Early diagnosis and intervention curative effect evaluation method for children with autism

Abstract

The invention discloses an early diagnosis and intervention curative effect evaluation method for children with autism, and belongs to the technical field of medical diagnosis. The method specifically comprises the following steps: S1, obtaining historical case big data: collecting the historical case big data of an autistic truly diagnosed child and a normally developed child and corresponding diagnosis results and intervention diagnosis and treatment records, and associating to form a historical database; s2, historical data preprocessing: performing cleaning, standardization and missing value processing on historical case big data, and associating and integrating the historical case big data into a structured sample set; by collecting behavior, physiology, environment and family multi-dimensional data, mining an internal mode of the data in combination with depth features, and then utilizing a multi-modal fusion diagnosis model for analysis, dependence on subjective experience of doctors and a single scale is reduced; the classification accuracy of the model is optimized through cross validation, a diagnosis report contains core abnormal features and a judgment basis, manual recheck is supported, and the risks of missed diagnosis and misdiagnosis are further reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical diagnosis, in particular to an autism early diagnosis and intervention effect evaluation method. BACKGROUND

[0002] Autism, also a pervasive developmental disorder, common disease types include apathy, initiative but strange and passive. Genetic and environmental factors are the main causes of interaction, studies have shown that the heritability can be as high as 80%-90%, in addition, including brain structure and function abnormalities, biological mechanisms and adverse factors during pregnancy, etc. will also affect the onset of autism. Most patients begin to appear social interaction disorders, communication disorders, and obvious interest, stereotyped and repetitive behavior patterns in childhood. Autism generally begins to show symptoms before the age of 3, such as almost no initiative to initiate social interaction, due to the peculiar form of social interaction, their behavior patterns are unique and stereotyped, and autism patients will show difficulty in understanding others' emotions and thoughts, and lack of social rules. In addition, 70% of patients may have one or more comorbid mental disorders. However, the existing autism early diagnosis and intervention effect evaluation technology has the following disadvantages: strong subjectivity in diagnosis: the existing diagnosis mostly relies on the observation and scale evaluation of clinicians (such as ABC scale and CARS scale), which is greatly influenced by subjective factors such as doctor's experience and observation time, and is prone to misdiagnosis or missed diagnosis, especially for young children and atypical symptoms, the accuracy is difficult to guarantee. SUMMARY

[0003] The purpose of the present application is to provide an autism early diagnosis and intervention effect evaluation method, which can solve the problem of strong subjectivity in autism early diagnosis and intervention effect evaluation technology.

[0004] Technical scheme: In order to solve the above technical problems, according to one aspect of the present application, more specifically, an autism early diagnosis and intervention effect evaluation method, specifically comprising the following steps:

[0005] S1, obtaining historical case big data: collecting historical case big data of autism diagnosed children and normally developed children and corresponding diagnosis results and intervention diagnosis and treatment records, and associating to form a historical database;

[0006] S2, historical data preprocessing: cleaning, standardizing and missing value processing of the historical case big data, and associating and integrating into a structured sample set;

[0007] S3, extracting deep features: extracting basic features, associated features and prediction features from the preprocessed data, extracting deep features that can represent the internal pattern of the data through multi-level feature engineering and data mining technology, and forming a deep feature set with unified dimensions;

[0008] S4, constructing a multi-modal fusion model: constructing a diagnosis model by a convolutional neural network + recurrent neural network + Transformer, constructing an intervention scheme generation model using a reinforcement learning framework, and constructing an efficacy evaluation model based on the diagnosis model;

[0009] S5, collecting suspected child data: collecting real-time big data of suspected children according to the same standardized specifications as historical data collection for suspected children with autism;

[0010] S6, early diagnosis of suspected children: pre-processing and deep feature extraction of suspected children's big data, inputting the diagnosis model to output risk levels, and generating a report containing core abnormal features;

[0011] S7, personalized intervention and dynamic evaluation: generating an intervention plan, setting short-term, medium-term, and long-term evaluation nodes, collecting data to evaluate efficacy, and optimizing the plan if it does not meet the standards.

[0012] Further, in step S1, the case big data includes: behavior data, physiological data, environmental and family data, diagnosis and efficacy correlation historical data;

[0013] Behavior data: objective records of children in a pre-set standardized evaluation scene, including video images and basic behavior indicators extracted therefrom;

[0014] Physiological data: physiological signals and time series records collected by medical-grade wearable devices, including autonomic nervous function indicators and sleep structure indicators;

[0015] Environmental and family data: information obtained through standardized scales and structured interviews, including family parenting style, parent cognitive level, and intervention cooperation degree;

[0016] Diagnosis and efficacy correlation data: corresponding sample clinical diagnosis basis, specific schemes used during intervention, and efficacy evaluation results at each stage;

[0017] All data are associated through individual unique identification to form a complete historical database containing original records, intermediate indicators, and final results.

