Machine learning-based huge child risk prediction method and system

Abstract

The invention discloses a huge child risk prediction method based on machine learning, and the method comprises the following steps: S1, multi-modal clinical data acquisition, S2, data preprocessing, S3, multi-algorithm model training and optimization, S4, model evaluation and selection: evaluating the performance of each model based on the accuracy rate, the precision rate, the recall rate, the AUC and a calibration curve, selecting a model with optimal comprehensive performance as a target prediction model; and S5, risk prediction and explanation: inputting a to-be-predicted sample into the target prediction model, outputting a huge child risk prediction result, and quantifying the contribution degree of each feature to the prediction result through a SHapley addition explanation (SHAP) method. Through deep fusion of machine learning and clinical multi-modal data, the technical problems that an existing method is single in dimension, insufficient in interpretability and poor in basic level adaptability are solved, and a solution which is high in precision, explainable and easy to operate is provided for huge child risk prediction.

Description

Technical Field

[0001] The present invention relates to the field of medical health information, and in particular to a method and system for predicting the risk of macrosomia based on machine learning. Background Art

[0002] The delivery of a macrosomia (a newborn with a birth weight ≥ 4000g) not only significantly increases the risk of birth injuries such as postpartum hemorrhage and perineal lacerations in the mother, but may also cause brachial plexus injury, asphyxia, and even death in the newborn, while also posing a hidden danger of metabolic diseases to the long-term health of the mother and baby. According to epidemiological statistics, the incidence of macrosomia in my country has been increasing year by year (currently about 7%-10%), and the prevention and control situation is becoming increasingly severe due to factors such as overnutrition during pregnancy, increased incidence of gestational diabetes mellitus (GDM), and advanced maternal age. Accurately predicting the risk of macrosomia is of great clinical significance for optimizing prenatal monitoring programs, reducing cesarean section rates, and improving perinatal outcomes.

[0003] Predicting macrosomia is crucial for optimizing delivery methods and reducing maternal and fetal complications. Current mainstream clinical methods rely on ultrasound indicators combined with traditional statistical models. These methods suffer from accuracy bottlenecks (±15% error rate), limitations of single-dimensional analysis (ignoring the synergistic effects of maternal characteristics and fetal growth), linear assumptions that fail to capture nonlinear relationships, and strong operator dependency (a 35% missed diagnosis rate in obese pregnant women). Furthermore, multi-source heterogeneous data (maternal characteristics, fetal ultrasound, comorbidities, etc.) are not effectively integrated, and traditional machine learning models (such as neural networks) lack interpretability, leading to low clinical confidence. Furthermore, the potential of integrating multimodal clinical data (maternal BMI, uterine height, fetal ultrasound indicators, gestational diabetes, etc.) has not been fully explored. While attempts have been made to optimize traditional machine learning models, they suffer from a lack of interpretability (such as the "black box" problem of neural networks) and inadequate feature interaction analysis. With the advancement of machine learning technology, its application in medical data mining and predictive modeling is becoming increasingly widespread. There is an urgent need to establish efficient and intelligent macrosomia prediction methods to assist clinical decision-making. Summary of the Invention

[0004] The present invention aims to provide a method and system for predicting the risk of macrosomia based on machine learning. Through the deep integration of machine learning and clinical multimodal data, it solves the technical problems of existing methods such as single dimension, insufficient interpretability and poor grassroots adaptability, and provides a high-precision, interpretable and easy-to-operate solution for the prediction of macrosomia risk.

[0005] A method for predicting the risk of macrosomia based on machine learning, comprising the following steps:

[0006] S1. Multimodal clinical data acquisition: Maternal characteristics, fetal ultrasound indicators, and pregnancy complications were collected. Maternal characteristics included admission abdominal circumference, uterine height, pre-pregnancy BMI, admission BMI, age, parity, height, pre-pregnancy weight, and admission weight. Fetal ultrasound indicators included fetal abdominal circumference, head circumference, femoral length, and maximum amniotic fluid depth. Pregnancy complications included gestational diabetes.

