System for predicting risk of delayed nausea and vomiting caused by child chemotherapy

Abstract

The invention discloses a risk prediction system for delayed nausea and vomiting caused by children chemotherapy, and belongs to the technical field of medical data processing. The system comprises a data acquisition module, a sample construction module, a model construction module and a risk analysis module. The sample construction module is used for processing the child patient clinical data through a genetic programming algorithm to generate extended feature data, and generating a model sample data set according to the extended feature data and the child patient clinical data; and the model construction module is used for training the artificial neural network model through the model sample data set. According to the invention, the clinical data of the child patient is acquired in real time, the current model sample data set is generated, and the risk prediction is carried out through the artificial neural network, so that an instant and accurate risk assessment result is provided, and a reliable decision support tool is provided for medical staff.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data processing, and specifically to a risk prediction system for delayed nausea and vomiting caused by chemotherapy in children. Background Art

[0002] CINV (Chemotherapy-induced nausea and vomiting) is a significant adverse reaction faced by child cancer patients receiving chemotherapy. Even with antiemetic regimens to relieve the symptoms, approximately 33%-60% of child patients still experience delayed CINV (reactions occurring 24 hours or more after chemotherapy). Delayed CINV is particularly distressing as it affects the patient's treatment compliance and overall quality of life, and increases the burden on the healthcare system. In children, the subjective nature of nausea, combined with communication challenges, leads to underreporting and suboptimal management of CINV. Therefore, CINV remains one of the most concerning side effects for patients and parents.

[0003] The pathophysiological mechanism of CINV is multifactorial and involves complex interactions between the central nervous system and the gastrointestinal tract, and the pathophysiological mechanism of CIN (Chemotherapy-induced nausea) is particularly difficult to understand. Current guidelines for the management of CINV in children are mainly based on the emetic potential of chemotherapy drugs and do not fully consider individual patient factors. Risk factors for CINV in adults have been clearly defined, and risk prediction models and scoring systems have been developed based on large-scale, multi-center studies. These scoring systems provide valuable clinical guidance by considering the emetic potential of chemotherapy and individual patient factors, and help optimize the management of CINV. However, in the pediatric population, these risk factors may vary significantly and are often not fully explored in the existing literature. Recent studies have shown that factors such as older age, higher body mass index, hematological malignancies, poor control of acute CINV, and motion sickness significantly affect the risk of delayed CINV. However, these studies are heterogeneous, and there is very little research on prediction models for delayed CINV in children in the prior art.

[0004] Machine learning is a technology that uses algorithms to learn from data and make predictions or decisions, and continuously optimizes model performance through self-optimization. Its advantages lie in processing large and complex datasets, quickly identifying patterns and trends, and providing higher accuracy and efficiency than traditional methods. In pediatric medicine, machine learning has been applied in fields such as disease diagnosis, treatment plan optimization, and risk prediction. However, in existing machine learning research, the prediction models for delayed CINV in children still have deficiencies, such as insufficient consideration of individual differences and single data sources. Summary of the Invention

[0005] To solve the problems existing in the above-mentioned background technology, the present invention aims to develop and verify a machine learning-based prediction model to evaluate the risk of delayed CINV in pediatric cancer patients, provide a decision-making basis for optimizing the prevention strategy of delayed CINV, thereby reducing its incidence and related adverse effects, and ultimately improving the quality of life of pediatric patients.

[0006] The present invention specifically adopts the following technical solutions: A risk prediction system for delayed nausea and vomiting caused by chemotherapy in children, comprising a data collection module, a sample construction module, a model construction module, and a risk analysis module; The data collection module is used to standardize the collection of clinical data of children and perform data preprocessing; The sample construction module is used to generate extended feature data by processing the clinical data of children through a genetic programming algorithm, and generate a model sample data set according to the extended feature data and the clinical data of children; The model construction module is used to train an artificial neural network model through the model sample data set; The risk analysis module is used to input the real-time obtained clinical data of children into the sample construction module to obtain the current model sample data set, input the current model sample data set into the artificial neural network model to obtain a risk prediction result, and obtain the prediction contribution of each input feature through the current model sample data set and the risk prediction result.

