Autism spectrum disorder screening application system

Abstract

The invention relates to an autism spectrum disorder screening application system which comprises a front-end user interaction device and a background data processing server. The front-end user interaction device comprises a child information management module; the scale screening module is used for providing at least one autism screening scale for the user to select and fill, and generating a scale screening score after receiving a scale answer filled by the user; the photo screening module is used for receiving at least one child front face image uploaded by the user, preprocessing the face image and then calling a photo analysis model to generate a photo screening score; the screening process of the front-end user interaction device comprises the steps that after a user selects a target child through the child information management module, operation of the scale screening module and the photo screening module is completed in sequence, and the background data processing server generates a comprehensive score and a risk level for scale screening scores and photo screening scores through a weighted voting mechanism. Through the application system, accurate ASD pre-diagnosis classification can be quickly provided.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to an application system for screening autism spectrum disorder. Background Technology

[0002] Autism Spectrum Disorder (ASD), also known as autism, is a neurodevelopmental disorder that affects social interaction, communication, and behavior. Autism is primarily caused by genetic and environmental factors; however, its pathogenesis remains unclear. Furthermore, due to the high heterogeneity of its phenotype and etiology, early diagnosis of autism is challenging. The Autism and Developmental Disorders Surveillance Network (ADDM) of the U.S. Centers for Disease Control and Prevention (CDC) reports that approximately 1 in 36 children exhibit symptoms associated with autism, with an average age of diagnosis of about 4 years, although parents and caregivers often notice abnormalities as early as age 2.

[0003] Currently, the diagnosis of autism mainly relies on observation and the experience of clinicians. This method is time-consuming, costly, subjective, prone to misdiagnosis, and highly dependent on professional medical personnel. Furthermore, the average age of diagnosis for children with autism is currently over 4 years old, which is not conducive to early intervention and treatment. Therefore, developing effective early screening applications for autism is crucial for timely diagnosis and intervention. Summary of the Invention

[0004] To address the aforementioned shortcomings, this invention proposes an autism spectrum disorder screening application system. This invention integrates multiple modalities of data screening methods, which not only provide accurate pre-diagnostic classification but also improve the efficiency and accessibility of the screening process, offering more timely intervention and support to affected children.

[0005] The technical solution adopted in this invention is an autism spectrum disorder screening application system, including a front-end user interaction device and a back-end data processing server, wherein the front-end user interaction device and the back-end data processing server establish a communication connection through a network. The front-end user interaction device includes: The child information management module is used to receive and store information on at least one child input by the user, and the child information is used for subsequent personalized screening and assessment. The scale screening module is used to provide at least one autism screening scale for users to select and fill in, and to generate a scale screening score after receiving the user's scale answers. The photo screening module is used to receive at least one frontal facial image of a child uploaded by the user, preprocess the facial image, and then call the photo analysis model to generate a photo screening score. The screening process of the front-end user interaction device is as follows: After the user selects the target child through the child information management module, the user completes the operation of the scale screening module and the photo screening module in sequence. The back-end data processing server uses a weighted voting mechanism to generate a comprehensive score and risk level based on the scale screening score and the photo screening score.

[0006] Preferably, the autism screening scale provided by the scale screening module includes at least the self-developed Chinese Children's Scale, which covers children aged 2 to 12 years, with one specific scale for each year of age, and a 12-year-old specific scale for children aged 13 to 18 years. Each of the self-developed Chinese Children's Scales contains 10 to 15 questions, each with a different weight value. The total score ranges from 0 to 100 points, with scores above 80 points indicating high risk, 60 to 79 points indicating medium risk, and 0 to 59 points indicating low risk.

[0007] Preferably, the self-developed Chinese children's scale is developed based on the DeepSeek-R1-Distill-Qwen-14B base model and is fine-tuned under supervision using multiple types of data; the multiple types of data include one or more of the following: academic papers, medical textbooks, multiple mental illness screening scales, or Chinese children's medical consultation dialogue datasets.

[0008] Preferably, the preprocessing step of the photo screening module is: to uniformly process the facial images uploaded by users to a standard size; The architecture of the photo analysis model includes: The feature extraction layer receives a standard-sized facial image, extracts the feature information of the standard-sized facial image, and outputs an image feature map; The spatial attention module is communicatively connected to the feature extraction layer, receives the image feature map, enhances attention to key regions in the image feature map, and outputs spatial attention features. The facial landmark detection layer receives a standard-sized facial image, detects landmarks in key areas of a child's face within the standard-sized facial image, encodes the landmarks, and outputs a feature vector. The global-local cross-fusion relationship converter encoder is communicatively connected to the spatial attention module and the facial landmark detection layer, respectively. It receives the spatial attention features and the feature vector, and performs global-local feature collaborative merging on the spatial attention features and the feature vector through a preset fusion mechanism to output dual-stream fused features. A multi-feature interaction network is communicatively connected to the global-local cross-fusion relationship converter encoder, receives the dual-stream fusion features, performs attention calculation on the dual-stream fusion features to filter effective features related to autism spectrum disorder identification, and outputs comprehensive features; The classification layer is connected to the multi-feature interaction network, receives the comprehensive features, classifies the comprehensive features into autism spectrum disorder and typical development, and generates and outputs the photo screening score.

[0009] Preferably, the standard size of the facial image is 224×224 pixels; the deep learning network used in the feature extraction layer is a deep separable convolutional neural network, a 50-layer residual network, or a lightweight convolutional neural network; the size of the image feature map output by the feature extraction layer is 256×7×7.

[0010] Preferably, the attention calculation process of the multi-feature interaction network includes: calculating the contribution weight of each feature dimension in the dual-stream fusion feature to the identification of autism spectrum disorder, dynamically retaining high-contribution features and filtering low-contribution features according to the contribution weight; at the same time, establishing an information interaction channel between different feature dimensions in the dual-stream fusion feature, integrating cross-dimensional feature information, and generating the comprehensive feature.

