A double scale AI intelligent mapping and scoring system independent of development environment

Abstract

The application discloses a double scale AI intelligent mapping and scoring system independent of a development environment, relates to the technical field of child development assessment, and comprises an independent AI engine layer, a double scale mapping library, a three-level deep learning mapping module, a dynamic weight model, a three-AI verification module, a self-iteration module, a data storage layer and a hardware adaptation layer. The double scale AI intelligent mapping and scoring system independent of the development environment is decoupled from the development environment and hardware devices through the independent AI engine, can be flexibly deployed on multiple platforms and reused, and greatly reduces the deployment and maintenance costs of institutions; the double scale collaborative assessment mode is adopted, and a low-sensitivity adaptation mechanism designed for special children is matched, so that the actual needs of child development assessment can be fully met, the behavior data collection quality is effectively improved, and the problems of strong subjectivity and single assessment dimension of traditional assessment are improved from the root.

Description

Technical Field

[0001] This invention relates to the field of child development assessment technology, specifically a dual-scale AI intelligent mapping and scoring system independent of the development environment. Background Technology

[0002] Autism spectrum disorder (ASD) is a neurodevelopmental disorder characterized by impairments in social communication and repetitive, stereotyped behaviors. Early and accurate assessment is crucial for improving the effectiveness of intervention and rehabilitation. Currently, assessments of childhood ASD mainly rely on manual scales, which suffer from low assessment efficiency, high subjectivity of results, and insufficient standardization. Specifically: The system is deeply coupled with the development environment and tightly bound to specific development frameworks and hardware devices, making it impossible to reuse across multiple operating systems and hardware platforms. This results in high deployment costs and difficult maintenance for various organizations. Furthermore, the assessment is limited to a single scale, failing to consider the inherent connection between social communication and repetitive stereotyped behaviors, leading to a significant bias in the assessment results. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a dual-scale AI intelligent mapping and scoring system that is independent of the development environment, thus solving the problems mentioned in the background section.

[0004] To achieve the above objectives, the present invention is implemented through the following technical solution: a dual-scale AI intelligent mapping and scoring system independent of the development environment, including an independent AI engine layer, a dual-scale mapping library, a three-level deep learning mapping module, a dynamic weight model, a triple AI verification module, a self-iterative module, a data storage layer, and a hardware adaptation layer. Each module cooperates with each other to complete the collection of children's behavioral data, mapping of behavioral features, calculation of intelligent scores, verification of evaluation results, and iterative optimization of system parameters in sequence. The independent AI engine layer adopts a standardized RESTful API interface with a cross-platform adaptation protocol to achieve complete decoupling from the development environment and hardware devices. The system supports Windows, Linux and Android operating systems and can be adapted to RK3588 and Jetson Orin Nano 8G Super hardware platforms to ensure independent and stable engine operation and achieve platform reuse. The dual-scale mapping library contains two types of assessment tools: a social communication ability scale (containing 6 indicators: eye contact, joint attention, language expression, emotion recognition, social interaction initiation, and instruction response) and a repetitive and stereotyped behavior scale (containing 4 indicators: repetitive actions, stereotyped language, single interest focus, and ritualistic behavior). It designs exclusive low-sensitivity adaptation rules for children with autism spectrum disorder, which include lowering the sound effect trigger threshold during interaction in game scenarios, simplifying animation interaction forms, and extending task response time. The low-sensitivity adaptation parameters are determined based on sensory sensitivity experimental data of 500 children aged 3-12 with ASD, and assessment thresholds are preset for three age groups: three to six years old, seven to nine years old, and ten to twelve years old. The three-level deep learning mapping module uses a hybrid CNN and LSTM model to extract behavioral features, performs primary mapping using a cosine similarity algorithm, and performs intermediate mapping using a multivariate linear regression algorithm. If the cosine similarity of the primary mapping is ≥0.7, it proceeds to the intermediate mapping stage; otherwise, it returns to the data collection stage for supplementary data collection. Finally, it combines with the triple AI verification module to output the final mapping result.