[0007] As a preferred solution, the clinical data of children at least includes age, gender, height, number of chemotherapy days, whether the chemotherapy regimen includes cisplatin, anticipatory CINV, acute CINV, previous CINV control effect, number of entertainment items, static activities, and expectations for CINV.

[0008] As a preferred solution, the data collection module includes a missing value processing module, a standardization coding module, and a balancing processing module; the missing value processing module is configured to perform median filling and mode filling processing on the clinical data of children; the standardization coding module is configured to perform standardization and dummy variable coding processing on the clinical data of children; the balancing processing module is configured to identify minority class samples in the clinical data of children, synthesize new samples based on the analysis results of the minority class samples, and fuse them into the clinical data of children.

[0009] As a preferred solution, the balancing processing module includes a sample recognition module, a nearest neighbor matching module, and a sample synthesis module; the sample recognition module is used to recognize minority class samples from the clinical data of children; the nearest neighbor matching module is used to match the nearest neighbor sample set of each minority class sample by calculating the Euclidean distance; the sample synthesis module is used to synthesize new samples according to each minority class sample and its nearest neighbor sample set, and fuse the new samples into the clinical data of children.

[0010] As a preferred solution, processing the clinical data of children through a genetic programming algorithm to generate extended feature data, and generating a model sample data set according to the extended feature data and the clinical data of children, includes the steps: Initializing the population: randomly generating a plurality of different individuals based on the clinical data of children, and each individual represents a feature combination formula; Evaluating the fitness of individuals: using each individual to generate new features; adding the generated features to the clinical data of children to form an extended feature data set; using the extended feature data set to train a logistic regression model, evaluating the prediction accuracy of the logistic regression model corresponding to each individual, and calculating the fitness value of each individual; Generating a new population: determining a candidate population for genetic operations according to the fitness value and the selection algorithm; performing genetic operations on the candidate population to obtain new individuals, and updating the population according to the new individuals; Iterative optimization: repeatedly executing the evaluation of the fitness of individuals and the generation of a new population until the termination condition is reached; Feature combination output: generating extended feature data according to the feature combination of the individual with the highest fitness in the last generation population; generating a model sample data set according to the extended feature data and the clinical data of children.

[0011] As a preferred solution, training an artificial neural network model through the model sample data set includes the steps: Dividing the model sample data set into a training set and a validation set; Setting the architecture of the artificial neural network model; the artificial neural network model includes an input layer, a hidden layer, and an output layer; the number of nodes in the input layer is equal to the number of features of the model sample data set; the output layer includes one node for outputting the risk of delayed CINV; Using the training set and the validation set to optimize the weights and biases of the neural network.

[0012] As a preferred solution, obtaining the prediction contribution of each input feature through the current model sample data set and the risk prediction result includes the steps: Setting a reference value for each input feature of the current model sample data set; inputting the reference values corresponding to all input features into the artificial neural network model to obtain a reference prediction result; Select the predicted sample instances to be analyzed, and generate all feature subsets of the predicted sample instances through mathematical combination; input each feature subset into the artificial neural network model to obtain the prediction results; Calculate the marginal contribution of each feature in the feature subset among all feature subsets according to the prediction results; dynamically configure weights according to the size of the feature subset; Perform weighted summation and normalization on the marginal contributions of each feature in all subset environments to obtain the prediction contribution of the feature.

[0013] As a preferred solution, the marginal contribution is expressed as: ; Wherein, represents the marginal contribution; represents the feature subset; i represents the feature number.

[0014] As a preferred solution, the weight is expressed as: ; Wherein, M is the total number of features; represents the feature subset; represents the combination number of selecting k features from M - 1 features; The prediction contribution is expressed as: ; Wherein, i represents the feature number; represents the prediction contribution of feature i; represents all feature subsets that do not contain feature i.

