Child obstructive sleep apnea prediction system

Abstract

The invention provides a children obstructive sleep apnea prediction system which comprises a data screening and preprocessing module, a feature selection and construction module, a children OSA prediction model construction module, a model prediction module and a prediction result analysis module. According to the system, based on conventional physical examination data, sleep questionnaire data, blood oxygen monitoring data and other data which are easy to obtain and not prone to interference, the robustness of biological signal collection is improved, meanwhile, initial features are established through an antagonism training strategy, and the reliability of a prediction model is improved; different input features are established through independent or random combination, the importance of each input feature in predicting children OSA is combined, an optimal basic model and hyper-parameters are determined, and a children OSA prediction model is established, so that the prediction accuracy of children OSA is remarkably improved, SHAP analysis is provided to meet the decision support requirements of clinical diagnosis and treatment, and the clinical diagnosis and treatment efficiency is improved. The early screening of the obstructive sleep apnea disease of children is realized.

Description

Technical Field

[0001] This invention relates to the field of sleep apnea prediction, and more particularly to a prediction system for obstructive sleep apnea in children. Background Technology

[0002] Obstructive sleep apnea (OSA) in children is a disorder caused by repeated partial or complete obstruction of the upper airway during sleep. This obstruction not only interferes with normal ventilation and sleep structure in children but also triggers a series of pathophysiological changes. Compared with adults, children's OSA has the following characteristics: 1. The etiology of children's OSA is complex, the symptoms are diverse, and the manifestations of the disease are varied and ambiguous; the symptoms are not as pronounced as in adults. 2. The symptoms and causes of children's OSA change dynamically during their growth and development. Therefore, common screening tools, such as sleep scales, are not very effective in screening for children's OSA.

[0003] The gold standard for diagnosing childhood obstructive apnea (OSA) is polysomnography (PSG). PSG collects physiological signals such as electroencephalogram (EEG), electrooculogram (EOG), mandibular electromyography (EMG), electrocardiogram (ECG), nasal and oral airflow, respiratory movements, and blood oxygen saturation during sleep to obtain the obstructive apnea-hypopnea index (OAHI). An OAHI > 1 breath / hour is diagnostic of OSA. However, the diagnostic rate of childhood OSA is low because hospitals have a limited number of PSG devices, the operation is complex, and the number of qualified sleep technicians is insufficient to meet the diagnostic needs of all patients. Furthermore, PSG channels are numerous and prone to dislodgement, requiring overnight supervision by professional technicians. Marking PSG signals also takes at least 3 hours, consuming significant personnel costs and resulting in high PSG costs, placing a considerable medical burden on families and society. Therefore, designing a simple, portable, low-cost, and highly accurate childhood OSA prediction system is crucial for rapidly screening high-risk OSA patients and reducing the social burden.

[0004] Although a variety of portable OSA prediction systems and devices have been proposed, there are still some limitations: (1) Existing prediction systems rely on complex detection equipment. For example, the invention patent with publication number CN117064333A discloses a primary screening device for obstructive sleep apnea-hypopnea syndrome, including a network training module and an OSAHS primary screening module. By constructing a deep learning model based on EEG, ECG and respiratory signals, it realizes sleep staging, awakening judgment and respiratory event detection. However, it still relies on complex multimodal physiological signals. The equipment is expensive and complicated to operate, making it difficult to achieve home-based screening.

[0005] (2) Existing prediction systems have poor robustness under non-ideal conditions. For example, the invention patent with publication number CN111696575A discloses a hypopnea and apnea detection and identification system based on a hybrid neural network model. The hybrid neural network model is used to classify the patient's normal snoring, hypopnea snoring and apnea snoring to obtain the AHI index. However, it is greatly affected by environmental noise and is not suitable for OSA patients without snoring symptoms.

[0006] (3) Most prediction systems are based on deep learning, so the prediction results have a “black box” problem, and the interpretability of the etiology of obstructive sleep apnea in children is poor. For example, the invention patent with publication number CN117084691 discloses a method for predicting the risk level of sleep apnea based on electrocardiogram signals. A convolutional neural network model was established based on electrocardiogram signals to achieve OSA risk level prediction. However, neural networks generally have a “black box” problem, the model has weak interpretability, and it is difficult to meet the decision support needs of clinical diagnosis and treatment.

[0007] In summary, this study focuses on screening for obstructive sleep apnea in children, reducing reliance on complex monitoring equipment, enabling large-scale and rapid screening, improving the robustness of screening, and providing interpretability of the prediction system to meet the decision support needs of clinical diagnosis and treatment. It aims to construct a prediction system that combines optimal algorithms and hyperparameters to achieve early screening for obstructive sleep apnea in children. Summary of the Invention

[0008] The purpose of this invention is to provide a childhood obstructive sleep apnea prediction system. This prediction system focuses on children and uses easily accessible and less susceptible data sets, such as routine physical examination data, sleep questionnaire data, and blood oxygen monitoring data, as its sample set. It employs machine learning algorithms, combined with feature engineering techniques and nested cross-validation, to select the optimal algorithm and hyperparameters. Furthermore, it establishes a childhood OSA prediction model through random combinations of features, enabling early screening of childhood obstructive sleep apnea. The system also provides a SHAP analysis module to visualize feature contribution, meeting the decision support needs of clinical diagnosis and treatment.

[0009] An embodiment provides a childhood obstructive sleep apnea prediction system, comprising: The data filtering and preprocessing module is used to collect children's basic information, physical examination information, sleep questionnaire data and blood oxygen saturation, and to filter the collected data to obtain initial features including basic information, physical examination information, sleep-related symptoms and blood oxygen saturation, and to preprocess the initial features. The feature selection and construction module takes the preprocessed initial features as input and whether OSA is present as output, establishes a univariate logistic regression equation to filter out features less than a threshold, constructs new features based on features related to growth and development, and adds noise perturbation to the new features to obtain the noisy features. The module for building a child OSA prediction model comprises three sub-modules: an optimal algorithm identification sub-module, which uses noisy features as input and trains the model using nested cross-validation and grid search to identify the optimal algorithm; an optimal hyperparameter identification sub-module, which uses the optimal algorithm and noisy features as input and trains the model using cross-validation and Bootstrap to identify the hyperparameters with the best AUC; and a feature combination identification sub-module, which uses noisy features as input, randomly combines the input features, and builds multiple child OSA prediction sub-models based on the optimal algorithm and optimal hyperparameters. The sub-models are then trained using the feature combination with the best AUC and the optimal hyperparameters to construct the final child OSA prediction model. The model prediction module is used to input basic information, physical examination information, sleep-related symptoms and blood oxygen saturation into the final childhood OSA prediction model, and output the probability of having OSA, predicting whether a child has OSA.