[0011] Preferably, the training and validation process of the photo analysis model includes: training the model based on a large open-source autism facial image dataset; and validating the trained model using facial photos extracted from a Chinese social media autism children video dataset.

[0012] Preferably, the facial landmark detection layer uses a facial feature point detection network to detect facial landmarks; the landmarks of the key facial regions include contour points and feature points corresponding to the eyes, nose, and mouth; the feature vector output by the facial landmark detection layer has a size of 256×7×7.

[0013] Preferably, in the weighted voting mechanism, the weight ratio of the scale screening score to the photo screening score is 1:1.

[0014] Preferably, the front-end user interaction device further includes: The video screening module is used to receive at least one 10-30 second video of a child's face uploaded by the user, and after preprocessing the facial video, call the video analysis model to generate a video screening score. If the combined risk level of the scale screening score and the photo screening score is high, the user can choose to activate the video screening module. The background data processing server will then combine the combined score and the video screening score to generate the final screening result.

[0015] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention discloses an independent autism rapid screening application system based on static photos and facial videos. Users upload standard facial photos and videos of their children, and the system can call the pre-deployed facial analysis model in the cloud server to automatically identify autism-related characteristic indicators, achieving one-click rapid screening. The operation is simple and convenient. This technology is time-saving and has a low threshold. Parents can conduct preliminary screening anytime and anywhere through mobile devices such as smartphones, providing a more convenient and efficient solution for the early detection of autism.

[0016] (2) This invention integrates multiple screening methods such as scale screening, photo screening, and video screening into the same application platform, which effectively improves the accuracy and efficiency of autism screening and provides users with a comprehensive and convenient early identification solution for autism. It can capture children's behavioral and physiological characteristics from different angles, enhance the reliability and comprehensiveness of screening results, and make the screening process more efficient and user-friendly.

[0017] (3) This invention develops highly accurate scales, photo and video processing models for Chinese children to ensure the localization and applicability of the screening tools. By using a localized Chinese social media video dataset of children with autism for training and optimization, and conducting professional physician assessment tests and small-scale internal testing in tertiary hospitals, these models can more accurately identify the risk of autism in Chinese children, thereby providing more accurate screening results and helping parents and medical professionals to detect and intervene in autism early. Attached Figure Description

[0018] The present invention will now be described in detail with reference to the embodiments and accompanying drawings, wherein: Figure 1 This illustrates the overall framework of the autism spectrum disorder screening application system; Figure 2 The application's homepage is displayed; Figure 3 The application's screening function interface is shown; Figure 4 The application's information interface is displayed; Figure 5 The app's "My" interface is displayed; Figure 6 The application's login screen is shown; Figure 7 The app's "Add Child" interface is shown; Figure 8 The application's child information management interface is shown; Figure 9 The application's scale screening list interface is displayed; Figure 10 The application's scale screening details page interface is shown; Figure 11 The interface showing the scale assessment results of the application is displayed; Figure 12 The application's photo screening interface is shown; Figure 13 The application's video screening interface is shown; Figure 14 The application's information content interface is displayed; Figure 15 The application's background login interface is shown; Figure 16 The application's scale management interface is shown; Figure 17 The scale editing interface of the application is shown; Figure 18 The application's photo screening configuration interface is shown; Figure 19 The application's video screening configuration interface is shown; Figure 20 The application's user management interface is shown; Figure 21 The application's information management interface is shown; Figure 22 The application's information editing interface is shown; Figure 23 The application's homepage carousel management interface is shown; Figure 24 The application's administrator management interface is shown; Figure 25 The application's APP version management interface is shown; Figure 26 The application's screening results display interface is shown; Figure 27 The application's information content display interface is shown; Figure 28 The application's background scale management interface is shown; Figure 29 The application's evaluation record management interface is shown. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar components or components having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0020] To address the problems that autism diagnosis relies heavily on observation and the experience of clinicians, which is time-consuming, costly, subjective, prone to misdiagnosis, and highly dependent on professional medical personnel, this invention proposes an autism spectrum disorder screening application system. This system can integrate multiple modal data screening methods, providing accurate pre-diagnosis classification, improving the efficiency and accessibility of the screening process, and offering more timely intervention and support for affected children.

[0021] In one embodiment, an autism spectrum disorder screening application system includes a front-end user interaction device and a back-end data processing server. The front-end user interaction device and the back-end data processing server establish a communication connection through a network. Through a closed-loop process of child information filing, multimodal screening, and back-end data fusion, efficient and accurate initial screening of autism spectrum disorder (ASD) is achieved.

[0022] The front-end user interaction device includes a child information management module, a scale screening module, and a photo screening module. (See also...) Figure 1 The front end generates requirements or data through user operations and transmits them to the back end; the back end processes the data based on business logic (such as calling screening algorithms and matching scale rules) and interacts with the database to complete data reading and writing; finally, the processing results (such as screening scores and information content) are fed back to the front end, realizing a complete closed loop of multimodal autism screening, information management, and popular science services.

[0023] The child information management module receives and stores information input by the user for at least one child. This child information is used for subsequent personalized screening and assessment. See details... Figures 7-8 After logging into the app, users can add multiple children for autism screening and enjoy convenient family autism management services. First, users need to complete basic information for each child, including key data such as name, gender, date of birth (for matching age-appropriate scales), and medical information. The system also provides optional fields such as family history of mental illness and the child's past mental illness history to collect more comprehensive health background information. A history of autism in three generations of immediate family members can be used as a reference weight for risk prediction, while a past mental illness history is used to rule out interference from other diseases. For example, if a child has a family history of autism, the system can appropriately increase the weight of the photo screening score when calculating the overall score, enhancing sensitivity to high-risk individuals. In this way, the application not only provides an efficient screening tool but also establishes a comprehensive health record for the child.

