uniformly calculated according to the actual timestamp. For example, if the behavioral data shows abnormalities at 60.2 seconds, 60.8 seconds, 61.4 seconds and 62.1 seconds, then first read the time difference between adjacent abnormal time positions. 60.8 seconds minus 60.2 seconds gives 0.6 seconds, 61.4 seconds minus 60.8 seconds gives 0.6 seconds, and 62.1 seconds minus 61.4 seconds gives 0.6 seconds.If the preset continuous judgment interval is 1 second, then none of the above positions exceed 1 second, and therefore can be identified as the same continuous abnormal interval. This continuous judgment interval can be determined through historical normal children's data. For example, by collecting normal assessment records of 100 children and statistically analyzing the time interval between adjacent effective sampling positions of the same feature, if 95% of the effective continuous sampling intervals do not exceed 1 second, then 1 second can be set as the maximum allowable interval for continuous temporal positions. Simultaneously, to distinguish different degrees of duration, three levels of duration can be set. For example, continuous abnormality duration less than 2 seconds is marked as short-term abnormality, 2 seconds but less than 5 seconds is marked as continuous abnormality, and 5 seconds or more is marked as long-term continuous abnormality. The aforementioned 2 seconds and 5 seconds can be set through historical abnormality records. For example, by collecting 100 assessment records confirmed to have obvious developmental abnormalities and statistically analyzing them... After calculating the duration of abnormal features, it was found that most short-term behavioral fluctuations lasted less than 2 seconds, while more stable abnormal states typically lasted more than 5 seconds. Therefore, 2 seconds and 5 seconds were used as the boundary for stage division. For example, if a child's sustained gaze time in the attention control dimension is continuously below the target range from 60 to 65 seconds, forming a continuous abnormal coverage time of 5 seconds, it can be identified as a long-term persistent abnormality. If a child's language response time is abnormal only from 80 to 81 seconds, the coverage time is approximately 1 second, which can be identified as a short-term abnormality. The abnormality persistence data formed in the above way includes continuous start time, continuous end time, continuous abnormal time length, number of continuous abnormal locations, and persistence level. It can distinguish between occasional single-point deviations and persistent abnormal manifestations from a time dimension, avoiding the formation of a strong abnormality representation simply because a single sampling point exceeds the reference range. The target feature range and actual feature value corresponding to each unmet feature are read, and the corresponding deviation interval is determined according to the direction of deviation. For example, if the target feature range for sustained gaze time is 4 to 12 seconds, and the actual feature value is 3 seconds, then the actual feature is below the lower limit of the target range of 4 seconds. Subtracting 3 seconds from 4 seconds gives 1 second, so the deviation interval for this feature is determined to be 1 second below the lower limit of the target range. If the actual feature value is 2 seconds, then subtracting 2 seconds from 4 seconds gives 2 seconds, so the deviation is greater than when the actual value is 3 seconds. If the actual feature value is 14 seconds, then subtracting 12 seconds from 14 seconds gives 2 seconds, so a deviation interval of 2 seconds above the upper limit of the target range is formed. For features that are in the middle of the target range but close to the boundary, for example, if the actual sustained gaze time is 4.3 seconds and the target range is 4 to 12 seconds, then subtracting 4 seconds from 4.3 seconds gives 0.3 seconds, indicating that it is 0 seconds away from the lower limit of the target range.A 3-second deviation is not considered a failed feature because it did not deviate from the target range. For physiological features, such as a target heart rate range of 70 to 110 beats per minute, an actual heart rate of 125 beats per minute is calculated by subtracting 110 from 125, resulting in 15 beats per minute, thus determining it as 15 beats per minute above the target upper limit. If the actual heart rate is 62 beats per minute, subtracting 62 from 70 results in 8 beats per minute, thus determining it as 8 beats per minute below the target lower limit. For the number of assessment errors, such as a target range of 0 to 1 error, an actual number of errors of 3 is calculated by subtracting 1 from 3, resulting in 2 errors, forming a deviation of 2 errors above the target range. Furthermore, to avoid different features due to unit differences... Since differences make direct comparison impossible, the degree of deviation is graded according to the allowable deviation range of each feature. For example, if the allowable deviation range for sustained fixation time is set to 1 second, the allowable deviation range for heart rate is set to 10 beats / minute, and the allowable deviation range for the number of assessment errors is set to 1 time, then a deviation of 1 second in sustained fixation time can be marked as a Level 1 deviation; a deviation of 2 seconds can be marked as a Level 2 deviation; and a deviation of 3 seconds or more can be marked as a Level 3 deviation. For heart rate, a deviation of 5 to 10 beats / minute below or above the reference boundary can be set as a Level 1 deviation, 10 to 20 beats / minute as a Level 2 deviation, and more than 20 beats / minute as a Level 3 deviation. These grading thresholds can be based on historical normal and abnormal samples. The distribution can be set up, for example, collecting normal data from 100 children of the same age and historical assessment data from 50 children with obvious abnormal characteristics. The differences in various characteristics exceeding the target range are then calculated, and the ranges in which normal data and abnormal data are mainly concentrated are observed to determine the boundaries of different levels. For example, in the normal sustained gaze time data of 100 children, the maximum short-term deviation exceeding the target lower limit is 1 second, while the majority of abnormal samples show sustained deviations of 2 to 3 seconds. Therefore, 1 second can be used as the first-level boundary, 2 seconds as the second-level boundary, and 3 seconds as the third-level boundary. It is important to note that the degree of deviation is not directly added together with behavioral, physiological, and assessment characteristics, but rather determined separately according to... Each unit and reference range is recorded. For example, if a child's sustained gaze time is less than the lower limit of 2 seconds, while their heart rate is higher than the upper limit of 15 beats / minute, and the number of assessment errors is higher than the upper limit of 2, then deviation data are generated for behavioral characteristics deviating by 2 seconds, physiological characteristics deviating by 15 times / minute, and assessment characteristics deviating by 2 times, respectively. The direction of deviation and characteristic type are retained, ultimately forming the deviation interval, deviation value, and deviation level for each unmet characteristic. This allows for the specification of the abnormality degree of unmet characteristics, avoiding only a rough representation as abnormal or normal. Furthermore, since each characteristic uses its own reference range and unit for deviation representation, the problem of meaninglessly summing seconds, times / minute, and number of errors directly does not occur. Read the developmental ability dimension identifiers corresponding to each unachieved feature. For example, map sustained gaze duration, task retention time, and number of attention target switching to the attention control ability dimension; map the number of active words, language response time, and sentence expression length to the language expression ability dimension; and map the number of active gazes, number of interactive responses, and duration of co-attention to the social interaction ability dimension. Then, for the same developmental ability dimension, read the degree of abnormality persistence and deviation. For example, in the attention control ability dimension, if sustained gaze duration is continuously abnormal between 60 and 65 seconds, with an abnormal duration of 5 seconds, and deviation values ​​of 2 seconds, 1.5 seconds, 1 second, 1.5 seconds, 1 second, and 0.5 seconds respectively, then this set of numbers can be... The abnormal features are categorized as follows: duration 5 seconds, 6 consecutive abnormal locations, maximum deviation 