[0007] S2. Data preprocessing: Perform one-hot encoding on the original data obtained in step S1 to process categorical variables, use the SMOTE algorithm to balance positive and negative samples, and divide the data into training and test sets;

[0008] S3. Multi-algorithm model training and optimization: We build classification models using five machine learning algorithms: random forest, XGBoost, CatBoost, K-nearest neighbor, and multilayer perceptron. We optimize model parameters through hyperparameter grid search and five-fold cross-validation.

[0009] S4. Model evaluation and selection: Evaluate the performance of each model based on accuracy, precision, recall, AUC, and calibration curves, and select the model with the best overall performance as the target prediction model;

[0010] S5. Risk prediction and interpretation: The samples to be predicted are input into the target prediction model, and the macrosomia risk prediction results are output. The contribution of each feature to the prediction result is quantified using the SHapley additive interpretation (SHAP) method.

[0011] The optimized model with the best comprehensive performance in step S4 is the XGBoost model, and its optimal hyperparameters are 'max_depth':6, 'n_estimators':100.

[0012] In the optimized step S5, the SHAP method evaluates the feature importance by calculating the average SHAP value of each feature, and determines that fetal abdominal circumference, admission BMI, head circumference, uterine height, and maternal abdominal circumference are key predictive features.

[0013] The optimized ratio of the training set to the test set is 8:2, and an independent external validation set is randomly selected to verify the generalization ability of the model.

[0014] A macrosomia risk prediction system based on machine learning includes a data acquisition module for collecting maternal characteristics, fetal ultrasound indicators, and pregnancy complications of pregnant women; a data preprocessing module containing a one-hot encoding unit and a sample balancing unit for processing categorical variables and balancing positive and negative samples; a model training module that integrates random forest, XGBoost, CatBoost, K-nearest neighbor, and multilayer perceptron algorithms, and optimizes model parameters through grid search and cross-validation; a model evaluation module that evaluates model performance based on accuracy, precision, recall, AUC, and calibration curves, and outputs the optimal model; and a prediction and interpretation module that generates risk prediction results based on the optimal model and provides feature contribution explanations through the SHAP method.

[0015] Optimized, the parameter optimization range of the XGBoost model in the model training module includes max_depth of 3-10 and n_estimators of 50-200.

[0016] The data collected by the data acquisition module are optimized and extracted through a structured electronic medical record system, including clinical data of 3233 Han Chinese pregnant women who have delivered singleton cephalic fetuses.

[0017] Optimally, the risk prediction results output by the prediction and interpretation module are connected to the clinical decision support system for the formulation of personalized delivery plans.

[0018] How this application works:

[0019] This method integrates multi-dimensional clinical data and uses machine learning algorithms to build an interpretable prediction model to accurately assess the risk of macrosomia. The specific workflow is as follows:

[0020] (1) Multimodal data collection and feature engineering

[0021] Data collection: 15 maternal characteristics (e.g., abdominal circumference at admission, uterine height, pre-pregnancy BMI, age, and parity), fetal ultrasound parameters (e.g., abdominal circumference, head circumference, and femur length), and comorbidity data (e.g., gestational diabetes) were collected. These data are directly related to fetal growth and development (e.g., abdominal circumference reflects body fat accumulation, and admission BMI is associated with overnutrition during pregnancy).

[0022] Data value: Breaking through the limitations of traditional single ultrasound weight estimation, it captures complex biological mechanisms through cross-dimensional feature combinations (such as the genetic association between maternal height and fetal head circumference).

[0023] (2) Data preprocessing and sample balance

[0024] One-hot encoding: Converts categorical variables (such as parity or gestational diabetes status) into numerical data, allowing machine learning models to identify nonlinear relationships.

