Newborn congenital heart disease screening method based on machine learning

Abstract

The invention discloses a neonatal congenital heart disease screening method based on machine learning, and the method comprises the following steps: 1, designing a questionnaire according to literatures related to adverse birth result risk factors, collecting infant cardiac ultrasound examination data, and then preprocessing the data; step 2, randomly dividing the data obtained in the step 1 into a training set and a test set, and performing dimension reduction processing on features by using lASSO regression; and step 3, based on the features obtained in the step 2, training a prediction model by using machine learning algorithms such as a random forest, a support vector machine, a lightweight gradient elevator, logistic regression and extreme gradient lifting. According to the method, the model is optimized through combination of multi-step data processing and various machine learning algorithms, related features can be comprehensively and scientifically screened, the accuracy of the prediction model is improved, a more reliable and effective method is provided for screening the neonatal congenital heart disease, early-stage accurate screening is facilitated, and the screening efficiency and quality are improved.

Description

Technical Field

[0001] The present invention relates to a screening method, and more particularly to a method for screening congenital heart disease in newborns based on machine learning. Background Art

[0002] Currently, screening methods for congenital heart disease (CHD) in newborns primarily include pulse oximetry (POX) screening, cardiac auscultation, and echocardiography, but each of these methods has significant limitations. POX screening is widely used due to its noninvasive and rapid nature, but it is primarily designed to detect hypoxemia associated with critical CHD and is less effective in screening children with CHD who do not have hypoxemia. A meta-analysis of approximately 600,000 newborns showed an overall sensitivity of 72.2% for POX screening alone. Furthermore, POX screening is not suitable for high altitudes due to the significantly increased false-positive rate caused by hypoxic environments. Cardiac auscultation, a common first-line screening measure, requires extensive training and clinical experience to accurately identify heart murmurs. Furthermore, auscultation results are significantly influenced by the subjective judgment of the auscultator, resulting in a high rate of missed diagnoses. Although echocardiography is the gold standard for CHD diagnosis, it relies on the experience of the screener and the quality of the ultrasound equipment. Its high cost and complexity make it unsuitable for large-scale screening. These limitations indicate that existing screening methods have problems such as poor adaptability, low cost-effectiveness, and high rates of missed diagnosis and misdiagnosis in practical applications. There is an urgent need for a screening method that is adaptable, independent of professional pediatricians, and cost-effective.

[0003] Therefore, the current state-of-the-art approach of using machine learning to assist in screening can effectively address the aforementioned technical issues. However, this approach has significant limitations. Mainstream models rely on data such as fetal electrocardiograms and ultrasound images, which require specialized equipment to collect. Ultrasound diagnosis itself is susceptible to operator experience and equipment resolution, limiting the model's generalization ability. For example, maternal obesity and fetal position can reduce ultrasound image quality, thereby affecting the stability of model predictions. Proteomics analysis requires high-end equipment such as mass spectrometers, resulting in high costs per test and limited widespread adoption in low-income areas. The "black box" nature of deep learning models still hinders clinical trust. For example, the CNN's process for extracting local features from ultrasound images lacks intuitive explanation, which can affect physicians and patients' acceptance of prediction results. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the purpose of the present invention is to provide a method for screening for congenital heart disease in newborns based on machine learning, which collects questionnaire survey data and then combines it with machine learning, and the generalization ability of the model is not easily limited.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for screening newborns for congenital heart disease based on machine learning, comprising the following steps: Step 1: A questionnaire was designed based on the literature on risk factors for adverse birth outcomes. The questionnaire was distributed to pregnant women in the hospital for completion, and the responses were collected. Infants and mothers were matched in a 1:3 ratio based on infant sex and maternal age. Echocardiographic data from the infants were collected and preprocessed. Step 2: The data obtained in step 1 are randomly divided into a training set and a test set. The training set is used to train the prediction model, and the test set is used to evaluate the performance of the prediction model. The features of the training set and the test set are then extracted. The features with missing data and excessive collinearity are excluded from the obtained features. LASSO regression is then applied to select relevant features from the extracted features. Step three: Based on the features obtained in step two, the prediction model is trained using the random forest algorithm, support vector machine algorithm, optical gradient boosting machine algorithm, logistic regression algorithm, and extreme gradient boosting algorithm in machine learning algorithms. The models are all cross-validated with 5-fold on the same training set and stratified sampling is used to avoid data distribution shift. Then, a combination of grid search and manual fine-tuning is used to obtain the final hyperparameters to optimize the prediction model.