[0005] Furthermore, the dynamic weighting model adjusts the weights of the dual scales and each sub-indicator in real time based on three types of parameters: the age of the assessment subject, behavioral characteristics (frequency of stereotyped actions, degree of social response delay), and assessment scenario (hospital clinical scenario, community screening scenario). The adjustment coefficient is also adjusted in the community screening scenario. A value of 0.55 is used in clinical diagnostic scenarios. The value is 0.65. This is an adjustment factor for age and disease characteristics; The triple AI verification module includes three verification mechanisms that work together to verify the evaluation results. The first mechanism is rule base verification, which verifies the compliance of the mapping results based on the pre-set rules of the dual scales. The second mechanism is sample library comparison, which calls no less than 10,000 sets of children's behavior sample library to carry out similarity matching verification. The third mechanism is real-time scene verification, which combines real-time data of skeletal points, voice, and vision collected by hardware to complete cross-verification.

[0006] Furthermore, the self-iterative module operates based on an incremental learning algorithm, receives clinical assessment feedback data and final assessment results, and optimizes the dual-scale mapping rules, weight model parameters, and validation thresholds through a loss function. The data storage layer uses the AES-256 encryption algorithm to store data locally in CBC mode. The key is kept by a designated person in the institution. Only clinicians are authorized to access the data through identity verification. Only behavioral characteristic data is stored. Raw image and video privacy data are not stored. The data storage period is no less than three years, which complies with the requirements of the Regulations on the Protection of Minors on the Internet. The hardware adaptation layer is compatible with 4K ultra-high-definition cameras, 8-array microphones, 160-point skeletal motion capture devices, and children's haptic feedback wristbands. It collects multiple types of data, including visual (facial expressions, body movements, eye movements), voice (response time, vocabulary, tone), and motion (skeletal point coordinates, social distance), enabling the collection and analysis of multiple types of data.

[0007] Furthermore, the specific workflow of the system is as follows: The first step is data collection: collecting visual, speech, and somatosensory multimodal behavioral feature data of children in gamified scenarios through the hardware adaptation layer; The second step is engine loading: the independent AI engine layer completes cross-platform startup and calls the core data of three types of dual-scale mapping libraries: age threshold, scale indicator standard feature vector, and low-sensitivity adaptation parameters. The third step, three-level mapping scoring: The three-level deep learning mapping module completes feature extraction, primary mapping and intermediate mapping in sequence. The dynamic weight model adjusts the index weights in sync and outputs a preliminary evaluation score. The fourth step is triple verification: the triple AI verification module performs cross-validation on the preliminary score to eliminate abnormal data and biased results; Step 5, Results Output: The system outputs the scores for each item on the dual scale, the total score, and the risk level assessment results for autism spectrum disorder. Step 6, Iterative Optimization: The self-iterative module receives clinical feedback data and automatically optimizes system parameters.

[0008] Furthermore, the CNN and LSTM hybrid model completes the behavioral feature extraction by including the following steps: CNN is used to extract local features from behavioral data. The calculation formula is as follows: in, For convolution output features, It is the ReLU activation function. For convolution kernel weights, To input behavioral data, For bias terms; LSTM is used to extract temporal features from behavioral data. The calculation formula is as follows: in, This represents the hidden state at the current moment. This represents the hidden state at the previous time step. For time-series input features, It is the LSTM weight matrix. LSTM bias vector; By concatenating the two types of features, a hybrid feature is obtained: in, This represents the feature vector after mixing.

[0009] Furthermore, the cosine similarity algorithm completes the initial mapping, calculating the degree of matching between behavioral features and scale indicators using cosine similarity, as shown in the formula: in, For the extracted behavioral feature vector, This is the standard feature vector of the scale indicators. A higher value indicates a better feature match. The multivariate linear regression algorithm completes the intermediate mapping, and by constructing a dynamic calculation model of feature weights, it combines the child's age and behavioral performance to achieve a precise mapping. The formula is as follows: in, For intermediate mapping score, to For regression coefficients, to The independent variables are age and behavioral characteristics. This is the error term.

[0010] Furthermore, the weight calculation method in the dynamic weight model is as follows: in, To assess the total weight, This is a moderating factor for age and symptoms; the value range for children with autism spectrum disorder is 0.5-0.7. As a weighting factor for social communication metrics, Weights for repetitive and stereotyped behaviors.