[0018] Further, in step S2, when pre-processing the behavior data, frame-level segmentation is performed on the video images, and abnormal behavior annotations caused by device shaking and occlusion are corrected through human-machine collaboration. Outlier detection is performed on the basic behavior indicators, and short-time missing data is supplemented through interpolation. Unstructured behavior descriptions are converted into quantitative scores;

[0019] In the preprocessing of physiological data, the sliding window method is used for smoothing the continuous time series physiological data to filter out the peak noise caused by motion interference; the Z-score standardization algorithm is used to convert different physiological indicators to a unified scale; the time series data with a missing rate lower than the preset value is filled with the mean value of the adjacent period, and the samples with a missing rate not lower than the preset value are marked and processed separately;

[0020] In the preprocessing of environmental and family data, qualitative descriptions are converted into quantitative values; scale data are subjected to reliability test, invalid items are removed, and missing values are processed by multiple imputation method;

[0021] Based on the unique identification of individuals, the preprocessed behavior data, physiological data, environmental and family data, diagnosis and efficacy correlation history data are associated into structured samples, invalid samples are removed through logical verification, and finally a historical sample set directly used for feature extraction is formed.

[0022] Further, the step S3 specifically comprises the following steps:

[0023] S31, basic features: fine processing is performed on the preprocessed behavior data, physiological data, environmental and family data, statistical transformation including coefficient of variation, peak frequency, diurnal difference, distribution characteristic analysis including quantile, skewness, and high-dimensional indicators are generated;

[0024] S32, correlation features: through cross-modal correlation analysis techniques including Pearson correlation coefficient, partial least squares regression, the internal correlation patterns between behavior, physiology, environment and family dimensions are quantified, including linear correlation, conditional correlation;

[0025] S33, prediction features: through time series prediction models including but not limited to long short-term memory network (LSTM) and gradient boosting regression tree, the historical data is trend extrapolated, the indicators that can predict long-term results are extracted, including the future change slope of behavior indicators, the sensitivity coefficient of physiological indicators to intervention response, the contribution weight of family factors to long-term prognosis, etc.;

[0026] S34, redundant features are removed through feature selection algorithm, and finally a deep feature set including basic features, correlation features and prediction features is formed, and standardized processing is performed to ensure the comparability between features.

[0027] Further, in step S4, the diagnostic model takes the deep features as input, uses a convolutional neural network to extract behavior-related feature space patterns, a recurrent neural network to extract physiological-related feature timing patterns, and a Transformer to extract the global association of environmental features and three-level features. The model outputs the autism risk level and the corresponding probability by dynamically allocating weights to the three-level features through an attention mechanism. The model training uses cross-validation, and the optimization goal is to maximize the classification accuracy.

[0028] The intervention scheme generation model takes the diagnostic results and deep features as input, uses a reinforcement learning structure, takes the intervention strategy parameters as the action space, and calculates the weighted sum of the behavior improvement rate, the physiological index compliance rate, and the family adaptation degree as the reward function. The optimal intervention strategy is generated through a policy gradient algorithm.

[0029] The efficacy evaluation model takes real-time data and deep features during the intervention process as input, adds a deep feature difference analysis module before and after the intervention based on the diagnostic model, calculates the standardized difference between the deep features before and after the intervention, and outputs the behavior improvement rate based on the basic features, the physiological-behavior system improvement degree based on the associated features, and the long-term prognosis trend based on the predicted features. The model has built-in evaluation thresholds to automatically determine significant improvement, improvement, and no improvement.

[0030] Further, in step S5, the suspected child big data includes:

[0031] Behavior data: In a standardized evaluation scene consistent with historical data, behavior videos are recorded through high-definition cameras, and consistent basic behavior indicators are extracted synchronously.

[0032] Physiological data: Wearable devices of the same model as the historical data are used to collect physiological indicators, and child daily activity logs are recorded during the collection period.

[0033] Environmental and family data: Through standardized electronic questionnaires and structured interviews consistent with historical data, information including family environment, parent cognition, and intervention cooperation willingness is obtained.

[0034] Further, in step S6, the suspected child big data preprocessing strictly follows the preprocessing specifications of step S2, and the behavior, physiological, environmental, and family data of the suspected child are cleaned, standardized, and integrated. The deep feature extraction is based on the deep feature extraction rules of step S3, and the deep features are extracted from the preprocessed suspected child data. The deep features of the suspected child are input into the constructed multi-modal fusion diagnostic model, and the risk level and corresponding probability are output. When the high-risk probability is not lower than the preset value, it is determined as suspected autism. The report includes the risk level and determination basis, core abnormal features, and abnormal pattern analysis. The report is pushed to the clinician and parent end simultaneously, supporting manual review and supplementary explanation.