[0025] SMOTE algorithm: To address the sample imbalance problem caused by the low incidence of macrosomia (4.9%), it synthesizes minority class samples (oversampling) to balance the distribution of positive and negative samples, avoiding the model's bias towards the majority class (non-macrosomia).

[0026] Dataset division: The training set and test set are divided into 8:2 ratios, and an independent external validation set is retained to ensure the generalization ability of the model.

[0027] (3) Multi-algorithm comparison and model optimization

[0028] Algorithm selection: Five algorithms, including random forest, XGBoost, and CatBoost, are used. XGBoost, a gradient boosting tree algorithm, controls overfitting through regularization terms, supports parallel computing and missing value processing, and is suitable for high-dimensional clinical data modeling.

[0029] Hyperparameter optimization: Through grid search (e.g., max_depth = 3-10, n_estimators = 50-200 for XGBoost) and five-fold cross-validation, the optimal parameter combination (e.g., max_depth = 6, n_estimators = 100 for XGBoost) is determined to improve the model's fit on the training set.

[0030] (4) Model evaluation and optimization

[0031] Performance metrics: Accuracy, precision, recall, AUC (area under the receiver operating characteristic (ROC) curve), and calibration curves were used for comprehensive evaluation. The XGBoost model achieved an AUC of 0.9986 and an accuracy of 0.973 on the test set, significantly outperforming traditional methods (e.g., ultrasound weight estimation with an AUC of approximately 0.85).

[0032] Decision basis: The XGBoost model with the best overall performance was selected as the final prediction model. Its advantage is that it can automatically capture nonlinear interactions between features (such as the synergistic effect of fetal abdominal circumference and gestational diabetes).

[0033] (5) Risk prediction and explainability analysis

[0034] Prediction output: 15 features of the sample to be tested are input, and the model calculates the risk probability through nonlinear combination, and the threshold is used to determine whether the baby is at high risk of macrosomia.

[0035] SHAP interpretation: Utilizes SHapley additive interpretation technology to quantify the contribution of each feature to the prediction result (e.g., fetal abdominal circumference has the highest average SHAP value, indicating that it is a core risk factor). Bar charts / scatter plots are used to visualize the direction and intensity of feature influence, helping doctors understand the model's decision logic.

[0036] The system achieves full process automation from data collection to clinical decision-making through modular design, with hardware and software working together to support prediction tasks.

[0037] (1) Hardware Architecture

[0038] Data collection layer: clinical data is acquired in real time through the hospital electronic medical record (EMR) system interface and stored in a structured database (such as MySQL) to ensure data integrity and timeliness.

[0039] Computing layer: Based on servers or cloud platforms equipped with Intel CPUs and NVIDIA GPUs, it runs machine learning model training and inference tasks, and uses parallel computing to accelerate the iterative optimization of algorithms such as XGBoost.

[0040] Output layer: Displays prediction results and feature explanation reports through the web or mobile interface, and connects to the hospital's clinical decision support system (CDSS).

[0041] (2) Software module collaboration

[0042] Data acquisition module: extracts data from EMR according to preset fields (such as the 15 features as described in claim 1), supports manual re-entry and batch import, and automatically filters unqualified samples (such as non-Han nationality, multiple pregnancies). Preprocessing module: One-hot encoding unit: binary encoding of classification fields such as "number of pregnancies" and "gestational diabetes"; sample balancing unit: calls the SMOTE algorithm API to generate a balanced training data set. Model training module: Algorithm integration: encapsulates 5 algorithms in the Scikit-learn library and provides a visual parameter adjustment interface; automatic tuning: generates a parameter combination matrix through grid search, runs five-fold cross-validation in parallel, and records the performance indicators of each model. Evaluation and prediction module: Comparative analysis: automatically generates ROC curves and confusion matrices, and recommends the optimal model (such as XGBoost); real-time inference: receives sample data to be predicted, calls the trained model file (.pkl format) to output the risk probability, and triggers the SHAP explanation calculation. Interpretation and interaction module:

[0043] Feature attribution: Calculate the SHAP value of each feature based on the SHAP library to generate a global feature importance ranking (such as fetal abdominal circumference > admission BMI > head circumference) and a single-sample local explanation; Visualization report: Displays prediction results and key feature contributions in the form of charts, which can be downloaded or printed by doctors.