[0010] In one embodiment, the childhood obstructive sleep apnea prediction system further includes a prediction result analysis module, which includes an overall SHAP analysis submodule and an age-stratified SHAP analysis submodule. The overall SHAP analysis submodule is used to calculate the average Shapley value of each input feature on the child OSA sample based on the final child OSA prediction model using the SHAP algorithm, and to analyze whether each input feature is an overall risk factor for child OSA and its corresponding weight. The age-stratified SHAP analysis submodule is used to stratify children by age and build OSA prediction models for each child based on the age stratification results. and The SHAP algorithm was used to analyze the child OSA prediction model. and Feature importance analysis was conducted to analyze the main OSA risk factors in children across age strata and to identify the main characteristics of OSA in children across each age stratum.

[0011] In one embodiment, the data filtering and preprocessing module filters the collected data to obtain initial features including basic information, physical examination information, sleep-related symptoms, and blood oxygen saturation. This includes: establishing mapping rules based on the collected sleep questionnaire data, mapping the results of each question in the sleep questionnaire data to binary symptoms, with the mapping result being 0 or 1, representing the presence or absence of the symptoms, respectively; using the mapped symptoms as sleep-related symptoms, and combining them with the basic information, physical examination information, sleep questionnaire data, and blood oxygen saturation to form the initial features.

[0012] In one embodiment, the binary symptoms to be mapped are snoring, sleep apnea, mouth breathing, restless sleep, bedwetting, daytime sleepiness, attention deficit or hyperactivity, recurrent awakenings, and emotional and behavioral abnormalities.

[0013] In one embodiment, the mapping rule established based on the collected sleep questionnaire data includes: The sleep questionnaire data used were OSA-18 and PSQ children's sleep questionnaire. The score for each question in OSA-18 was 0-6, and the score for each question in PSQ was 0 and 1. For symptoms If OSA-18 includes the corresponding issue Ai The mapping rules are calculated as follows: , In the formula, Indicating the first in OSA-18 The original scores for each question This represents the median score of all samples for that question; For symptoms If the PSQ contains the corresponding question Bj The mapping rules are calculated as follows: , In the formula, Indicates the first in PSQ The original scores for each question; Final symptoms The condition for its existence is that the above binary classification mapping result for any related problem is 1, and the calculation formula is as follows: .

[0014] Furthermore, the initial features include: (1) Basic information: age, gender; (2) Physical examination information: height, weight, BMI, growth and development status, neck circumference, waist circumference, hip circumference, tonsil grading, among which growth and development status includes: thin, normal, overweight or obese. (3) Sleep-related symptoms: snoring, sleep apnea, mouth breathing, restless sleep, bedwetting, daytime sleepiness, attention deficit or hyperactivity, recurrent awakening, and abnormal emotional and behavioral patterns; (4) Blood oxygen saturation: percentage of sleep time with blood oxygen saturation <95%, <90%, <85%, <80%, <75% and <70%, percentage of sleep time with blood oxygen saturation ≤95%, ≤92%, ≤90%, ≤85%, ≤80%, ≤75% and ≤70%, average blood oxygen saturation during NREM sleep, average blood oxygen saturation during REM sleep, average blood oxygen saturation during sleep, oxygen saturation reduction rate of >=3% per hour during sleep, oxygen saturation reduction rate of >=4% per hour during sleep, lowest blood oxygen saturation during NREM sleep, lowest blood oxygen saturation during REM sleep, lowest blood oxygen saturation, and decrease in average oxygen saturation.

[0015] In one embodiment, the data filtering and preprocessing module performs initial feature preprocessing, including data standardization, missing value imputation, and data balancing. The data standardization is used to divide the initial features into categorical features and continuous features. The categorical features include: gender, growth and development status, tonsil grade, and various symptoms in sleep-related symptoms. The rest are continuous features. For continuous features, Z-Score is used for data standardization. The missing value imputation is used to fill in data in the initial features with a missing ratio of less than 30% by employing the k-nearest neighbor algorithm; The data balancing is used to synthesize minority class samples using a synthetic minority oversampling method, for each minority class sample In the feature space, find the Euclidean distance. of Based on the nearest neighbor samples, and according to the upsampling factor, from Randomly select a sample from its nearest neighbors and compare it with... Synthesize new samples together and repeat. This time, negative and positive samples from OSA were obtained to achieve data balance.

[0016] In one embodiment, the feature selection and construction module filters out features below a threshold by establishing a univariate logistic regression equation, including: using the univariate logistic regression equation to filter out variables significantly related to OSA from the preprocessed initial features, calculated using the following formula: , In the formula, Indicates the independent variable Below, dependent variable The probability of the dependent variable, where the dependent variable is... The definition is: when OSA is positive, When OSA is negative, , and The regression coefficients obtained from training reflect the independent variables. For dependent variable The impact; Based on the calculated p-values ​​of each variable in logistic regression, the significance of each variable on the result is determined, and features with p values ​​less than the threshold are selected. The selected features are considered to be significantly correlated with OSA.

[0017] More preferably, the threshold value is set to 0.05.

[0018] In one embodiment, in the feature selection and construction module, growth and development-related features are obtained through the following methods: Design matrices containing linear, quadratic, and intercept terms for age, respectively. A vector containing predictor variables: height, weight, neck circumference, waist circumference, and hip circumference. and coefficient matrix Modeling the changes of each predictor variable with age, expressed as: ,in Residual vector; Estimate the coefficient matrix using the least squares method. Then, the corrected residual vector is obtained. , is used to represent the corrected features of each predictor variable after removing the influence of age, resulting in a combination of features related to growth and development.

[0019] Furthermore, the modified feature vector includes: modified height, modified weight, modified neck circumference, modified waist circumference, and modified hip circumference.

[0020] In one embodiment, the feature selection and construction module selects features with p less than a threshold and features related to growth and development to construct new features, including: (1) Basic information: age, gender; (2) Physical examination information: height, weight, neck circumference, waist circumference, hip circumference; (3) Sleep-related symptoms: snoring, sleep apnea, mouth breathing, bedwetting, and recurrent awakenings; (4) Blood oxygen saturation: Percentage of sleep time with blood oxygen saturation <95% and <90%, percentage of sleep time with blood oxygen saturation ≤95%, ≤92%, and ≤90%, average blood oxygen saturation during NREM sleep, average blood oxygen saturation during REM sleep, average blood oxygen saturation during sleep, oxygen saturation reduction rate of >=3% per hour during sleep, oxygen saturation reduction rate of >=4% per hour during sleep, lowest blood oxygen saturation during NREM sleep, lowest blood oxygen saturation during REM sleep, lowest blood oxygen saturation, and decrease in average oxygen saturation.

[0021] In one embodiment, adding noise perturbation to a new feature to obtain a noisy feature includes: Binomial distribution perturbations were added to gender and sleep-related symptom characteristics; For continuous features, add Gaussian noise perturbation.