2 seconds, and average deviation approximately 1.25 seconds. The average deviation value is obtained by first adding 2, 1.5, 1, 1.5, 1, and 0.5 to get 7.5, and then dividing 7.5 by 6 to get 1.25. If the number of evaluation errors in the same attention control ability dimension is 2, 2, 3, and 2 times respectively between 62 and 65 seconds, with a target upper limit of 1 time, the corresponding deviation values ​​are 1, 1, 2, and 1 times respectively. Adding 1, 1, 2, and 1 gives 5, and then dividing 5 by 4 gives 1.25 times. Therefore, the evaluation feature can be defined as: duration 4 time locations, maximum deviation 2 times, and average deviation 1.The system represents 25 instances of abnormal features. If a capability dimension contains both behavioral and evaluation features, the deviations from these two different units are not directly summed. Instead, the duration of behavioral abnormality, the degree of behavioral deviation, the duration of evaluation abnormality, and the degree of evaluation deviation are stored separately. Furthermore, the temporal overlap between these two features is recorded. For example, if a behavioral abnormality lasts 60 to 65 seconds and an evaluation abnormality lasts 62 to 65 seconds, the overlap is 62 to 65 seconds, covering a 3-second time range. For physiological features, if the heart rate deviates above the reference range between 60 and 62 seconds, it is separately categorized as a physiological abnormality for that capability dimension. Synchronous representations can then be determined based on the common temporal positions of different features within the same capability dimension. For example, if attentional behavioral abnormalities occur between 60 and 65 seconds and evaluation error abnormalities occur between 62 and 65 seconds, their continuous overlap between 62 and 65 seconds can be recorded as a common abnormal time interval for that capability dimension. To ensure a unified structure for the abnormal feature representations, fixed fields can be set, including capability dimension, abnormal feature type, abnormal start time, abnormal end time, and abnormal duration. The data includes the number of consecutive abnormal locations, maximum deviation, average deviation, deviation direction, behavioral abnormality information, physiological abnormality information, and assessment abnormality information. These fields can be configured through data structure during system establishment and filled in item by item according to the data generated in the aforementioned steps, rather than regenerating the original data. For example, the attention control ability dimension can ultimately form an abnormal feature representation, including behavioral abnormality for 5 seconds, maximum low deviation for 2 seconds, assessment abnormality for 4 seconds, maximum high deviation twice, overlap of the two for 3 seconds, and no continuous deviation in physiological characteristics; the language expression ability dimension includes language response time abnormality for 4 seconds, maximum high deviation for 1.5 seconds, and abnormal number of active words for 3 seconds; the motor coordination ability dimension includes action completion time abnormality for 6 seconds, maximum high deviation for 2 seconds, etc. Abnormal features from different sources are aggregated according to developmental ability dimensions, while maintaining the units and abnormal attributes of different data types. This can elevate a single abnormal feature to a multi-feature abnormal representation under a specific developmental ability dimension, and reflect whether different types of features appear together in the same time period through time overlap information. The abnormal characteristics of each developmental ability dimension were identified. For example, in the target child's attention control ability dimension, abnormal behavior lasting 5 seconds, maximum deviation of 2 seconds, abnormal assessment lasting 4 seconds, maximum deviation of 2 times, and common abnormal duration of 3 seconds were identified. In the language expression ability dimension, abnormal language response time lasting 4 seconds, maximum deviation of 1.5 seconds, and abnormal number of spontaneous words lasting 3 seconds were identified. In the social interaction ability dimension, abnormal active gaze lasting 2 seconds, maximum deviation of 1 time, and abnormal interaction response lasting 3 seconds were identified. In the motor coordination ability dimension, abnormal action completion time lasting 6 seconds, maximum deviation of 2 seconds was identified. In the emotion regulation ability dimension, abnormal emotion recovery time lasting 5 seconds, maximum deviation of 2 seconds was identified. The data is represented in 3 seconds; then, the above dimensions are aggregated according to a preset data organization order. This preset order can be arranged according to social interaction, attention control, language expression, motor coordination, and emotion regulation, or according to the order in which the anomalies occur. For example, the earliest 60-second attention anomaly is recorded first, followed by the 80-second language anomaly and the 100-second motor coordination anomaly. It is important to note that the aggregation process does not simply add the anomaly values ​​from different dimensions together. Instead, it preserves the independent set of anomaly features for each dimension, along with their timing and deviation information. For example, the deviation units for the attention control ability dimension are seconds and the number of occurrences, while the deviation units for the language expression ability dimension might be seconds and the number of words. The two are not... Since numerical values ​​can be directly added, dimension identifiers are used for differentiation. To facilitate retrieval, anomaly records are established for each developmental ability dimension. For example, if a dimension has at least one unachieved feature that lasts for more than 2 seconds and deviates beyond the corresponding first-level threshold, the dimension is marked as having a persistent abnormal representation. If two or more different types of features appear consecutively within the same time interval, a multi-feature synchronous abnormal representation identifier is added. For example, if behavioral abnormalities in the attention control ability dimension last from 60 seconds to 65 seconds, and assessment abnormalities last from 62 seconds to 65 seconds, then they share a common time interval of 62 to 65 seconds, which can be considered a multi-feature synchronous abnormality for that dimension. Identification; the above identification rules can be established through historical data. For example, the duration of normal short-term fluctuations and the occurrence of multiple features simultaneously can be statistically analyzed from 100 children's assessment records. If there are few cases in the normal data where two different types of abnormal features overlap continuously for more than 3 seconds, then 3 seconds can be used as a reference condition for identifying synchronous abnormalities. Finally, the abnormal feature representations, abnormal time intervals, duration, degree of deviation, direction of deviation, and multi-feature synchronization information of each ability dimension are collected to form complete developmental abnormality representation data. For example, attention control ability is identified as continuous abnormality for 5 seconds, maximum low deviation for 2 seconds, and overlap of behavioral and assessment abnormalities for 3 seconds; language expression ability is identified as continuous abnormality for 4 seconds and maximum high deviation for 1 second.The system generates structured results, including 5 seconds of continuous abnormality and 6 seconds of maximum high deviation in motor coordination ability, while retaining the index relationship between each result and the original trigger feature set, dimensional feature data, deviation identification data, and abnormal feature sequence. This allows for subsequent tracing back to the specific time location and original data source. The system unifies the scattered abnormality persistence, deviation degree, and multimodal abnormality features under multiple developmental ability dimensions into structured developmental abnormality representation data, without losing independent information from different ability dimensions and feature types. This enables further determination of the developmental abnormality category and degree based on specific ability dimensions, abnormality persistence, deviation degree, and multi-feature synchronization, thus forming a hierarchical processing chain from original multimodal data to local abnormalities, then to continuous abnormalities, and finally to multidimensional abnormality representations.