[0044] Beneficial effects of this application:

[0045] (1) Multimodal data integration: Breaking through the limitations of traditional single ultrasound indicators or maternal characteristics, for the first time, 15 cross-dimensional clinical characteristics (such as uterine height, abdominal circumference, gestational diabetes, etc.) were integrated to build a "maternal-fetal" joint prediction system;

[0046] (2) Algorithm comparison and interpretability: Through a systematic comparison of five machine learning algorithms, a high-performance XGBoost model was selected, and SHAP interpretation technology was introduced to solve the "black box" problem of traditional machine learning and enhance clinical trust;

[0047] (3) Sample balance and generalization ability: The SMOTE algorithm was used to deal with the sample imbalance problem caused by the low incidence of macrosomia (4.9%), and stratified validation was performed to ensure the applicability of the model in the Han Chinese population with singleton cephalic presentation.

[0048] (4) Data foundation: a clear data collection scope (such as objective indicators in electronic medical records) and preprocessing process (one-hot encoding, SMOTE), which can realize automated data processing through existing medical information systems; Algorithms and tools: algorithm deployment is realized using mature machine learning libraries such as Scikit-learn, hyperparameter grid search and five-fold cross-validation are conventional optimization methods, and computing resources can be achieved through ordinary workstations or cloud computing; Clinical validation: the model performance is verified based on 3233 real clinical data (test set accuracy is 0.973, AUC=0.9986), which has a real-world application foundation. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A flowchart of this application;

[0050] Figure 2 Corresponding diagrams of five models;

[0051] Figure 3 It is a decision curve analysis diagram;

[0052] Figure 4 This is the risk prediction result diagram. DETAILED DESCRIPTION

[0053] The following is further described in detail through specific implementation methods:

[0054] A method for predicting the risk of macrosomia based on machine learning, comprising the following steps: Figure 1, clinical delivery data of 3233 Han pregnant women from 2024 to 2025 were collected from a tertiary hospital, including basic characteristics, test indicators, fetal indicators, etc.; missing value processing, outlier removal and SMOTE method were used to achieve data balance. 1.2 Feature selection and dimensionality reduction: Lasso regression was used to screen key predictive variables, and the final variable set was retained based on clinical significance. 1.3 Model construction and training: Prediction models were constructed using machine learning algorithms such as random forest (RF), XGBoost, CatBoost, K nearest neighbor (KNN) and multilayer perceptron (MLP). 1.4 Model evaluation and selection: Compare the R of each model on the validation set. 2 Value, area under the ROC curve (AUC), decision curve (DCA) and calibration curve, and select the best performing model as the final model. Model interpretability analysis: The SHAP method is introduced to evaluate the contribution of each feature to the model prediction results and provide a personalized explanation basis. 1.5 Model validation: In order to comprehensively verify the generalization ability and predictive performance of the constructed machine learning model in practical applications, after the model training and internal validation are completed, 20 external validation samples are further introduced for independent testing to ensure the objectivity and authenticity of the external validation. 1.6 Result display and output: A graphical user interface (GUI) is constructed using the trained prediction model, including a model training interface, a test interface and a prediction result visualization interface for medical staff to operate and use.

[0055] The experimental environment consists of two parts: hardware equipment and software tools.

[0056] The experiment runs on a computer equipped with a 13th Gen Intel(R) Core(TM) i5-13600K CPU and an NVIDIA GeForce RTX4060 GPU to ensure computational efficiency and model training speed. In terms of the software environment, the Windows operating system is used, and the Python environment is managed through Anaconda, with Python 3.9 being the primary Python version used. Core libraries include scikit-learn (for building random forest models), NumPy and Pandas (for data processing), Matplotlib and Seaborn (for data visualization), and code writing, debugging, and result analysis are performed in the Jupyter Notebook interactive environment. This hardware and software configuration supports efficient model training and optimization, ensuring experimental stability and reproducibility.

