An autism early screening rehabilitation training system based on AI intelligent identification

Abstract

The application discloses an autism early screening rehabilitation training system based on AI intelligent identification, and particularly relates to the technical field of medical rehabilitation; the system is characterized in that: a time sequence alignment is performed on continuous interactive videos, continuous interactive voices and task operation records of children, a multi-round interactive behavior sequence is formed, and a corresponding task exposure history sequence is constructed; a gaze change, a head turning, a facial response, a voice response and an operation time delay are extracted to form an original social attention feature sequence; the original social attention feature sequence is sequentially matched with the task exposure history sequence, a task-induced adaptation component is separated, and feature correction is completed; an autism risk auxiliary screening value is generated based on the corrected social attention feature sequence, and a rehabilitation training task is determined in combination with the task exposure history sequence.

Description

Technical Field

[0001] This invention relates to the field of medical rehabilitation technology, and more specifically, to an AI-powered intelligent recognition system for early screening and rehabilitation training of autism. Background Technology

[0002] Existing methods for autism risk screening and rehabilitation training typically involve presenting stimulating content, collecting video, audio, and operational records of children during interactions, and extracting behavioral characteristics from these records to form risk screening results or adjust training content.

[0003] During continuous interaction, the same or similar stimuli may be presented multiple times, and children will also receive rewards, prompts, or retry feedback. Children's familiarity with the task rules, stimuli, and feedback methods will affect their subsequent behavioral performance. Existing methods usually directly compare changes in behavioral characteristics across different rounds of interaction, making it difficult to distinguish between adaptive changes caused by task repetition, feedback experience, and task familiarity, and changes in children's social attention. This results in a lack of a consistent basis for comparing behavioral characteristics across rounds. Summary of the Invention

[0004] In order to overcome the above-mentioned deficiencies of the prior art, embodiments of the present invention provide an AI-intelligent recognition-based early screening and rehabilitation training system for autism to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] An AI-powered early screening and rehabilitation training system for autism, comprising:

[0007] Interactive acquisition module: Acquires continuous interactive videos, continuous interactive voices, and task operation records of children, and aligns them according to the timing of stimulus presentation and feedback triggering to form a multi-round interactive behavior sequence;

[0008] Exposure building module: Based on stimulus identifiers, number of stimulus repetitions, intervals between adjacent stimuli, and feedback trigger records, construct a task exposure history sequence corresponding to a multi-round interaction behavior sequence;

[0009] Feature extraction module: Extracts gaze changes, head turning, facial responses, voice responses, and operation delays from multi-turn interaction behavior sequences to form the original social attention feature sequence;

[0010] Adaptive Separation Module: Performs sequence registration between the original social attention feature sequence and the task exposure history sequence, and separates the task-induced adaptation component that changes synchronously with the task exposure history sequence;

[0011] Feature correction module: Corrects the original social attention feature sequence based on the task-induced adaptation component to obtain a corrected social attention feature sequence that can be compared across rounds;

[0012] Screening and training decision module: Generates autism risk auxiliary screening value based on the corrected social attention feature sequence, and determines rehabilitation training tasks based on the autism risk auxiliary screening value and the task exposure history sequence.

[0013] In a preferred embodiment, continuous interactive video, continuous interactive voice, and task operation records of the child are acquired, and then aligned temporally according to the stimulus presentation time and feedback trigger time to form a multi-round interactive behavior sequence, specifically:

[0014] During the execution of the interactive task, the child's continuous interactive video, continuous interactive voice and task operation record are acquired simultaneously, and the start acquisition time, stimulus presentation time and feedback trigger time in the child's continuous interactive video, continuous interactive voice and task operation record are marked respectively.

[0015] The children's continuous interactive videos, continuous interactive voices, and task operation records are aligned chronologically according to the timing of stimulus presentation and feedback triggering to form a multi-round interactive behavior sequence.

[0016] In a preferred embodiment, a task exposure history sequence corresponding to a multi-round interaction behavior sequence is constructed based on stimulus identifiers, stimulus repetition counts, adjacent stimulus intervals, and feedback trigger records. Specifically:

[0017] Read task operation records aligned with stimulus presentation times from multi-round interaction behavior sequences, and extract stimulus identifiers corresponding to stimulus presentation times and feedback trigger records corresponding to feedback trigger times;

[0018] The number of stimulus repetitions is counted according to the order of appearance of stimulus labels in the multi-round interactive behavior sequence, and the interval between adjacent stimuli is determined based on the presentation time of two adjacent stimuli.

[0019] Stimulus identifiers, number of stimulus repetitions, intervals between adjacent stimuli, and feedback trigger records are written into the same timeline according to the timing of stimulus presentation, thus constructing a task exposure history sequence corresponding to the multi-round interaction behavior sequence.

[0020] In a preferred embodiment, gaze changes, head turning, facial responses, voice responses, and operation delays are extracted from the multi-turn interaction behavior sequence to form the original social attention feature sequence, specifically:

[0021] Interaction segments are divided using the stimulus presentation time in a multi-round interactive behavior sequence as the boundary.

[0022] Extract gaze changes, head turning, and facial responses from each interactive segment of a child's continuous interactive video;

[0023] Extracting voice responses from each interactive segment of a child's continuous interactive speech;

[0024] Extract the operation latency within each interaction segment from the task operation record;

[0025] The original social attention feature sequence is formed by arranging the interaction segments in the sequence of multiple rounds of interaction behavior.

[0026] In a preferred embodiment, the original social attention feature sequence is sequence registered with the task exposure history sequence to separate the task-induced adaptation component that changes synchronously with the task exposure history sequence, specifically:

[0027] The original social attention feature sequence and the task exposure history sequence are registered along the same time axis to obtain the social attention features and task exposure records corresponding to each interaction segment.

[0028] Based on the changes in the number of stimulus repetitions, the interval between adjacent stimuli, and the direction of change in feedback trigger records, we screened the changes in social attention features that were synchronous with the task exposure history sequence, and used the changes in social attention features as the task-induced adaptation component.

[0029] In a preferred embodiment, the original social attention feature sequence is corrected based on the task-induced adaptation component to obtain a corrected social attention feature sequence that can be compared across rounds, specifically as follows:

[0030] Based on the correspondence between the original social attention feature sequence and the task-induced adaptation component on the same time axis, the feature terms to be corrected for gaze changes, head turning, facial response, voice response and operation delay in each interaction segment are determined.

[0031] Subtract the corresponding task-induced adaptation component from the feature terms to be corrected for each interaction segment, and write the subtracted feature terms to be corrected into the corresponding positions in the original social attention feature sequence to obtain a cross-round comparable corrected social attention feature sequence.