[0035] In a preferred embodiment, based on developmental abnormality representation data, and combined with the degree of abnormality persistence, deviation, and corresponding developmental ability dimensions, the developmental abnormality category and degree are determined to generate a child's developmental status assessment result, including: S701. Obtain the abnormal feature representations corresponding to each developmental ability dimension from the developmental abnormality representation data, and determine the abnormal feature level based on the degree of abnormality persistence and deviation corresponding to the abnormal feature representations. S702. Classify the abnormality category of the target child according to the abnormality level corresponding to each developmental ability dimension; S703. Determine the degree of abnormality corresponding to the developmental abnormality category based on the level of abnormality characteristics and the degree of abnormality persistence. S704. Collect the developmental abnormality categories, abnormality degrees and corresponding developmental ability dimensions to generate the child's developmental status assessment results.

[0036] As described in steps S701 to S704 above, abnormal feature representations corresponding to each developmental ability dimension in the developmental abnormality representation data are read. For example, abnormal feature representations corresponding to the attention control ability dimension are read, including information such as sustained gaze time below the target range, number of assessment errors above the target range, duration of behavioral abnormality of 5 seconds, duration of assessment abnormality of 4 seconds, and 3-second overlap between behavioral abnormality and assessment abnormality. Subsequently, the degree of abnormality and deviation corresponding to the abnormal feature representation are obtained respectively. For example, sustained gaze time is continuously below the target range from 60 to 65 seconds, the continuous coverage time is 5 seconds, the maximum deviation is 2 seconds, and the number of assessment errors is... The number of consecutive deviations exceeding the target range between 62 and 65 seconds, with a continuous coverage time of 4 seconds and a maximum deviation of 2 times; to convert the duration and deviation of different characteristics into a unified abnormal characteristic level, an abnormal characteristic level judgment rule can be established in advance. This rule can be determined through historical children's assessment data. For example, historical assessment records of 100 children at the same developmental stage can be collected, including 80 children with normal development and 20 children with obvious developmental abnormalities. The duration and deviation of the abnormal characteristics of each child can be statistically analyzed. First, the duration can be statistically analyzed according to three intervals: less than 2 seconds, 2 to 5 seconds, and less than or equal to 5 seconds. Then, the deviation can be statistically analyzed. The degree of deviation is statistically analyzed according to three levels: Level 1, Level 2, and Level 3. For example, regarding sustained gaze duration, in short-term fluctuations in normal children, 95% of abnormal durations are less than 2 seconds, while durations exceeding 5 seconds are mainly concentrated in abnormal characterization samples. Therefore, less than 2 seconds can be classified as short duration, 2 to 5 seconds as medium duration, and 5 seconds and above as high duration. For the degree of deviation, it is graded according to the deviation threshold corresponding to the feature type. For example, the allowable deviation range for sustained gaze duration is 1 second. When the actual deviation does not exceed 1 second, it is determined as Level 1 deviation; when it exceeds 1 second but does not exceed 2 seconds, it is determined as Level 2 deviation; and when it exceeds 2 seconds, it is determined as Level 3 deviation. The deviation is classified as Level 3. For example, if the actual sustained fixation time is 3 seconds and the target lower limit is 4 seconds, subtracting 3 seconds from 4 seconds gives 1 second, which is Level 1 deviation. If the actual sustained fixation time is 2 seconds, subtracting 2 seconds from 4 seconds gives 2 seconds, which is Level 2 deviation. If the actual sustained fixation time is 1 second, subtracting 1 second from 4 seconds gives 3 seconds, which is Level 3 deviation. For the number of assessment errors, if the target range is 0 to 1, then if the actual number is 2, subtracting 1 from 2 gives 1, which is Level 1 deviation. If the actual number is 3, subtracting 1 from 3 gives 2, which is Level 2 deviation. If the actual number is 4, subtracting 1 from 4 gives 3, which is Level 3 deviation.Since duration and deviation have different meanings, this step does not directly add the number of seconds and the number of occurrences. Instead, it first forms the duration level and deviation level separately, and then determines the level of the abnormal feature according to preset combination rules. For example, if the duration level is low and the deviation level is low, it is determined as a level 1 abnormal feature; if the duration level is medium or the deviation level is medium, it is determined as a level 2 abnormal feature; if at least one of the duration level and deviation level reaches a high level and the other is not lower than a medium level, it is determined as a level 3 abnormal feature. For example, if a sustained gaze abnormality lasts for 5 seconds and the maximum deviation is 2 seconds, then the duration level reaches a high level and the deviation level reaches a level 2, so the sustained gaze abnormality can be determined as a level 3 abnormal feature. If the response time of another language only lasts for 2 seconds and the maximum deviation is 0.8 seconds, then it can be determined as a level 3 abnormal feature. The abnormal features are classified as either Level 1 or Level 2, with the level determined according to pre-established combination rules. Furthermore, for cases where multiple abnormal features exist within the same developmental ability dimension, each abnormal feature is graded separately, rather than being directly merged into a single level. For example, the attention control ability dimension can be divided into three levels: Level 3 for sustained gaze abnormality, Level 2 for assessment error abnormality, and Level 1 for task retention time abnormality. Each level is indexed to its original time position, deviation value, and duration. By simultaneously considering both the duration and magnitude of the abnormality, short-term, small-amplitude normal fluctuations are distinguished from long-term, large-amplitude persistent abnormalities, while different features retain their own units and data meanings, thus forming a clearly defined abnormal feature level. The abnormal feature set is read and categorized according to developmental ability dimensions. For example, if a target child exhibits a level 3 abnormality in sustained gaze time, a level 2 abnormality in task retention time, and a level 2 abnormality in the number of assessment errors in the attention control dimension; a level 1 abnormality in language response time in the language expression dimension; and a level 1 abnormality in active gaze time in the social interaction dimension, then abnormal feature sets for attention control, language expression, and social interaction are formed respectively. Subsequently, a classification rule for developmental abnormalities is established. This rule can be configured by child healthcare professionals based on the developmental ability dimension classification results and corrected using historical assessment data. For example, in 100 historical assessment records... The assessment records are used to statistically analyze the combinations of abnormal features in each ability dimension. If a certain ability dimension has at least two different types of abnormal features, and at least one of them reaches level two or above, and these abnormal features overlap for at least 2 seconds, then that ability dimension is identified as the corresponding major developmental abnormality category. For example, the attention control ability dimension has level three sustained gaze abnormality and level two assessment error abnormality, and the two overlap for 3 seconds between 62 and 65 seconds. Since there are two different types of abnormal features, at least one of which reaches level two or above, and the common abnormality time is at least 3 seconds, the attention control abnormality category can be identified. If the language expression ability dimension has only one feature that lasts for 1 second and deviates from 0...A level 1 abnormality lasting 5 seconds can be temporarily recorded as a mild abnormality in language expression, rather than forming a separate primary abnormality category, since it does not meet the criteria for category formation. If the social interaction ability dimension has level 2 abnormalities in active gaze and level 2 abnormalities in interactive response, with durations of 4 seconds and 3 seconds respectively, and a common time position of 2 seconds, then a social interaction abnormality category can be determined. For multiple ability dimensions that simultaneously meet the criteria for category formation, multiple developmental abnormality categories can be output simultaneously. For example, if the target child simultaneously forms an attention control abnormality category and a social interaction abnormality category, the two categories are not directly merged into one disease category; instead, their respective developmental ability dimensions and abnormality feature sets are retained. Furthermore, to avoid forming a category solely based on the number of abnormalities due to a large amount of collected data for a particular ability dimension, four conditions can be set in the classification rules: the number of different feature types, the abnormality level, the duration, and the time overlap. For example, at least two different feature types reaching level 2 or above, continuous abnormal time reaching 3 seconds, and at least two... The common