[0057] Clinical data of pregnant women admitted to the hospital between April 2024 and March 2025 were collected. The inclusion criteria were as follows: (1) singleton, Han ethnicity, cephalic presentation, and 3233 women who had delivered met the inclusion criteria. The exclusion criteria were as follows: (1) non-Han ethnicity; (2) non-cephalic presentation; (3) women who did not cooperate with follow-up. Finally, the data of 2586 women were included in the training set, and the data of 647 women were included in the test set for in vitro validation. This study was retrospective and all data were anonymized, so the requirement for patient informed consent could be waived. The dataset was divided into 8:2 parts using the holdout method to ensure that the model had sufficient data support during training, reduce the risk of model overfitting, and provide more objective and accurate performance evaluation results, thereby improving the practical application value of the model. After the grid search algorithm and five-fold cross-validation, the optimal parameters of each model were determined, as shown in Table 1 below.

[0058] Table 1

[0059]

[0060]

[0061] After optimizing the hyperparameters for each model, this article retrained the five machine learning models using their respective optimal parameters. The test set data was then fed into the trained models, and various evaluation metrics on the test set were calculated, including accuracy, precision, and recall. The results are summarized in the table below. Comparative analysis reveals that the performance of the various models on the test set varies, reflecting the differences in sample learning and generalization capabilities of the different algorithms.

[0062] Judging from the average values ​​of the evaluation indicators, the XGBoost model has the best comprehensive performance. The average value of the three indicators reached 0.973, ranking first among all models, indicating that it is the most stable in balancing the recognition of positive and negative samples and reducing the misjudgment rate. XGBoost has the advantages of a gradient boosting framework in ensemble learning. It can optimize the model structure by gradually fitting the residuals, effectively improving the expression and generalization capabilities of complex features. In summary, XGBoost performed the most robustly in this study, taking into account both high precision and low misjudgment, and is suitable for high-reliability prediction of clinical features. The performance of the MLP model is relatively weak, and its classification ability can be improved in the future by increasing the size of training samples, optimizing the network structure, or adopting transfer learning. See Table 2 below:

[0063] Table 2

[0064]

[0065] Down Figure 2The confusion matrix plots for five machine learning models on the test set are presented to evaluate their performance in identifying samples from different categories. Analysis of the confusion matrices reveals that the five models generally outperformed category 1 (big-headed children) in identifying samples from category 0 (non-big-headed children). The K-nearest neighbor (KNN) algorithm performed best in identifying samples from category 1, achieving 100% prediction accuracy. This means that no cases of misclassifying category 1 as category 0 occurred in the test set, demonstrating its high sensitivity to this category. However, the KNN model's accuracy for category 0 was relatively low, with a high misclassification rate, indicating some discrepancies in its ability to distinguish samples from different categories, and further improvement is needed in terms of model stability.

[0066] In contrast, the overall prediction accuracy of the multilayer perceptron (MLP) neural network was relatively low, with a particularly high misclassification rate for category 1. This phenomenon may be attributed to the limited number of training samples, which prevented the model from fully learning the characteristic distribution of various samples during training, thereby affecting the model's generalization ability. Neural network models are sensitive to data volume and feature diversity. When samples are insufficient or features are unrepresentative, they are prone to overfitting or underfitting, thereby reducing predictive performance. Overall, although some models perform well in specific categories, further data expansion, feature optimization, or integration strategies are needed to improve the model's robustness and balance in multi-category sample recognition.