[0032] In a preferred embodiment, an autism risk auxiliary screening value is generated based on a corrected social attention feature sequence, and a rehabilitation training task is determined based on the autism risk auxiliary screening value and the task exposure history sequence, specifically as follows:

[0033] The gaze changes, head turns, facial responses, voice responses, and operation delays corresponding to each interaction segment in the corrected social attention feature sequence are compared with the preset screening feature range to generate autism risk auxiliary screening values.

[0034] Stimulus identifiers, number of repetitions, intervals between adjacent stimuli, and feedback trigger records are retrieved from the task exposure history sequence. Combined with autism risk screening values, the order of stimulus presentation, task difficulty, and feedback triggering methods are determined to identify rehabilitation training tasks.

[0035] The technical effects and advantages of the AI-powered intelligent recognition autism early screening and rehabilitation training system of this invention are as follows:

[0036] By temporally aligning children's continuous interactive videos, continuous interactive voice recordings, and task operation records, a correspondence can be established between behavioral data from different sources at the moments of stimulus presentation and feedback triggering. By constructing a task exposure history sequence, the number of stimulus repetitions, the interval between adjacent stimuli, and feedback triggering records are correlated with multi-round interactive behavior sequences. Gazing changes, head turning, facial responses, voice responses, and operation delays are extracted from these sequences to form a continuously comparable original social attention feature sequence. Through sequence registration, task-evoked adaptation components that change synchronously with the task exposure history are separated, and these components are used to correct the original social attention feature sequence, reducing interference from repetitive stimuli, feedback reinforcement, and task familiarization on cross-round behavioral representations. Based on the corrected social attention feature sequence, autism risk auxiliary screening values ​​are generated, and rehabilitation training tasks are determined in conjunction with the task exposure history sequence, achieving a continuous connection between auxiliary screening data and training task adjustments. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the structure of an AI-powered intelligent recognition system for early screening and rehabilitation training of autism according to the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0039] Example 1

[0040] Figure 1 This invention discloses an AI-powered intelligent recognition system for early screening and rehabilitation training of autism, comprising:

[0041] Interactive acquisition module: Acquires continuous interactive videos, continuous interactive voices, and task operation records of children, and aligns them according to the timing of stimulus presentation and feedback triggering to form a multi-round interactive behavior sequence;

[0042] Exposure building module: Based on stimulus identifiers, number of stimulus repetitions, intervals between adjacent stimuli, and feedback trigger records, construct a task exposure history sequence corresponding to a multi-round interaction behavior sequence;

[0043] Feature extraction module: Extracts gaze changes, head turning, facial responses, voice responses, and operation delays from multi-turn interaction behavior sequences to form the original social attention feature sequence;

[0044] Adaptive Separation Module: Performs sequence registration between the original social attention feature sequence and the task exposure history sequence, and separates the task-induced adaptation component that changes synchronously with the task exposure history sequence;

[0045] Feature correction module: Corrects the original social attention feature sequence based on the task-induced adaptation component to obtain a corrected social attention feature sequence that can be compared across rounds;

[0046] Screening and training decision module: Generates autism risk auxiliary screening value based on the corrected social attention feature sequence, and determines rehabilitation training tasks based on the autism risk auxiliary screening value and the task exposure history sequence.

[0047] Acquire continuous interactive videos, continuous interactive voice recordings, and task operation records of children, and align them chronologically according to the stimulus presentation time and feedback trigger time to form a multi-turn interactive behavior sequence, including:

[0048] During the execution of the interactive task, the child's continuous interactive video, continuous interactive voice and task operation record are acquired simultaneously, and the start acquisition time, stimulus presentation time and feedback trigger time in the child's continuous interactive video, continuous interactive voice and task operation record are marked respectively.

[0049] Specifically, the interactive task execution process refers to the continuous process in which a child receives stimulus content, generates a gaze response, generates a speech response, generates an operational response, and receives feedback content within the same interactive flow. Continuous interactive video for children refers to a set of image frames continuously captured from the start to the end of the interactive task, including at least the child's facial area, the child's head area, and the relative positional changes between the child and the stimulus content. Continuous interactive speech for children refers to audio sampling data continuously captured from the start to the end of the interactive task, including at least the child's speech responses and prompts in the interactive task environment. Task operation records refer to event records generated and written during the interactive task execution process, including events such as the presentation of stimulus content, events such as the child touching or clicking, events such as the child completing or interrupting the task, and events such as the triggering of feedback content.

[0050] The continuous interactive video, continuous interactive voice, and task operation records for children all use the same timing benchmark to mark time information. This benchmark is established from the zero moment generated at the start of the interactive task. Each frame in the continuous interactive video corresponds to the frame acquisition time, each audio sample in the continuous interactive voice corresponds to the audio sampling time, and each event in the task operation record corresponds to the event occurrence time.

[0051] When the interactive task enters the executable state, the capture of continuous interactive video and voice recordings of the child, as well as the writing of task operation records, are initiated simultaneously. The capture time of the first frame of the continuous interactive video, the sampling time of the first audio segment of the continuous interactive voice recording, and the occurrence time of the task start event in the task operation record are recorded. The initial capture time is used to define the starting point for the continuous interactive video, continuous interactive voice recording, and task operation record to enter the same timeline. When the capture time of the first frame of the continuous interactive video, the sampling time of the first audio segment of the continuous interactive voice recording, and the occurrence time of the task start event in the task operation record are consistent, these three times are collectively marked as the initial capture time.

[0052] The children's continuous interactive videos, continuous interactive voices, and task operation records are aligned in time according to the timing of stimulus presentation and feedback triggering to form a multi-round interactive behavior sequence.

[0053] Specifically, the stimulus presentation moment refers to the point in time when the stimulus content begins to be perceptibly presented to the child. The stimulus content includes, but is not limited to, image stimuli, video stimuli, sound stimuli, facial stimuli, toy image stimuli, and command voice stimuli. The stimulus presentation moment is determined by the time of occurrence of the event in the task operation record where the stimulus content begins to be presented. When the stimulus content includes both visual and auditory presentation, the stimulus presentation moment is determined by the earlier of the visual presentation event and the auditory playback event. The feedback trigger moment refers to the point in time when the feedback content begins to be triggered. The feedback content includes, but is not limited to, reward sounds, reward animations, prompts, task progression prompts, and task retry prompts. The feedback trigger moment is determined by the time of occurrence of the feedback content triggering event in the task operation record. When the feedback content is triggered by the child's operation result, the feedback trigger moment is determined by the time of occurrence of the feedback content triggering event in the task operation record.