[0011] Furthermore, the loss function calculation formula in the self-iterative module is as follows: in, This is the total loss value. For clinically accurate scoring, For the system to predict scores, This is an L2 regularization term used to prevent the model from overfitting.

[0012] Furthermore, the encryption calculation formula in the data storage layer is as follows: in, For encrypted ciphertext, For AES-256 encryption functions, A 256-bit encryption key. This refers to plaintext behavioral feature data to be stored.

[0013] This invention provides a dual-scale AI intelligent mapping and scoring system independent of the development environment, which has the following beneficial effects: 1. This dual-scale AI intelligent mapping and scoring system, independent of the development environment, decouples itself from the development environment and hardware devices through an independent AI engine. It can be flexibly deployed and reused on multiple platforms, significantly reducing the deployment and maintenance costs for institutions. It adopts a dual-scale collaborative assessment mode, coupled with a low-sensitivity adaptation mechanism designed for children with special needs, which can fully meet the actual needs of children's developmental assessment, effectively improve the quality of behavioral data collection, and fundamentally improve the problems of strong subjectivity and single assessment dimensions in traditional assessments.

[0014] 2. This dual-scale AI intelligent mapping and scoring system, independent of the development environment, is equipped with a triple AI verification mechanism and a self-iterative optimization module. It can continuously optimize the assessment logic and improve the reliability of the assessment results, making the assessment process more stable and the results more accurate. At the same time, it adopts compliant privacy protection and data storage solutions to ensure the safety of children's data throughout the process. It can be adapted to various assessment scenarios and has strong practicality and promotional value. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the steps of a dual-scale AI intelligent mapping and scoring system independent of the development environment, as described in this invention. Detailed Implementation

[0016] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0017] like Figure 1 As shown, the present invention provides a technical solution: a dual-scale AI intelligent mapping and scoring system independent of the development environment, including an independent AI engine layer, a dual-scale mapping library, a three-level deep learning mapping module, a dynamic weight model, a triple AI verification module, a self-iterative module, a data storage layer, and a hardware adaptation layer. Each module cooperates with each other to complete the collection of children's behavioral data, behavioral feature mapping, intelligent score calculation, evaluation result verification, and system parameter iterative optimization in sequence. The independent AI engine layer adopts a standardized RESTful API interface with a cross-platform adaptation protocol to achieve complete decoupling from the development environment and hardware devices. The system supports Windows, Linux and Android operating systems and can be adapted to RK3588 and Jetson Orin Nano 8G Super hardware platforms to ensure independent and stable engine operation and achieve platform reuse. The dual-scale mapping library contains two types of assessment tools: the social communication ability scale and the repetitive stereotyped behavior scale. It designs exclusive low-sensitivity adaptation rules for children with autism spectrum disorder, including lowering the sound effect trigger threshold during interaction in game scenarios, simplifying animation interaction forms, and extending task response time. It also presets assessment thresholds for three age groups: three to six years old, seven to nine years old, and ten to twelve years old. The dual-scale mapping library is designed according to children's developmental patterns, with preset assessment thresholds for three age groups: 3-6 years old (lower age), 7-9 years old (middle age), and 10-12 years old (older age). It covers two major scales: social communication skills and repetitive stereotyped behaviors. Simultaneously, it matches low-sensitivity adaptation parameters for the corresponding age groups, with the specific standards as follows: I. Assessment Thresholds and Indicator Standards for Children Aged 3-6 (Younger Age Group) (a) Social Communication Skills Scale Eye contact: Each contact lasts for ≥1 second, with ≥3 active eye contact times per minute; Note: A player must initiate at least two such actions in a single game. Non-verbal / verbal responses: Response time ≤ 8 seconds, accuracy ≥ 60%; Social interaction initiation: ≥1 time initiated in a single game session; Language expression: ≥5 effective words per session; Emotion recognition: It can recognize two basic emotions: happiness and anger. (ii) Repetitive Stereotyped Behaviors Scale Repetitive movements: ≤2 times per minute; Stereotype: ≤3 times per game; Single-interest focus: ≤5 minutes per session; Ritualistic behaviors: ≤2 times per session.