[0035] Further, in step S7, the risk level of the suspected child, the core abnormal features and the deep features are input into the intervention scheme generation model to generate a personalized intervention scheme including a child training plan, a family guidance strategy and a phased goal. Three core evaluation nodes are set according to the intervention process. The short-term evaluation focuses on the immediate response of behavior, the medium-term evaluation evaluates the physiological-behavioral improvement, and the long-term evaluation judges the long-term prognosis trend. The interval between each node is set to be flexibly adjusted. Real-time data of the child is collected at each node to generate real-time deep features, which are input into the efficacy evaluation model to output multi-dimensional evaluation results: behavior improvement rate, physiological-behavioral improvement degree and long-term prognosis trend. If the result of any evaluation dimension is improved, the scheme optimization process is automatically triggered: the core features that are not improved are located by the efficacy evaluation model, and the intervention scheme generation model is called to regenerate the optimization strategy. The optimized scheme needs to retain the modules that have been verified effective in the original scheme.

[0036] Beneficial effects: reduce the subjectivity of diagnosis and improve accuracy: by collecting multi-dimensional data of behavior, physiology, environment and family, combining deep feature mining of internal patterns, and using multi-modal fusion diagnosis model analysis, the dependence on doctor's subjective experience and single scale is reduced; the model is optimized by cross-validation to improve the classification accuracy, and the diagnosis report contains core abnormal features and judgment basis, which supports artificial review and further reduces the risk of missed diagnosis and misdiagnosis. Strengthen the pertinence of intervention scheme: based on the diagnosis results and deep features, an intervention scheme generation model is constructed using a reinforcement learning framework, with “behavior improvement rate + physiological index compliance rate + family adaptation degree” as the reward function, to generate a personalized scheme including a child training plan and a family guidance strategy, avoiding the limitations of general schemes and adapting to different core abnormal features and family environments of children.

[0037] Perfect the efficacy evaluation system and realize the intervention-evaluation closed loop: the efficacy evaluation model combines the deep feature difference analysis before and after intervention to evaluate from multiple dimensions of “behavior improvement rate, physiological-behavioral improvement degree and long-term prognosis trend”, covering short-term, medium-term and long-term dynamic nodes; when not meeting the standard, the core problem is automatically located and the scheme is optimized, retaining effective modules, solving the problem of single evaluation dimension and the disconnection between intervention and evaluation in the prior art, and improving the effectiveness of intervention. Ensure data quality and comparability: standardized data collection specifications (such as unified evaluation scene and equipment model) are used throughout the process, and data is cleaned through human-machine collaborative correction, Z-score standardization, multiple imputation and other preprocessing methods to ensure the consistency of historical data and suspected child data, providing a reliable data basis for feature extraction, model training and diagnosis evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is a method flow diagram. DETAILED DESCRIPTION

[0039] In order to make the technical scheme of the present application clearer, the present application will be further described in detail below in combination with the drawings and specific embodiments.

[0040] Embodiment 1

[0041] Clinical application embodiment of the First Comprehensive Children's Hospital in a certain city

[0042] I. Preparation for implementation

[0043] Implementing subject: Development and Behavior Department of the First Comprehensive Children's Hospital in a certain city, equipped with 3 child psychological assessors, 2 data engineers, and 5 clinical doctors;

[0044] Equipment and tools: high-definition camera (resolution 1080P, frame rate 30fps, brand: Hikvision DS-2CD3T56WD-I5), medical-grade wearable bracelet (supports heart rate variability and skin conductance detection, sampling frequency 1 / hour, battery life 24 hours, brand: Lepu H10), electronic questionnaire system (developed based on WeChat mini program, containing 8 environmental and family core indicators), GPU server (NVIDIA A100, memory 64GB, framework: PyTorch 1.13).

[0045] II. Implementation steps

[0046] Step 1: Obtain historical case big data

[0047] Data source: Collect historical case data from January 2019 to December 2023, a total of 2000 cases, including 1000 cases of autism diagnosed children (age 1-3 years old, 680 males and 320 females, all meeting the DSM-5 autism diagnosis standard), and 1000 cases of normally developed children (age 1-3 years old, 520 males and 480 females, confirmed by Developmental Assessment Scale to have no developmental abnormalities);

[0048] Data content:

[0049] Behavioral data: Each child contains behavioral video under standardized scenes (parent-child interaction for 15 minutes, independent play for 10 minutes, group adaptation for 5 minutes), and 12 basic behavior indicators (such as social interaction frequency, eye contact duration, repetitive stereotyped behavior frequency, etc.) extracted therefrom;

[0050] Physiological data: 7-day continuous physiological data collected by wearable bracelet, including heart rate variability (normal range 30-50ms), skin conductance (normal range 1-5μS), sleep latency (normal range 15-30 minutes), and 6 other indicators;

[0051] Environmental and family data: Standardized scales filled by parents (Cronbach's a = 0.85) and 15-minute structured interview records, including 8 indicators such as parenting frequency (length of parent-child interaction per week), parents' autism cognition score (1-5 points), intervention compliance (1-5 points), etc.

[0052] Diagnosis and efficacy correlation data: DSM-5 diagnosis compliance of diagnosed children, intervention plan (such as social training duration, behavior correction frequency), efficacy evaluation records at each stage (2 weeks, 8 weeks, 16 weeks) (such as ABC scale score changes, physiological indicator compliance);

[0053] Data correlation: Correlate the above multi-dimensional data with "child case ID" as the unique identifier to form a complete historical database stored in the hospital data center.