[0024] The scale screening module provides users with at least one autism screening scale to select and complete. After receiving the user's answers, it generates a screening score. After the user completes the form, the front end transmits the data of each question's answer to the back end. The back end then automatically calculates the screening score based on the scale's preset weighting rules (e.g., "avoiding eye contact" has a higher weight than "language delay" for a particular question), thus avoiding errors from manual scoring.

[0025] The photo screening module receives at least one frontal facial image of a child uploaded by a user. After preprocessing the facial image, it calls a photo analysis model to generate a photo screening score. The photo screening module can extract key facial features (such as eye contours, mouth shape, and facial symmetry) and generate a photo screening score that reflects the correlation between the child's facial features and ASD.

[0026] The screening process of the front-end user interaction device is as follows: After the user selects the target child through the child information management module, the user completes the operation of the scale screening module and the photo screening module in sequence. The back-end data processing server uses a weighted voting mechanism to generate a comprehensive score and risk level based on the scale screening score and the photo screening score.

[0027] The core function of the backend is multimodal data fusion, which integrates scale screening scores and photo screening scores through a weighted voting mechanism. Based on the complementary nature of the two screening methods, the scale score reflects "behavioral habits," while the photo score reflects "physiological facial features." Their weights can be set to 1:1 to balance the integration of children's behavioral information provided by the scale and the disease-related information implied in the facial photos, ultimately generating a comprehensive score and achieving effective information balance. The backend automatically generates a comprehensive risk level based on the weighted calculation result (e.g., 80 points from the scale + 70 points from the photo, resulting in a weighted score of 75), comparing it to preset risk thresholds (e.g., above 80 points - high risk, 60-79 points - medium risk, 0-59 points - low risk), and then feeds this information back to the frontend for users to view.

[0028] The process design follows the principle of low cognitive load, allowing users to complete it without requiring specialized knowledge. Launch the app and select the target child through the child information management module (if it is the first time using the app, enter the child's information first). Enter the scale screening module, select the appropriate scale according to the child's age, and complete the form as prompted (the front end verifies the required fields in real time to avoid omissions). The system will automatically redirect to the photo screening module, where users can upload a clear image of the child's face (the system provides shooting guidelines, such as "sufficient lighting and no obstructions"). The front end synchronously transmits the questionnaire answer data and the preprocessed image data to the back end. The backend completes the weighted calculation within 10-30 seconds and returns "overall score + risk level + brief interpretation" to the frontend (e.g., "overall score 75 points, medium risk, it is recommended to go to a professional medical institution for a follow-up examination within 1 month").

[0029] This application system, through its front-end and back-end collaborative architecture, multimodal screening, and personalized adaptation design, not only solves the core pain points of traditional autism screening such as insufficient professional resources, low accuracy, and complex operation, but also realizes the connection between initial screening in the home setting and follow-up examination in professional institutions. It provides a practical technical solution for the early detection and intervention of ASD, and has both clinical practical value (improving screening efficiency and accuracy) and user value (lowering the threshold for use).

[0030] In one embodiment, the front-end user interaction device further includes: The video screening module receives at least one 10-30 second video of a child's face uploaded by the user. After preprocessing the video, it calls a video analysis model to generate a video screening score. Similar to the photo screening module, the video screening module requires the user to upload a video of the child's face taken in a natural environment. The video analysis model analyzes the uploaded data and returns an assessment score on whether the child has autism. Based on the assessment score, the application provides personalized suggestions to the user, guiding them on whether they should seek professional medical help at a hospital for more in-depth diagnosis and intervention.

[0031] If the combined risk level of the scale screening score and the photo screening score is high, the user can choose to activate the video screening module. The background data processing server will then combine the combined score and the video screening score to generate the final screening result.

[0032] Scale screening allows parents and healthcare professionals to quickly assess a child's social, communication, and behavioral patterns; photo screening uses computer vision algorithms to analyze a child's facial features to identify potential signs of autism; and video screening achieves more accurate early autism screening by analyzing a child's comprehensive facial dynamic features. The combination of multiple screening methods provides a more flexible screening tool, reducing the risk of misdiagnosis and missed diagnosis. In this embodiment, scale screening and photo screening are mandatory, while video screening is optional. In this way, the present invention provides a comprehensive, efficient, and user-friendly solution for autism screening, thereby improving the outcomes of early identification and intervention and having a positive impact on the long-term development of children with autism.

[0033] In one embodiment, the autism screening scale provided by the scale screening module includes at least a self-developed Chinese children's scale, which covers children aged 2 to 12 years, with one specific scale for each year of age, and a 12-year-old-specific scale for children aged 13 to 18 years. Each self-developed Chinese children's scale contains 10 to 15 questions, each with a different weight value. The total score ranges from 0 to 100 points, with scores above 80 points indicating high risk, 60 to 79 points indicating medium risk, and 0 to 59 points indicating low risk.

[0034] The core of this embodiment is to design a dedicated scale system to meet the localized and precise needs of screening for autism spectrum disorder (ASD) in Chinese children. By linking the system with the front-end children's information management module and the back-end scoring logic, a complete screening chain is formed, from age-appropriateness, question answering, weighted scoring to risk classification.

[0035] The golden period for the onset and intervention of ASD symptoms is between 2 and 12 years old. During this stage, children's cognitive, language, and social skills develop rapidly, with significant differences in developmental milestones for each year of age (e.g., at age 2, it is necessary to assess whether they can point to objects to express needs; at age 6, it is necessary to assess whether they can participate in group games and follow rules; and at age 12, it is necessary to assess whether they can understand complex social situations). Age-specific scales can avoid the underestimation of younger children and the misjudgment of older children caused by the traditional "uniform scale for ages 3-6" scale. The developmental characteristics of adolescents aged 13-18 are highly similar to those of 12-year-old children (social logic and language abilities have tended to stabilize), eliminating the need for designing new scales, simplifying system maintenance, and ensuring screening consistency.