[0006] As a further improvement of the present invention, the step of evaluating the contribution of the features in the prediction model is further included, specifically: applying the SHAP algorithm to evaluate the contribution of each feature in the prediction model based on the machine learning algorithm, and displaying the results in descending order.

[0007] As a further improvement of the present invention, the specific method of applying LASSO regression to select relevant features from the extracted features in step 2 is: first, the extracted features are normalized, and then a LASSO regression model is constructed, and the regularization parameter λ is set, and the LASSO model is fitted on the training set. When log (λ) reaches the minimum mean square error, the features remaining after optimization and deletion are selected as the final extracted features.

[0008] As a further improvement of the present invention, the specific method of excluding features with missing data and excessive collinearity from the acquired features in step 2 is: performing Spearman correlation analysis, selecting features with high correlation coefficients, and deleting features with weak correlation with the result features.

[0009] As a further improvement of the present invention, the specific method of using the random forest algorithm, support vector machine algorithm, lightweight gradient boosting machine algorithm, logistic regression algorithm and extreme gradient boosting algorithm in the machine learning algorithm to train the prediction model in step three is: the models of the five machine algorithms are cross-validated on the same training set and stratified sampling is adopted, and then the different models are trained with the optimal hyperparameters to obtain the models and the AUC values, sensitivity, specificity, positive predictive value, negative predictive value and F1 values ​​of the models are comprehensively compared on the test set. Finally, the model based on the extreme gradient boosting algorithm is obtained as the best model, and the model AUC value is 0.7.

[0010] Beneficial Effects of the Present Invention: Compared to existing technologies, this present invention integrates questionnaire data using a machine learning algorithm (XGBOOST algorithm) to construct a non-invasive CHD screening model with an AUC exceeding 0.7. This model overcomes the following drawbacks of traditional methods: 1) Reduces reliance on specialized equipment. Eliminating the need for high-cost equipment such as echocardiography or fetal electrocardiography, the present invention requires only clinical indicators collected through routine questionnaires (such as history of pregnancy infections and family medical history), addressing the challenges of mass adoption of existing technologies due to the limited availability of these equipment. 2) Eliminates the influence of operator experience. Through standardized questionnaire data input and automated model analysis, the present invention avoids subjective bias in cardiac auscultation and the reliance of ultrasound diagnosis on operator experience. 3) Enhances model interpretability and clinical credibility. SHAP values ​​address the "black box" problem. The SHAP (SHapley Additive exPlanations) algorithm quantifies the contribution of each risk factor (e.g., the weight of gestational diabetes is 12.3%), making predictions traceable and verifiable, thereby increasing the adoption of model results by both physicians and patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 Schematic diagram of the LASSO regression curve, showing (A) the relationship between the regression coefficient and the logarithm (λ); (B) the MSE and log(λ) curves. λmin represents the cutoff point where the MSE takes its minimum value, and λ1se represents the point where the MSE takes 1× the standard error. Figure 2 Schematic diagram of ROC curve comparison of machine learning models for congenital heart disease prediction; Figure 3 Schematic diagram of the performance decision curve analysis of the XGBoost model on the test set; Figure 4 Schematic diagram of important features based on shaply values. In the figure, (A) SHAP summary bar chart. (B) SHAP summary dot chart; Figure 5 is the SHAP dependency graph. DETAILED DESCRIPTION

[0012] The present invention will be further described below with reference to the embodiments shown in the accompanying drawings.