[0022] Furthermore, the features of the added binomial distribution perturbation include: (1) basic information: gender; (2) sleep-related symptoms: snoring, sleep apnea, mouth breathing, bedwetting, and recurrent awakenings; The features of adding Gaussian noise perturbation include: (1) basic information: age; (2) physical examination information: height, weight, neck circumference, waist circumference, hip circumference, corrected height, corrected weight, corrected neck circumference, corrected waist circumference, corrected hip circumference; (3) blood oxygen saturation: percentage of sleep time with blood oxygen saturation <95% and <90%, percentage of sleep time with blood oxygen saturation ≤95%, ≤92%, and ≤90%, average blood oxygen saturation during NREM sleep, average blood oxygen saturation during REM sleep, average blood oxygen saturation during sleep, oxygen saturation reduction rate of >=3% per hour during sleep, oxygen saturation reduction rate of >=4% per hour during sleep, minimum blood oxygen saturation during NREM sleep, minimum blood oxygen saturation during REM sleep, minimum blood oxygen saturation, and decrease in average oxygen saturation.

[0023] In one embodiment, the optimal algorithm identification submodule uses the noisy features as input, trains using nested cross-validation and grid search algorithms, and identifies the optimal algorithm, including: Using the noisy features as input, select an algorithm that can handle both continuous and categorical features; Based on the selected algorithm, internal cross-validation is applied in each fold of the external cross-validation, and grid search is combined to obtain the hyperparameters that perform best in the internal cross-validation. The best hyperparameters are then applied to the external cross-validation to select the best algorithm.

[0024] Furthermore, the selected algorithms include: extreme gradient boosting, support vector machine, random forest, lightweight gradient boosting machine, linear discriminant analysis, and K nearest neighbors.

[0025] Furthermore, the hyperparameters of the selected algorithm are tuned using grid search to identify the optimal algorithm, including: Extreme gradient boosting involves adjusting the tree depth, learning rate, number of trees, random sampling ratio, proportion of features randomly selected for each tree, and hyperparameters of the L2 regularization term. Adjusting the regularization parameter and the hyperparameters of the RBF kernel in support vector machines; In random forests, hyperparameters such as the number of trees, tree depth, minimum number of samples required for node splits, minimum number of samples required for leaf nodes, and maximum number of features are adjusted. In a lightweight gradient booster, hyperparameters are adjusted for the number of trees, tree depth, learning rate, number of leaf nodes, and minimum number of samples in a leaf node. In linear discriminant analysis, hyperparameters such as the solution method, shrinkage coefficient, class prior probability, and number of feature dimensions after dimensionality reduction are adjusted. In K-Nearest Neighbors, the number of nearest neighbors, the weight allocation method, and the hyperparameters of the nearest neighbor search algorithm are adjusted. Evaluate the AUC, accuracy, specificity, sensitivity, precision, and F1 score of the hyperparameter-tuned algorithm to identify the algorithm with the best AUC.

[0026] In one embodiment, the optimal hyperparameter identification submodule, using the noisy features as input, employs standard cross-validation and Bootstrap training to identify hyperparameters with the optimal AUC, including: The data is divided into training and test sets, using the noise-added features as input. Based on the training set, grid search and 10-fold cross-validation are used to search for the set of hyperparameters corresponding to the best algorithm. Based on the test set, the Bootstrap sampling method is used to train a set of hyperparameters. The AUC, accuracy, specificity, sensitivity, precision and F1 score of each set of hyperparameters are evaluated, and the hyperparameter with the best AUC is identified.

[0027] In one embodiment, the input features are randomly combined, and multiple child OSA prediction sub-models are established based on the optimal algorithm and optimal hyperparameters, including: Based on the noisy features of each input and considering the importance of each input feature in predicting childhood OSA, six feature combinations are established. ~ and the corresponding child OSA prediction sub-model ~ , where feature combination Includes basic information and physical examination information; feature combinations Including sleep-related symptoms; characteristic combinations Including blood oxygen saturation; characteristic combinations Includes basic information, physical examination information, and sleep-related symptoms; characteristic combinations Including sleep-related symptoms and blood oxygen saturation; characteristic combination This includes basic information, physical examination information, sleep-related symptoms, and blood oxygen saturation.

[0028] By constructing and training a child OSA prediction model using different features individually or in combination, we can gain a clearer understanding of the importance of each dimension or combination of features in predicting child OSA, which helps identify the contribution of features or feature combinations to model performance. Furthermore, in some cases, using all features may lead to an overly complex model, especially with limited data. Feature combination testing can help find a simpler yet still effective model, thereby improving the model's interpretability and generalization ability.

[0029] In one embodiment, the age-stratified SHAP analysis submodule stratifies children by age to obtain samples of children aged 3-7 and 7-14.

[0030] Compared with the prior art, the beneficial effects of the present invention include at least the following: (1) The present invention is based on children’s physical examination data, sleep questionnaire data and blood oxygen monitoring data. The data acquisition is simple, avoiding the use of expensive and complicated monitoring equipment. The cost of the prediction system is greatly reduced, and the operation is simple. It does not require professional technicians to perform complicated equipment operation.

[0031] (2) By selecting feature data that is easy to collect and not easily affected by environmental factors, the robustness of biological signal acquisition is improved; at the same time, adversarial training strategy is used to artificially inject noise and missing values ​​to establish initial features, which improves the robustness and reliability of the prediction model.

[0032] (3) Based on the initial features, different input features are established by individual or random combination. The importance of each input feature in predicting children’s OSA is combined, and the best algorithm and best hyperparameter identification are used to construct a prediction model adapted to children’s OSA, which significantly improves the prediction accuracy of children’s OSA and realizes early screening of obstructive sleep apnea in children.

[0033] (4) Based on the constructed pediatric OSA prediction model, SHAP analysis is further used to visualize the feature contribution, enhance the interpretability of the model, enhance the effectiveness of decision support, and meet the decision support needs of clinical diagnosis and treatment. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0035] Figure 1 This is a schematic diagram of the obstructive sleep apnea prediction system for children provided by the present invention.

[0036] Figure 2 This is a schematic diagram of the prediction result analysis module provided by the present invention.

[0037] Figure 3 The SHAP analysis diagrams for all samples in the overall SHAP analysis submodule provided for this embodiment.

[0038] Figure 4 The SHAP analysis diagram of the 3-7 year old samples in the age-stratified SHAP analysis submodule provided for the example.

[0039] Figure 5 The SHAP analysis diagram of the 7-14 year old sample in the age-stratified SHAP analysis submodule provided in the example. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and given in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0041] This invention provides a childhood obstructive sleep apnea prediction system, such as... Figure 1 As shown, the system includes the following modules: data filtering and preprocessing module, feature selection and construction module, children's OSA prediction model construction module, model prediction module, and prediction result analysis module. The first three modules complete the model development, and the fourth and fifth modules realize the model application.