[0036] Each customized scale contains 10-15 questions, keeping the number short (5-8 minutes to complete) and information-dense to prevent parents from abandoning the screening due to cumbersome procedures. The questions focus on high-frequency life scenarios for Chinese children, such as: family interaction scenarios ("Do they actively share toys or fun stories with their parents?", "Do they respond promptly when called by name?"); school / kindergarten scenarios ("Do they participate in classmates' games instead of playing alone?", "Do they understand simple instructions from teachers?"); and culturally appropriate scenarios ("Can they have simple conversations with relatives and friends during family gatherings such as Chinese New Year and birthdays?", different from non-mainstream scenarios like "community parties" in foreign scales), ensuring parents can answer accurately based on real-life observations and reducing "judgment bias caused by unfamiliar scenarios."

[0037] The weighting design breaks away from the traditional scale's crude model of 1 point per item with equal weight. It assigns higher weight to core symptom questions and lower weight to non-specific symptom questions. For example, high-weighted questions (10-15 points) directly relate to symptoms of core ASD disorders, such as avoidance of eye contact (e.g., not looking at the other person when speaking), delayed language development (e.g., not being able to say simple words at age 2, or not being able to express complete sentences at age 6), and repetitive stereotyped behaviors (e.g., repeatedly clapping, spinning, and being difficult to interrupt). Low-weighted questions (3-5 points) are symptoms not specific to ASD and may overlap with other developmental problems, such as occasional picky eating, sensitivity to specific sounds, and reluctance to try new toys. This weighting difference ensures that the scoring focuses on key ASD characteristics, reducing the interference of non-core symptoms on the results.

[0038] The scale's total score is set from 0 to 100, corresponding to three risk levels. High risk (80 points or above): Indicates that the child has multiple core symptoms of ASD, and the frequency / severity of the symptoms meets the clinical initial screening positive criteria; Medium risk (60-79 points): Indicates that the child has some ASD-related symptoms, but has not reached the positive threshold, requiring "follow-up examination within 1-3 months + continuous observation of symptom changes"; Low risk (0-59 points): Indicates that the child's current developmental performance is consistent with typical developmental disorder (TD) children, and "simultaneous monitoring during routine annual physical examinations" is recommended. This grading standard can effectively avoid over-treatment or missed diagnosis.

[0039] Furthermore, the autism screening scales provided by the scale screening module also include international standard scales, including: the Infant Autism Screening Scale (10-item version, Q-CHAT-10) for infants around 18 months of age, the revised version of the Childhood Autism Rating Scale (including follow-up, M-CHAT-R / F) for infants aged 16–30 months, the Social Communication Questionnaire (SCQ) for children aged 4 years and older, the Autism Spectrum Quotient (AQ-Child) for children aged 4–11 years, the AQ-Adolescent for adolescents aged 12–15 years, and the AQ-Adult for adults aged 16 years and older.

[0040] In one embodiment, the self-developed Chinese children's scale is based on the DeepSeek-R1-Distill-Qwen-14B base model, and undergoes supervised fine-tuning using multiple types of data. These multiple types of data include one or more of the following: academic papers, medical textbooks, various mental illness screening scales, or Chinese children's online consultation datasets. The 14B-parameter base model possesses sufficient medical knowledge understanding and language generation logic, enabling it to accurately parse ASD diagnostic criteria (such as DSM-5 and ICD-11) and convert them into screening questions. Simultaneously, the Distill version, through model compression technology, reduces the model size by 40%-50% while retaining over 90% of core performance, making it compatible with mobile deployments for autism screening apps (avoiding slow loading and memory consumption issues caused by excessively large models). Supervised fine-tuning is the core of this solution, aiming to teach the 14B base model how to generate scales that meet the needs of ASD screening for Chinese children. Four types of core data are selected based on requirements, each playing a different training role.

[0041] Academic papers, such as core journal articles in the field of ASD diagnosis and screening in the past 5 years, are used to enable the model to grasp the latest clinical consensus on ASD and ensure that the scale questions are aligned with medical diagnostic standards.

[0042] Authoritative medical textbooks, such as "Child and Adolescent Psychiatry," are used to solidify the basic medical logic of the model, such as "language skills that a 2-year-old child should master" and "social interaction standards for a 6-year-old child," to avoid generating problems that do not conform to developmental patterns.

[0043] Multiple mental illness screening scales, such as the aforementioned international standard scales, are used to allow the model to learn the scale's question structure (such as closed-ended question design and symptom dimension classification), while also identifying localization defects of international scales to avoid replicating incompatible content.

[0044] Chinese children's medical consultation dialogue datasets, such as consultation records from pediatric / child psychiatric departments of top-tier hospitals in China and interview data from community children's health centers, are used for core localization data. This allows the model to extract high-frequency symptom descriptions and real-life scenarios (such as "kindergarten group activities" and "family meals") that Chinese parents are concerned about, and generate questions that parents can understand and observe.

[0045] In one embodiment, the preprocessing step of the photo screening module is as follows: the facial images uploaded by users are uniformly processed to a standard size, and the standardized preprocessing eliminates data differences. Specifically, the standard size of the facial images is such as 224×224 pixels, which is suitable for the input requirements of mainstream deep learning networks. This not only preserves the detailed features of key facial regions (eyes, nose, mouth, facial contours) (such as eye spacing and mouth shape), but also avoids increased model computation time due to excessively large sizes.