time position of each abnormal feature must meet the above conditions to be identified as a major developmental abnormality category. These conditions can be set using historical data. For example, by reviewing 100 assessment records, if it is found that the duration of a single abnormal feature is less than 2 seconds, it is mostly within the normal range, while two or more consecutive abnormal features exceeding 3 seconds show higher stability. In this case, 2 features and 3 seconds are used as reference conditions for category formation. The final developmental abnormality category data is formed, such as the attention control abnormality category, corresponding attention control ability dimension, sustained gaze level 3 abnormality, assessment error level 2 abnormality, and common abnormality of 3 seconds; and the social interaction abnormality category, corresponding social interaction ability dimension, active gaze level 2 abnormality, interactive response level 2 abnormality, and common abnormality of 2 seconds. Instead of directly determining the abnormality category based on a single indicator, multiple abnormal features are combined and categorized according to developmental ability dimensions, so that the category formation has temporal continuity and multi-feature support, while retaining the situation where multiple ability dimensions are abnormal at the same time, improving the completeness and interpretability of the child's developmental status assessment results. The system acquires the set of abnormal feature levels and the duration of abnormality for each developmental abnormality category. For example, the attention control abnormality category includes Level 3 sustained gaze abnormality and Level 2 assessment error abnormality. Sustained gaze abnormality lasts for 5 seconds, assessment error abnormality lasts for 4 seconds, and the combined duration of both abnormalities is 3 seconds. Subsequently, the system classifies abnormalities according to preset abnormality severity judgment rules. These rules can be established based on historical children's assessment data and assessment standards developed by child healthcare professionals. For example, Level 1, Level 2, and Level 3 abnormality severity can be set. Level 1 abnormality severity corresponds to abnormal features primarily at Level 1 and a duration of less than 3 seconds. Level 2 abnormality severity corresponds to at least one Level 2 abnormal feature and a duration of 3 seconds but less than 5 seconds. Level 3 abnormality severity... There should be at least one level-three abnormal feature or at least two level-two or higher abnormal features, and the duration should be more than 5 seconds. The above range can be determined using historical assessment data. For example, by collecting assessment data from 100 children, and statistically analyzing the duration of normal short-term fluctuations, mild abnormalities, and persistent abnormalities, if most normal short-term fluctuations are less than 2 seconds, general abnormalities are mainly concentrated between 2 and 5 seconds, and persistent obvious abnormalities are mainly more than 5 seconds, then 3 seconds and 5 seconds can be used as reference boundaries for grading the degree of abnormality. In actual judgment, for example, in the category of attentional control abnormalities, sustained gaze abnormality is level three, lasting 5 seconds continuously, and assessment error abnormality is level two, lasting 4 seconds continuously. Because of the presence of level-three abnormal features and the predominantly persistent abnormality... If the duration reaches 5 seconds, the abnormality level of the attention control abnormality category can be determined to be level three. If the language response time in another language expression abnormality category is only level two and lasts for 3 seconds, its abnormality level can be determined to be level two. If the active gaze abnormality and interaction response abnormality are both level one in the social interaction abnormality category, and both abnormalities last for 2 seconds, it can be determined to be level one. For cases where there are multiple abnormal features in the same category but with inconsistent levels, the abnormality level of the category can be determined by the highest abnormal feature level and the duration. For example, if there are three abnormal features of level one, two, and three, first read the highest level as level three, then check the duration of the level three abnormal feature. If the level three abnormality only appears for 1 second and then... If the system returns to normal, but a level 2 anomaly persists for 6 seconds, the final category of anomaly severity will be determined as level 2 according to preset rules, rather than simply determining it as level 3 because of the presence of a single level 3 anomaly. A condition for high-level confirmation is set that a high-level anomaly lasts for at least 3 seconds. For example, if a level 3 anomaly characteristic lasts for less than 3 seconds and there are no other persistent level 2 or higher anomalies, it is only recorded as a short-term high deviation, and the category of anomaly severity is not directly increased. If a level 3 anomaly lasts for more than 3 seconds, it enters the high-level judgment stage; if a level 3 anomaly lasts for 5 seconds, it is directly determined as a level 3 anomaly. The specific thresholds for the above rules can be set based on historical data during the system establishment phase, and the anomaly level, duration boundary, category formation condition, and update time can be saved in the data table.Combining the level of abnormal features with the duration of time to determine the degree of category abnormality can avoid the overall degree of abnormality being overestimated due to a single instantaneous high deviation feature. At the same time, it can obtain a higher degree of abnormality representation for cases with long duration and multiple features that are abnormal together, making the final evaluation results more stable. The process involves retrieving all developmental abnormality categories for the target child, such as attention control abnormalities, social interaction abnormalities, and language expression abnormalities. The severity of each abnormality is then determined, for example, level three for attention control, level two for social interaction, and level one for language expression. Each category is then linked to its corresponding developmental ability dimension: attention control abnormality, attention control ability dimension, level three abnormality; social interaction abnormality, social interaction ability dimension, level two abnormality; and language expression abnormality, language expression ability dimension, level one abnormality. Finally, the abnormal characteristics, duration, deviation, and time interval of each category are used as the basis for the analysis. The supporting information of the results is correlated. For example, the category of attention control abnormality is associated with abnormal sustained gaze for 60 to 65 seconds, abnormal assessment error for 62 to 65 seconds, maximum deviation of 2 seconds, and common abnormal time of 3 seconds; the category of social interaction abnormality is associated with abnormal active gaze for 70 to 73 seconds and abnormal interaction response; the category of language expression abnormality is associated with abnormal language response time for 80 to 82 seconds. In order to have a unified data structure for the final results, a record structure for the child's developmental status assessment results can be established in the system, including the assessment time, child's developmental stage, developmental abnormality category, corresponding developmental ability dimension, degree of abnormality, duration of abnormality, main abnormality characteristics, main deviation direction, and corresponding original data index. The assessment time is derived from the current data. The timestamps for the start and end of the assessment task are used, and the raw data index is linked by the child's assessment record number, feature time location, and feature type. For example, a final record of a child's developmental status assessment result could be: Attention control ability dimension shows an abnormality category, level three, mainly manifested as sustained gaze time continuously below the target range for 5 seconds and assessment error count continuously exceeding the target range for 4 seconds, with maximum deviations of 2 seconds and 2 errors respectively, both types of abnormalities lasting for 3 seconds. Another record could be: Social interaction ability dimension shows an abnormality category, level two, mainly manifested as abnormal active gaze and interaction response lasting for 2 seconds. If a certain ability dimension does not show any abnormalities that meet the category conditions... The assessment results will record that no clear abnormality has been formed, without forcibly generating an abnormality category. For multiple assessment time points, they can also be arranged according to the assessment time to obtain the status changes of the same child in different assessment cycles. For example, if the degree of attention control abnormality is level three in the first assessment and it becomes level two after subsequent training or retesting, the system can retain the corresponding time and original abnormality features of the two assessment results instead of overwriting the historical records. The final assessment result of the child's developmental status can be output to the assessment interface used by child health care personnel, and at the same time retain the abnormality category, abnormality degree, corresponding developmental ability dimension, and abnormal feature information supporting the result, so that staff can view the specific time of abnormality occurrence and the original multimodal data.By providing assessment results that specify the type, severity, and developmental capacity dimensions of developmental abnormalities, rather than directly providing disease diagnoses, this approach is more suitable as supplementary information for children's developmental screening, health assessments, and subsequent manual review.