[0042] In this embodiment, the development and application of a child OSA prediction model are illustrated as an example: In the data screening and preprocessing module, basic information about children, physical examination data, sleep questionnaire data, and blood oxygen saturation data are collected first, and the data are preprocessed, including blank value filling, data standardization, and data balancing.

[0043] In the feature selection and construction module, a univariate logistic regression equation is established with the above features as input and whether or not the patient has OSA as output. Features with p<0.05 are selected as input features for the model. Height, weight, neck circumference, waist circumference, and hip circumference are corrected according to age to obtain corrected height, corrected weight, corrected neck circumference, corrected waist circumference, and corrected hip circumference.

[0044] In the children's OSA prediction model building module, nested cross-validation and grid search algorithms are used to select the best-performing model from 6 machine learning models. Based on this model, ordinary cross-validation and Bootstrap are used to select the best feature combination from 6 feature combinations to build the model most suitable for this children's OSA dataset.

[0045] In the model prediction module, a user's basic information, physical examination information, sleep-related symptoms, and blood oxygen saturation are input into the prediction model to predict whether the user has OSA.

[0046] In the prediction results analysis module, SHAP analysis is performed on the entire dataset to understand the overall importance of characteristics in the child population and their impact on the results, facilitating children's prevention of common OSA risk characteristics. The module also segments children by age group to reveal how the importance of each characteristic indicator changes with age. Specifically: 1. Data filtering and preprocessing module, which is used to collect children's basic information, physical examination information, sleep questionnaire data and blood oxygen saturation, and filter the collected data to obtain initial features including basic information, physical examination information, sleep-related symptoms and blood oxygen saturation, and preprocess the initial features.

[0047] (1) Data filtering and mapping The Pediatric Adenotonsillectomy Trial for Snoring (PATS) and the Childhood Adenotonsillectomy Trial (CHAT) databases were used. These two databases compared behavioral, quality of life, and sleep monitoring data in children aged 3–14 years before and after adenoidectomy and tonsillectomy. This example incorporates basic information, physical examination data, sleep questionnaire data, and blood oxygen saturation data from the databases.

[0048] The sleep questionnaires used were the OSA-18 and the PSQ (Pediatric Sleep Questionnaire). In the OSA-18, each question was scored from 0 to 6, and in the PSQ, each question was scored between 0 and 1. Mapping rules were used to map the results of each question in the sleep questionnaire to binary symptoms. The binary symptoms to be mapped were snoring, sleep apnea, mouth breathing, restless sleep, bedwetting, daytime sleepiness, attention deficit or hyperactivity, recurrent awakenings, and emotional and behavioral abnormalities. The mapping result was 0 or 1, where 0 indicates the absence of the symptom and 1 indicates the presence of the symptom. The specific mapping rules were as follows: For symptoms If OSA-18 includes the corresponding issue Ai The mapping rules are calculated as follows: , In the formula, Indicating the first in OSA-18 The original scores for each question This represents the median score of all samples for this problem. The median score for the OSA-18 problem in this embodiment is shown in Table 1: Table 1 Median of OSA-18 Problems

[0049] For symptoms If the PSQ contains the corresponding question Bj The mapping rules are calculated as follows: , In the formula, Indicates the first in PSQ The original scores for each question; Ultimately, symptoms The condition for its existence is that the above binary classification mapping result for any related problem is 1, and the calculation formula is as follows: .

[0050] The initial features obtained from the preliminary database screening are divided into four dimensions. The symptoms mapped using the above method constitute the third dimension, namely sleep-related symptoms. These, along with basic information, physical examination information, sleep questionnaire data, and blood oxygen saturation, form the initial features, including: 1) Basic information: age, gender; 2) Physical examination information: height, weight, BMI, growth and development status, neck circumference, waist circumference, hip circumference, tonsil grade, among which growth and development status includes: underweight, normal, overweight or obese. 3) Sleep-related symptoms: snoring, sleep apnea, mouth breathing, restless sleep, bedwetting, daytime sleepiness, attention deficit or hyperactivity, recurrent awakenings, and abnormal mood and behavior; 4) Blood oxygen saturation: Percentage of sleep time with blood oxygen saturation <95%, <90%, <85%, <80%, <75%, and <70%; percentage of sleep time with blood oxygen saturation ≤95%, ≤92%, ≤90%, ≤85%, ≤80%, ≤75%, and ≤70%; average blood oxygen saturation during NREM sleep; average blood oxygen saturation during REM sleep; average blood oxygen saturation during sleep; oxygen saturation reduction rate per hour >=3% during sleep; oxygen saturation reduction rate per hour >=4% during sleep; lowest blood oxygen saturation during NREM sleep; lowest blood oxygen saturation during REM sleep; lowest blood oxygen saturation; and decrease in average oxygen saturation.

[0051] (2) Data preprocessing Data standardization: Initial features were divided into categorical features and continuous features. Categorical features included: gender, growth and development status, tonsil grade, and various symptoms from sleep-related symptoms. The remaining features were continuous features. For continuous features, Z-score data standardization was performed. , In the formula, The original data values ​​for continuous features. Continuous characteristics The mean, Continuous characteristics The standard deviation.

[0052] Missing value imputation: The database contains 1358 samples, of which 325 contain missing values. The k-nearest neighbor algorithm is used to impute data where the missing value ratio in the initial features is less than 30%. Specifically, for samples containing missing values... Calculate its comparison with all other samples The distance is calculated as follows: , In the formula, Indicates sample The missing feature dimensions of China and Africa Indicates sample Chinese characteristics The value, Indicates sample Chinese characteristics The value, That is, a sample With sample Euclidean distance on the non-missing feature dimension; Next, for the samples Find the closest distance, that is smallest Let the number of samples (neighbors) be denoted as the set: , In the formula, Representation and Sample The closest One sample, This means A set of samples; Finally, assuming the sample In the There are missing values ​​in each feature, using middle The first neighbor's Fill in the eigenvalues:

[0053] That is, for the sample of Features, calculate its The neighbor The average of the feature values; for categorical features, calculate its... The neighbor The mode of eigenvalues, where Indicates sample No. The value of each feature, belong In this embodiment, all of the above... Take 5.

[0054] Data balancing: After missing value imputation, all data are standardized again. Synthetic Minority Oversampling Technique (SMOTE) is used to synthesize minority class samples, achieving a balance between the number of minority and majority class samples. Specifically: For each minority class sample in the dataset In the feature space, find the Euclidean distance. of The nearest neighbor samples. Based on the upsampling factor. ,from Randomly select a sample from the nearest neighbors and compare it with... Synthesize new samples together and repeat. The synthesis formula is as follows: , In the formula, Minority class The first sample Each attribute value It is a random number between [0,1]. It is a sample The One nearest neighbor sample; Representative at and A new sample synthesized from these samples. In this embodiment, Take 5. Take 1.