[0067] Down Figure 3 The ROC (Receiver Operating Characteristic) curves and their corresponding AUC (Area Under the Curve) values ​​corresponding to the five models are shown to evaluate the comprehensive performance of the five machine learning models in classification tasks. It can be clearly seen from the figure that the overall distribution of the ROC curves of the five models is relatively ideal, and most curves are close to the upper left corner of the image, which means that the model has a higher true positive rate (True Positive Rate) and a lower false positive rate (False Positive Rate) under different classification thresholds, that is, it shows a stronger classification ability. The AUC value is an important indicator to measure the comprehensive discrimination ability of the model. The value range is between 0.5 and 1. The closer it is to 1, the closer the model is to a perfect classifier. The experimental results show that the AUC values ​​of all models are at a high level, reflecting the excellent performance of the models in distinguishing positive and negative samples. The AUC value of the CatBoost model is as high as 0.998, which is almost close to the theoretical optimal state, indicating that it has excellent robustness and recognition accuracy on the test set. Other models such as XGBoost and RandomForest also performed well, with AUC values ​​above 0.95, reflecting strong robustness.

[0068] In a decision curve analysis (DCA), the net benefits of different models varied with the decision threshold. The results showed that RandomForest, XGBoost, CatBoost, and KNN achieved significantly higher net benefits than the "All" and "None" strategies across most threshold ranges, demonstrating their superior decision-making value in practical applications. MLP, on the other hand, exhibited negative net benefits at high thresholds, suggesting its unsuitability for high-confidence scenarios and limited clinical or practical application value.

[0069] In the calibration curve analysis, the predicted probabilities of RandomForest and CatBoost are closer to the ideal "perfect calibration line," demonstrating good calibration capabilities. However, the curves of XGBoost and MLP show large fluctuations and deviations from the true probabilities, indicating that their predictions are less reliable. Overall, RandomForest and CatBoost excel in both decision value and calibration performance, making them more trustworthy classification models.

[0070] By comprehensively comparing the visualizations of the SHAP analysis results (including histograms and scatter plots), we found that feature AC consistently exhibited the highest feature importance in three mainstream classification models: Random Forest (RF), XGBoost, and CatBoost, with its SHAP contribution consistently remaining above 0.35. This indicates that regardless of how the model structure changes, the AC variable has the most significant impact on the model's prediction of the "big head" results. In the scatter plot, feature AC's points are distributed widely and densely, indicating that it not only makes a significant contribution to the local model but also has a strong influence on the majority of samples. This "inter-model consistency" strengthens our confidence in the value of the AC variable.

[0071] In contrast, the features "Weight at Admission," "Gravidity," and "Diabetes" generally exhibit lower SHAP contributions, with fewer points and shorter lengths in the scatter plots. This indicates that these features have weak predictive power in the existing model and contribute little to the final classification result.

[0072] From a data science perspective, this result is a statistical correlation automatically learned by the model based on the training samples, which can effectively capture the underlying patterns and nonlinear relationships in the data. However, when we compare it with existing clinical knowledge, we may find a certain degree of inconsistency. Traditional medical research often emphasizes pregnancy number and history of diabetes as important risk factors, which are highly valuable in clinical decision-making. Therefore, the "deviation" between model output and clinical experience deserves further discussion and dialectical analysis:

[0073] 1. The randomness of model results and the data-driven nature

[0074] Machine learning models rely on data samples for training. If certain features are under-distributed across the sample, have insignificant fluctuations, or suffer from measurement bias, their contributions may be underestimated. For example, in the current sample set, the "Diabetes" label may be incomplete, have inconsistent recording standards, or its impact may be masked by the AC feature. Therefore, model results are not absolute, but rather represent the "best explanation given the current dataset."

[0075] 2. The systematic nature and limitations of clinical experience

[0076] Clinical experience, derived from years of accumulation and mechanistic reasoning, can also be prone to averaging, making it difficult to capture the complex interactions between individual data. While AC is not prominently emphasized in traditional medical research, in data-driven models, it may reflect a potential, multifactorial composite indicator. This suggests that we should focus on the discovery and validation of novel indicators in clinical research.