[0046] In addition to size standardization, preprocessing can also include steps such as image quality optimization to further improve the accuracy of feature extraction. Examples include: face alignment: using face detection algorithms to locate key facial points (such as the eyes and nose tip), adjusting tilted images to a frontal pose, and avoiding feature distortion caused by shooting angle; background removal: preserving the facial area (cropped to 224×224), removing cluttered backgrounds (such as furniture and clothing), and reducing interference from irrelevant information on the model; and pixel normalization.

[0047] The architecture of the photo analysis model includes: a feature extraction layer, a spatial attention module, a facial landmark detection layer, a global-local cross-fusion relationship converter encoder, a multi-feature interaction network, and a classification layer.

[0048] The feature extraction layer receives a standard-sized facial image, extracts its feature information, and outputs an image feature map. This transforms the pixel information of the original image into a machine-understandable feature vector, providing foundational data for subsequent attention enhancement and fusion. The deep learning network used in the feature extraction layer is a deep separable convolutional neural network, a 50-layer residual network, or a lightweight convolutional neural network. The size of the image feature map output by the feature extraction layer is, for example, 256×7×7.

[0049] The spatial attention module, communicatively connected to the feature extraction layer, receives the image feature map, enhances the attention of key regions within the image feature map, and outputs spatial attention features. This module enhances key region features through attention weight calculation. For each spatial location (each pixel in a 7×7 grid) in the image feature map, an attention weight is calculated. The weight value is positively correlated with the ASD (Advanced Surface Distance) of the corresponding facial region (e.g., 0.8-0.9 for the periorbital region and 0.2-0.3 for the cheek region). The feature map is then multiplied element-wise by the attention weights to obtain the spatial attention features, thus amplifying the features of key regions and suppressing features of redundant regions. The spatial attention module allows the model to focus on key areas, reduces interference from irrelevant information, and improves the targeting of subsequent feature fusion.

[0050] The facial landmark detection layer receives a standard-sized facial image, detects landmarks in key areas of a child's face within the image, encodes these landmarks, and outputs a feature vector. While the feature extraction layer focuses on "texture features" (such as skin texture), this layer supplements facial features with "structural dimensions" (such as eye spacing and mouth contour), creating a complementary effect. Specifically, the facial landmark detection layer uses a facial landmark detection network to detect facial landmarks; the landmarks in key facial areas include contour points and feature points corresponding to the eyes, nose, and mouth; the feature vector output by the facial landmark detection layer has a size of, for example, 256×7×7.

[0051] The global-local cross-fusion relationship converter encoder is communicatively connected to the spatial attention module and the facial landmark detection layer, respectively. It receives the spatial attention features and the feature vector, and performs global-local feature collaborative merging of the spatial attention features and the feature vector through a preset fusion mechanism. This combines detailed local features with broader facial features, outputting a dual-stream fused feature. Breaking away from traditional simple stitching fusion methods, it achieves deep fusion through a global and local dual-attention mechanism, solving the problem of the separation between texture features and structural features. The local multi-head self-attention layer focuses on "local detail association," such as fusing local association between periorbital texture features and interocular distance structure features to capture the combined anomaly of "excessively wide interocular distance + stiff periorbital texture." The global multi-head self-attention layer focuses on "global overall association," such as fusing global association between facial contour structure features and full-face texture features to determine the overall features of "abnormal facial proportions + stiff facial expression." The collaborative merging merges the local fusion results with the global fusion results through "residual connections," preserving the accuracy of local details while taking into account the integrity of global features, achieving a 1+1>2 effect of "texture + structure" features, allowing the model to identify ASD facial anomalies from both detail and overall dimensions.

[0052] A multi-feature interaction network, communicatively connected to the global-local cross-fusion relationship converter encoder, receives the dual-stream fused features, performs attention calculations on the dual-stream fused features to filter effective features relevant to autism spectrum disorder (ASD) identification, and outputs comprehensive features. The multi-feature interaction network further refines features, reducing the interference of irrelevant information on classification results and improving the model's noise resistance. Specifically, the attention calculation process of the multi-feature interaction network may include: calculating the contribution weight of each feature dimension in the dual-stream fused features to ASD identification; dynamically retaining high-contribution features and filtering low-contribution features based on the contribution weights; simultaneously, establishing information interaction channels between different feature dimensions in the dual-stream fused features to achieve cross-dimensional feature information integration, ultimately generating the comprehensive features. This enhances the model's ability to analyze subtle changes in facial morphology, texture, and other ASD indicators, improves the model's robustness to different facial expressions and conditions, and significantly improves screening accuracy.

[0053] The classification layer, communicatively connected to the multi-feature interaction network, receives the comprehensive features, classifies the comprehensive features into autism spectrum disorder and typical developmental characteristics, and generates and outputs the photo screening score. The classification layer transforms abstract feature vectors into user-understandable screening scores, providing the core basis for the system's comprehensive risk assessment in terms of image dimensions.

[0054] In one embodiment, the training and validation process of the photo analysis model includes: training the model based on the large open-source autism facial image dataset AFID; and validating the trained model using facial photos extracted from a Chinese social media autism children video dataset.

[0055] The dataset contains video data of 182 Chinese children aged 6 months to 15 years with autism engaging in natural activities. The data primarily comes from major Chinese social media platforms. Facial images of each child were cropped to validate the model's performance. Testing on these 182 facial photos of Chinese children showed an accuracy rate of 86%, demonstrating the model's effectiveness. The photo screening results provide recommendations based on risk level (low, medium, high). If the analysis suggests a high probability of a child having autism, the application will recommend further video screening for a more comprehensive assessment.

[0056] In one embodiment, the front-end user interaction device further includes: a user login module, used to receive account information entered by the user to complete the login, or to receive registration information of a new user to create a personal account.