[0054] Temporal alignment refers to uniformly converting the frame acquisition time of each frame in a child's continuous interactive video, the audio sampling time of each audio sample in a child's continuous interactive speech, and the event occurrence time of each event in the task operation record into relative time compared to the initial acquisition time, and then arranging the child's continuous interactive video, child's continuous interactive speech, and task operation record according to relative time. During temporal alignment, the stimulus presentation time serves as the time anchor point for the start of stimulus presentation in the interactive task, and the feedback trigger time serves as the time anchor point for the start of feedback triggering in the interactive task. Video frames located between two adjacent stimulus presentation times in the child's continuous interactive video, audio sample data located between two adjacent stimulus presentation times in the child's continuous interactive speech, and event records located between two adjacent stimulus presentation times in the task operation record are grouped into the same interactive segment. The feedback trigger time is written into the interactive segment containing the feedback trigger time according to relative time.

[0055] The multi-turn interactive behavior sequence consists of interactive segments arranged chronologically. Each segment corresponds to a stimulus presentation moment and includes a continuous interactive video segment, a continuous interactive speech segment, a task operation record segment, and a feedback trigger moment corresponding to the stimulus presentation moment. The continuous interactive video segment consists of video frames between two adjacent stimulus presentation moments; after the last stimulus presentation moment, the continuous interactive video segment consists of video frames from the last stimulus presentation moment to the end of the interactive task. The continuous interactive speech segment consists of audio sample data between two adjacent stimulus presentation moments; after the last stimulus presentation moment, the continuous interactive speech segment consists of audio sample data from the last stimulus presentation moment to the end of the interactive task. The task operation record segment consists of event records between two adjacent stimulus presentation moments; after the last stimulus presentation moment, the task operation record segment consists of event records from the last stimulus presentation moment to the end of the interactive task. If an interactive segment contains multiple feedback trigger moments, these multiple feedback trigger moments are written to the corresponding interactive segment in chronological order; if an interactive segment does not contain any feedback trigger moments, the feedback trigger moment field in the corresponding interactive segment is marked with a null value. The resulting multi-round interactive behavior sequence uses interactive segments as the basic arrangement objects, and retains the stimulus presentation time, feedback trigger time, continuous interactive video segments of children, continuous interactive voice segments of children, and task operation record segments within each interactive segment.

[0056] Based on stimulus identifiers, stimulus repetition counts, intervals between adjacent stimuli, and feedback trigger records, a task exposure history sequence corresponding to a multi-round interaction behavior sequence is constructed, including:

[0057] Read task operation records aligned with stimulus presentation times from multi-round interaction behavior sequences, and extract stimulus identifiers corresponding to stimulus presentation times and feedback trigger records corresponding to feedback trigger times;

[0058] Specifically, reading task operation records aligned with stimulus presentation times refers to reading task operation record segments within each interaction segment according to the order of interaction segments in a multi-round interactive behavior sequence, and identifying event records related to the start of stimulus presentation and the start of feedback triggering from these task operation record segments. The event records in the task operation record segments are arranged relative to the initial acquisition time, so each event record corresponds to an interaction segment in the multi-round interactive behavior sequence.

[0059] The stimulus identifier corresponding to the stimulus presentation time is extracted from the event record of the stimulus content's initial presentation in the task operation log. The stimulus identifier is used to distinguish the stimulus content presented to the child during the interactive task execution. The stimulus identifier can be formed by both stimulus type information and stimulus content information. Stimulus type information distinguishes whether the stimulus content belongs to an image stimulus, video stimulus, sound stimulus, human face stimulus, toy image stimulus, or command voice stimulus. Stimulus content information distinguishes different stimulus contents under the same stimulus type information. The stimulus identifier can be formed using characters, numbers, or a combination of characters and numbers. For example, a human face stimulus can be set as a combined identifier containing the human face category and human face number; a command voice stimulus can be set as a combined identifier containing the command voice category and command voice number. The event record of the stimulus content's initial presentation in the task operation log includes the stimulus presentation time and the stimulus identifier. When reading the task operation log, the stimulus presentation time and the stimulus identifier are written as the same stimulus presentation record into the intermediate record of the task operation log.

[0060] The feedback trigger record corresponding to the feedback trigger moment is extracted from the event record corresponding to the time the feedback content trigger event occurs in the task operation record. The feedback trigger record indicates whether the child received feedback content within the interactive segment, and includes at least the feedback trigger moment and the feedback content type. The feedback content type is used to distinguish between reward sounds, reward animations, prompt voices, task advancement prompts, or task retry prompts. If multiple feedback trigger moments exist within the same interactive segment, multiple feedback trigger records are generated according to the chronological order of the multiple feedback trigger moments, and all multiple feedback trigger records are written to the corresponding position within the same interactive segment. If no feedback trigger moment exists within the same interactive segment, a null value marker is written to the corresponding position within the same interactive segment. The null value marker indicates that no feedback content was triggered within the same interactive segment.

[0061] The number of stimulus repetitions is counted according to the order of appearance of stimulus labels in the multi-round interactive behavior sequence, and the interval between adjacent stimuli is determined based on the presentation time of two adjacent stimuli.

[0062] Specifically, the number of stimulus repetitions is counted based on the sequential appearance of stimulus markers in the multi-round interaction sequence. When counting stimulus repetitions, the first interaction segment in the multi-round interaction sequence is used as the starting point, and the stimulus markers corresponding to each interaction segment are read sequentially according to their order in the sequence. For the currently read stimulus marker, the number of times it has appeared in previous interaction segments is retrieved, and this number is incremented by 1 to determine the stimulus repetition count for the current interaction segment. If the currently read stimulus marker has not appeared in previous interaction segments, the stimulus repetition count for the current interaction segment is recorded as 1. Through this statistical method, each appearance of the same stimulus marker in the multi-round interaction sequence receives an increasing stimulus repetition count, and different stimulus markers are counted independently.

[0063] The adjacent stimulus interval is determined based on the presentation times of two adjacent stimuli. Two adjacent stimulus presentation times refer to the stimulus presentation times corresponding to two temporally adjacent interaction segments in a multi-turn interaction sequence. To determine the adjacent stimulus interval, the stimulus presentation times corresponding to the current interaction segment and the stimulus presentation times corresponding to the adjacent interaction segments preceding the current interaction segment are read, and the time difference between these two times is calculated as the adjacent stimulus interval for the current interaction segment. For the first interaction segment in a multi-turn interaction sequence, since there are no preceding adjacent interaction segments, the adjacent stimulus interval for the first interaction segment is marked as null. The null mark indicates that there is no calculable preceding adjacent stimulus interval for the first interaction segment.