[0013] 1 , the method for screening newborns for congenital heart disease based on machine learning in this embodiment specifically includes the following steps: Step one involves designing a questionnaire based on literature on risk factors for adverse birth outcomes. This questionnaire is distributed to pregnant women in hospitals, and the responses are subsequently collected. Simultaneously, infants and mothers are matched based on infant sex and maternal age in a 1:3 ratio. Echocardiographic data from the infants is then collected and preprocessed. This multi-dimensional data collection step, compared to existing single-screening methods, provides more comprehensive information, reduces the risk of missed or misdiagnosed diagnoses due to incomplete data, and improves screening accuracy.

[0014] In step 2, the data acquired in step 1 is randomly divided into a training set and a test set. The training set is used to train the prediction model, and the test set is used to evaluate the prediction model's performance. Features are then extracted from the training and test sets, and features with missing data and excessive collinearity are excluded from the extracted features. The specific method for excluding missing data and excessive collinearity is to conduct a Spearman correlation analysis, select features with high correlation coefficients, and exclude features with weak correlation with the resulting features. LASSO regression is then applied to select relevant features from the extracted features. As shown in Figure 1, the specific method is as follows: first, the extracted features are normalized, then a LASSO regression model is constructed with a set regularization parameter λ. The LASSO model is fitted on the training set. When log(λ) achieves the minimum mean squared error, the remaining features after optimization and pruning are selected as the final extracted features. This step effectively improves the model's generalization ability through appropriate data feature processing, addressing the generalization limitation of mainstream models in the existing art due to improper data processing.

[0015] In step three, based on the features obtained in step two, the prediction model was trained using the machine learning algorithms Random Forest, Support Vector Machine, Lightweight Gradient Boosting, Logistic Regression, and Extreme Gradient Boosting. A combination of grid search and manual fine-tuning was then used to obtain the final hyperparameters, thereby optimizing the prediction model. This step, through multi-algorithm training and hyperparameter optimization, improved the model's performance, avoiding the poor adaptability and low cost-effectiveness of existing models due to a single algorithm or unoptimized parameters. In the process of pre-training the model using the five learning algorithms, the optimal hyperparameters were used to train different models and the AUC values, sensitivity, specificity, positive predictive value, negative predictive value, and F1 values ​​of the models were comprehensively compared on the test set (30% of the data). The XGBOOST-based model was ultimately selected as the best model, with an AUC value of 0.7.

[0016] Furthermore, it also includes the step of evaluating the contribution of features in the prediction model. The specific approach is: applying the SHAP algorithm to evaluate the contribution of each feature in the prediction model based on the machine learning algorithm and displaying it in descending order. Figure 4 As shown in the figure, this step makes the model results more interpretable, solving the problem in the existing technology that the "black box" characteristics of deep learning models hinder clinical trust.

[0017] In summary, this solution overcomes many problems existing in existing screening methods, such as poor adaptability, low cost-effectiveness, high rates of missed and misdiagnoses, limited model generalization capabilities, and "black box" characteristics that hinder clinical trust, through comprehensive data collection methods, scientific data feature processing, training of multiple algorithms, hyperparameter optimization, and evaluation of feature contributions. It demonstrates greater adaptability and cost-effectiveness, while improving screening accuracy and model interpretability.

[0018] The following specific examples are provided in this embodiment to further illustrate the advanced nature of the method for screening newborn congenital heart disease based on machine learning of the present invention. Specifically, the following four steps are performed: 1. Data Processing: From September 2022 to July 2024, a multicenter, hospital-based, case-control study was conducted at three hospitals (Jiaxing Maternal and Child Health Hospital, Dongyang Maternal and Child Health Hospital, and Ninghai Maternal and Child Health Hospital). Participants were recruited from pregnant women delivering at these three collaborating hospitals. All newborns participating in the study underwent cardiac ultrasound examinations performed by professional echocardiographers. All pregnant women were surveyed for risk factors during pregnancy by trained personnel using a self-designed questionnaire. The case group consisted of mothers whose newborns were diagnosed with congenital heart disease by echocardiography. In addition, the control group included mothers whose newborns were diagnosed without heart defects. Infants and mothers were matched in a 1:3 ratio by infant sex and maternal age.