[0055] After data balancing, there were 813 OSA-negative and 813 OSA-positive samples. The characteristics of the samples were as follows: 1) Basic information: age, gender; 2) Physical examination information: height, weight, BMI, growth and development status (underweight, normal, overweight, obese), neck circumference, waist circumference, hip circumference, tonsil classification; 3) Sleep-related symptoms: snoring, sleep apnea, mouth breathing, restless sleep, bedwetting, daytime sleepiness, attention deficit or hyperactivity, recurrent awakenings, and abnormal mood and behavior; 4) Blood oxygen saturation: Percentage of sleep time with blood oxygen saturation <95%, <90%, <85%, <80%, <75%, <70%; percentage of sleep time with blood oxygen saturation ≤95%, ≤92%, ≤90%, ≤85%, ≤80%, ≤75%, ≤70%; average blood oxygen saturation during NREM sleep; average blood oxygen saturation during REM sleep; average blood oxygen saturation during sleep; oxygen saturation reduction rate per hour >=3% during sleep; oxygen saturation reduction rate per hour >=4% during sleep; lowest blood oxygen saturation during NREM sleep; lowest blood oxygen saturation during REM sleep; lowest blood oxygen saturation; and decrease in average oxygen saturation.

[0056] 2. Feature selection and construction module: This module takes the preprocessed initial features as input and whether the user has OSA as output, establishes a univariate logistic regression equation to screen features that are less than a threshold, constructs new features based on features related to growth and development, and adds noise perturbation to the new features to obtain the noisy features.

[0057] (1) Feature Filtering Using univariate logistic regression, variables significantly associated with OSA are selected from the preprocessed initial features obtained from the data filtering and preprocessing module: , In the formula, Indicates the independent variable Below, dependent variable The probability of the dependent variable, where the dependent variable is... The definition is: when OSA is positive, When OSA is negative, , and The regression coefficients obtained from training reflect the independent variables. For dependent variable The impact; Based on the calculated p-values ​​of each variable in logistic regression, the significance of each variable on the results is determined, and features with p < 0.05 are selected, indicating that the selected features are significantly correlated with OSA.

[0058] The final effective features obtained from the screening, namely the variables that are significantly associated with OSA, include: 1) Basic information: age, gender; 2) Physical examination information: height, weight, neck circumference, waist circumference, hip circumference; 3) Sleep-related symptoms: snoring, sleep apnea, mouth breathing, bedwetting, and recurrent awakenings; 4) Blood oxygen saturation: Percentage of sleep time with blood oxygen saturation <95% and <90%, percentage of sleep time with blood oxygen saturation ≤95%, ≤92%, and ≤90%, average blood oxygen saturation during NREM sleep, average blood oxygen saturation during REM sleep, average blood oxygen saturation during sleep, oxygen saturation reduction rate of >=3% per hour during sleep, oxygen saturation reduction rate of >=4% per hour during sleep, lowest blood oxygen saturation during NREM sleep, lowest blood oxygen saturation during REM sleep, lowest blood oxygen saturation, and decrease in average oxygen saturation.

[0059] (2) Feature construction For growth and development-related characteristics, multinomial regression is used to address the feature shift problem caused by nonlinear changes in children's growth and development. Design includes age The matrix of first-order terms, second-order terms, and intercept terms : , The design includes the predictor variable height. ,weight Neck circumference ,waistline Hip circumference vector : , Design the coefficient matrix (each column corresponds to the regression coefficient of a variable): , Based on height For example, This represents the age intercept value obtained from the regression. The coefficient of the linear term representing age. The coefficient of the quadratic term representing age.

[0060] The changes of each variable with age are expressed as follows: , in Residual vector: , Estimate the coefficient matrix using the least squares method. Afterwards, the corrected residual vector for: , in, This refers to the vector of corrected variables after removing the influence of age from each growth and development variable, including: corrected height, corrected weight, corrected neck circumference, corrected waist circumference, and corrected hip circumference.

[0061] The constructed features include: 1) Basic information: age, gender; 2) Physical examination information: height, weight, neck circumference, waist circumference, hip circumference, corrected height, corrected weight, corrected neck circumference, corrected waist circumference, corrected hip circumference; 3) Sleep-related symptoms: snoring, sleep apnea, mouth breathing, bedwetting, and recurrent awakenings; 4) Blood oxygen saturation: Percentage of sleep time with blood oxygen saturation <95% and <90%, percentage of sleep time with blood oxygen saturation ≤95%, ≤92%, and ≤90%, average blood oxygen saturation during NREM sleep, average blood oxygen saturation during REM sleep, average blood oxygen saturation during sleep, oxygen saturation reduction rate of >=3% per hour during sleep, oxygen saturation reduction rate of >=4% per hour during sleep, lowest blood oxygen saturation during NREM sleep, lowest blood oxygen saturation during REM sleep, lowest blood oxygen saturation, and decrease in average oxygen saturation.

[0062] (3) Injection of artificial noise Introducing noise can improve the robustness of the model. In this embodiment, random noise was added to the dataset. For gender and sleep-related symptom features, binomial distribution perturbation was added; for continuous features, Gaussian noise perturbation was added.

[0063] Features with added binomial distribution perturbation include: (1) basic information: gender; (2) sleep-related symptoms: snoring, sleep apnea, mouth breathing, bedwetting, and recurrent awakenings. The calculation formula is: , In the formula, Represents a binary feature. This represents the binary feature after adding a binomial distribution.

[0064] The features of adding Gaussian noise perturbation include: (1) basic information: age; (2) physical examination information: height, weight, neck circumference, waist circumference, hip circumference, corrected height, corrected weight, corrected neck circumference, corrected waist circumference, corrected hip circumference; (3) blood oxygen saturation: percentage of sleep time with blood oxygen saturation <95% and <90%, percentage of sleep time with blood oxygen saturation ≤95%, ≤92%, and ≤90%, average blood oxygen saturation during NREM sleep, average blood oxygen saturation during REM sleep, average blood oxygen saturation during sleep, oxygen saturation reduction rate of >=3% per hour during sleep, oxygen saturation reduction rate of >=4% per hour during sleep, lowest blood oxygen saturation during NREM sleep, lowest blood oxygen saturation during REM sleep, lowest blood oxygen saturation, and average oxygen saturation decrease. The calculation formula is as follows: , In the formula, Indicates continuous features, This represents the continuous feature after adding Gaussian noise.