[0018] (iii) Low sensitivity adaptation threshold: sound effects reduced by 30%, animation within 15 frames, response time 10 seconds; II. Assessment Thresholds and Indicator Standards for 7-9 Year Olds (Middle Age Group) (a) Social Communication Skills Scale Eye contact: ≥1.5 seconds per contact, ≥4 times per minute; Note: ≥3 times in a single game; Non-verbal / verbal responses: Response time ≤ 6 seconds, accuracy ≥ 70%; Social interaction initiation: ≥2 times per session; Language expression: Output ≥ 3 complete sentences per session; Emotion recognition: It can recognize three emotions: happy, angry, and sad. (ii) Repetitive Stereotyped Behaviors Scale Repetitive actions: ≤ 1 time per minute; Stereotype: ≤2 times per game; Single-interest focus: ≤8 minutes per session; Ritualistic behavior: ≤ 1 time per session.

[0019] (iii) Low sensitivity adaptation threshold: sound effects reduced by 20%, animation within 20 frames, response time 7 seconds; III. Assessment Thresholds and Indicator Standards for 10-12 Year Olds (Older Age Group) (a) Social Communication Skills Scale Eye contact: ≥2 seconds per contact, ≥5 times per minute; Note: ≥4 times in a single game; Non-verbal / verbal responses: Response time ≤ 4 seconds, accuracy ≥ 80%; Social interaction initiation: ≥3 times per session; Language expression: Able to complete full-length dialogues; Emotion recognition: Capable of recognizing four or more complex emotions; (ii) Repetitive Stereotyped Behaviors Scale Repetitive movements: ≤0.5 times per minute; Stereotype: ≤1 time per game; Single-interest focus: ≤10 minutes per session; Ritualistic behavior: Not mandatory for individual events; (iii) Low sensitivity adaptation threshold: sound effects reduced by 10%, animation within 25 frames, response time 5 seconds; When the system starts, it automatically identifies the child's age and loads the corresponding age-group indicator standards, scoring thresholds, and low-sensitivity parameters. All thresholds are the core judgment basis for the dual-scale mapping and scoring, and also the basic verification standard for triple AI verification. The three-level deep learning mapping module uses a hybrid CNN and LSTM model to extract behavioral features, performs primary mapping through the cosine similarity algorithm, and enters intermediate mapping when the similarity is ≥0.7. The intermediate mapping is completed by relying on the multivariate linear regression algorithm, and finally outputs the final mapping result by combining the triple AI verification module. The hybrid CNN and LSTM model completes behavioral feature extraction through the following steps: CNN is used to extract local features from behavioral data. The calculation formula is as follows: in, For convolution output features, It is the ReLU activation function. For convolution kernel weights, To input behavioral data, For bias terms; LSTM is used to extract temporal features from behavioral data. The calculation formula is as follows: in, This represents the hidden state at the current moment. This represents the hidden state at the previous time step. For time-series input features, It is a weight matrix. Bias vector; By concatenating the two types of features, a hybrid feature is obtained: in, This represents the feature vector after mixing; The cosine similarity algorithm completes the initial mapping, calculating the degree of matching between behavioral features and scale indicators using cosine similarity. The formula is as follows: in, For the extracted behavioral feature vector, This is the standard feature vector of the scale indicators. A higher value indicates a better feature match. The multivariate linear regression algorithm completes the intermediate mapping, and by constructing a dynamic calculation model of feature weights, it combines the child's age and behavioral performance to achieve a precise mapping. The formula is as follows: in, For intermediate mapping score, to For regression coefficients, to The independent variables are age and behavioral characteristics. This is the error term; The dynamic weighting model adjusts the weights of the two scales and each sub-indicator in real time based on three types of parameters: the age of the assessment subject, behavioral characteristics, and assessment scenario. The weights in the dynamic weight model are calculated as follows: in, To assess the total weight, This is a moderating factor for age and symptoms; the value range for children with autism spectrum disorder is 0.5-0.7. As a weighting factor for social communication metrics, Weighting of indicators for repetitive and stereotyped behaviors; The triple AI verification module includes three verification mechanisms that work together to verify the evaluation results. The first mechanism is rule base verification, which verifies the compliance of the mapping results based on the pre-set rules of the dual scales. The second mechanism is sample library comparison, which calls no less than 10,000 sets of children's behavior sample library to carry out similarity matching verification. The third mechanism is real-time scene verification, which combines real-time data of skeletal points, voice, and vision collected by hardware to complete cross-verification. Specifically, the system first performs rule base verification. It activates the dual-scale preset rule base and compares the results generated by AI intelligent mapping with the standard specifications in the rule base. For example, if the dual-scale specifies that the frequency of language expression of children in a certain age group should be within a specific range, the system will automatically check whether the language expression data in the mapping results conforms to the rule. If there is a deviation, an early warning will be triggered, indicating that there may be an assessment abnormality. Then, a sample database comparison is performed. The database containing over 10,000 sets of children's behavior samples is called up, and a deep learning algorithm is used to match the behavioral data of the current assessment subject with the sample database data. The system will calculate the similarity score of each behavioral indicator. If the overall similarity is lower than the preset threshold, it indicates that the behavioral pattern of the assessment subject is significantly different from that of the regular samples, and the accuracy of the assessment results needs to be further verified. Finally, real-time scene verification is implemented, synchronously accessing real-time data such as skeletal points, voice, and vision collected by hardware devices, and cross-validating them with the results of AI mapping and the first two verifications. For example, if