[0064] Stimulus identifiers, number of stimulus repetitions, intervals between adjacent stimuli, and feedback trigger records are written into the same timeline according to the timing of stimulus presentation, thus constructing a task exposure history sequence corresponding to the multi-round interaction behavior sequence.

[0065] Specifically, the same timeline uses the initial acquisition time as time zero, and each stimulus presentation time as a time anchor in the task exposure history sequence. Each time anchor corresponds to one task exposure record, and each task exposure record includes the corresponding stimulus presentation time, the corresponding stimulus identifier, the corresponding number of stimulus repetitions, the corresponding interval between adjacent stimuli, and the corresponding feedback trigger record. When writing to the same timeline, all task exposure records are arranged from earliest to latest according to the stimulus presentation time, and then a one-to-one correspondence is maintained between the task exposure records and the interaction segments in the multi-round interaction behavior sequence.

[0066] The task exposure history sequence consists of task exposure records arranged according to the timing of stimulus presentation. The task exposure history sequence corresponds to the multi-turn interaction behavior sequence, meaning that each task exposure record in the task exposure history sequence corresponds to an interaction segment in the multi-turn interaction behavior sequence, and the order of the task exposure records in the task exposure history sequence is consistent with the order of the interaction segments in the multi-turn interaction behavior sequence.

[0067] The sequence of multi-turn interaction behaviors is used to extract gaze changes, head turning, facial responses, voice responses, and operation delays to form the original social attention feature sequence, including:

[0068] Interaction segments are divided using the stimulus presentation time in a multi-round interactive behavior sequence as the boundary.

[0069] Specifically, dividing interaction segments based on the stimulus presentation times in a multi-turn interaction sequence means determining the start and end points of each interaction segment sequentially according to the stimulus presentation times already written in the multi-turn interaction sequence. The start point of each interaction segment is the corresponding stimulus presentation time, and the end point of each interaction segment is the next stimulus presentation time that is temporally adjacent; for the last interaction segment, the end point is the end time of the interaction task. After completing the segment division, each interaction segment includes a continuous interactive video segment, a continuous interactive voice segment, and a task operation record segment. Each interaction segment also corresponds to a task exposure record in the task exposure history sequence.

[0070] Extract gaze changes, head turning, and facial responses from each interactive segment of a child's continuous interactive video;

[0071] Specifically, when extracting gaze changes within each interactive segment of a child's continuous interactive video, the continuous interactive video segment of the child is read according to the frame acquisition time. The child's facial region, eye region, and stimulus content presentation area are determined in each frame of the continuous interactive video segment. The child's facial region is determined by the positional relationship of the facial contour, eyes, nose, and mouth in the image. The eye region is determined by the position of the left and right eyes within the child's facial region. The stimulus content presentation area is determined by the position of the stimulus content in the task operation recording segment. Gazing changes are used to represent changes in the child's gaze relative to the stimulus content presentation area. Gazing changes are determined by the relative relationship between the center position of the child's eye region, the opening and closing state of the eye region, the orientation of the child's facial region, and the position of the stimulus content presentation area. If the child's gaze moves from outside the stimulus content presentation area to within the stimulus content presentation area in a continuous image frame, a gaze entry change is recorded; if the child's gaze moves from within the stimulus content presentation area to outside the stimulus content presentation area in a continuous image frame, a gaze exit change is recorded; if the child's gaze remains within the stimulus content presentation area in a continuous image frame, the gaze duration is accumulated according to the frame acquisition time.

[0072] When extracting head turning within each interactive segment of a child's continuous interactive video, the child's head orientation is determined based on the geometric relationships between the positions of the left and right eyes, nose, and mouth in the child's facial region. Head turning represents the deflection state and changes in the child's head relative to the direction of the stimulus content. Head turning is determined jointly by the lateral offset, longitudinal offset, and changes in the area of ​​the child's facial region. If the center position of the child's facial region continuously shifts relative to the center position of the stimulus content's presentation area, the head turning direction is recorded according to the direction of the offset; if the area of ​​the child's facial region continuously decreases within the interactive segment, the change in the child's head moving away from the direction of the stimulus content's presentation is recorded; if the area of ​​the child's facial region continuously increases within the interactive segment, the change in the child's head moving closer to the direction of the stimulus content's presentation is recorded.

[0073] Facial responses are used to represent changes in facial expressions or mouth states that occur in a child's facial area after the stimulus is presented. Facial responses are determined by changes in eye opening and closing, eyebrow position, and mouth opening and closing, and are written into the interaction segment according to the time of the first appearance of the facial response and the duration of the facial response.

[0074] Extracting voice responses from each interactive segment of a child's continuous interactive speech;

[0075] Specifically, the continuous interactive speech segments of the child within each interactive segment are read according to the audio sampling time, and then divided into continuous audio windows. The length of the continuous audio window is determined based on the sampling rate of the child's continuous interactive speech and the shortest identifiable duration of the speech response in the interactive task; for example, the continuous audio window can be set to 200 milliseconds to 500 milliseconds. Speech responses represent detectable vocal behaviors produced by the child after the stimulus presentation time. When determining a speech response, the volume energy and zero-crossing changes within the continuous audio window are calculated, and the silent audio window before the stimulus presentation time is read. An environmental volume energy baseline range is formed based on the average volume energy and volume energy fluctuation range within the silent audio window. If the volume energy within the continuous audio window is higher than the environmental volume energy baseline range, and the continuous audio window is located after the stimulus presentation time, then the continuous audio window is marked as a candidate speech response window. If the interval between adjacent candidate speech response windows does not exceed a preset merging interval, then adjacent candidate speech response windows are merged into the same speech response. The preset merging interval is determined based on the allowed pause time of the speech response in the interactive task; for example, it can be set to 300 milliseconds to 800 milliseconds.

[0076] Extract the operation latency within each interaction segment from the task operation record;

[0077] Specifically, the task operation record segment within each interactive segment is read, and events indicating that the child touches or clicks, or completes or interrupts the task, are identified within the task operation record segment. Operation delay represents the time difference between the stimulus presentation moment and the child's first occurrence of the target operation. The target operation is determined according to the task requirements of the interactive task and includes, but is not limited to, touching the stimulus content presentation area, clicking a specified graphic, selecting a graphic corresponding to a specified sound, and dragging a specified graphic to the target position. For an interactive segment, if an event corresponding to the target operation exists in the task operation record segment, the timing of the event corresponding to the target operation is subtracted from the stimulus presentation moment of the interactive segment to obtain the operation delay corresponding to the interactive segment. If no event corresponding to the target operation exists in the task operation record segment, the operation delay corresponding to the interactive segment is written to a null value marker, which indicates that no target operation was detected within the interactive segment.