[0015] As a preferred solution, the risk prediction system for delayed nausea and vomiting caused by pediatric chemotherapy of the present invention further includes a visualization module; the visualization module is used to construct a contribution bar chart according to the prediction contributions of all input features of any sample in the current model sample dataset; it is also used to statistically calculate the absolute mean or frequency proportion of the prediction contributions of all input features of all samples in the current model sample dataset, and generate a global feature importance ranking.

[0016] Compared with the prior art, the present invention has the following beneficial effects: Through the sample construction module and the model construction module, the present invention utilizes the high-quality extended features generated by genetic programming and the modeling ability of the artificial neural network to avoid the inefficiency and subjectivity of manual feature engineering, so as to achieve high-precision prediction of the risk of delayed nausea and vomiting; by obtaining the clinical data of children in real time and generating the current model sample dataset, and performing risk prediction through the trained artificial neural network, so as to provide an instant and accurate risk assessment result, thereby providing a reliable decision-making support tool for medical staff.

[0017] In the embodiments of the present invention, genetic programming algorithm is used to process the clinical data of children to generate an extended feature dataset, and a logistic regression model is used as a fitness evaluation tool for the extended feature dataset, taking into account both the computational efficiency of generating the extended feature dataset and the evaluation accuracy. Moreover, the generated extended feature dataset has a clear mathematical meaning, which is convenient for clinicians to understand and apply.

[0018] In the embodiments of the present invention, the risk analysis module performs exhaustive feature subset combination and dynamic weight allocation to meet the fairness of contribution allocation; and for the artificial neural network model, it can effectively process multi-variables and non-linear relationships, enabling the capture of complex interaction effects inside the model; by comprehensively evaluating the individual and interactive contributions of features, the model can provide more reliable predictions and decision support. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0020] Figure 1 It is a schematic structural diagram of a risk prediction system for delayed nausea and vomiting caused by chemotherapy in children according to an embodiment of the present invention; Figure 2 It is a schematic flowchart of generating extended feature data by processing the clinical data of children with genetic programming algorithm and generating a model sample dataset according to the extended feature data and the clinical data of children according to an embodiment of the present invention; Figure 3 It is a schematic flowchart of obtaining the prediction contribution of each input feature according to the current model sample dataset and the risk prediction result according to an embodiment of the present invention; Figure 4 It is a schematic diagram of the visualization result of the user input form according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention. It should be noted that relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations.

[0022] Please refer to Figure 1 , the present invention provides a risk prediction system for delayed nausea and vomiting caused by chemotherapy in children, including a data acquisition module, a sample construction module, a model construction module, and a risk analysis module; The data acquisition module is used to standardize the collection of clinical data of children and perform data preprocessing; The sample construction module is used to generate extended feature data by processing the clinical data of children through a genetic programming algorithm, and generate a model sample data set based on the extended feature data and the clinical data of children; The model construction module is used to train an artificial neural network model through the model sample data set; The risk analysis module is used to input the clinically obtained real-time data of children into the sample construction module to obtain the current model sample data set, input the current model sample data set into the artificial neural network model to obtain a risk prediction result, and obtain the prediction contribution of each input feature through the current model sample data set and the risk prediction result.

[0023] Through the sample construction module and the model construction module, the present invention utilizes the high-quality extended features generated by genetic programming and the modeling ability of artificial neural networks to avoid the inefficiency and subjectivity of manual feature engineering, so as to achieve high-precision prediction of the risk of delayed nausea and vomiting; by obtaining the clinical data of children in real time and generating the current model sample data set, and performing risk prediction through the trained artificial neural network, to provide an immediate and accurate risk assessment result, thus providing a reliable decision support tool for medical staff.