[0019] This example uses a questionnaire survey. The questionnaire was designed by the research team based on previous literature on risk factors for adverse birth outcomes. It includes questions about parental demographic information, body measurements, genetic predisposition, maternal health during pregnancy, prenatal exposures, and prenatal supplement intake. The questionnaire also includes questions about viral infections, smoking, alcohol and caffeine intake in the three months before the last menstrual period and during pregnancy, lifestyle, medication use during pregnancy, and mood and sleep status. Overall, the questionnaire surveys 12 key aspects related to risk factors and early warning patterns for congenital heart disease in newborns.

[0020] 2. Feature Extraction: In this study, a total of 1488 valid responses were collected. After excluding cases of other birth defects or preterm births, a total of 1348 questionnaires were included, including 365 participants in the CHD group and 983 participants in the control group. These data were randomly divided into a training set (70%) and a test set (30%). The training set, consisting of 945 participants (462 males and 483 females), was used to train the prediction model, while 403 participants (201 males and 202 females) were included in the test set for performance evaluation. Table 1 describes the baseline characteristics of participants in the case and control groups. There were no significant differences in demographic characteristics between the CHD and non-CHD groups. Given that multicollinearity can reduce model accuracy, we performed Spearman correlation analysis to select features with high correlation coefficients (r>0.7) and removed variables with weak correlations with the outcome variable. Because the data were missing more than 25%, the feature of average annual household income was excluded to prevent model bias caused by excessive missing values.

[0021] After excluding features with missing data and excessive collinearity, 105 features remained. Since too many features may lead to overfitting in machine learning models, we applied LASSO regression to select relevant features from the available variables. The number of features was reduced to 21 when log(λ) achieved the minimum mean squared error and further reduced to 3 when it reached one standard error (one SE) above the minimum distance. This was obtained using 10-fold internal cross-validation with LASSO ( Figure 1 A, B). We selected the point that achieved the minimum mean square error (optimal lambda value = 0.017). By applying LASSO regression, we ultimately selected 21 variables, including family history of heart disease, family history of chronic disease, family history of diabetes, viral infection in the 3 months before the last menstrual period, viral infection during pregnancy, thyroid disease diagnosed during pregnancy, medication use during pregnancy, antiviral drug use during pregnancy, vitamin D supplementation, iron supplementation, calcium supplementation, exposure to pesticides and agricultural chemicals during pregnancy, exposure to microwaves during pregnancy, mood state during mid-pregnancy, maternal age, paternal ethnicity, paternal occupation, maternal prepregnancy BMI, mid-pregnancy multivitamin use, maternal weight gain during pregnancy, and paternal education level.

[0022] Table 1: Comparison of demographic characteristics between children with and without congenital heart disease.

[0023] 3. Model Construction: In this study, five machine learning algorithms widely used for disease prediction were used to train a CHD prediction model. These five machine learning algorithms included random forest (RF), support vector machine (SVM), lightweight gradient boosting machine (lightGBM), logistic regression (LR), and extreme gradient boosting (XGBoost). To optimize the prediction model, a combination of grid search and manual fine-tuning was applied to obtain the final hyperparameters.

[0024] We selected five machine learning models (SVM, RF, LR, XGBoost, and LightGBM) to develop a prediction model for neonatal CHD using 21 selected features. Table 2 shows the performance of all models in the internal test set. The XGBoost-based model achieved the best overall performance among the five models, with an area under the curve (AUC) of 0.7, followed by LightGBM (AUC: 0.64), LR (AUC: 0.65), RF (AUC: 0.65), and SVM (AUC: 0.55). The optimal parameters for these five models are detailed in Supplementary Table 2. Figure 2 The ROC curves of the five CHD prediction models in the test group are shown. The DCA results of the test set show that the XGBoost-based CHD prediction model shows a positive net benefit within a certain range of probability thresholds ( Figure 3 ).