[0078] The original social attention feature sequence is formed according to the chronological order of the interaction segments in the multi-round interaction behavior sequence;

[0079] Specifically, the original social attention feature sequence consists of multiple original social attention feature records, each corresponding to an interaction segment. Each original social attention feature record includes the stimulus presentation time corresponding to the interaction segment, the gaze change corresponding to the interaction segment, the head turn corresponding to the interaction segment, the facial response corresponding to the interaction segment, the voice response corresponding to the interaction segment, and the operation delay corresponding to the interaction segment. The arrangement order of the original social attention feature sequence is consistent with the arrangement order of the interaction segments in the multi-turn interaction behavior sequence, and each original social attention feature record in the original social attention feature sequence maintains the same stimulus presentation time correspondence with a task exposure record in the task exposure history sequence.

[0080] Sequence registration is performed between the original social attention feature sequence and the task exposure history sequence to separate the task-induced adaptation component that changes synchronously with the task exposure history sequence, including:

[0081] The original social attention feature sequence and the task exposure history sequence are registered along the same time axis to obtain the social attention features and task exposure records corresponding to each interaction segment.

[0082] Specifically, sequence registration uses the stimulus presentation time as a common index, and pairs the original social attention feature records and task exposure records with the same stimulus presentation time into the same registration record, forming a set of registration records arranged according to the stimulus presentation time.

[0083] During sequence registration, one original social attention feature record from the original social attention feature sequence is read, and a task exposure record with the same stimulus presentation time is searched in the task exposure history sequence. If a task exposure record with the same stimulus presentation time is found, the gaze change, head turn, facial response, voice response, and operation delay from the original social attention feature record are written into the same registration record along with the stimulus identifier, stimulus repetition count, adjacent stimulus interval, and feedback trigger record from the task exposure record. If no task exposure record with the same stimulus presentation time exists in the task exposure history sequence, no corresponding registration record is generated, and the original social attention feature record is retained as an unregistered record. If a task exposure record with a null feedback trigger record exists in the task exposure history sequence, the null feedback trigger record is written as the feedback trigger record content of the corresponding registration record, and the corresponding registration record is not deleted.

[0084] The social attention features and task exposure records corresponding to each interaction segment are determined by the registration record set. The social attention features in each registration record include gaze changes, head turning, facial responses, vocal responses, and operational delays. The task exposure records in each registration record include stimulus identifiers, stimulus repetition counts, intervals between adjacent stimuli, and feedback trigger records. The registration record set is arranged from earliest to latest stimulus presentation time, allowing adjacent registration records to be used to calculate changes in social attention features and task exposure records. Changes in social attention features refer to the differences between the same type of social attention features in adjacent registration records; these differences can manifest as increases, decreases, appearance, disappearance, replacement, or no change in state. Changes in task exposure records refer to the differences between the same type of task exposure records in adjacent registration records; the direction of these changes characterizes the trend of the task exposure history sequence between adjacent registration records.

[0085] Based on the changes in the number of stimulus repetitions, the interval between adjacent stimuli, and the direction of change in feedback triggering records, we screened the changes in social attention features that were synchronous with the task exposure history sequence and used the changes in social attention features as the task-induced adaptation component.

[0086] Specifically, the direction of change in the number of stimulus repetitions is determined by comparing the number of stimulus repetitions in two adjacent registration records. When the number of stimulus repetitions in the current registration record is greater than the number of stimulus repetitions in the previous adjacent registration record, the direction of change is increasing; when the number of stimulus repetitions in the current registration record is equal to the number of stimulus repetitions in the previous adjacent registration record, the direction of change is unchanged; when the number of stimulus repetitions in the current registration record is less than the number of stimulus repetitions in the previous adjacent registration record, the direction of change is resetting. The direction of change in the interval between adjacent stimuli is determined by comparing the intervals between adjacent stimuli in two adjacent registration records. When the interval between adjacent stimuli in the current registration record is greater than the interval between adjacent adjacent registration records, the direction of change is lengthening; when the interval between adjacent stimuli in the current registration record is less than the interval between adjacent adjacent registration records, the direction of change is shortening; when the interval between adjacent stimuli in the current registration record is equal to the interval between adjacent adjacent registration records, the direction of change is unchanged. The direction of change of the feedback trigger record is determined by comparing the feedback trigger records in two adjacent registration records. When the feedback trigger record changes from a null value marker to a non-null value marker, the direction of change of the feedback trigger record is "appearance"; when the feedback trigger record changes from a non-null value marker to a null value marker, the direction of change of the feedback trigger record is "disappearance"; when the feedback content type in the feedback trigger record changes, the direction of change of the feedback trigger record is "replacement"; when the feedback content type in the feedback trigger record does not change, the direction of change of the feedback trigger record is "unchanged".

[0087] When screening for changes in social attention features that are synchronous with the task exposure history sequence, two adjacent registration records are used as the screening objects, and the changes in social attention features such as gaze change, head turning, facial response, speech response, and operation delay between the two adjacent registration records are calculated separately. For numerical forms of gaze duration, facial response duration, speech response duration, and operation delay, the changes in social attention features are obtained by subtracting the values ​​in the previous adjacent registration records from the values ​​in the current registration record. For state forms of gaze entry change, gaze exit change, head turning direction, and feedback content type, the state difference results are obtained by comparing the state in the current registration record with the state in the previous adjacent registration records. If the changes in social attention features occur in the same pair of adjacent registration records where the stimulus repetition count, the interval between adjacent stimuli, or the feedback trigger record changes in time sequence, then the changes in social attention features are recorded as candidate synchronous changes.

[0088] Candidate synchronization changes need to be categorized based on changes in task exposure records. If the number of stimulus repetitions increases in the adjacent registration pair corresponding to a candidate synchronization change, the candidate synchronization change is classified as a repetitive stimulus-related change. If the interval between adjacent stimuli increases or decreases in the adjacent registration pair corresponding to a candidate synchronization change, the candidate synchronization change is classified as a stimulus interval-related change. If a feedback trigger record appears, disappears, or is replaced in the adjacent registration pair corresponding to a candidate synchronization change, the candidate synchronization change is classified as a feedback trigger-related change. If the same candidate synchronization change corresponds to at least two of the repetitive stimulus-related changes, stimulus interval-related changes, and feedback trigger-related changes, then the same candidate synchronization change is saved according to its corresponding category, and the corresponding stimulus presentation time and stimulus identifier are retained.