there is a logical conflict between the voice expression data and skeletal movement data of the evaluation object (such as active language expression but a withdrawn limb movement), the system will re-analyze the data, correct any possible misjudgments, and ensure that the evaluation results can truly reflect the behavior in the actual scene. Example of the triple AI verification module: Taking a 3-year-old child with autism spectrum disorder in a gamified assessment scenario of catching a groundhog as an example, the triple AI verification module performs layer-by-layer verification on the preliminary score after three-level mapping and dynamic weight calibration in the order of rule base verification → sample library comparison → real-time scenario verification. The specific implementation process is as follows: Rule base verification: The system automatically loads the preset rule base of the dual scale for children aged 3-6, and compares the results of the four core indicators—eye contact, joint attention, repetitive movements, and voice response—generated by AI mapping with the standard thresholds in the rule base one by one. Verification of social communication rules: The mapping results show that the number of active video attention per minute is 2, which is lower than the threshold of "≥3 times" in the rule base; the number of times joint attention is initiated is 1, which is lower than the threshold of "≥2 times" in the rule base; the voice response accuracy is 55%, which is lower than the threshold of "≥60%" in the rule base. The system marks the above three items as rule deviation items. Verify repetitive and stereotyped behavior rules: The mapping results show that the repetitive action occurs 3 times per minute, which is higher than the rule base threshold of "≤2 times". The system marks it as a rule deviation item. The system generates a rule verification report, marking four deviation indicators, triggering a preliminary warning, and prompting that it needs to proceed to the next level of comparison and verification, and does not directly accept the preliminary score; Sample database comparison: The system calls upon a sample database of at least 10,000 behavioral standards of 3-year-old children (including samples of healthy children and children with ASD), and uses a deep learning algorithm (an existing algorithm) to perform a full-domain matching of the current child's behavioral characteristics with the sample database data. Feature matching: The four features collected in this study—eye avoidance, lack of co-attention, frequent stereotyped movements, and slow speech response—were compared with the feature vectors of 3-year-old ASD children in the sample database. The overall similarity was 0.89, which is higher than the preset threshold of 0.70. Exclusion of differences: The similarity with the healthy children's sample was only 0.31, which is far below the threshold, thus excluding abnormal situations such as "data collection error" and "behavioral record error"; The current child's behavior pattern is confirmed to be highly consistent with the standard sample of a 3-year-old child with ASD. The deviation items in the rule base verification are real behavioral manifestations, rather than system calculation errors. The deviation marks are retained and the process proceeds to the third layer of verification. Real-time scene verification: The system synchronously retrieves skeletal point, voice, and visual data collected in real time by the hardware, and performs logical cross-validation with the results of the first two layers of verification to ensure that the score closely matches real-world scene behavior. Skeletal somatosensory data verification: Real-time capture of children's social distance consistently exceeding the safe range, lack of coordination and synchronization in limb movements, and prolonged withdrawal posture, which is completely consistent with the mapping results of "social communication deficit"; Voice data verification: Real-time voice data shows that the average response time of children to commands is 11 seconds and there is no active voice interaction, which is completely consistent with the mapping result of "slow voice response and low accuracy". Visual data verification: Real-time visual data shows that children consistently avoid eye contact and fixate without a specific target, which is completely consistent with the mapping result of "insufficient eye contact"; Conflict verification: No logical conflicts such as "active vocal expression but physical withdrawal" or "normal gaze but no shared attention" were found, and there is no need to recalculate the mapping results; All three layers of verification passed, confirming the authenticity and validity of the preliminary score, eliminating abnormal data interference, and outputting the final calibrated evaluation score. After triple AI verification, the system confirms that the scores for social communication skill deficiencies and repetitive stereotyped behaviors are true and credible. The final scores are submitted to the results output stage, and the verification logs are stored in the data storage layer to provide a basis for subsequent system iterations. The self-iterative module operates based on an incremental learning algorithm, receiving clinical assessment feedback data and final assessment results, providing feedback every 100 cases or... When the value is ≥0.1, the parameters are optimized by optimizing the dual-scale mapping rule, weight model parameters and validation threshold through the loss function; The formula for calculating the loss function in the self-iterative module is: in, This is the total loss value. For clinically accurate scoring, For the system to predict scores, This is an L2 regularization term used to prevent the model from overfitting; The data storage layer uses the AES-256 encryption algorithm to store data locally, saving only behavioral feature data and not storing original image or video privacy data. The data storage period is no less than three years, which complies with the requirements of the Regulations on the Protection of Minors on the Internet. The encryption calculation formula in the data storage layer is: in, For encrypted ciphertext, For AES-256 encryption functions, A 256-bit encryption key. Plaintext behavioral feature data to be stored; The hardware adaptation layer is compatible with 4K ultra-high-definition cameras, 8-array microphones, 160-point skeletal motion capture devices, and children's haptic feedback wristbands, enabling the collection and analysis of visual (eye tracking, facial expressions), voice (response time, tone of voice), and motion (skeletal points, social distance) data.