[0025] Table 2: Prediction performance of five machine learning models 4. Interpretable Models To better understand the contribution of features in the CHD model and identify key risk factors, we applied the SHAP algorithm to evaluate the contribution of each feature in the XGBoost-based CHD prediction model. Figure 4 The top ten risk factors in the model on the training set are shown. The contribution of each feature to the model is calculated using the average SHAP value and displayed in descending order. The top ten features include viral infection during pregnancy, maternal BMI before delivery, viral infection within three months before the last menstrual period, medication use during pregnancy, father's job, emotional stability during the second trimester, maternal age, weight gain during pregnancy, iron supplementation, and family history of diabetes ( Figure 4 A, B). In addition, SHAP dependency plots help understand how individual features affect the output of the predictive model. Figure 5 The actual values ​​of maternal prenatal BMI and maternal age, along with their SHAP values, are shown. A positive SHAP value corresponds to a relationship between the prediction model and the positive class. For example, a maternal age > 35 years or a prenatal BMI > 30 would result in a SHAP value greater than 0, pushing the prediction toward the CHD class.

[0026] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the concept of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for screening newborns for congenital heart disease based on machine learning, characterized by: The steps include: Step 1: A questionnaire was designed based on the literature on risk factors for adverse birth outcomes. The questionnaire was distributed to pregnant women in the hospital for completion, and the responses were collected. Infants and mothers were matched in a 1:3 ratio based on infant sex and maternal age. Echocardiographic data from the infants were collected and preprocessed. Step 2: The data obtained in step 1 are randomly divided into a training set and a test set. The training set is used to train the prediction model, and the test set is used to evaluate the performance of the prediction model. The features of the training set and the test set are then extracted. Among the features obtained, features with missing data and excessive collinearity are excluded, and LASSO regression is applied to select relevant features from the extracted features. Step three: Based on the features obtained in step two, the prediction model is trained using the random forest algorithm, support vector machine algorithm, lightweight gradient boosting algorithm, logistic regression algorithm, and extreme gradient boosting algorithm in machine learning algorithms. The models are all cross-validated with 5-fold cross validation on the same training set and stratified sampling is used; then a combination of grid search and manual fine-tuning is used to obtain the final hyperparameters to optimize the prediction model.

2. The method for screening newborns for congenital heart disease based on machine learning according to claim 1, characterized in that: The method also includes a step of evaluating the contribution of the features in the prediction model, specifically: applying the SHAP algorithm to evaluate the contribution of each feature in the prediction model based on the machine learning algorithm and displaying the results in descending order.

3. The method for screening newborns for congenital heart disease based on machine learning according to claim 2, characterized in that: The specific method of applying LASSO regression to select relevant features from the extracted features in the step 2 is: first, the extracted features are normalized, then a LASSO regression model is constructed, and the regularization parameter λ is set, and the LASSO model is fitted on the training set. When log (λ) reaches the minimum mean square error, the features remaining after optimization and deletion are selected as the final extracted features.

4. The method for screening newborns for congenital heart disease based on machine learning according to claim 3, characterized in that: The specific method of excluding features with missing data and excessive collinearity from the features obtained in step 2 is: performing Spearman correlation analysis, selecting features with high correlation coefficients, and deleting features with weak correlation with the result features.

5. The method for screening newborns for congenital heart disease based on machine learning according to any one of claims 1 to 4, characterized in that: In the step 3, the specific method of using the random forest algorithm, support vector machine algorithm, lightweight gradient boosting machine algorithm, logistic regression algorithm and extreme gradient boosting algorithm in the machine learning algorithm to train the prediction model is as follows: the models of the five machine algorithms are cross-validated on the same training set with 5-fold cross validation and stratified sampling, and then the different models are trained with the optimal hyperparameters to obtain the models and the AUC values, sensitivity, specificity, positive predictive value, negative predictive value and F1 values ​​of the models are comprehensively compared on the test set. Finally, the model based on the extreme gradient boosting algorithm is obtained as the best model, and the model AUC value is 0.7.

Images