[0054] Step 2: Historical data preprocessing

[0055] Behavior data preprocessing:

[0056] Video frame segmentation: Segment behavior video by 0.5 seconds per frame, generating 1.2 million frames of images;

[0057] Abnormal correction: Correct 2.3% of abnormal behavior annotations caused by device jitter and child obstruction through "machine initial labeling + manual review" (3 evaluators cross-verification);

[0058] Outliers and quantification: Remove 1.5% of behavior indicator outliers using the 3σ principle, supplement 0.8% of short-term missing data using interpolation method, and convert non-structured descriptions such as "few times of initiating social interaction" into 1-5 point quantitative scores;

[0059] Physiological data preprocessing:

[0060] Smoothing and noise reduction: Smooth physiological time series data using a 1-hour sliding window to filter out 3.1% of peak noise caused by motion interference;

[0061] Standardization and missing values: Standardize physiological indicators using Z-score algorithm (mean 0, standard deviation 1), fill in adjacent period mean for time series data with missing rate <3%, and mark 28 samples with missing rate ≥3% separately (excluded in subsequent feature extraction);

[0062] Environmental and family data preprocessing:

[0063] Quantification and verification: Convert qualitative descriptions such as "high compliance" and "moderate cognition" into 1-5 point quantitative values, perform reliability and validity tests on scale data, and remove 2.1% of invalid items (such as logically inconsistent answers);

[0064] Missing value processing: 1.2% of missing values were handled using multiple imputation method (based on sample characteristics of the same age group and family type);

[0065] Data integration and verification: After correlating the pre-processed multi-dimensional data, 1.5% of invalid samples were removed through logical verification (such as "sleep duration > 24 hours" and "negative behavior indicators"), and finally a structured sample set of 1970 examples was formed.

[0066] Step 3: Extracting deep features

[0067] 1: Basic feature extraction:

[0068] Behavioral dimension: Calculate 12 refined indicators such as "eye contact duration coefficient of variation (standard deviation, mean)" and "active language interval mean";

[0069] Physiological dimension: Calculate 12 indicators such as "heart rate variability diurnal difference (daytime mean - nighttime mean)" and "skin conductance response peak frequency (number of peak occurrences in 24 hours / 24)";

[0070] Environment and family dimension: Calculate 10 indicators such as "parental intervention execution rate fluctuation value (standard deviation of execution frequency per week)" and "monthly average of family education frequency"; a total of 34 basic features;

[0071] 2: Correlation feature extraction:

[0072] Calculate 10 linear correlation features using Pearson correlation coefficient, such as "negative correlation coefficient between repetitive stereotyped behavior frequency and heart rate variability" and "positive correlation coefficient between social interaction frequency and family education frequency";

[0073] Calculate 8 conditional correlation features using partial least squares regression, such as "conditional correlation coefficient between social interaction frequency and skin conductance response when parental cognitive score > 3"; a total of 18 correlation features;

[0074] 3: Prediction feature extraction:

[0075] Predict 6 behavioral and physiological prediction indicators based on LSTM model, such as "trend slope of social interaction frequency change after 1 week" and "probability of meeting heart rate variability standard after 2 weeks";

[0076] Calculate 4 prognosis prediction indicators based on gradient boosting regression tree, such as "contribution weight of family intervention cooperation degree to behavior improvement rate after 3 months" and "impact coefficient of parental cognitive score on physiological indicator recovery speed"; a total of 10 prediction features;

[0077] 4: Feature optimization: Variance Inflation Factor (VIF) was used to screen features (VIF < 5), and 2 redundant basic features were removed, finally forming a 50-dimensional deep feature set, and Z-score standardization was used to ensure feature comparability.

[0078] Step 4: Construction of multi-modal fusion model

[0079] Diagnosis model construction:

[0080] Architecture: "CNN (ResNet50) + RNN (LSTM) + Transformer" hybrid architecture was adopted, CNN extracted spatial features of behavior video frames, RNN extracted dynamic features of physiological time series data, and Transformer modeled global association between environmental features and three-level features;

[0081] Weight distribution: Through attention mechanism, the weight of three-level features was allocated (basic features 0.3, associated features 0.4, and predicted features 0.3);

[0082] Training and optimization: 5-fold cross-validation was used, and classification accuracy was used as the optimization objective, after training, the model accuracy on the validation set reached 95.2%, and the high-risk judgment accuracy was 94.8%;

[0083] Intervention scheme generation model construction:

[0084] Framework: Based on deep reinforcement learning (DQN algorithm), "risk level + deep features" as state input, "training duration (10-30 minutes / day), training frequency (1-3 times / day), parent guidance frequency (1-2 times / week)" as action space;

[0085] Reward function: Reward value = 0.4 x behavior improvement rate + 0.3 x physiological index compliance rate + 0.3 x family adaptation degree (full score is 1), the reward function of the model generated after training converges to 0.82;

[0086] Efficacy evaluation model construction:

[0087] Module addition: "Intervention before and after feature difference analysis module" was added based on the diagnosis model, and the feature standardization difference (difference, baseline standard deviation) was calculated;

[0088] Evaluation index: Output "behavior improvement rate (change mean of basic features)" "physiological-behavioral collaborative improvement degree (recovery rate of associated features)" "long-term prognosis trend (change slope of predicted features)", built-in evaluation threshold (based on historical effective samples 95% confidence interval): improvement rate ≥20% for "significant improvement", 5%-20% for "improvement", and <5% for "no improvement".