[0020] Taking the assessment of a 3-year-old child with ASD as an example; the hardware includes a 43-inch 1080P smart screen, a 160-point skeletal motion sensing system, a 4K camera, an 8-array microphone, and a children's tactile wristband; the operating system is Android 8.0 or above, the hardware platform is the RK3588 embedded AI platform, and the independent AI engine layer is started through the RESTful API interface, which is completely decoupled from the development environment and can run independently locally; One hundred 3-year-old children diagnosed with ASD were selected as the assessment sample. All samples met the DSM-5 diagnostic criteria for autism spectrum disorder and had no other serious neurodevelopmental complications. The system's exclusive gamified assessment scenario, a collaborative interactive game of catching a groundhog, was used. The scenario was loaded with low-sensitivity adaptation rules for ASD in children aged 3-6: the sound effect trigger threshold was reduced by 30%, the animation interaction was simplified to 15 frames, and the task response time was extended to 10 seconds. The hardware adaptation layer completes self-testing of all data collection devices, automatically identifies the assessment subject as 3 years old, loads age-appropriate assessment thresholds and indicator standards for children aged 3-6, and enables ASD low-sensitivity adaptation rules. During the child's participation in the "Catch the Groundhog" game, the hardware devices simultaneously collect the following behavioral data, without storing any original images or video privacy data: Motion data: social distance, synchronization rate of body movements, and skeletal point movement trajectory; Visual data: duration of eye contact, number of times actively focused on the screen, facial expression features; Voice data: command response duration, voice response frequency, and pronunciation clarity; The independent AI engine layer starts across the Android platform, loading the indicator standards and feature vector library of the social communication ability scale and the repetitive stereotyped behavior scale; Feature extraction: A CNN model is used to extract local behavioral features, and an LSTM model is used to extract temporal behavioral features. The features are then concatenated to obtain a hybrid behavioral feature vector. Primary mapping: The matching degree between behavioral features and scale indicators is calculated using the cosine similarity algorithm. If the matching degree is ≥0.7, it proceeds to intermediate mapping. Intermediate mapping: The scoring is mapped using a multivariate linear regression algorithm combined with the age characteristics of 3-year-old children; The system identifies and assesses 3-year-old children with ASD, adjusting for age and symptoms. Set the weights to 0.6, with the social communication indicator weighted at 0.6 and the repetitive stereotyped behavior indicator weighted at 0.4. Complete the weight calibration and output the preliminary score. Rule base verification: Verify the scoring logic against the ASD assessment rules for children aged 3-6; Sample database comparison: Similarity matching was performed by calling a sample database of 10,000 sets of behavioral standards of 3-year-old children; Real-time scene verification: Combine skeletal points, voice, and visual real-time data for cross-validation and eliminate abnormal scores; Finally, the system outputs sub-scores for social communication, sub-scores for repetitive and stereotyped behaviors, a comprehensive score, and an ASD risk warning level. The data storage layer uses AES-256CBC encryption mode to store the assessment data, with a storage period of ≥3 years, retaining only behavioral characteristic data. In addition, it collects manual assessment feedback data from 30 clinical pediatric developmental and behavioral physicians. The self-iterative module optimizes the low-sensitivity adaptation rules and indicator weight parameters through incremental learning algorithms to complete the system iteration. Based on the above description, this invention achieves decoupling from the development environment and hardware devices through an independent AI engine, enabling flexible deployment and reuse across multiple platforms and significantly reducing the deployment and maintenance costs for institutions. It adopts a dual-scale collaborative assessment mode, coupled with a low-sensitivity adaptation mechanism designed for children with special needs, which can fully meet the actual needs of children's developmental assessment, effectively improve the quality of behavioral data collection, and fundamentally improve the problems of strong subjectivity and single assessment dimensions in traditional assessments. The system is equipped with a triple AI verification mechanism and a self-iterative optimization module, which can continuously optimize the assessment logic and improve the reliability of the assessment results, making the assessment process more stable and the results more accurate. At the same time, it adopts a compliant privacy protection and data storage solution to ensure the safety of children's data throughout the process. It can be adapted to various assessment scenarios and has strong practicality and promotional value.