[0065] The final constructed noisy features include: 1) Basic Information: Age; 2) Physical examination information: height, weight, neck circumference, waist circumference, hip circumference, corrected height, corrected weight, corrected neck circumference, corrected waist circumference, corrected hip circumference; 3) Blood oxygen saturation: Percentage of sleep time with blood oxygen saturation <95% and <90%, percentage of sleep time with blood oxygen saturation ≤95%, ≤92%, and ≤90%, average blood oxygen saturation during NREM sleep, average blood oxygen saturation during REM sleep, average blood oxygen saturation during sleep, oxygen saturation reduction rate of >=3% per hour during sleep, oxygen saturation reduction rate of >=4% per hour during sleep, lowest blood oxygen saturation during NREM sleep, lowest blood oxygen saturation during REM sleep, lowest blood oxygen saturation, and decrease in average oxygen saturation.

[0066] 3. The child OSA prediction model construction module includes: an optimal algorithm identification submodule, which uses noisy features as input and trains using nested cross-validation and grid search to identify the optimal algorithm; an optimal hyperparameter identification submodule, which, based on the optimal algorithm and using noisy features as input, uses cross-validation and Bootstrap training to identify the hyperparameters with the best AUC; and a feature combination identification submodule, which uses noisy features as input, randomly combines the input features, and builds multiple child OSA prediction sub-models based on the optimal algorithm and optimal hyperparameters. The sub-models are then trained using the feature combination with the best AUC and the optimal hyperparameters to construct the final child OSA prediction model.

[0067] (1) Optimal Algorithm Recognition Submodule In this embodiment, six mainstream supervised learning algorithms capable of handling both continuous and categorical features were used to build models, and the one that performed best on the sample dataset was selected. These six models are: Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Random Forest (RF), Lightweight Gradient Boosting Machine (LightGBM), Linear Discriminant Analysis (LDA), and K-Nearest Neighbors (KNN).

[0068] To select the most suitable learning algorithm for a specific classification task, nested cross-validation (nCV) was used. By adjusting hyperparameters, nCV can select the most suitable model for the specific classification task. Specifically, based on the 1626 samples with artificially injected noise, each fold of the outer cross-validation (CV) uses the corresponding training set's inner cross-validation (CV) and a grid search to obtain the hyperparameters that perform best in the inner CV. These hyperparameters are then applied to the outer CV to obtain the algorithm's best performance on the dataset, thus identifying the optimal algorithm.

[0069] Compared to standard computer vision (CV), nCV provides a more accurate estimate of a model's true generalization ability, offering a more rigorous and reliable framework for model selection and performance evaluation. During training, nCV with 10-fold internal and external performance was used. The list of hyperparameters requiring grid search tuning is as follows: XGBoost: tree depth, learning rate, number of trees, random sampling ratio, proportion of features randomly selected for each tree, and L2 regularization term; SVM: Regularization parameters and RBF kernel parameters; RF: number of trees, tree depth, minimum number of samples required for node splitting, minimum number of samples required for leaf nodes, and maximum number of features; LightGBM: number of trees, tree depth, learning rate, number of leaf nodes, and minimum number of samples in a leaf node; LDA: The solution method of LDA, the shrinkage coefficient, the class prior probability, and the number of feature dimensions after dimensionality reduction; KNN: the number of nearest neighbors, weight allocation method, and nearest neighbor search algorithm.

[0070] In this embodiment, the results of the six basic machine learning models are shown in Table 2: Table 2 Performance of 6 basic machine learning models

[0072] (2) Optimal hyperparameter identification submodule After the above training process, the best-performing basic machine learning model on this dataset is selected as the "winning algorithm". In this embodiment, the "winning algorithm" is XGBoost. To obtain the "winning algorithm"... The "superior hyperparameters" We trained using additional regular cross-validation and evaluated the performance using Bootstrap sampling on the test set.

[0073] Specifically, the dataset contains 1626 samples with artificially injected noise. It is first divided into training and test sets in an 8:2 ratio. On the training set, grid search and 10-fold cross-validation are used to find the optimal parameters. The list of hyperparameters searched includes tree depth, learning rate, number of trees, random sampling ratio, proportion of randomly selected features per tree, and L2 regularization. After obtaining the optimal hyperparameters, Bootstrap sampling, repeated 1000 times with replacement and drawing 100% of the samples each time, is used on the test set to evaluate the area under the curve (AUC), accuracy, specificity, sensitivity, precision, and F1 score of the basic machine learning model.

[0074] (3) Feature combination recognition submodule Based on the noisy features of each input, to compare the importance of features from different dimensions in predicting childhood OSA, this embodiment combines features from four dimensions (basic information, physical examination information, sleep-related symptoms, and blood oxygen saturation). Each combination contains 1 to 4 dimensions, and a prediction model is built for each combination, resulting in 6 sub-models for predicting childhood OSA. The combinations are as follows: Feature combination Basic information, physical examination information. The trained child OSA prediction sub-model is: .

[0075] Feature combination Sleep-related symptoms. The trained child OSA prediction sub-model is as follows: .

[0076] Feature combination Blood oxygen saturation. The trained sub-model for predicting OSA in children is: .

[0077] Feature combination Basic information, physical examination information, and sleep-related symptoms. The trained child OSA prediction sub-model is as follows: .

[0078] Feature combination Sleep-related symptoms and blood oxygen saturation. The trained sub-model for predicting pediatric OSA is: .

[0079] Feature combination Basic information, physical examination information, sleep-related symptoms, and blood oxygen saturation. The trained sub-model for predicting pediatric OSA is as follows: .

[0080] The performance of the child OSA prediction sub-model trained with the six feature combinations is shown in Table 3.

[0081] Table 3. Performance of the 6 feature combinations

[0082] A higher AUC indicates a better ability of the model to distinguish between different classes of samples. In this embodiment, the model that performs best is... The AUC on the test set is as high as 0.85, proving that... The sub-model exhibits superior performance, demonstrating that the feature combination selected in this embodiment can effectively predict childhood OSA.

[0083] Select the feature combination with the best AUC, i.e., feature combination The model, along with its corresponding "superior hyperparameters," was then retrained using XGBoost on the entire dataset—specifically, on the dataset containing 1626 samples with artificially injected noise—to obtain the final child OSA prediction model. .

[0084] 4. Model prediction module, which is used to input basic information, physical examination information, sleep-related symptoms and blood oxygen saturation into the final childhood OSA prediction model, and output the probability of having OSA, predicting whether the child has OSA.

[0085] In this embodiment, the user's basic information, physical examination information, sleep-related symptoms, and blood oxygen saturation are input into the final pediatric OSA prediction model. In this context, the system predicts whether a user has OSA, thus enabling the prediction of OSA.

[0086] 5. Prediction Result Analysis Module like Figure 2 As shown, the prediction results analysis module includes an overall SHAP analysis submodule and an age-stratified SHAP analysis submodule. By using SHAP analysis, the contribution of model features can be visualized, the weights of key indicators can be verified, the interpretability of the model can be improved, and the needs of clinical decision support can be met.

