Children autism auxiliary screening device based on machine learning

Abstract

The invention provides a machine learning-based child autism auxiliary screening device, which comprises a video playing mechanism for playing animations or video clips, and an electroencephalogram signal acquisition mechanism for acquiring electroencephalogram signals when children watch the animations or video clips, and the central processing mechanism is used for outputting a conclusion whether the tested child is an autistic child or not after being processed by a built-in functional module according to the electroencephalogram signal output by the electroencephalogram signal acquisition mechanism when the child to be screened watches the animation or the video clip. In the using process, the device is easy and convenient to operate, particularly suitable for screening the infantile autism of children, rapid in judgment result output, extremely high in accuracy rate, low in price, easy to maintain and replace and suitable for large-area popularization in medical institutions, and the blank of the device in the prior art is filled up.

Description

Technical Field

[0001] The present invention relates to the technical field of medical diagnosis auxiliary devices, and in particular to a child autism auxiliary screening device based on machine learning. Background Art

[0002] Autism is a relatively serious neurodevelopmental disorder. Patients usually have social difficulties, delayed language development, stereotyped and repetitive behaviors, and other symptoms. According to UN data, as of 2021, more than 70 million people in the world are suffering from autism. About 1 in every 160 children suffers from the disease.

[0003] Up to now, there is no specific medicine to treat autism. Most patients are intervened through education guidance, rehabilitation training and other means to improve abnormal symptoms. "Early detection, early intervention" has always been the golden rule recognized in the autism circle. The earlier the intervention treatment, the higher the probability of recovery of the patient. However, factors such as insufficient public awareness, lack of social communication ability of patients, unclear causes of the disease, and heterogeneity of spectrum symptoms make "early detection, early intervention" difficult.

[0004] Poor communication and language skills are one of the characteristics of autism. A considerable number of autistic children have delayed language development. They rarely initiate conversations, often talk to themselves, ignore the person they are talking to, or only understand what others say literally. In terms of semantics, autistic children can only understand the figurative meaning of language, but not the abstract and extended meaning of language. They have significant differences in the way they process semantic understanding from ordinary people. Therefore, language communication disorders have also become one of the important criteria for clinical diagnosis of autism. Electroencephalogram (EEG) is an electrophysiological monitoring method for non-invasively recording brain electrical activity. Electrodes are placed along the scalp, and then multiple electrodes placed on the scalp are used to record the spontaneous electrical activity of the brain over a period of time. Previous studies have shown that the core pathophysiological feature of ASD (Autistic Spectrum Disorder) is the abnormality of the brain's oscillation rhythm. This abnormality can be captured by EEG. Therefore, using EEG as the original input for diagnosing autism is more reliable to a certain extent.

[0005] At present, auxiliary screening for autism in children mostly uses scales, eye movements, images and other features as the basis for discrimination, and uses computational models to integrate single-modal or multi-modal features for screening and judgment, which has the following limitations: First, the steps for collecting patient data are cumbersome, requiring autistic children to cooperate for a long time; second, the screening program is not efficient enough, the whole process cycle is long, and it is difficult to ensure real-time performance; third, the screening accuracy is low, and the discrimination results are of little reference value to both doctors and patients. For the above reasons, there is an urgent need for an efficient and high-precision auxiliary screening device for autism in children to help doctors complete preliminary screening and carry out targeted intervention treatment as soon as possible. Summary of the invention

[0006] The purpose of the present invention is to solve the problems existing in the prior art and provide a new child autism auxiliary screening device based on machine learning that fills the gap in the prior art.

[0007] The technical solution of the present invention is: the child autism auxiliary screening device based on machine learning of the present invention has the following structural features: it includes a video playing mechanism for playing cartoons or video clips, an electroencephalogram signal acquisition mechanism for collecting electroencephalogram signals when children watch cartoons or video clips, and a central processing mechanism for outputting a conclusion of whether the tested child is an autistic child after processing the electroencephalogram signals of the children to be screened when watching cartoons or video clips output by the electroencephalogram signal acquisition mechanism through a built-in functional module.