[0057] In one embodiment, the front-end user interaction device further includes an information module for displaying autism-related science popularization content and application operation guidelines. The information module is used to educate parents about the symptoms and harms of autism, promote the benefits of early autism screening, and raise family awareness of early screening and treatment of autism, thereby reducing the social harm and economic losses caused by autism. In the information dissemination module, staff will regularly update relevant knowledge and the latest research progress on autism, such as intervention and treatment, prognosis and support, aiming to help parents better understand and cope with autism through professional science popularization education and support.

[0058] In one embodiment, the backend data processing server includes: The scale management module is used by the administrator to add, remove, edit, and maintain information on the autism screening scales in the scale screening module. The photo screening configuration module is used by the administrator to configure the parameters and interface information of the photo analysis model called by the photo screening module. The video screening configuration module is used by the administrator to configure the parameters and interface information of the video analysis model called by the video screening module. The user management module is used by administrators to view and approve user information and manage user status. The information management module is used by administrators to add, delete, and modify content within the information module.

[0059] For application administrators, this invention features a comprehensive backend management platform. This platform integrates core functions such as data management, business logic processing, and security control, enabling administrators to efficiently and conveniently complete several key operations, including user permission management, real-time data monitoring and analysis, dynamic adjustment of the mobile UI, flexible configuration of functional modules, and rapid publishing and updating of information content. Through a carefully designed interactive process and a visual interface, the platform significantly simplifies daily management tasks, greatly improves management efficiency, and ensures timely maintenance and updates of application content. This innovative design not only enhances the system's robustness and usability but also effectively extends the application's lifecycle, providing strong technical support for the application's sustainable development.

[0060] The application system of this invention can effectively utilize various types of children's behavioral data, including scales, photos, and videos. By comprehensively employing scale screening, photo screening, and video screening, it can effectively capture the behavioral characteristics and physiological responses of children with autism from different perspectives. This overcomes the problems of traditional single-scale screening methods, such as crude methods, low accuracy, insufficient standardization, inability to comprehensively capture the complex behavioral characteristics of children with autism, numerous interfering factors, technical limitations, and insufficient reporting of psychological measurement attributes. It achieves early, efficient, accurate, and comprehensive screening for autism. Simultaneously, this invention also emphasizes user-friendly interface design and data privacy and security, ensuring easy use even for non-professionals and protecting users' health data.

[0061] In a more specific embodiment, the application provided by the present invention is centered on its user-friendly front-end interface, which is divided into four main parts: "Home", "Screening", "Information", and "My Account", as shown below. Figure 2 As shown, the homepage showcases the application's core functions, including quick access to the screening module and historical assessments, as well as informational content about autism and an application operation guide, ensuring users can quickly access any information and functions they need, thus enhancing user experience and ease of use. Figure 3 The screening interface integrates multiple screening methods, including scale screening, photo screening, and video screening, to meet the needs of different users and provide a more comprehensive assessment. For example... Figure 4 The information interface aims to provide users with the latest information, research progress, and educational strategies regarding autism, helping them better understand autism. For example... Figure 5 The "My" interface allows users to create and update a child's personal profile, including basic information and medical history. Through this comprehensive and meticulous design, the application aims to provide a holistic solution for the early identification and management of autism.

[0062] Regarding the core front-end functionality: User login, such as... Figure 6Users log in to the application by entering their username and password. First-time users need to register to create a personal account. After logging in, users will be directly taken to the autism screening homepage. Complete child information is required, such as... Figures 7-8 After logging into the app, users can add multiple children for autism screening and enjoy convenient family autism management services. First, users need to complete basic information for each child, including name, gender, date of birth, and family history of autism. This information is used for personalized screening and assessment. During subsequent screenings, users must select the added children before proceeding. In addition, the app provides optional fields such as family history of mental illness and the child's past mental illness history to collect more comprehensive health background information. In this way, the app not only provides an efficient screening tool but also establishes a comprehensive health record for each child.

[0063] Regarding scale screening. For example... Figures 9-11 The screening scales cover a variety of internationally recognized standardized autism assessment scales. In addition, taking into account the needs of different cultural and linguistic backgrounds, this application has specially customized a series of scales for Chinese children. Figure 9 The system intuitively displays two sample scales that have been uploaded. Users can choose the appropriate scale based on the child's age and other specific conditions. After completing the scale and clicking submit, the application will return the results of this assessment, which users can view in "Historical Assessments." Photo screening, such as... Figure 12 In the photo screening module, users can take a clear, unobstructed frontal photo of their child, or upload a suitable photo from their phone's album. This module specifically emphasizes the need for an unobstructed frontal photo to ensure the photo analysis model can accurately capture the child's facial features. The uploaded photo is sent to the server's photo analysis model, which uses deep learning and image processing techniques for in-depth analysis. After analysis, the model returns a quantified score representing the child's probability of having autism. Video screening, such as... Figure 13 In this module, users are guided to take photos of their child's face using a smartphone or other camera device, or upload videos of the child's face from their album. This process aims to capture a comprehensive view of the child's face. The video analysis model will deeply analyze the temporal expression changes, micro-expression features, and facial movement patterns in these videos. Through multimodal fusion, explicit spatial guidance, and efficient bidirectional temporal modeling, it effectively extracts short-term micro-expressions and long-term behavioral patterns related to ASD diagnosis from facial videos, assessing the child's autism risk.

[0064] Regarding information, such as... Figure 14In the information module, users can not only view a series of popular science articles about autism, including its symptoms, diagnosis process, treatment methods, and family support strategies, to enhance their understanding of autism, but also find detailed operation guides for the application. These guides cover how to effectively use core functions such as scale screening, photo screening, and video screening, as well as how to accurately interpret screening results and receive follow-up guidance.