[0077] 3. Complementarity between Data Science and Clinical Medicine

[0078] Truly high-quality medical decision-making should be based on the integration of data science and clinical experience. Model outputs should be interpreted and utilized as "decision-making tools" rather than replacing clinical judgment. Similarly, clinical experts should also value the "new data-driven clues" provided by the model and use them as a basis for further mechanistic research or clinical trial validation.

[0079] Clinical data verification

[0080] In order to fully verify the generalization ability and predictive performance of the constructed machine learning model in practical applications, this paper further introduced 20 external validation samples for independent testing after the model training and internal validation were completed, including 10 "big-headed" samples and 10 "non-big-headed" samples. All samples are derived from real clinical data and did not participate in model training to ensure the objectivity and authenticity of the external validation. Five machine learning models, including random forest (RF), XGBoost, CatBoost, KNN and multi-layer perceptron (MLP), were used to predict external samples. The input features included variables with high contribution in the previous SHAP analysis. The prediction results are shown in the following table, which show good classification performance as a whole, as shown in Table 3:

[0081] Table 3

[0082]

[0083] The prediction accuracy of non-big-head samples was 100%, and none of the five models made any misjudgments in this type of samples, indicating that the models had a relatively sufficient learning effect on the feature distribution of the "normal" category and exhibited extremely high specificity. There were slight misjudgments in the big-head samples, and some models had certain differences in the identification of big-head samples. Among them, the CatBoost model performed the most stably, with a comprehensive prediction accuracy of 90% in both types of samples, demonstrating strong robustness and generalization ability.

[0084] This validation result further confirms the effectiveness and feasibility of our machine learning models. Despite a small amount of error in the prediction of the large-scale sample, overall performance shows that each model accurately captures the inherent patterns of most target features. In particular, even without external data in the training, the model maintains a high recognition rate, which is of great significance for the practical deployment and promotion of the model.

[0085] GUI interface construction

[0086] Aiming at the problem of clinical macrosomia risk prediction, a macrosomia risk prediction software based on machine learning was designed. This software can support the training and comparison of multiple machine learning models, allowing users to upload Excel data files for model training, provide risk prediction functions, and use trained models to evaluate new data. Medical researchers upload historical data containing maternal indicators and pregnancy outcomes. The system automatically trains multiple models and recommends the best model. Clinicians can then use the model to conduct risk assessments on the indicators of new patients. This software is written in Python, and the writing platform is PyCharm. The software interface is designed using the PyQt5 library in Python. The library used for the neural network is PyTorch, and the modeling method uses the SKlearn library. Use the best model recommended by the machine learning modeling and testing interface for prediction. Enter the clinical data of the pregnant woman and click the "Predict" button to determine whether there is a risk of macrocephaly and the probability of macrocephaly. See Figure 4 shown.

[0087] This application has the following innovations:

[0088] (1) Multimodal data integration: Breaking through the limitations of traditional single ultrasound indicators or maternal characteristics, for the first time, 15 cross-dimensional clinical characteristics (such as uterine height, abdominal circumference, gestational diabetes, etc.) were integrated to build a "maternal-fetal" joint prediction system;

[0089] (2) Algorithm comparison and interpretability: Through a systematic comparison of five machine learning algorithms, a high-performance XGBoost model was selected, and SHAP interpretation technology was introduced to solve the "black box" problem of traditional machine learning and enhance clinical trust;

[0090] (3) Sample balance and generalization ability: The SMOTE algorithm was used to deal with the sample imbalance problem caused by the low incidence of macrosomia (4.9%), and stratified validation was performed to ensure the applicability of the model in the Han Chinese population with singleton cephalic presentation.

[0091] (4) Data foundation: a clear data collection scope (such as objective indicators in electronic medical records) and preprocessing process (one-hot encoding, SMOTE), which can realize automated data processing through existing medical information systems; Algorithms and tools: algorithm deployment is realized using mature machine learning libraries such as Scikit-learn, hyperparameter grid search and five-fold cross-validation are conventional optimization methods, and computing resources can be achieved through ordinary workstations or cloud computing; Clinical validation: the model performance is verified based on 3233 real clinical data (test set accuracy is 0.973, AUC=0.9986), which has a real-world application foundation.