[0008] A further solution is that the functional modules built into the above-mentioned central processing mechanism include: A word segmentation module is used to perform word segmentation on the lines involved in the cartoon or video clip being played, so that each word in the lines corresponds to a corresponding time segment in the cartoon or video clip; The pre-trained language model is used to process the dialogue text after the word segmentation module to obtain the context-related semantic embedding vector of each word; A data preprocessing module, used to preprocess the EEG signals collected and sent by the EEG signal acquisition mechanism and extract target segment EEG signals; The first regression model fits the semantic embedding vector of each word to the EEG data of each channel and each word in the corresponding time when the known normal children watch the set cartoon or video clip, so as to obtain the semantic processing mode of the normal children; The second regression model fits the semantic embedding vector of each word to the EEG data of each channel and each word in the corresponding time when the autistic children watch the set cartoons or video clips, so as to obtain the semantic processing mode that represents the autistic children; A screening judgment module is used to input the electroencephalogram signal of the child to be screened when watching the set cartoon or video clip into the first and second regression models after being processed by the data preprocessing module, and then respectively calculate the distance between the two sets of semantic embedding vectors obtained and the semantic embedding vectors of the first and second regression models generated by the dialogue text. If the calculated semantic embedding vector is closer to the first regression model, the child to be screened is determined to be a normal child; if the calculated semantic embedding vector is closer to the second regression model, the child to be screened is determined to be an autistic child; The result output module is used to output and display the judgment result output by the screening judgment module.

[0009] A further solution is: the above-mentioned word segmentation module uses the jieba word segmentation tool to perform word segmentation, and manually adjusts the word segmentation results.

[0010] A further solution is: the above-mentioned pre-trained language model is a pre-trained ELMo language model in which the dimension of the semantic embedding vector is set to 1024 dimensions.

[0011] A further solution is that both the first regression model and the second regression model are constructed based on regularized linear models optimized by ridge regression.

[0012] A further solution is that the above-mentioned data preprocessing module performs preprocessing on the EEG signal specifically including: ① Filtering: Perform 0.5-30HZ bandpass filtering and 48-52HZ trap-pass filtering on the collected signal; ② Remove the skirt electrodes on the face in the collected signal; ③ Segmentation: segment the collected signal with 1 second as the segment unit; ④ Interpolation of electrodes with excessive artifacts; ⑤ Eliminate the segments with too large artifacts and record the eliminated segments as empty; ⑥ Independent component analysis: remove physiological artifacts; ⑦Re-reference: average reference; ⑧Splicing EEG signal fragments; ⑨Animation clip extraction: extract EEG signals according to the video annotation time; ⑩Take the average of the EEG signals for each word segment.

[0013] A further solution is that the video playing mechanism is an electronic device that can play cartoons or videos, and when in use, the video playing mechanism plays the cartoons or video clips for 2 to 5 minutes.

[0014] A further solution is: the above-mentioned EEG signal acquisition mechanism includes a conductive paste applied on the child's scalp and an electrode cap worn on the child's head when in use.

[0015] The present invention has positive effects: (1) When the device of the present invention is used, it can almost instantly give a high-accuracy conclusion on whether the child to be tested has autism symptoms, with high screening efficiency and accuracy, thus filling the gap in such devices in the prior art. (2) When the device of the present invention is used, the child to be screened only needs to apply conductive paste on the scalp and wear an electrode cap to watch a cartoon or video clip for about 2 minutes, so it is particularly suitable for screening children for autism. (3) The present invention uses a pre-trained language model to capture the semantic information of the lines in the cartoon, and uses a high-temporal-resolution EEG model as the input of the machine learning model to perform regression prediction on the semantic information to screen children for autism. This is an innovative way to achieve efficient, accurate and convenient screening of children for autism. In addition, the hardware of the device of the present invention is based on existing components, is low-cost, easy to maintain and replace, and is suitable for large-scale promotion in medical institutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic block diagram of the overall structure of the present invention; Figure 2 A schematic diagram of the semantic processing method using EEG signals in the present invention; Figure 3 It is a schematic diagram of the working process of the data processing module of the present invention; Figure 4 It is a schematic diagram of the principle of screening normal children and autistic children when the present invention is used. DETAILED DESCRIPTION

[0017] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0018] (Example 1) See Figures 1 to 4 The child autism auxiliary screening device based on machine learning in this embodiment is mainly composed of a video playback mechanism, an EEG signal acquisition mechanism and a central processing mechanism.

[0019] The video playing mechanism is used to play cartoons or video clips that children are interested in, and the playing time is 2 to 5 minutes. The video playing mechanism can adopt electronic devices that can play cartoons or videos in the prior art.

[0020] The electroencephalogram signal acquisition mechanism is used to collect electroencephalogram signals when children watch cartoons or video clips. In this embodiment, the electroencephalogram signal acquisition mechanism includes a conductive paste applied to the child's scalp and an electrode cap worn on the child's head. The conductive paste and the electrode cap cooperate to collect electroencephalogram signals, which is a mature existing technology and will not be described in detail.

[0021] The central processing mechanism is used to output the conclusion of whether the tested child is autistic after processing the EEG signal output by the EEG signal acquisition mechanism when the child to be screened watches the cartoon or video clip through the built-in functional module. The hardware of the central processing mechanism can adopt desktop computers, laptop computers, servers, etc. in the existing technology.