[0037] Please see the appendix Figure 2 As shown, the present invention also provides an intelligent assessment system for children's developmental disorders based on multimodal data, used in the above-mentioned intelligent assessment method for children's developmental disorders based on multimodal data, comprising: The dimensional feature module is used to acquire multimodal data of the target child in a preset scenario, divide the multimodal data into developmental stages to obtain multimodal time series data, and combine preset rules to acquire dimensional feature data corresponding to each developmental ability dimension. The abnormality triggering module determines the location of physiological changes and abnormal behavior based on dimensional feature data, and determines the developmental abnormality trigger point based on the time overlap interval between the locations of physiological changes and abnormal behavior. The trigger feature module determines the abnormal observation time window based on the developmental abnormality trigger point, and extracts the target dimension features corresponding to the abnormal observation time window from the dimension feature data to obtain the trigger feature set; The deviation identification module, based on the trigger feature set, compares the dimensional feature data under the same capability dimension and generates deviation identification data for the corresponding capability dimension. The module for unachieved features obtains the target feature range corresponding to each developmental ability dimension and developmental stage based on preset reference data, compares the trigger feature set and deviation identification data with the target feature range set, obtains the unachieved features for each developmental ability dimension, and forms an abnormal feature sequence in chronological order. The anomaly characterization module determines the degree of persistence and deviation of unachieved features based on the anomaly feature sequence, and obtains developmental anomaly characterization data by combining the unachieved features corresponding to different developmental ability dimensions. The developmental assessment module determines the type and severity of developmental abnormalities based on data representing developmental abnormalities, combined with the degree of abnormality persistence, deviation, and corresponding developmental ability dimensions, and generates assessment results of the child's developmental status.