[0065] Regarding the backend. For example... Figure 15 The application's backend management system is comprehensively designed, divided into multiple functional modules to meet different management needs: the scale listing and delisting management module allows administrators to dynamically update and maintain autism screening scales; the photo screening configuration module enables administrators to adjust and optimize photo screening algorithms to improve the accuracy of facial feature analysis; the video screening configuration module is used to manage and configure video analysis models to ensure the efficiency and reliability of video analysis; in addition, the backend also includes user management, information management, version management, and other modules, which together constitute a powerful backend support system.

[0066] (1) Scale Management See Figures 16-17 In the backend questionnaire management module, administrators are given flexible management capabilities, allowing them to add and remove questionnaires and edit questionnaire content. Administrators can choose to manually enter questions directly in the question field, or, for efficiency, download standardized templates, edit questions and options within the templates, and then upload them all at once. The questionnaires contain various question types, including single-choice, multiple-choice, and true / false questions. Each questionnaire not only records detailed basic information such as version number, questionnaire type, usage conditions, and questionnaire status, but also includes technical information such as the questionnaire's analysis application interface address, ensuring that the questionnaire can not only be used by users on the front end, but also seamlessly integrate with the analysis system in the back end, enabling effective data management and analysis.

[0067] (2) Photo screening configuration See Figure 18 In this section, administrators configure the photo screening function by setting the name, analysis interface application address, and function status of the photo screening. Finally, they can click the "Save" button to make the configuration effective. By adding new photo analysis configurations, modifying existing configurations, or deleting outdated configurations, the system can better connect to different photo analysis interfaces and meet diverse application needs.

[0068] (3) Video screening configuration See Figure 19Similar to photo screening configurations, administrators can also adjust and manage various video analytics settings in the video screening configuration. Administrators can add new video analytics models, modify parameters of existing models, and delete or disable configurations that are no longer applicable.

[0069] (4) User Management See Figure 20 In the user management module, administrators are granted permissions to view and review users' basic information, and to modify user status and reset passwords as needed. Furthermore, administrators can define user types, which helps to segment user groups, enabling more precise data management and service customization, and also facilitates the later export of user data for analysis.

[0070] (5) Information Management See Figures 21-23 In the information management module, information publishers can add, delete, modify, and categorize information, including its status and type. They can add new science articles and user guides, delete outdated or irrelevant information based on the latest research or user feedback, and promptly modify and update existing content. Furthermore, publishers can set information statuses, such as pushing certain information to the homepage to improve its visibility. They can also categorize information and set different types, helping users find content of interest more quickly and creating a clearer and more organized information structure.

[0071] In addition to managing news feeds, news publishers also need to manage the homepage carousel. Through a well-designed carousel, the application can showcase the latest autism-related news, upcoming health lectures and support group events, user reviews and success stories, as well as new features and improvements. These carousels not only attract user attention but also quickly convey the application's core values ​​and latest developments, allowing users to learn about the latest features and important information immediately, thus increasing user engagement and satisfaction.

[0072] (6) Administrator Management See Figure 24In the administrator management module, this application has meticulously designed a multi-layered role system, including super administrators, sub-super administrators, ordinary administrators, data administrators, and information publishers, to achieve refined permission management and responsibility allocation. Sub-super administrators have broad permissions and are primarily responsible for handling critical tasks such as version updates and administrator assignments; ordinary administrators are responsible for managing users and scales, configuring screening functions, and handling the application's daily maintenance; data administrators focus on managing sensitive data such as user data, children's information, and assessment records, ensuring data security and accuracy; and information publishers are responsible for adding, editing, and publishing information content, maintaining its updates and quality. Through this role division, various roles are assigned to different administrators, who can efficiently handle different modules of the application's backend according to their responsibilities and permissions, thereby improving the operational efficiency and professionalism of the entire backend management system.

[0073] (7) Version Management See Figure 25 When an application fixes a security vulnerability or adds / removes certain features, it needs to be updated to ensure all users receive the best user experience and security. This application employs a forced update strategy to ensure all users can promptly upgrade to the latest version. In this way, the application can continuously provide stable, reliable, and up-to-date services while protecting users from potential security threats. The forced update mechanism also helps maintain consistency in the user interface and user experience, ensuring all users enjoy the latest and most complete features.

[0074] This embodiment provides a multifunctional rapid intelligent screening platform for autism behavior, integrating multiple functions such as autism popularization and promotion, autism scale screening, autism facial photo screening, autism video screening, and back-end management. It provides a versatile, easy-to-use, accurate, and convenient early screening platform for autism.

[0075] This invention provides an autism screening platform adapted to the Chinese local context. The facial photo screening and video screening AI models used are developed based on a dataset of autism children's videos from Chinese social media. Compared with similar autism screening AI models developed abroad, the localized tool has higher accuracy. The selection of screening scales also fully considers Chinese cultural characteristics and family patterns, and designs scale tools that are more in line with local needs.

[0076] This invention belongs to the field of mobile intelligent platform technology and can be used on systems such as iOS and Android. It also involves JAVA language and database technology.

[0077] In one embodiment, the present invention has been developed and deployed to a cloud server. All three screening modules have been tested and verified on multiple volunteers. The system's backend management system and information management system have also undergone systematic testing. Figures 26-29 It can provide visual support for the display of test results and the operation of background functions in this process, indicating that the technical solution of the present invention has practical feasibility and effectiveness, and that the functions of each module and the overall operation of the system meet the design expectations.

[0078] In the description of this specification, the use of terms such as "Embodiment 1," "this embodiment," or "in one embodiment" indicates that the specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example; moreover, the specific features, structures, materials, or characteristics described may be combined in any appropriate manner in one or more embodiments or examples.