[0087] (1) Overall SHAP analysis submodule, which is used to calculate the average Shapley value of each input feature on the child OSA sample based on the final child OSA prediction model using the SHAP algorithm, and analyze whether each input feature is an overall risk factor for child OSA and its corresponding weight.

[0088] Regarding the final child OSA prediction model obtained above The SHAP algorithm was used to analyze all training samples to obtain the average Shapley value of each feature across all samples. This objectively evaluated whether each feature was an overall risk factor for childhood OSA and what its weight was. Its SHAP value Defined as: , in, It is the set of all features. For feature set The total number of characteristics; yes The middle does not contain features Any subset of (i.e., S⊆N\{d}), For feature subset The number of features; It is a feature subset The corresponding model prediction value, In a subset Add features based on The model predictions after that.

[0089] Using SHAP values ​​allows for an objective understanding of how each feature contributes to the model's predictions, thereby improving the model's interpretability.

[0090] like Figure 3 The results of SHAP analysis on all samples are shown. The vertical axis represents each feature, arranged from top to bottom in descending order of weight. Red indicates that the feature has a positive impact on the model's prediction, meaning that increasing the value of the feature will increase the predicted value, making the model more inclined to predict positive results; blue indicates the opposite. The features, from top to bottom, are: oxygen saturation reduction rate of >=4% per hour during sleep, oxygen saturation reduction rate of >=3% per hour during sleep, mean oxygen saturation during REM sleep, percentage of sleep time with oxygen saturation <95%, snoring, repeated awakenings, percentage of sleep time with oxygen saturation ≤95%, corrected waist circumference, corrected height, height, lowest oxygen saturation during REM sleep, hip circumference, decrease in mean oxygen saturation, corrected neck circumference, corrected hip circumference, weight, gender, waist circumference, mean oxygen saturation during sleep, and neck circumference.

[0091] As shown in the diagram, various indicators of blood oxygen saturation are important predictors of childhood OSA; snoring and recurrent awakenings are the main symptoms of childhood OSA and require attention from parents; excessive waist circumference and excessively short height are the main signs of childhood OSA.

[0092] (2) Age-stratified SHAP analysis submodule, which is used to stratify children by age and build children's OSA prediction models based on the age stratification results. and The SHAP algorithm was used to analyze the child OSA prediction model. and Feature importance analysis was conducted to analyze the main OSA risk factors in children across age strata and to identify the main characteristics of OSA in children across each age stratum.

[0093] Due to growth and development, the risk factors for OSA in children may change at different ages. To more specifically explain the risk factors for OSA in children of different ages, this embodiment, based on the above steps, stratifies by age and establishes two predictive models for samples aged 3-7 and 7-14 years. and And using the SHAP algorithm, respectively for and Feature importance analysis was performed to explain and compare the main OSA risk factors in children of two age groups, and to more accurately identify which features are more important in a particular age group.

[0094] Figure 4 This is the result of SHAP analysis on a sample aged 3-7 years. The vertical axis represents each feature, arranged from top to bottom in descending order of weight. Red indicates that the feature has a positive impact on the model's prediction, meaning that increasing the value of the feature will increase the predicted value, making the model more inclined to predict positive results; blue indicates the opposite. The features, from top to bottom, are: oxygen saturation reduction rate of >=3% per hour during sleep, oxygen saturation reduction rate of >=4% per hour during sleep, hip circumference, snoring, corrected height, mean blood oxygen saturation during sleep, percentage of sleep time with blood oxygen saturation ≤95%, corrected neck circumference, decrease in mean oxygen saturation, mouth breathing, corrected waist circumference, lowest blood oxygen saturation during NREM sleep, age, mean blood oxygen saturation during REM sleep, weight, height, percentage of sleep time with blood oxygen saturation <95%, corrected weight, repeated awakenings, and waist circumference.

[0095] As the diagram shows, compared to the general population, mouth breathing is an additional major symptom of OSA in children aged 3-7; the younger the child, the higher the risk of developing OSA.

[0096] Figure 5This is the result of SHAP analysis on a sample aged 7-14 years. The vertical axis represents each feature, arranged from top to bottom in descending order of weight. Red indicates that the feature has a positive impact on the model's prediction, meaning that increasing the value of the feature will increase the predicted value, making the model more inclined to predict positive results; blue indicates the opposite. The features, from top to bottom, are: oxygen saturation reduction rate of >=4% per hour during sleep, percentage of sleep time with oxygen saturation <95%, oxygen saturation reduction rate of >=3% per hour during sleep, mean oxygen saturation during REM sleep, height, sex, waist circumference, bedwetting, neck circumference, lowest oxygen saturation during NREM sleep, corrected waist circumference, recurrent awakenings, lowest oxygen saturation during REM sleep, percentage of sleep time with oxygen saturation ≤95%, mean oxygen saturation during sleep, corrected hip circumference, weight, decrease in mean oxygen saturation, corrected weight, and corrected height.

[0097] As the diagram shows, males are more likely to have OSA among children aged 7-14.

[0098] In summary, the childhood obstructive sleep apnea prediction system provided by this invention incorporates easily accessible data such as sleep questionnaires, physical examinations, and blood oxygen saturation data. It is low-cost, simple to operate, and suitable for rapid screening across various dimensions of features. By selecting easily collected feature data that is less susceptible to environmental interference, the robustness of biosignal acquisition is improved. Simultaneously, an adversarial training strategy is used to artificially inject noise and missing values ​​to establish initial features, enhancing the robustness and reliability of the prediction model. Based on the initial features, different input features are established individually or in random combinations. Combining the importance of each input feature in predicting childhood OSA, and through optimal algorithm and hyperparameter identification, a prediction model adapted to childhood OSA is constructed, significantly improving the prediction accuracy of childhood OSA and enabling early screening of childhood obstructive sleep apnea. Finally, based on the constructed childhood OSA prediction model, SHAP analysis is further used to visualize feature contribution, enhancing model interpretability and the effectiveness of decision support, meeting the decision support needs of clinical diagnosis and treatment.

[0099] Furthermore, the terms "upper," "lower," "inner," "outer," "front," and "rear" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of components and steps described in these embodiments do not limit the scope of the invention. Of course, the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of the invention. All equivalent changes or modifications made to the structures, features, and principles described in the claims of this invention should be included within the scope of the claims of this invention.