[0038] As described above, the dimensional feature module acquires multimodal data of the target child in a preset scenario. The multimodal data includes behavioral data, physiological data, and assessment data. Among them, behavioral data can be acquired through camera equipment, motion acquisition equipment, or interactive devices to obtain information such as changes in the target child's gaze, changes in movement, attention duration, language response, and interactive behavior during the test. Physiological data can be acquired through heart rate sensors, body temperature sensors, or other physiological parameter acquisition devices to obtain the target child's physiological response information at the corresponding time and location. Assessment data is obtained through child development scales, test item records, or interactive evaluation results. The developmental stage of the target child is determined based on the behavioral data and assessment data, and the behavioral data, physiological data, and assessment data are time-calibrated according to a unified time benchmark so that multimodal data with different acquisition frequencies and different data formats can be mapped to the same time position. Then, according to the preset developmental ability dimension classification rules, the multimodal data are respectively classified into the social interaction ability dimension, attention control ability dimension, language expression ability dimension, motor coordination ability dimension, and emotion regulation ability dimension, thereby forming dimensional feature data corresponding to each developmental ability dimension. After obtaining dimensional feature data, the anomaly triggering module does not directly judge anomalies based on a single behavioral indicator. Instead, it determines the locations of physiological changes and behavioral anomalies from the physiological feature sequences and behavioral feature sequences of the corresponding developmental ability dimensions. The location of physiological changes is used to characterize the location where a physiological feature shows a significant change between adjacent time positions. For example, if the heart rate changes from 87 beats / minute to 95 beats / minute during continuous sampling, and the change reaches a preset change threshold, the corresponding time position is determined as the location of physiological change. The location of behavioral anomalies is used to characterize the location where a behavioral feature exceeds the reference change range for the corresponding developmental stage. For example, if the target child's continuous gaze time is lower than the reference range for the same developmental stage, the corresponding time position is determined as the location of behavioral anomalies. The physiological change locations and behavioral anomaly locations are matched according to a unified time benchmark. When the two types of locations overlap within a preset time interval, the corresponding overlapping location is determined as the developmental anomaly trigger point. This avoids misjudging anomalies based solely on a single physiological or behavioral indicator. Instead, it uses the synchronous changes over time from different data sources to determine the key observation locations for subsequent observation. The trigger feature module extends forward and backward by a preset duration based on the developmental abnormality trigger point to form an abnormality observation time window. It then extracts behavioral, physiological, and assessment features within this abnormality observation time window from the dimensional feature data. For example, if the developmental abnormality trigger point is determined at the 60th second, it can extend forward by 3 seconds and backward by 5 seconds to form an abnormality observation time window from the 57th to the 65th second. Subsequently, it extracts features such as continuous gaze duration, action response, physiological parameters, and test evaluation results within this time window and categorizes them according to the developmental ability dimension to which each feature belongs, forming a trigger feature set. By setting the abnormality observation time window, the system not only focuses on the abnormality trigger point itself but also obtains continuous change information before and after the abnormality occurs, thereby preserving the contextual features in the occurrence, development, and continuation of the abnormality. The deviation identification module compares the trigger feature set with the dimensional feature data under the corresponding developmental ability dimension, and obtains reference feature data under the same developmental stage and the same ability dimension by combining preset reference data. For example, for the attention control ability dimension of children aged 3 to 4, reference ranges such as continuous gaze time of 4 to 12 seconds and task maintenance time of 6 to 15 seconds can be obtained from the historical assessment data of children at the same stage after desensitization. For physiological features, the corresponding normal variation range can be obtained from the historical physiological data collected by children at the same stage. The module compares the features according to the feature type and time position, and judges whether the difference between the actual feature and the reference feature exceeds the preset difference range. When the continuous gaze time is 3 seconds and the lower limit of the corresponding reference range is 4 seconds, and the difference between the two reaches the preset difference threshold, the time position and the corresponding feature are marked as deviation positions. When multiple features under the same ability dimension show deviations at similar time positions, deviation identification data of the corresponding ability dimension is formed, and the original multimodal data is converted into structured deviation information with time position, ability dimension and abnormal direction. The unachieved feature module uses preset reference data to establish target feature ranges corresponding to developmental stages and developmental ability dimensions. It matches the trigger feature set and deviation identification data with the target feature range set. Based on each developmental ability dimension, it first determines the corresponding target feature range according to the developmental stage of the target child. Then, it aligns the actual feature with the target feature range according to feature type, time position, and ability dimension. When the actual feature is outside the target feature range, it further confirms whether the feature belongs to an abnormal feature that needs continuous attention based on the deviation identification data. For example, if the target child's continuous gaze time during the period from 60 seconds to 65 seconds is 3 seconds, 2.5 seconds, 3 seconds, 2.5 seconds, 3 seconds, and 3.5 seconds, respectively, while the lower limit of the target feature range is 4 seconds, then none of the above time positions are within the corresponding target feature range. If the number of test errors increases from the normal 0 to 1 times to 2 to 3 times during the period from 62 seconds to 65 seconds, multiple unachieved features under the attention control dimension can be further formed. These unachieved features are arranged in chronological order to form an abnormal feature sequence, transforming scattered abnormal data into abnormal change information with temporal continuity. The anomaly representation module determines the consecutive occurrence intervals of each unachieved feature according to developmental ability dimensions, feature types, and temporal locations. It also determines the duration of the anomaly based on the length of the consecutive occurrences. For example, if a certain attentional feature fails to enter the target feature range from the 60th to the 65th second, its anomaly duration is 5 seconds. If another anomaly only appears once at the 70th second, it can be considered a short-term anomaly. The module then determines the degree of deviation of the unachieved feature from the target feature range. For instance, if the sustained gaze time is 3 seconds and the lower limit of the target range is 4 seconds, it is 1 second below the lower limit; if the sustained gaze time is 2 seconds, it is 2 seconds below the lower limit. For test errors... If the number of errors increases from the normal upper limit of 1 to 3, it exceeds the target range upper limit by 2. Then, the degree of abnormality and deviation are classified according to the developmental ability dimension. Abnormalities of multiple features within the same time period are aggregated. For example, if there are abnormalities in the sustained gaze time and the number of test errors in the attention control ability dimension, and the two have a common time interval between the 62nd and 65th seconds, then developmental abnormality representation data containing abnormal duration, deviation degree, abnormal feature type and common time interval are formed. Thus, the final evaluation basis is no longer a single abnormal value, but a comprehensive abnormal representation that includes time persistence, deviation degree and common changes of multiple features. The developmental assessment module categorizes abnormal features across different developmental ability dimensions based on developmental abnormality representation data. It then determines the type and severity of the developmental abnormality by considering the persistence, deviation, and the developmental ability dimension to which the abnormal feature belongs. For example, when the attention control ability dimension exhibits both persistent gaze deficit and an increase in test errors, and both abnormalities occur concurrently over a consecutive time period, it can be categorized as an attention control abnormality. Similarly, when the language expression ability dimension shows a persistent deviation from the target feature range in language response duration, word completeness, and language test evaluation results at the same developmental stage, it can be categorized as a language expression abnormality. Regarding the severity of the abnormality, it can be further defined... Abnormal features are graded based on their duration and degree of deviation. For example, a duration of less than 2 seconds and a low degree of deviation is classified as a low degree of abnormality, a duration of 2 to 5 seconds or a moderate degree of deviation is classified as a moderate degree of abnormality, and a duration of more than 5 seconds and a large degree of deviation is classified as a high degree of abnormality. Finally, the determined developmental abnormality categories, abnormalities, corresponding developmental ability dimensions, and abnormal features that form the assessment results are collected to generate a child developmental status assessment result. This achieves a step-by-step processing from multimodal raw data, abnormality triggering, local temporal feature extraction, deviation confirmation, abnormality persistence characterization to developmental status assessment.