[0092] The above is only an embodiment of the present invention, and the common knowledge such as the specific structure and characteristics of the scheme is not described in detail here. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A method for predicting the risk of macrosomia based on machine learning, characterized in that: The steps include: S1. Multimodal clinical data acquisition: Maternal characteristics, fetal ultrasound indicators, and pregnancy complications were collected. Maternal characteristics included admission abdominal circumference, uterine height, pre-pregnancy BMI, admission BMI, age, parity, height, pre-pregnancy weight, and admission weight. Fetal ultrasound indicators included fetal abdominal circumference, head circumference, femoral length, and maximum amniotic fluid depth. Pregnancy complications included gestational diabetes. S2. Data preprocessing: Perform one-hot encoding on the original data obtained in step S1 to process categorical variables, use the SMOTE algorithm to balance positive and negative samples, and divide the data into training and test sets; S3. Multi-algorithm model training and optimization: We build classification models using five machine learning algorithms: random forest, XGBoost, CatBoost, K-nearest neighbor, and multilayer perceptron. We optimize model parameters through hyperparameter grid search and five-fold cross-validation. S4. Model evaluation and selection: Evaluate the performance of each model based on accuracy, precision, recall, AUC, and calibration curves, and select the model with the best overall performance as the target prediction model; S5. Risk prediction and interpretation: The samples to be predicted are input into the target prediction model, and the macrosomia risk prediction results are output. The contribution of each feature to the prediction result is quantified using the SHapley additive interpretation (SHAP) method.

2. The method for predicting the risk of macrosomia based on machine learning according to claim 1, characterized in that: The model with the best comprehensive performance in step S4 is the XGBoost model, and its optimal hyperparameters are 'max_depth':6, 'n_estimators':

100.

3. The method for predicting macrosomia risk based on machine learning according to claim 1, characterized in that: The SHAP method described in step S5 evaluates feature importance by calculating the average SHAP value of each feature, and determines fetal abdominal circumference, admission BMI, head circumference, uterine height, and maternal abdominal circumference as key predictive features.

4. The method for predicting macrosomia risk based on machine learning according to claim 1, characterized in that: The ratio of the training set to the test set is 8:2, and an independent external validation set is randomly selected to verify the generalization ability of the model.

5. A macrosomia risk prediction system based on machine learning, obtained according to any one of claims 1 to 4, characterized in that: include: Data acquisition module: used to collect maternal characteristics, fetal ultrasound indicators and pregnancy complications data of pregnant women; Data preprocessing module: includes a one-hot encoding unit and a sample balancing unit, which are used to process categorical variables and balance positive and negative samples; Model training module: Integrates random forest, XGBoost, CatBoost, K-nearest neighbor, and multilayer perceptron algorithms, and optimizes model parameters through grid search and cross-validation; Model evaluation module: Evaluates model performance based on accuracy, precision, recall, AUC, and calibration curve, and outputs the optimal model; Prediction and interpretation module: Generates risk prediction results based on the optimal model and provides feature contribution explanations through the SHAP method.

6. The macrosomia risk prediction system based on machine learning according to claim 5, characterized in that: The parameter optimization ranges of the XGBoost model in the model training module include max_depth of 3-10 and n_estimators of 50-200.

7. The macrosomia risk prediction system based on machine learning according to claim 6, characterized in that: The data collected by the data acquisition module are structured and extracted from the electronic medical record system, and include clinical data of 3233 Han Chinese pregnant women who have delivered singleton cephalic fetuses.

8. The macrosomia risk prediction system based on machine learning according to claim 7, characterized in that: The risk prediction results output by the prediction and interpretation module are connected to the clinical decision support system for the formulation of personalized delivery plans.

Images