[0022] The central processing mechanism of this embodiment has the following built-in functional modules: 1. The word segmentation module is used to segment the lines involved in the cartoons or video clips selected for children to watch using the existing jieba word segmentation tool, manually adjust the word segmentation results, and mark the selections of the cartoons or video clips according to the word segmentation results, so that each word in the lines corresponds to a corresponding time segment in the cartoon or video clip.

[0023] 2. A pre-trained language model is used to process the dialogue text after segmentation by the segmentation module to obtain the context-related semantic embedding vector of each word. In this embodiment, the pre-trained language model uses a pre-trained ELMo (Embeddings from Language Models) language model to obtain the context-related semantic embedding vector of each word, and the dimension of the semantic embedding vector is set to 1024 dimensions.

[0024] 3. Data preprocessing module, used to preprocess the EEG signals collected and sent by the EEG signal acquisition mechanism and extract the target EEG signal. Figure 3 The data preprocessing module performs preprocessing on the EEG signal including: ① Filtering: Perform 0.5-30HZ bandpass filtering and 48-52HZ trap-pass filtering on the collected signal; ② Remove the skirt electrodes on the face in the collected signal; ③ Segmentation: segment the collected signal with 1 second as the segment unit; ④ Interpolation of electrodes with excessive artifacts; ⑤ Eliminate the segments with too large artifacts and record the eliminated segments as empty; ⑥ Independent component analysis: remove physiological artifacts; ⑦Re-reference: average reference; ⑧Splicing EEG signal fragments; ⑨Animation clip extraction: extract EEG signals according to the video annotation time; ⑩Take the average of the EEG signals for each word segment.

[0025] It should be noted that the methods of each sub-step in this step are mature existing technologies in the field, so the specific methods of each sub-step will not be described in detail.

[0026] 4. The first regression model uses the existing regularized linear model optimized by ridge regression to fit the semantic embedding vector of each word to the EEG data of each channel and each word in the corresponding time when the known normal children watch the set cartoon or video clip, so as to obtain the first regression model that characterizes the semantic processing mode of normal children; that is, the first regression model is established by collecting the EEG signals of the known normal children when watching the set cartoon or video clip and fitting the semantic embedding vectors of the corresponding words in the lines.

[0027] 5. The second regression model uses the existing regularized linear model optimized by ridge regression to fit the semantic embedding vector of each word to the EEG data of each channel and each word in the corresponding time when the known autistic children watch the set cartoons or video clips, so as to obtain the second regression model that characterizes the semantic processing mode of autistic children; that is, the second regression model is established by collecting the EEG signals of the known autistic children when watching the set cartoons or video clips and fitting the semantic embedding vectors of the corresponding words in the lines.

[0028] 6. A screening judgment module is used to input the EEG signals of the screened children when they watch the set cartoons or video clips into the first and second regression models respectively after being processed by the data preprocessing module to obtain two sets of semantic embedding vectors, and then calculate the distances between the two sets of semantic embedding vectors and the semantic embedding vectors of the first and second regression models generated by the dialogue texts. If the calculated semantic embedding vector is closer to the first regression model, the screened child is determined to be a normal child; if the calculated semantic embedding vector is closer to the second regression model, the screened child is determined to be an autistic child.

[0029] 7. A result output module is used to output and display the judgment result output by the screening judgment module. The result output module is composed of a display screen.

[0030] The method of using the child autism auxiliary screening device based on machine learning in this embodiment is briefly described as follows: When used for the first time, a group of confirmed autistic children and normal children were asked to watch a 3-minute excerpt from the cartoon "Peppa Pig" in a quiet state, and the EEG data of the autistic children and normal children participating in the test were collected. The first and second regression models were established in the central processing agency as described above.

[0031] During normal use, the conductive paste of the electroencephalogram signal acquisition mechanism is applied on the scalp of the child to be screened and an electrode cap for detection is worn on the child's head. Then the video playback mechanism is turned on to play a clip of the cartoon "Peppa Pig" for a set time of 2 to 5 minutes. The electroencephalogram signal acquisition mechanism sends the collected electroencephalogram signals of the child watching the cartoon to the central processing mechanism. The electroencephalogram data processed by the data preprocessing module of the central processing mechanism are respectively input into the first and second regression models to obtain two corresponding sets of semantic embedding vectors. The screening and judgment module respectively calculates the distance between the two sets of semantic embedding vectors and the semantic embedding vectors of the first and second regression models generated from the dialogue text. If the semantic embedding vector calculated by the screening is closer to the first regression model, the child under examination is determined to be a normal child. If the calculated semantic embedding vector is closer to the second regression model, the child under examination is determined to be an autistic child.