[0089] Step 5: Collect suspected children data

[0090] Sample selection: 50 suspected autism children (1.2-3 years old, 32 males and 18 females, all positive in M-CHAT screening scale) were selected from January to March 2024;

[0091] Data collection:

[0092] Behavioral data: 30-minute behavior videos (15 minutes of parent-child interaction, 10 minutes of independent play, and 5 minutes of group adaptation) were recorded in a standardized assessment room (temperature 23℃, noise 35dB, illumination 400lux), and 12 basic behavior indicators were extracted;

[0093] Physiological data: Wear the same type of wearable bracelet as the historical data, collect physiological data for 7 consecutive days, and parents record the child's daily activity log (such as 8:00-12:00 daytime activities, 19:00-7:00 sleep) simultaneously;

[0094] Environmental and family data: parents fill out standardized electronic questionnaires through WeChat mini-program and participate in 15-minute structured interviews to obtain 8 environmental and family indicators.

[0095] Step 6: Early diagnosis of suspected children

[0096] Data preprocessing: Strictly follow the preprocessing specifications of step S2, clean and standardize the data of 50 suspected children, and there is no invalid sample;

[0097] Deep feature extraction: Based on the rules of step S3, 50-dimensional deep features of 50 children are generated;

[0098] Diagnosis results: Input the deep features into the diagnosis model, output the risk level: 28 cases of high risk (probability ≥0.8, judged as "suspected autism"), 15 cases of medium risk (probability 0.5-0.8), and 7 cases of low risk (probability <0.5);

[0099] Report generation: The diagnosis report contains risk level (such as "high risk, probability 0.86"), core abnormal features (such as "repetitive stereotyped behavior frequency and heart rate variability negative correlation coefficient -0.62, higher than normal range -0.3~-0.1"), and abnormal pattern analysis (such as "mainly associated with behavior-physiological abnormalities, family intervention compliance meets the standard"), which is pushed to the clinician and parent end simultaneously, and there is no adjustment after the doctor reviews.

[0100] Step 7: Individualized intervention and dynamic evaluation

[0101] Intervention plan generation:

[0102] High-risk 28 cases: 16 cases of "insufficient social interaction", generate "20 minutes of daily parent-child role-playing training (1 time / day) + 2 times of parent guidance per week"; 22 cases of "repeated stereotyped behavior prominent", generate "15 minutes of daily alternative behavior training (puzzle, building blocks, 2 times / day) + parent daily check-in reminders";

[0103] Medium-risk 15 cases: generate "15 minutes of daily social game training (1 time / day) + 1 time of parent guidance per week";

[0104] Evaluation nodes: short-term (2nd week of intervention), medium-term (8th week), long-term (16th week);

[0105] Evaluation results:

[0106] Short-term evaluation: Among the 28 high-risk children, 18 cases had a behavior improvement rate of ≥10% (such as social interaction frequency increasing from 2 times / day to 4 times / day), and 15 cases had a physiological-behavioral coordination improvement degree of ≥40%; for the 7 children who did not improve (improvement rate <5%), the optimized solution was "increasing training duration to 25 minutes / day";

[0107] Medium-term evaluation: Among the 28 high-risk children, 25 cases had a behavior improvement rate of ≥20%, 23 cases had a coordination improvement degree of ≥60%, and 24 cases had a long-term prognosis trend of "good"; for the 3 children with low family adaptation, the parent guidance class was adjusted to "one-on-one online guidance";

[0108] Long-term evaluation: Among the 28 high-risk children, 26 cases had a behavior improvement rate of ≥30%, 27 cases had a coordination improvement degree of ≥80%, 28 cases had a prognosis trend of "good", and the intervention effectiveness was 100%; 15 medium-risk children were converted to low-risk, and 7 low-risk children had no developmental abnormalities.

[0109] III. Implementation Conclusion

[0110] This example builds a model through 2000 historical data to achieve precise diagnosis of 50 suspected children (high-risk judgment accuracy rate 94.8%), and after 16 weeks of intervention, the high-risk children intervention effectiveness is 100%, verifying the feasibility and effectiveness of the method.

[0111] Example 2:

[0112] Application of community screening in a certain provincial maternal and child health hospital

[0113] I. Implementation Preparation

[0114] Implementation subject: A certain provincial maternal and child health hospital community health service center, equipped with 4 child health doctors, 1 data specialist, and 2 social workers;

[0115] Devices and tools: high-definition camera (resolution 1080P, frame rate 30fps, brand: Dahua DH-IPC-HFW5249E), portable medical-grade wearable bracelet (supports heart rate variability detection, sampling frequency 1 / hour, battery life 36 hours, brand: YUWELLYH-106), electronic questionnaire system (based on Alipay applet development), cloud server (Alibaba Cloud ECS, GPU: T4, memory 32GB, framework: TensorFlow2.10).