[0021] The embodiments of the present invention are given for illustrative and descriptive purposes only, and are not intended to be exhaustive or to limit the invention to the forms disclosed. Many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described in order to better illustrate the principles and practical application of the invention, and to enable those skilled in the art to understand the invention and to design various embodiments with various modifications suitable for a particular purpose.

Claims

1. A dual-scale AI intelligent mapping and scoring system independent of the development environment, characterized in that: It includes an independent AI engine layer, a dual-scale mapping library, a three-level deep learning mapping module, a dynamic weight model, a triple AI verification module, a self-iterative module, a data storage layer, and a hardware adaptation layer. Each module works together to complete the collection of children's behavioral data, mapping of behavioral features, calculation of intelligent scores, verification of assessment results, and iterative optimization of system parameters. The independent AI engine layer adopts a standardized RESTful API interface with a cross-platform adaptation protocol to achieve complete decoupling from the development environment and hardware devices. The system supports Windows, Linux and Android operating systems and can be adapted to RK3588 and Jetson Orin Nano 8G Super hardware platforms to ensure independent and stable engine operation and achieve platform reuse. The dual-scale mapping library contains two types of assessment tools: the social communication ability scale and the repetitive stereotyped behavior scale. It designs exclusive low-sensitivity adaptation rules for children with autism spectrum disorder, including lowering the sound effect trigger threshold during interaction in game scenarios, simplifying animation interaction forms, and extending task response time. It also presets assessment thresholds for three age groups: three to six years old, seven to nine years old, and ten to twelve years old. The three-level deep learning mapping module uses a hybrid CNN and LSTM model to extract behavioral features, performs primary mapping through the cosine similarity algorithm, completes intermediate mapping based on the multivariate linear regression algorithm, and finally outputs the final mapping result by combining the triple AI verification module.

2. The dual-scale AI intelligent mapping and scoring system independent of the development environment as described in claim 1, characterized in that: The dynamic weighting model adjusts the weights of the dual scales and each sub-indicator in real time based on three types of parameters: the age of the assessment subject, behavioral characteristics, and assessment scenario. The triple AI verification module includes three verification mechanisms that work together to verify the assessment results. The first mechanism is rule base verification, which verifies the compliance of the mapping results based on the pre-set rules of the dual scales. The second mechanism is sample database comparison, which calls on a sample database of no less than 10,000 children’s behaviors to conduct similarity matching verification. The third item is real-time scene verification, which involves cross-validation by combining real-time data of skeletal points, voice, and vision collected by the hardware.