[0039] And, a smart assessment terminal for children's developmental disorders based on multimodal data, including: One or more processors; A storage device on which one or more programs are stored; When one or more programs are executed by one or more processors, the one or more processors implement an intelligent assessment method for children's developmental disorders based on multimodal data.

[0040] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.

Claims

1. A method for intelligent assessment of developmental disorders in children based on multimodal data, characterized in that, include: Acquire multimodal data of target children in preset scenarios, divide the multimodal data into developmental stages to obtain multimodal time series data, and combine preset rules to obtain dimensional feature data corresponding to each developmental ability dimension; Based on dimensional feature data, the locations of physiological changes and behavioral abnormalities are determined, and the trigger points for developmental abnormalities are determined based on the time overlap between the locations of physiological changes and behavioral abnormalities. Based on the developmental abnormality trigger point, the abnormality observation time window is determined, and the target dimension features corresponding to the abnormality observation time window are extracted from the dimensional feature data to obtain the trigger feature set; Based on the trigger feature set, the dimensional feature data are compared under the same capability dimension to generate the deviation label data of the corresponding capability dimension; Based on preset reference data, the target feature range corresponding to each developmental ability dimension and developmental stage is obtained. The trigger feature set and deviation identification data are compared with the target feature range set to obtain the unachieved features of each developmental ability dimension and form an abnormal feature sequence in chronological order. Based on the abnormal feature sequence, the degree of abnormality and deviation of the unachieved features are determined, and combined with the unachieved features corresponding to different developmental ability dimensions, developmental abnormality characterization data are obtained. Based on developmental abnormality representation data, and combined with the degree of abnormality persistence, degree of deviation and corresponding developmental ability dimensions, the developmental abnormality category and degree of abnormality are determined, and the child's developmental status assessment results are generated.

2. The intelligent assessment method for developmental disorders in children based on multimodal data according to claim 1, characterized in that, Acquire multimodal data of the target child in a preset scenario, divide the multimodal data into developmental stages to obtain multimodal time-series data, and combine it with preset rules to obtain dimensional feature data corresponding to each developmental ability dimension, including: Acquire multimodal data of the target children in a preset scenario, including behavioral data, physiological data, and assessment data; The multimodal data were divided into developmental stages based on behavioral and assessment data, and the physiological data were allocated accordingly according to the developmental stage division results to obtain the stage datasets for the corresponding developmental stages. Based on the stage dataset, a unified time reference is constructed, and the behavioral data, physiological data, and evaluation data are time-calibrated to obtain multimodal time series data, which includes behavioral time series data, physiological time series data, and evaluation time series data. Based on multimodal time-series data, behavioral feature sequences, physiological feature sequences, and evaluation feature sequences are extracted from behavioral time-series data, physiological time-series data, and evaluation time-series data respectively to form multi-source feature sequences; Based on the preset developmental ability dimension division rules, behavioral feature sequences, physiological feature sequences and assessment feature sequences are assigned to the corresponding developmental ability dimensions. Among them, the developmental ability dimensions include social interaction ability dimension, attention control ability dimension, language expression ability dimension, motor coordination ability dimension and emotion regulation ability dimension, thus obtaining the dimension feature set corresponding to each developmental ability dimension. Based on the dimensional feature set, the behavioral feature sequence, physiological feature sequence, and assessment feature sequence under the same developmental ability dimension are integrated and processed to generate dimensional feature data corresponding to each developmental ability dimension. The dimensional feature data includes behavioral representation information, physiological response information, and assessment scale information under the corresponding developmental ability dimension.

3. The intelligent assessment method for developmental disorders in children based on multimodal data according to claim 1, characterized in that, Based on dimensional feature data, the locations of physiological changes and behavioral abnormalities are determined. Furthermore, based on the temporal overlap between these locations, developmental abnormality trigger points are identified, including: Extract physiological feature sequences corresponding to each developmental ability dimension from dimensional feature data, detect the change amplitude of physiological feature sequences according to adjacent temporal positions, and determine the temporal positions where the change amplitude exceeds the preset change threshold as physiological change positions. Extract the behavioral feature sequence corresponding to each developmental ability dimension from the dimensional feature data, and determine the temporal position where the behavioral feature exceeds the corresponding reference range of change as the behavioral abnormal position; The locations of physiological changes and abnormal behaviors are matched in chronological order to determine the time overlap between the two within a preset time interval. The temporal position that simultaneously corresponds to the location of physiological changes and the location of behavioral abnormalities within the time overlap interval is determined as the developmental abnormality trigger point.