[0089] After classifying candidate synchronization changes, consistency screening is performed on consecutive candidate synchronization changes under the same stimulus identifier. Consistency screening refers to reading the direction of change of the same type of social attention feature changes in consecutive registration records corresponding to the same stimulus identifier, and determining whether the direction of change of the same type of social attention feature changes continuously corresponds to the direction of change of the same type of task exposure record changes. Continuous correspondence means that in at least two adjacent registration record pairs under the same stimulus identifier, the direction of change of the same type of social attention feature changes corresponds to the direction of change of the same type of task exposure record changes. If, under the same stimulus identifier, the direction of change of gaze duration, head turning direction, facial response duration, voice response duration, or operation delay continuously corresponds to an increase in the number of stimulus repetitions, a change in the interval between adjacent stimuli, or a change in feedback triggering records, then the corresponding social attention feature change is determined to be a social attention feature change that changes synchronously with the task exposure history sequence. If the same type of social attention feature change appears only in a single interaction segment and does not form a continuous correspondence in adjacent interaction segments, then the corresponding social attention feature change is not determined to be a social attention feature change that changes synchronously with the task exposure history sequence.

[0090] The task-evoked adaptation component is formed by changes in social attention features that synchronize with the task exposure history sequence. The task-evoked adaptation component is written into the corresponding registration record according to the stimulus presentation time, retaining the corresponding social attention feature type, stimulus identifier, stimulus repetition count, adjacent stimulus interval, and feedback trigger record. The social attention feature types corresponding to the task-evoked adaptation component include gaze change correspondence type, head turning correspondence type, facial response correspondence type, voice response correspondence type, and operation delay correspondence type. The numerical content of the task-evoked adaptation component adopts the same data format as the corresponding social attention feature change, and the state content of the task-evoked adaptation component adopts the same state format as the corresponding social attention feature change. After writing is completed, each interaction segment generating a registration record can be mapped to the social attention features in the original social attention feature sequence, the task exposure record in the task exposure history sequence, and the selected task-evoked adaptation component via the same timeline.

[0091] The original social attention feature sequence is corrected based on the task-induced adaptation component to obtain a cross-round comparable corrected social attention feature sequence, including:

[0092] Based on the correspondence between the original social attention feature sequence and the task-induced adaptation component on the same time axis, the feature terms to be corrected for gaze changes, head turning, facial response, voice response and operation delay in each interaction segment are determined.

[0093] Specifically, a record of original social attention features is read from the original social attention feature sequence. Then, based on the stimulus presentation time in the original social attention feature record, a registration record with the same stimulus presentation time is searched in the registration record containing the task-evoked adaptation component. If a task-evoked adaptation component exists in the registration record with the same stimulus presentation time, then the corresponding gaze change, head turn, facial response, speech response, or operation delay is selected as the feature to be corrected from the original social attention feature record according to the social attention feature type corresponding to the task-evoked adaptation component. If no task-evoked adaptation component exists in the registration record with the same stimulus presentation time, then the gaze change, head turn, facial response, speech response, and operation delay in the original social attention feature record are not subtracted.

[0094] The features to be corrected are categorized into numerical and state-based features. Numerical features include gaze duration, facial response duration, speech response duration, and operation delay. State-based features include gaze entry changes, gaze exit changes, head turning direction, and facial response state. For state-based features, they are converted into subtractable numerical values ​​based on the number of times the state occurs or its duration within the same interaction segment, and then used as the feature to be corrected for subtraction. For example, gaze entry changes can be converted into numerical values ​​based on the number of times they occur within the interaction segment, head turning direction can be converted into numerical values ​​based on the duration of head turning direction, and facial response state can be converted into numerical values ​​based on the duration of facial response. The conversion method remains consistent throughout the execution of the same interactive task, ensuring that the same type of state-based feature uses the same data form across different interaction segments.

[0095] Subtract the corresponding task-induced adaptation component from the feature terms to be corrected for each interaction segment, and write the subtracted feature terms to be corrected into the corresponding positions in the original social attention feature sequence to obtain a cross-round comparable corrected social attention feature sequence.

[0096] Specifically, when subtracting the corresponding task-evoked adaptation component from the feature item to be corrected for each interaction segment, processing is performed according to the stimulus presentation time and social attention feature type. For the same interaction segment, if the feature item to be corrected is a gaze change, and the social attention feature type corresponding to the task-evoked adaptation component is the gaze change type, then the gaze change-corresponding task-evoked adaptation component is subtracted from the gaze change-corresponding task-evoked adaptation component. For the same interaction segment, if the feature item to be corrected is a head turn, and the task-evoked adaptation component corresponding to the task-evoked adaptation component is the head turn type, then the head turn-corresponding task-evoked adaptation component is subtracted from the head turn-corresponding task-evoked adaptation component. Facial response, voice response, and operation delay are subtracted separately according to the same correspondence. Subtraction is only performed when the feature item to be corrected and the task-evoked adaptation component have the same stimulus presentation time and the same social attention feature type; cross-subtraction is not performed between different social attention feature types.

[0097] During the subtraction process, if both the feature to be corrected and the task-induced adaptation component are numerical, the value of the feature to be corrected is subtracted from the value of the task-induced adaptation component to obtain the feature to be corrected after subtraction. If the feature to be corrected is obtained by transformation from a state form, and the task-induced adaptation component is formed by the social attention feature change of the corresponding state form, the same transformation method is used to obtain the values ​​of the feature to be corrected and the task-induced adaptation component, and then the subtraction process is performed. If the feature to be corrected is marked as null, the null mark is retained, and no subtraction process is performed. If the feature to be corrected after subtraction has a duration or frequency value less than zero, the feature to be corrected after subtraction is written as zero.

[0098] When writing the subtracted feature items to be corrected into the corresponding positions in the original social attention feature sequence, the writing position is located according to the stimulus presentation time and the social attention feature type of the feature item to be corrected in the original social attention feature record. The writing position is the field in the original social attention feature sequence that has the same stimulus presentation time and the same social attention feature type as the subtracted feature item to be corrected. Fixation changes, head turns, facial responses, voice responses, and operation delays that are not processed by subtraction retain their original values ​​or original states in the original social attention feature sequence. After writing all interaction segments, all written original social attention feature records are rearranged according to the stimulus presentation time to obtain a cross-round comparable corrected social attention feature sequence. Each corrected social attention feature record in the corrected social attention feature sequence still corresponds to an interaction segment and includes the stimulus presentation time corresponding to the interaction segment, the corrected fixation changes, the corrected head turns, the corrected facial responses, the corrected voice responses, and the corrected operation delays.