3. The dual-scale AI intelligent mapping and scoring system independent of the development environment as described in claim 1, characterized in that: The self-iterative module runs based on an incremental learning algorithm, receives clinical assessment feedback data and final assessment results, and optimizes the dual-scale mapping rules, weight model parameters and validation thresholds through a loss function. The data storage layer uses the AES-256 encryption algorithm to store data locally, saving only behavioral feature data and not storing original image or video privacy data. The data storage period is no less than three years, which complies with the requirements of the Regulations on the Protection of Minors on the Internet. The hardware adaptation layer is compatible with 4K ultra-high-definition cameras, 8-array microphones, 160-point skeletal motion capture devices, and children's haptic feedback wristbands, enabling the collection and analysis of multiple types of data, including visual, voice, and motion sensing data.

4. A dual-scale AI intelligent mapping and scoring system independent of the development environment according to any one of claims 1-3, characterized in that: The specific workflow of the system is as follows: The first step is data collection: collecting multimodal behavioral data of children in gamified scenarios through a hardware adaptation layer; The second step is engine loading: The independent AI engine layer completes cross-platform startup and calls relevant data from the dual-scale mapping library; The third step, three-level mapping scoring: The three-level deep learning mapping module completes feature extraction, primary mapping and intermediate mapping in sequence. The dynamic weight model adjusts the index weights in sync and outputs a preliminary evaluation score. The fourth step is triple verification: the triple AI verification module performs cross-validation on the preliminary score to eliminate abnormal data and biased results; Step 5, Results Output: The system outputs the scores for each item on the dual scale, the total score, and the risk level assessment results for autism spectrum disorder. Step 6, Iterative Optimization: The self-iterative module receives clinical feedback data and automatically optimizes system parameters.

5. The dual-scale AI intelligent mapping and scoring system independent of the development environment as described in claim 1, characterized in that: The CNN and LSTM hybrid model completes behavioral feature extraction through the following steps: CNN is used to extract local features from behavioral data. The calculation formula is as follows: in, For convolution output features, It is the ReLU activation function. For convolution kernel weights, To input behavioral data, For bias terms; LSTM is used to extract temporal features from behavioral data. The calculation formula is as follows: in, This represents the hidden state at the current moment. This represents the hidden state at the previous time step. For time-series input features, It is a weight matrix. Bias vector; By concatenating the two types of features, a hybrid feature is obtained: in, This represents the feature vector after mixing.

6. The dual-scale AI intelligent mapping and scoring system independent of the development environment as described in claim 1, characterized in that: The cosine similarity algorithm completes the initial mapping, calculating the degree of matching between behavioral features and scale indicators using cosine similarity. The formula is as follows: in, For the extracted behavioral feature vector, This is the standard feature vector of the scale indicators. A higher numerical value indicates a better feature match. The multivariate linear regression algorithm completes the intermediate mapping, and by constructing a dynamic calculation model of feature weights, it combines the child's age and behavioral performance to achieve a precise mapping. The formula is as follows: in, For intermediate mapping scores, to For regression coefficients, to The independent variables are age and behavioral characteristics. This is the error term.

7. The dual-scale AI intelligent mapping and scoring system independent of the development environment according to claim 2, characterized in that: The weight calculation method in the dynamic weight model is as follows: in, To assess the total weight, This is a moderating factor for age and symptoms; the value range for children with autism spectrum disorder is 0.5-0.

7. As a weighting factor for social communication metrics, Weights for repetitive and stereotyped behaviors.

8. The dual-scale AI intelligent mapping and scoring system independent of the development environment according to claim 3, characterized in that: The formula for calculating the loss function in the self-iterative module is as follows: in, This is the total loss value. For clinically accurate scoring, For the system to predict scores, This is an L2 regularization term used to prevent the model from overfitting.

9. A dual-scale AI intelligent mapping and scoring system independent of the development environment as described in claim 3, characterized in that: The encryption calculation formula in the data storage layer is: in, For encrypted ciphertext, For AES-256 encryption functions, A 256-bit encryption key. This refers to plaintext behavioral feature data to be stored.

Images