[0032] As can be seen from the foregoing, the child autism auxiliary screening device based on machine learning in this embodiment has the following characteristics: first, it is easy to use and harmless to children. The input data is a non-invasive EEG signal. Children only need to apply conductive paste on the scalp and wear an electrode cap before collecting data to complete data collection. During the data collection process, the patient only needs to keep the head relatively stable and watch a cartoon for about 2 minutes to obtain high-quality data, and based on the acquired EEG signal data, a highly accurate prediction result can be given instantly.

[0033] The above embodiments are descriptions of specific implementation methods of the present invention rather than limitations of the present invention. Technical personnel in the relevant technical field can make various changes and modifications to obtain corresponding equivalent technical solutions without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions should be included in the patent protection scope of the present invention.

Claims

1. A device for assisting screening of children's autism based on machine learning, characterized in that: It includes a video playing mechanism for playing cartoons or video clips, an electroencephalogram signal collecting mechanism for collecting electroencephalogram signals when children watch cartoons or video clips, and a central processing mechanism for outputting a conclusion on whether the tested child is an autistic child after processing the electroencephalogram signals output by the electroencephalogram signal collecting mechanism when the children to be screened watch cartoons or video clips through built-in functional modules.

2. The device for assisting screening of children's autism based on machine learning according to claim 1, characterized in that: The functional modules built into the central processing unit include: A word segmentation module is used to perform word segmentation on the lines involved in the cartoon or video clip being played, so that each word in the lines corresponds to a corresponding time segment in the cartoon or video clip; The pre-trained language model is used to process the dialogue text after the word segmentation module to obtain the context-related semantic embedding vector of each word; A data preprocessing module, used to preprocess the EEG signals collected and sent by the EEG signal acquisition mechanism and extract target segment EEG signals; The first regression model fits the semantic embedding vector of each word to the EEG data of each channel and each word in the corresponding time when the known normal children watch the set cartoon or video clip, so as to obtain the semantic processing mode of the normal children; The second regression model fits the semantic embedding vector of each word to the EEG data of each channel and each word in the corresponding time when the autistic children watch the set cartoons or video clips, so as to obtain the semantic processing mode that represents the autistic children; A screening judgment module is used to input the electroencephalogram signal of the child to be screened when watching the set cartoon or video clip into the first and second regression models after being processed by the data preprocessing module, and then respectively calculate the distance between the two sets of semantic embedding vectors obtained and the semantic embedding vectors of the first and second regression models generated by the dialogue text. If the calculated semantic embedding vector is closer to the first regression model, the child to be screened is determined to be a normal child; if the calculated semantic embedding vector is closer to the second regression model, the child to be screened is determined to be an autistic child; The result output module is used to output and display the judgment result output by the screening judgment module.

3. The device for assisting screening of children's autism based on machine learning according to claim 2, characterized in that: The word segmentation module uses the jieba word segmentation tool to perform word segmentation and manually adjusts the word segmentation results.

4. The device for assisting screening of children's autism based on machine learning according to claim 2, characterized in that: The pre-trained language model is a pre-trained ELMo language model in which the dimension of the semantic embedding vector is set to 1024 dimensions.

5. The device for assisting screening of children's autism based on machine learning according to claim 2, characterized in that: The first regression model and the second regression model are both constructed based on regularized linear models optimized by ridge regression.

6. The device for assisting screening of children's autism based on machine learning according to claim 2, characterized in that: The data preprocessing module performs preprocessing on the EEG signal specifically including: ① Filtering: Perform 0.5-30HZ bandpass filtering and 48-52HZ trap-pass filtering on the collected signal; ② Remove the skirt electrodes on the face in the collected signal; ③ Segmentation: segment the collected signal with 1 second as the segment unit; ④ Interpolation of electrodes with excessive artifacts; ⑤ Eliminate the segments with too large artifacts and record the eliminated segments as empty; ⑥ Independent component analysis: remove physiological artifacts; ⑦Re-reference: average reference; ⑧Splicing EEG signal fragments; ⑨Animation clip extraction: extract EEG signals according to the video annotation time; ⑩Take the average of the EEG signals for each word segment.

7. The device for assisting screening of children's autism based on machine learning according to claim 1, characterized in that: The video playing mechanism is an electronic device that can play cartoons or videos. When in use, the video playing mechanism plays the cartoons or video clips for 2 to 5 minutes.

8. The device for assisting screening of children's autism based on machine learning according to claim 1, characterized in that: The electroencephalogram signal collection mechanism comprises a conductive paste applied on the scalp of a child and an electrode cap worn on the child's head when in use.

Images