[0116] II. Implementation steps

[0117] Step 1: Obtain historical case big data

[0118] Data sources: Collect historical case data from January 2020 to March 2024 of the Provincial Maternal and Child Health Hospital, a total of 2500 cases, including 1200 cases of autism diagnosed children (age 1-3 years old, 820 boys and 380 girls, meeting the DSM-5 standard), 1300 cases of normally developed children (age 1-3 years old, 680 boys and 620 girls, confirmed normal by Denver Development Screening Test);

[0119] Data content:

[0120] Behavioral data: Each case contains behavioral video and 12 basic behavior indicators under standardized scenarios (parent-child interaction for 15 minutes, toy game for 10 minutes, group activity for 5 minutes);

[0121] Physiological data: 7-day continuous physiological data collected by wearable bracelet, including heart rate variability, skin electrical response, deep sleep proportion (normal range 20%-30%) and other 6 indicators;

[0122] Environmental and family data: Standardized scale (Cronbach's α = 0.88) and 10-minute structured interview records filled out by parents, including 8 indicators such as family living environment (1-5 points), parent intervention willingness (1-5 points);

[0123] Diagnosis and efficacy correlation data: Diagnosis basis, intervention plan, and efficacy records at each stage (such as CARS scale score changes) of diagnosed children;

[0124] Data correlation: Use "child health record ID" as the unique identifier to correlate multi-dimensional data and form a historical database.

[0125] Step 2: Historical data preprocessing

[0126] Behavioral data preprocessing: 0.5 seconds per frame to segment video, 3.1% abnormal annotation corrected by human-computer collaboration, 1.8% outliers removed by 3σ principle, 1.1% missing data supplemented by interpolation;

[0127] Physiological data preprocessing: 1-hour sliding window smoothing, Z-score standardization, missing rate <3% filled with mean, 35 samples with missing rate ≥3% marked for exclusion;

[0128] Environmental and family data preprocessing: qualitative to quantitative (1-5 points), 1.9% invalid items were removed for reliability test, and 1.5% missing values were processed by multiple imputation;

[0129] Integration verification: logical verification after association data, 1.7% invalid samples were removed to form a structured sample set of 2458 cases.

[0130] Step 3: Extract deep features

[0131] 1: Basic features: 12 behavioral dimensions (such as "active language interval standard deviation"), 12 physiological dimensions (such as "heart rate variability day-to-day fluctuation value"), 10 environmental and family dimensions (such as "parental intervention willingness week change value"), a total of 34 items;

[0132] 2: Correlation features: Pearson correlation coefficient calculation of 10 linear correlations (such as "social gaze duration and family interaction frequency positive correlation coefficient"), partial least squares regression calculation of 8 conditional correlations, a total of 18;

[0133] 3: Prediction features: LSTM prediction of 6 items (such as "2-week repeated stereotyped behavior frequency change slope"), gradient boosting regression tree calculation of 4 items (such as "family environment contribution weight to 6-month physiological index compliance rate"), a total of 10 items;

[0134] 4: Optimization: VIF <5 screening, 2 redundant features are removed to form a 50-dimensional deep feature set, and standardized processing is performed.

[0135] Step 4: Build a multi-modal fusion model

[0136] Diagnosis model: "CNN (MobileNetV3) + RNN (GRU) + Transformer" architecture, three-level feature weight (basic 0.3, correlation 0.4, prediction 0.3), 5-fold cross-validation, validation set accuracy 94.9%, high-risk accuracy 94.5%;

[0137] Intervention program generation model: DQN reinforcement learning framework, action space includes "training type (social, behavior correction), duration (15-25 minutes), guidance frequency (1-2 times / week)", reward function = 0.45 x behavior improvement rate + 0.25 x physiological compliance rate + 0.3 x family adaptation degree, convergence reward value 0.81;

[0138] Efficacy evaluation model: contains feature difference analysis module, evaluation threshold is the same as example 1.

[0139] Step 5: Collecting suspected children data

[0140] Sample selection: 60 suspected children with autism (1-3 years old, 38 males and 22 females, M-CHAT positive) from community screening from April to June 2024;

[0141] Data collection:

[0142] Behavioral data: Record 30-minute video in community assessment room (temperature 24℃, noise 38dB, illumination 350lux), extract 12 behavioral indicators;

[0143] Physiological data: Wear portable bracelet for continuous 7-day collection, parents record activity log;

[0144] Environmental and family data: Alipay app questionnaire + 10-minute interview.