[0099] Autism risk auxiliary screening values ​​are generated based on the corrected social attention feature sequence, and rehabilitation training tasks are determined based on the autism risk auxiliary screening values ​​and the task exposure history sequence, including:

[0100] The gaze changes, head turns, facial responses, voice responses, and operation delays corresponding to each interaction segment in the corrected social attention feature sequence are compared with the preset screening feature range to generate autism risk auxiliary screening values.

[0101] Specifically, when determining the preset screening feature range, gaze changes, head turning, facial responses, voice responses, and operation delays are read from the reference interaction data, and corresponding ranges are formed according to the minimum and maximum reference values ​​for each type of feature. If there are records in the reference interaction data where the duration of consecutive absence of continuous interactive video frames or continuous interactive voice sampling data exceeds the preset acquisition interruption time, or if there are null value markers in the reference interaction data, these records and null value markers are not included in the determination of the preset screening feature range. The preset acquisition interruption time is determined based on the shortest duration of a single interactive segment in the interactive task; for example, it can be set to half the shortest duration of a single interactive segment.

[0102] When comparing the gaze changes, head turns, facial responses, speech responses, and operation delays corresponding to each interaction segment in the corrected social attention feature sequence with the preset screening feature ranges, the data is processed one by one from the stimulus presentation time of the corrected social attention feature record to the latest time. For each corrected social attention feature record, the corrected gaze changes are compared with the corresponding gaze change range, the corrected head turns are compared with the corresponding head turns range, the corrected facial responses are compared with the corresponding facial responses range, the corrected speech responses are compared with the corresponding speech responses range, and the corrected operation delays are compared with the corresponding operation delay range. If a certain type of corrected feature value is within the corresponding preset screening feature range, the corresponding feature item is marked as not deviating; if a certain type of corrected feature value is less than the lower limit or greater than the upper limit of the corresponding preset screening feature range, the corresponding feature item is marked as deviating. If a certain type of corrected feature value is marked as null, the corresponding feature item is marked as missing. Missing marks are not counted in the number of deviation marks but are retained in the corrected social attention feature record.

[0103] The autism risk auxiliary screening value is generated based on the number of deviation markers and the number of features involved in the comparison. The number of features involved in the comparison refers to the total number of gaze changes, head turns, facial responses, vocal responses, and operational delays in the corrected social attention feature sequence that lack missing markers. The number of deviation markers refers to the number of features involved in the comparison that have deviation markers written in them. When generating the autism risk auxiliary screening value, the number of deviation markers is divided by the number of features involved in the comparison, and then converted to a percentage. If the number of features involved in the comparison is 0, no autism risk auxiliary screening value is generated, and a missing screening value marker is written in the corresponding interactive task.

[0104] Stimulus identifiers, number of repetitions, intervals between adjacent stimuli, and feedback trigger records are retrieved from the task exposure history sequence. Combined with autism risk auxiliary screening values, the order of stimulus presentation, task difficulty, and feedback triggering methods are determined to identify rehabilitation training tasks.

[0105] Specifically, after obtaining the autism risk screening value, stimulus identifiers, stimulus repetition counts, adjacent stimulus intervals, and feedback trigger records are retrieved from the task exposure history sequence. The retrieval method involves reading each task exposure record from earliest to latest according to the stimulus presentation time in the task exposure history sequence, and extracting the stimulus identifier, stimulus repetition count, adjacent stimulus interval, and feedback trigger record from each record. The stimulus identifier is used to identify the stimulus content that has been presented; the stimulus repetition count is used to determine the number of times the same stimulus content has appeared; the adjacent stimulus interval is used to determine the time difference between stimulus presentations of adjacent interactive segments; and the feedback trigger record is used to determine whether a reward sound, reward animation, prompt voice, task advancement prompt, or task retry prompt has been triggered within the interactive segment. After retrieval, the task exposure records in the task exposure history sequence are grouped according to the stimulus identifier, and the stimulus repetition count, adjacent stimulus interval, and feedback trigger record are retained within each group corresponding to the stimulus identifier.

[0106] When determining the stimulus presentation order using autism risk screening values, stimulus markers with higher and lower repetition frequencies are identified based on the task exposure history sequence. Stimulus markers with higher repetition frequencies are those located in the first interval after being sorted by the frequency of repetition during the same interactive task, while those located in the second interval are those located in the last interval. The first and second intervals are determined based on the number of stimulus markers during the same interactive task; for example, the top 30% of stimulus markers can be considered as those with higher repetition frequencies, and the bottom 30% as those with lower repetition frequencies. The arrangement of stimulus markers in the next rehabilitation training task is then determined based on the autism risk screening values. When the autism risk screening value is within the low-risk range, the stimulus presentation order prioritizes stimuli with lower repetition frequencies, and inserts stimulus markers corresponding to different stimulus types between adjacent interactive segments. When the autism risk screening value is within the intermediate-risk range, the stimulus presentation order alternates between stimuli with higher and lower repetition frequencies. When the autism risk screening value is within the high-risk range, the stimulus presentation order prioritizes stimuli that the child has already encountered and whose feedback trigger record is a non-empty value. The low-risk, intermediate-risk, and high-risk ranges are set according to percentage-based numerical intervals before the interactive task begins. For example, the low-risk range can be set to 0 to 30, the intermediate-risk range to 31 to 60, and the high-risk range to 61 to 100.

[0107] When determining task difficulty using autism risk screening values, the task difficulty is jointly determined by the number of stimulus content categories, the complexity of the target operation, and the frequency of cue voice appearances. The number of stimulus content categories refers to the number of stimulus types used in the same rehabilitation training task. The complexity of the target operation refers to the number of touch, click, selection, or drag actions the child needs to perform to complete the target operation. The frequency of cue voice appearances refers to the number of times cue voices are allowed to appear in the rehabilitation training task. When the autism risk screening value is in the low-risk range, the task difficulty is set to increase the number of stimulus content categories and increase the complexity of the target operation; when the autism risk screening value is in the intermediate-risk range, the task difficulty is set to maintain the current number of stimulus content categories and maintain the current complexity of the target operation; when the autism risk screening value is in the high-risk range, the task difficulty is set to decrease the number of stimulus content categories and decrease the complexity of the target operation. Each change in task difficulty is recorded in the task operation record, serving as a data source for constructing the next task exposure history sequence.