[0100] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A childhood obstructive sleep apnea prediction system, characterized in that, include: The data filtering and preprocessing module is used to collect children's basic information, physical examination information, sleep questionnaire data and blood oxygen saturation, and to filter the collected data to obtain initial features including basic information, physical examination information, sleep-related symptoms and blood oxygen saturation, and to preprocess the initial features. The feature selection and construction module takes the preprocessed initial features as input and whether OSA is present as output, establishes a univariate logistic regression equation to filter out features less than a threshold, constructs new features based on features related to growth and development, and adds noise perturbation to the new features to obtain the noisy features. The module for building a child OSA prediction model comprises three sub-modules: an optimal algorithm identification sub-module, which uses noisy features as input and trains the model using nested cross-validation and grid search to identify the optimal algorithm; an optimal hyperparameter identification sub-module, which uses the optimal algorithm and noisy features as input and trains the model using cross-validation and Bootstrap to identify the hyperparameters with the best AUC; and a feature combination identification sub-module, which uses noisy features as input, randomly combines the input features, and builds multiple child OSA prediction sub-models based on the optimal algorithm and optimal hyperparameters. The sub-models are then trained using the feature combination with the best AUC and the optimal hyperparameters to construct the final child OSA prediction model. The model prediction module is used to input basic information, physical examination information, sleep-related symptoms and blood oxygen saturation into the final childhood OSA prediction model, and output the probability of having OSA, predicting whether a child has OSA.

2. The childhood obstructive sleep apnea prediction system according to claim 1, characterized in that, The childhood obstructive sleep apnea prediction system also includes a prediction result analysis module, which includes an overall SHAP analysis submodule and an age-stratified SHAP analysis submodule. The overall SHAP analysis submodule is used to calculate the average Shapley value of each input feature on the child OSA sample based on the final child OSA prediction model using the SHAP algorithm, and to analyze whether each input feature is an overall risk factor for child OSA and its corresponding weight. The age-stratified SHAP analysis submodule is used to stratify children by age and build OSA prediction models for each child based on the age stratification results. and The SHAP algorithm was used to analyze the child OSA prediction model. and Feature importance analysis was conducted to analyze the main OSA risk factors in children across age strata and to identify the main characteristics of OSA in children across each age stratum.

3. The childhood obstructive sleep apnea prediction system according to claim 1, characterized in that, In the data filtering and preprocessing module, initial features including basic information, physical examination information, sleep-related symptoms, and blood oxygen saturation are obtained based on the collected data. This includes: establishing mapping rules based on the collected sleep questionnaire data, mapping the results of each question in the sleep questionnaire data to binary symptoms, with the mapping result being 0 or 1, representing the presence or absence of the symptoms, respectively; and using the mapped symptoms as sleep-related symptoms, which, together with the basic information, physical examination information, sleep questionnaire data, and blood oxygen saturation, constitute the initial features.

4. The childhood obstructive sleep apnea prediction system according to claim 3, characterized in that, The required binary symptoms to be mapped are snoring, sleep apnea, mouth breathing, restless sleep, bedwetting, daytime sleepiness, attention deficit or hyperactivity, recurrent awakenings, and abnormal emotional and behavioral patterns.

5. The childhood obstructive sleep apnea prediction system according to claim 3, characterized in that, Based on the collected sleep questionnaire data, the mapping rules established include: The sleep questionnaire data used were OSA-18 and PSQ children's sleep questionnaire. The score for each question in OSA-18 was 0-6, and the score for each question in PSQ was 0 and 1. For symptoms If OSA-18 includes the corresponding issue Ai The mapping rules are calculated as follows: ， In the formula, Indicating the first in OSA-18 The original scores for each question This represents the median score of all samples for that question; For symptoms If the PSQ contains the corresponding question Bj The mapping rules are calculated as follows: ， In the formula, Indicates the first in PSQ The original scores for each question; Final symptoms The condition for its existence is that the above binary classification mapping result for any related problem is 1, and the calculation formula is as follows: 。 6. The childhood obstructive sleep apnea prediction system according to claim 5, characterized in that, In the feature selection and construction module, features below a threshold are selected by establishing a univariate logistic regression equation. This includes using a univariate logistic regression equation to select variables significantly related to OSA from the preprocessed initial features. The calculation formula is as follows: ， In the formula, Indicates the independent variable Below, dependent variable The probability of the dependent variable, where the dependent variable is... The definition is: when OSA is positive, When OSA is negative, , and The regression coefficients obtained from training reflect the independent variables. For dependent variable The impact; Based on the calculated p-values ​​of each variable in logistic regression, the significance of each variable on the result is determined, and features with p values ​​less than the threshold are selected. The selected features are considered to be significantly correlated with OSA.

7. The childhood obstructive sleep apnea prediction system according to claim 6, characterized in that, In the feature selection and construction module, growth and development-related features are obtained through the following methods: Design matrices containing linear, quadratic, and intercept terms for age, respectively. A vector containing predictor variables: height, weight, neck circumference, waist circumference, and hip circumference. and coefficient matrix Modeling the changes of each predictor variable with age, expressed as: ,in Residual vector; Estimate the coefficient matrix using the least squares method. Then, the corrected residual vector is obtained. , is used to represent the corrected features of each predictor variable after removing the influence of age, thus obtaining the features related to growth and development.

8. The childhood obstructive sleep apnea prediction system according to claim 1, characterized in that, In the optimal algorithm recognition submodule, the noisy features are used as input, and nested cross-validation and grid search algorithms are used for training to identify the optimal algorithm, including: Using the noisy features as input, select an algorithm that can handle both continuous and categorical features; Based on the selected algorithm, internal cross-validation is applied in each fold of the external cross-validation, and grid search is combined to obtain the hyperparameters that perform best in the internal cross-validation. The best hyperparameters are then applied to the external cross-validation to select the best algorithm.

9. The childhood obstructive sleep apnea prediction system according to claim 8, characterized in that, In the optimal hyperparameter identification submodule, using the noisy features as input, and trained with standard cross-validation and Bootstrap, the hyperparameters with the optimal AUC are identified, including: The data is divided into training and test sets, using the noise-added features as input. Based on the training set, grid search and 10-fold cross-validation are used to search for the set of hyperparameters corresponding to the best algorithm. Based on the test set, the Bootstrap sampling method is used to train a set of hyperparameters. The AUC, accuracy, specificity, sensitivity, precision and F1 score of each set of hyperparameters are evaluated, and the hyperparameter with the best AUC is identified.

10. The childhood obstructive sleep apnea prediction system according to claim 9, characterized in that, The input features are randomly combined, and multiple child OSA prediction sub-models are established based on the optimal algorithm and optimal hyperparameters, including: Based on the noisy features of each input and considering the importance of each input feature in predicting children's OSA, six feature combinations are established. ~ and the corresponding child OSA prediction sub-model ~ , where feature combination Includes basic information and physical examination information; feature combinations Including sleep-related symptoms; characteristic combinations Including blood oxygen saturation; characteristic combinations Includes basic information, physical examination information, and sleep-related symptoms; characteristic combinations Including sleep-related symptoms and blood oxygen saturation; characteristic combination This includes basic information, physical examination information, sleep-related symptoms, and blood oxygen saturation.

Images