[0145] Step 6: Early diagnosis of suspected children

[0146] Preprocessing: No invalid samples;

[0147] Feature extraction: 50-dimensional deep features of 60 children;

[0148] Diagnosis results: 32 cases of high risk (probability ≥0.8), 18 cases of medium risk, and 10 cases of low risk;

[0149] Report: Clear abnormal features (such as "social interaction frequency and parent intervention willingness positive correlation coefficient 0.21, lower than normal range 0.4-0.6"), no adjustment by doctor review.

[0150] Step 7: Individualized intervention and dynamic assessment

[0151] Program generation: Among the 32 high-risk cases, 20 cases of "social inadequacy" use "18-minute daily parent-child game (1 time / day) + parent guidance 2 times / week", and 25 cases of "stereotyped behavior" use "12-minute daily replacement training (2 times / day)";

[0152] Evaluation nodes: Short-term (2 weeks), medium-term (8 weeks), and long-term (16 weeks);

[0153] Evaluation results:

[0154] Short-term evaluation: 22 cases of high-risk behavior improvement rate ≥10%, 19 cases of cooperative improvement degree ≥40%, and 8 cases of no improvement increased training frequency to 2 times / day;

[0155] Medium-term evaluation: 29 cases of high-risk behavior improvement rate ≥20%, 27 cases of cooperative improvement degree ≥60%, and 3 cases of low family adaptation degree adjusted guidance form to "offline guidance";

[0156] Long-term evaluation: 30 cases of high-risk behavior improvement rate ≥ 30%, 31 cases of synergistic improvement degree ≥ 80%, intervention efficiency 96.9%; 18 cases of medium risk to low risk, 10 cases of low risk without abnormality.

[0157] III. Implementation conclusion

[0158] In the community scenario, the model is constructed by 2500 historical data, the diagnosis accuracy of 60 suspected children is 94.5%, the intervention efficiency of high-risk children is 96.9%, and the applicability of the method in the primary medical scene is verified.

[0159] The above-described embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of protection of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for early diagnosis and intervention efficacy evaluation of autistic children, characterized in that, Specifically comprising the following steps: S1, obtaining historical case big data: collecting historical case big data of children with autism and normally developed children and corresponding diagnosis results and intervention diagnosis and treatment records, and associating to form a historical database; S2, historical data preprocessing: cleaning, standardizing and missing value processing of the historical case big data, and associating and integrating into a structured sample set; S3, extracting deep features: extracting basic features, associated features and prediction features from the preprocessed data, extracting deep features that can represent the internal pattern of the data through multi-level feature engineering and data mining technology, and forming a unified deep feature set; S4, constructing a multi-modal fusion model: constructing a diagnosis model by convolutional neural network + recurrent neural network + Transformer, constructing an intervention scheme generation model using reinforcement learning framework, and constructing an efficacy evaluation model based on the diagnosis model; S5, collecting suspected child data: collecting real-time big data of suspected children with autism according to the standardized specification consistent with the historical data collection; S6, early diagnosis of suspected children: preprocessing suspected child big data, extracting deep features, inputting the diagnosis model to output risk level, and generating a report containing core abnormal features; S7, personalized intervention and dynamic evaluation: generating an intervention plan, setting short-term, medium-term and long-term evaluation nodes, collecting data to evaluate efficacy, and optimizing the plan if it does not meet the standard.

2. The method for early diagnosis and intervention effect evaluation of autistic children according to claim 1, characterized in that: In the step S1, the case big data includes: behavior data, physiological data, environmental and family data, diagnosis and efficacy associated historical data; Behavior data: objective records of children in pre-set standardized evaluation scenarios, including video images and basic behavior indicators extracted therefrom; Physiological data: physiological signals and time series records collected by medical-grade wearable devices, including autonomic nervous function indicators and sleep structure indicators; Environmental and family data: information obtained through standardized scales and structured interviews, including family upbringing patterns, parent cognitive level and intervention cooperation degree; Diagnosis and efficacy associated data: clinical diagnosis basis of the sample, specific scheme used in the intervention process and efficacy evaluation results at each stage; All data are associated through individual unique identification to form a complete historical database containing original records, intermediate indicators and final results.

3. The method for early diagnosis and intervention effect evaluation of autistic children according to claim 2, characterized in that: In the step S2, when preprocessing the behavior data, the video images are frame-level segmented, and the abnormal behavior labels caused by device shaking and occlusion are corrected through human-machine cooperation; the basic behavior indicators are subjected to outlier detection, and the short-time missing data is supplemented by interpolation method; the unstructured behavior description is converted into quantitative scores; When preprocessing the physiological data, the continuous time series physiological data are smoothed by sliding window method to filter out the peak noise caused by motion interference; different physiological indicators are converted to a unified scale by Z-score standardization algorithm; the time series data with a missing rate lower than the preset value are filled with the mean value of adjacent periods, and the samples with a missing rate not lower than the preset value are marked and processed separately; During the preprocessing of the environmental and family data, qualitative descriptions are converted into quantitative values; the scale data is subjected to reliability and validity tests, invalid items are removed, and multiple imputation methods are used to handle missing values; Based on the individual unique identifier, the preprocessed behavior data, physiological data, environmental and family data, and diagnosis and efficacy correlation history data are associated into structured samples, invalid samples are removed through logical verification, and finally a historical sample set directly usable for feature extraction is formed.