4. The intelligent assessment method for developmental disorders in children based on multimodal data according to claim 1, characterized in that, Based on the developmental abnormality trigger point, an abnormality observation time window is determined, and the target dimension features corresponding to the abnormality observation time window are extracted from the dimensional feature data to obtain the trigger feature set, including: Obtain the trigger timing position corresponding to the developmental abnormality trigger point, and extend it by a preset time in the direction before and after the trigger timing position to obtain the abnormality observation time window; The dimensional feature data are filtered by time according to the anomaly observation time window to extract behavioral features, physiological features and evaluation features located within the anomaly observation time window. Based on the extracted behavioral features, physiological features, and assessment features corresponding to the developmental ability dimensions, the target dimension features corresponding to each developmental ability dimension are categorized to obtain the trigger feature set.

5. The intelligent assessment method for developmental disorders in children based on multimodal data according to claim 1, characterized in that, Based on the trigger feature set, dimensional feature data are compared under the same capability dimension to generate corresponding capability dimension deviation label data, including: The developmental ability dimension is obtained based on the trigger feature set, and the corresponding dimension feature data is obtained from the dimension feature data. Based on preset reference data, obtain reference feature data corresponding to the developmental ability dimension and developmental stage. The feature data in the trigger feature set is compared with the reference feature data within the corresponding time range to obtain feature difference information, which includes behavioral feature differences, physiological feature differences and evaluation feature differences. Based on the feature difference information, determine whether each feature in the feature difference information exceeds the corresponding preset difference range, and obtain the deviation determination result; Based on the deviation determination results, corresponding deviation label data for the developmental ability dimension is generated.

6. The intelligent assessment method for developmental disorders in children based on multimodal data according to claim 1, characterized in that, Based on preset reference data, the target feature range corresponding to each developmental ability dimension and developmental stage is obtained. The trigger feature set and deviation identification data are compared with the target feature range set to obtain the unachieved features of each developmental ability dimension. An abnormal feature sequence is formed in chronological order, including: Based on preset reference data, the target feature range corresponding to each developmental ability dimension is obtained, and a target feature range set is constructed; Based on the trigger feature set, behavioral features, physiological features and assessment features under each developmental ability dimension are extracted, and the corresponding abnormal feature items are screened out in combination with the deviation identification data to obtain the feature data to be compared for each developmental ability dimension. The feature data to be compared for each developmental ability dimension are matched with the target feature range set according to time position and feature type to obtain the feature alignment result; Based on the feature alignment results, obtain the difference information between each feature to be compared and the corresponding target feature range; Based on the difference information and the preset judgment threshold, each feature to be compared is judged to see whether it exceeds the target feature range, and the judgment result is corrected by combining the deviation identification data to obtain the feature deviation result. Based on the feature deviation results, feature items that do not fall within the corresponding target feature range are selected as unachieved features and formed into an abnormal feature sequence in chronological order.

7. The intelligent assessment method for developmental disorders in children based on multimodal data according to claim 1, characterized in that, Based on the abnormal feature sequence, the persistence and deviation of the unachieved features are determined, and combined with the unachieved features corresponding to different developmental capacity dimensions, developmental abnormality characterization data are obtained, including: Obtain the consecutive occurrence time sequence positions corresponding to each unachieved feature from the abnormal feature sequence, and determine the abnormal persistence degree of each unachieved feature based on the consecutive occurrence time sequence positions; Obtain the deviation range between the unachieved feature and the corresponding reference feature, and determine the degree of deviation for each unachieved feature; The degree of abnormality persistence and deviation are classified according to the developmental ability dimension to which the unachieved feature belongs, and the abnormal feature representation corresponding to each developmental ability dimension is obtained. The abnormal feature representations corresponding to each developmental ability dimension are aggregated to obtain developmental abnormality representation data.

8. The intelligent assessment method for developmental disorders in children based on multimodal data according to claim 1, characterized in that, Based on developmental abnormality representation data, and combined with the persistence of the abnormality, the degree of deviation, and the corresponding developmental ability dimensions, the category and degree of developmental abnormality are determined, generating a child's developmental status assessment result, including: Abnormal feature representations corresponding to each developmental ability dimension are obtained from the developmental abnormality representation data, and the abnormal feature level is determined according to the degree of abnormality persistence and deviation of the abnormal feature representations. The abnormality category of the target child is determined by classifying the abnormality feature level corresponding to each developmental ability dimension. The degree of abnormality corresponding to the developmental abnormality category is determined based on the level of abnormality characteristics and the degree of abnormality persistence. The developmental abnormality categories, abnormality severity, and corresponding developmental ability dimensions are compiled to generate a child development status assessment result.

9. A multimodal data-based intelligent assessment system for children's developmental disorders, applied to the multimodal data-based intelligent assessment method for children's developmental disorders as described in any one of claims 1 to 8, characterized in that, include: The dimensional feature module is used to acquire multimodal data of the target child in a preset scenario, divide the multimodal data into developmental stages to obtain multimodal time series data, and combine preset rules to acquire dimensional feature data corresponding to each developmental ability dimension. The abnormality triggering module determines the location of physiological changes and abnormal behavior based on dimensional feature data, and determines the developmental abnormality trigger point based on the time overlap interval between the locations of physiological changes and abnormal behavior. The trigger feature module determines the abnormal observation time window based on the developmental abnormality trigger point, and extracts the target dimension features corresponding to the abnormal observation time window from the dimension feature data to obtain the trigger feature set; The deviation identification module, based on the trigger feature set, compares the dimensional feature data under the same capability dimension and generates deviation identification data for the corresponding capability dimension. The module for unachieved features obtains the target feature range corresponding to each developmental ability dimension and developmental stage based on preset reference data, compares the trigger feature set and deviation identification data with the target feature range set, obtains the unachieved features for each developmental ability dimension, and forms an abnormal feature sequence in chronological order. The anomaly characterization module determines the degree of persistence and deviation of unachieved features based on the anomaly feature sequence, and obtains developmental anomaly characterization data by combining the unachieved features corresponding to different developmental ability dimensions. The developmental assessment module determines the type and severity of developmental abnormalities based on data representing developmental abnormalities, combined with the degree of abnormality persistence, deviation, and corresponding developmental ability dimensions, and generates assessment results of the child's developmental status.

10. A smart assessment terminal for children's developmental disorders based on multimodal data, characterized in that: include: One or more processors; A storage device on which one or more programs are stored; When one or more programs are executed by one or more processors, the one or more processors implement the intelligent assessment method for developmental disorders in children based on multimodal data as described in any one of claims 1 to 8.

Images