[0108] When determining the feedback triggering method based on autism risk screening values, the feedback triggering records in the task exposure history sequence are read, and the occurrence of the feedback content type in each interaction segment is determined. The feedback triggering method includes the feedback content type, the feedback triggering condition, and the feedback triggering time. The feedback content type is selected from reward sounds, reward animations, prompt voices, task progression prompts, and task retry prompts. The feedback triggering condition is determined based on whether the child completes the target operation, fails to complete the target operation, or exceeds the operation waiting time. The feedback triggering time is determined according to the time of occurrence of the event corresponding to the target operation or the end of the operation waiting time. When the autism risk screening value is in the low-risk range, the feedback triggering method is set to trigger a task progression prompt after the child completes the target operation; when the autism risk screening value is in the intermediate-risk range, the feedback triggering method is set to trigger a reward sound or reward animation after the child completes the target operation; when the autism risk screening value is in the high-risk range, the feedback triggering method is set to trigger a prompt voice or task retry prompt after the child fails to complete the target operation or exceeds the operation waiting time. The finalized rehabilitation training tasks include the stimulus presentation order, task difficulty, and feedback triggering method, and are recorded in the task operation records of subsequent interactive tasks.

[0109] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. An AI-powered intelligent recognition-based early screening and rehabilitation training system for autism, characterized in that, include: Interactive acquisition module: Acquires continuous interactive videos, continuous interactive voices, and task operation records of children, and aligns them according to the timing of stimulus presentation and feedback triggering to form a multi-round interactive behavior sequence; Exposure building module: Based on stimulus identifiers, number of stimulus repetitions, intervals between adjacent stimuli, and feedback trigger records, construct a task exposure history sequence corresponding to a multi-round interaction behavior sequence; Feature extraction module: Extracts gaze changes, head turning, facial responses, voice responses, and operation delays from multi-turn interaction behavior sequences to form the original social attention feature sequence; Adaptive Separation Module: Performs sequence registration between the original social attention feature sequence and the task exposure history sequence, and separates the task-induced adaptation component that changes synchronously with the task exposure history sequence; Feature correction module: Corrects the original social attention feature sequence based on the task-induced adaptation component to obtain a corrected social attention feature sequence that can be compared across rounds; Screening and training decision module: Generates autism risk auxiliary screening value based on the corrected social attention feature sequence, and determines rehabilitation training tasks based on the autism risk auxiliary screening value and the task exposure history sequence.

2. The AI-powered intelligent recognition autism early screening and rehabilitation training system according to claim 1, characterized in that, Acquire continuous interactive videos, continuous interactive voice recordings, and task operation records of children, and align them chronologically according to the stimulus presentation time and feedback trigger time to form a multi-turn interactive behavior sequence, specifically: During the execution of the interactive task, the child's continuous interactive video, continuous interactive voice and task operation record are acquired simultaneously, and the start acquisition time, stimulus presentation time and feedback trigger time in the child's continuous interactive video, continuous interactive voice and task operation record are marked respectively. The children's continuous interactive videos, continuous interactive voices, and task operation records are aligned chronologically according to the timing of stimulus presentation and feedback triggering to form a multi-round interactive behavior sequence.

3. The AI-powered intelligent recognition-based early screening and rehabilitation training system for autism according to claim 1, characterized in that, Based on stimulus identifiers, stimulus repetition counts, intervals between adjacent stimuli, and feedback trigger records, a task exposure history sequence corresponding to a multi-turn interaction behavior sequence is constructed, specifically as follows: Read task operation records aligned with stimulus presentation times from multi-round interaction behavior sequences, and extract stimulus identifiers corresponding to stimulus presentation times and feedback trigger records corresponding to feedback trigger times; The number of stimulus repetitions is counted according to the order of appearance of stimulus labels in the multi-round interactive behavior sequence, and the interval between adjacent stimuli is determined based on the presentation time of two adjacent stimuli. Stimulus identifiers, number of stimulus repetitions, intervals between adjacent stimuli, and feedback trigger records are written into the same timeline according to the timing of stimulus presentation, thus constructing a task exposure history sequence corresponding to the multi-round interaction behavior sequence.

4. The AI-powered intelligent recognition autism early screening and rehabilitation training system according to claim 1, characterized in that, The sequence of multi-turn interaction behaviors is used to extract gaze changes, head turning, facial responses, voice responses, and operation delays to form the original social attention feature sequence, specifically: Interaction segments are divided using the stimulus presentation time in a multi-round interactive behavior sequence as the boundary. Extract gaze changes, head turning, and facial responses from each interactive segment of a child's continuous interactive video; Extracting voice responses from each interactive segment of a child's continuous interactive speech; Extract the operation latency within each interaction segment from the task operation record; The original social attention feature sequence is formed by arranging the interaction segments in the sequence of multiple rounds of interaction behavior.

5. The AI-powered intelligent recognition-based early screening and rehabilitation training system for autism according to claim 1, characterized in that, The original social attention feature sequence is sequence registered with the task exposure history sequence to separate the task-induced adaptation component that changes synchronously with the task exposure history sequence. Specifically: The original social attention feature sequence and the task exposure history sequence are registered along the same time axis to obtain the social attention features and task exposure records corresponding to each interaction segment. Based on the changes in the number of stimulus repetitions, the interval between adjacent stimuli, and the direction of change in feedback trigger records, we screened the changes in social attention features that were synchronous with the task exposure history sequence, and used the changes in social attention features as the task-induced adaptation component.

6. The AI-powered intelligent recognition autism early screening and rehabilitation training system according to claim 1, characterized in that, The original social attention feature sequence is corrected based on the task-induced adaptation component to obtain a corrected social attention feature sequence that can be compared across rounds, specifically: Based on the correspondence between the original social attention feature sequence and the task-induced adaptation component on the same time axis, the feature terms to be corrected for gaze changes, head turning, facial response, voice response and operation delay in each interaction segment are determined. Subtract the corresponding task-induced adaptation component from the feature terms to be corrected for each interaction segment, and write the subtracted feature terms to be corrected into the corresponding positions in the original social attention feature sequence to obtain a cross-round comparable corrected social attention feature sequence.

7. The AI-powered intelligent recognition-based early screening and rehabilitation training system for autism according to claim 1, characterized in that, Autism risk auxiliary screening values ​​are generated based on the corrected social attention feature sequence, and rehabilitation training tasks are determined based on the autism risk auxiliary screening values ​​and the task exposure history sequence, specifically: The gaze changes, head turns, facial responses, voice responses, and operation delays corresponding to each interaction segment in the corrected social attention feature sequence are compared with the preset screening feature range to generate autism risk auxiliary screening values. Stimulus identifiers, number of repetitions, intervals between adjacent stimuli, and feedback trigger records are retrieved from the task exposure history sequence. Combined with autism risk screening values, the order of stimulus presentation, task difficulty, and feedback triggering methods are determined to identify rehabilitation training tasks.

Images