4. The method for early diagnosis and intervention effect evaluation of autistic children according to claim 3, characterized in that: The step S3 specifically includes the following steps: S31, basic features: fine processing is performed on the preprocessed behavior data, physiological data, environmental and family data, statistical transformations including coefficient of variation, peak frequency, and diurnal difference, distribution feature analysis including quantile and skewness, and high-dimensional indexes are generated; S32, correlation features: cross-modal correlation analysis techniques including Pearson correlation coefficient and partial least squares regression are used to quantify the internal correlation patterns between behavior, physiology, environment, and family dimensions, including linear correlation and conditional correlation; S33, prediction features: time series prediction models including but not limited to long short-term memory network (LSTM) and gradient boosting regression tree are used to extrapolate trends from historical data, and indexes that can predict long-term outcomes are extracted, including future change slope of behavior indexes, sensitivity coefficient of physiological indexes to intervention response, and contribution weight of family factors to long-term prognosis; S34, redundant features are removed through feature selection algorithms, and finally a deep feature set including basic features, correlation features, and prediction features is formed, and standardization processing is performed to ensure comparability between features.

5. The method for early diagnosis and intervention efficacy evaluation of autistic children according to claim 4, characterized in that: In the step S4, the diagnosis model takes the deep features as input, uses a hybrid architecture of convolutional neural network to extract behavior-related feature space patterns, recurrent neural network to extract physiology-related feature time series patterns, and Transformer to extract global correlations between environmental features and tertiary features, dynamically allocates weights to the tertiary features through an attention mechanism, outputs autism risk levels and corresponding probabilities, and uses cross-validation for model training, with the optimization goal being to maximize classification accuracy; The intervention scheme generation model takes the diagnosis results and deep features as input, uses a reinforcement learning structure, takes intervention strategy parameters as the action space, calculates the weighted sum of behavior improvement rate, physiological index compliance rate, and family adaptability as the reward function, and trains the optimal intervention strategy through a policy gradient algorithm; The efficacy evaluation model takes real-time data and deep features during the intervention process as input, adds a deep feature difference analysis module before and after the intervention based on the diagnosis model, calculates the standardized difference between the deep features before and after the intervention, outputs the behavior improvement rate based on the basic features, the physiological-behavior system improvement degree based on the correlation features, and the long-term prognosis trend based on the prediction features, and has built-in evaluation thresholds to automatically determine significant improvement, improvement, and no improvement.

6. The method for early diagnosis and intervention efficacy evaluation of autistic children according to claim 5, characterized in that: In the step S5, the suspected child big data includes: Behavior data: in a standardized evaluation scene consistent with historical data, behavior videos are recorded through high-definition cameras, and consistent basic behavior indexes with historical data are extracted synchronously; Physiological data: physiological indicators are collected by using the same type of wearable device as the historical data, and a daily activity log of the child is recorded during the collection period; Environmental and family data: information including family environment, parent cognition, and willingness to cooperate with intervention is obtained through standardized electronic questionnaires and structured interviews consistent with historical data.

7. The method for early diagnosis and intervention effect evaluation of autistic children according to claim 6, characterized in that: In step S6, the suspected child big data is preprocessed strictly following the preprocessing specifications in step S2, and the behavior, physiology, environment, and family data of the suspected child are cleaned, standardized, and integrated. Deep feature extraction is based on the deep feature extraction rules in step S3, and deep features are extracted from the preprocessed suspected child data. The deep features of the suspected child are input into the constructed multi-modal fusion diagnosis model, and the risk level and corresponding probability are output. When the high-risk probability is not less than the preset value, it is determined that the suspected child has autism. The report includes risk level, judgment basis, core abnormal features, and abnormal pattern analysis. The report is pushed to the clinician and parent end simultaneously to support manual review and supplementary explanation.

8. The method for early diagnosis and intervention efficacy evaluation of autistic children according to claim 7, characterized in that: In step S7, the risk level, core abnormal features, and deep features of the suspected child are input into the intervention scheme generation model to generate a personalized intervention scheme including child training plan, family guidance strategy, and phase target. Three core evaluation nodes are set according to the intervention process. Short-term evaluation focuses on immediate response of behavior, medium-term evaluation evaluates the improvement of physiology-behavior coordination, and long-term evaluation judges the long-term prognosis trend. The interval between nodes is set with flexible adjustment time. Real-time data of the child is collected at each node to generate real-time deep features, which are input into the efficacy evaluation model to output multi-dimensional evaluation results: behavior improvement rate, physiology-behavior coordination improvement degree, and long-term prognosis trend. If the result of any evaluation dimension is improvement, the scheme optimization process is automatically triggered: the core features that have not improved are located by the efficacy evaluation model, the intervention scheme generation model is called to regenerate the optimization strategy, and the optimized scheme needs to retain the verified effective modules in the original scheme.

Images