[0079] In the description of this specification, the terms "connection," "installation," "fixing," "setting," and "having" are interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0080] In the description of this specification, relational terms such as “first” and “second” are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase “comprising one…” does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0081] The above description of the embodiments is intended to enable those skilled in the art to understand and apply the technology of this invention. Those skilled in the art can easily make various modifications to these examples and apply the general principles described herein to other embodiments without creative effort. Therefore, this invention is not limited to the above embodiments. Modifications in the following situations should be within the scope of protection of this invention: ① New technical solutions implemented based on the technical solution of this invention and combined with existing common knowledge, where the technical effects of the new technical solution do not exceed the technical effects of this invention; ② Equivalent substitutions of some features of the technical solution of this invention using known technology, resulting in the same technical effects as those of this invention; ③ Extendable technical solutions based on the technical solution of this invention, where the substantive content of the extended technical solution does not exceed the technical solution of this invention; ④ Equivalent transformations made using the content of this specification and drawings, directly or indirectly applied to other related technical fields.

Claims

1. An application system for screening autism spectrum disorder, characterized in that, It includes a front-end user interaction device and a back-end data processing server, wherein the front-end user interaction device and the back-end data processing server establish a communication connection through a network; The front-end user interaction device includes: The child information management module is used to receive and store information on at least one child input by the user, and the child information is used for subsequent personalized screening and assessment. The scale screening module is used to provide at least one autism screening scale for users to select and fill in, and to generate a scale screening score after receiving the scale answers filled in by the users; The photo screening module is used to receive at least one frontal facial image of a child uploaded by the user, preprocess the facial image, and then call the photo analysis model to generate a photo screening score. The screening process of the front-end user interaction device is as follows: After the user selects the target child through the child information management module, the user completes the operation of the scale screening module and the photo screening module in sequence. The back-end data processing server uses a weighted voting mechanism to generate a comprehensive score and risk level based on the scale screening score and the photo screening score.

2. The system according to claim 1, characterized in that, The autism screening scales provided by the scale screening module include at least the self-developed Chinese Children's Scale, which covers children aged 2 to 12 years, with one specific scale for each year of age. Children aged 13 to 18 years are adapted to the 12-year-old specific scale. Each of the self-developed Chinese Children's Scales contains 10 to 15 questions, each with a different weight value. The total score ranges from 0 to 100 points, with scores above 80 points indicating high risk, 60 to 79 points indicating medium risk, and 0 to 59 points indicating low risk.

3. The system according to claim 2, characterized in that, The self-developed Chinese Children's Scale is based on the DeepSeek-R1-Distill-Qwen-14B base model and undergoes supervised fine-tuning through multiple types of data. These multiple types of data include one or more of the following: academic papers, medical textbooks, various mental illness screening scales, or Chinese children's medical consultation dialogue datasets.

4. The system according to any one of claims 1-3, characterized in that, The preprocessing step of the photo screening module is to uniformly process the facial images uploaded by users to a standard size; The architecture of the photo analysis model includes: The feature extraction layer receives a standard-sized facial image, extracts the feature information of the standard-sized facial image, and outputs an image feature map; The spatial attention module is communicatively connected to the feature extraction layer, receives the image feature map, enhances attention to key regions in the image feature map, and outputs spatial attention features. The facial landmark detection layer receives a standard-sized facial image, detects landmarks in key areas of a child's face within the standard-sized facial image, encodes the landmarks, and outputs a feature vector. The global-local cross-fusion relationship converter encoder is communicatively connected to the spatial attention module and the facial landmark detection layer, respectively. It receives the spatial attention features and the feature vector, and performs global-local feature collaborative merging on the spatial attention features and the feature vector through a preset fusion mechanism to output dual-stream fused features. A multi-feature interaction network is communicatively connected to the global-local cross-fusion relationship converter encoder, receives the dual-stream fusion features, performs attention calculation on the dual-stream fusion features to filter effective features related to autism spectrum disorder identification, and outputs comprehensive features; The classification layer is connected to the multi-feature interaction network, receives the comprehensive features, classifies the comprehensive features into autism spectrum disorder and typical development, and generates and outputs the photo screening score.

5. The system according to claim 4, characterized in that, The standard size of the facial image is 224×224 pixels; the deep learning network used in the feature extraction layer is a deep separable convolutional neural network, a 50-layer residual network, or a lightweight convolutional neural network; the size of the image feature map output by the feature extraction layer is 256×7×7.

6. The system according to claim 5, characterized in that, The attention calculation process of the multi-feature interaction network includes: calculating the contribution weight of each feature dimension in the dual-stream fusion feature to the identification of autism spectrum disorder; dynamically retaining high-contribution features and filtering low-contribution features according to the contribution weight; and simultaneously establishing information interaction channels between different feature dimensions in the dual-stream fusion feature, integrating cross-dimensional feature information, and generating the comprehensive feature.

7. The system according to claim 6, characterized in that, The training and validation process of the photo analysis model includes: training the model based on a large open-source autism facial image dataset; and validating the trained model using facial photos extracted from a Chinese social media autism children video dataset.

8. The system according to claim 7, characterized in that, The facial landmark detection layer uses a facial feature point detection network to detect facial landmarks; the landmarks of the key facial regions include contour points and feature points corresponding to the eyes, nose, and mouth; the feature vector output by the facial landmark detection layer has a size of 256×7×7.

9. The system according to any one of claims 5-8, characterized in that, In the weighted voting mechanism, the weight ratio of the scale screening score to the photo screening score is 1:

1.

10. The system according to any one of claims 1-3 or 5-8, characterized in that, The front-end user interaction device also includes: The video screening module is used to receive at least one 10-30 second video of a child's face uploaded by the user, and after preprocessing the facial video, call the video analysis model to generate a video screening score. If the combined risk level of the scale screening score and the photo screening score is high, the user can choose to activate the video screening module. The background data processing server will then combine the combined score and the video screening score to generate the final screening result.

Images