[0024] Specifically, the following embodiments are used to elaborate in detail on a risk prediction system for delayed nausea and vomiting caused by chemotherapy in children according to the present invention: A risk prediction system for delayed nausea and vomiting caused by chemotherapy in children, including a data acquisition module, a sample construction module, a model construction module, and a risk analysis module; The data acquisition module is used to standardize the collection of clinical data of children and perform data preprocessing. Among them, the data source of the data acquisition module is the clinical data of inpatients in the pediatric oncology department (aged 4 - 18 years), including demographic characteristics, chemotherapy regimens, records of acute CINV, etc. Inpatients in the pediatric oncology department aged between 4 and 18 years, diagnosed with malignant tumors and receiving intravenous chemotherapy, are recruited, and all patients or their agents signed a written informed consent at the time of enrollment. For inpatients in the pediatric oncology department meeting the above conditions, the exclusion criteria include: having cognitive or sensory impairments that prevent questionnaire completion, nausea and vomiting caused by underlying diseases (such as gastrointestinal diseases, increased intracranial pressure, etc.), and participating in other antiemetic drug clinical trials. The acquisition method of the above clinical data can be to record daily symptoms (through PeNAT score, NCI - CTCAE grading) using an electronic structured diary (WeChat / phone reminder).

[0025] The sample construction module is used to process the clinical data of children through a genetic programming algorithm to generate extended feature data, and generate a model sample data set based on the extended feature data and the clinical data of children. The model construction module is used to train an artificial neural network model through the model sample data set. The risk analysis module is used to input the clinically obtained data of children in real time into the sample construction module to obtain the current model sample data set, input the current model sample data set into the artificial neural network model to obtain a risk prediction result, and obtain the prediction contribution of each input feature through the current model sample data set and the risk prediction result.

[0026] In one embodiment, the clinical data of children is divided into 29 original input features, which at least include age, gender, height, number of chemotherapy days, whether the chemotherapy regimen includes cisplatin, anticipatory CINV, acute CINV, previous CINV control effect, number of entertainment items, static activities, and expectation of CINV.

[0027] Among them, some of the original input features can be set through prior knowledge or industry standards. Among them, cisplatin is a highly emetogenic chemotherapy drug. Whether the chemotherapy regimen includes cisplatin indicates whether the child has received a cisplatin-based chemotherapy regimen. The use of cisplatin is usually associated with a higher risk of CINV. The previous CINV control effect indicates the control situation of CINV in the child during previous chemotherapy cycles; if the previous CINV control was poor, it may mean that the child has a poor response to antiemetic drugs, increasing the risk of future CINV. Anticipatory CINV indicates that the child has CINV before this chemotherapy. The number of entertainment items indicates the number of entertainment activities the child participated in during chemotherapy. Entertainment activities can distract the child's attention and relieve anxiety, thus potentially reducing the incidence of CINV. Static activities indicate whether the child participated in static activities such as reading or painting. These activities may help relieve anxiety and distract attention, thus reducing the occurrence of CINV. The anticipation of CINV indicates the child's anticipation of whether CINV will occur and its severity during this chemotherapy cycle.

[0028] In this embodiment, the occurrence of acute CINV is an important predictor of delayed CINV (nausea and vomiting from 24 hours after chemotherapy to before the start of the next chemotherapy cycle), and there is a strong correlation between the two. Specifically, acute CINV is defined according to the PeNAT score and the vomiting score. Among them, the PeNAT score is the score reflecting the degree of nausea of the PeNAT, an assessment tool for chemotherapy-related nausea in children. A score of 3 or more indicates relatively severe nausea in the patient.

[0029] The vomiting score means that if the patient vomits 3 or more times within 24 hours, it is also regarded as an indication of acute CINV. This embodiment refers to the assessment method of CIV (Chemotherapy-induced vomiting) in the Common Terminology Criteria for Adverse Events (NCI-CTCAE) proposed by the National Cancer Institute of the United States, and quantifies the vomiting score into 5 levels, corresponding to "1-2 vomiting episodes within 24 hours", "3-5 vomiting episodes within 24 hours", "more than 6 vomiting episodes within 24 hours", "life-threatening situation requiring emergency treatment", and "death" respectively.

[0030] These features together constitute the input of the machine learning model to help predict the risk of delayed CINV in children after chemotherapy. By analyzing these features, the model can identify high-risk children and provide targeted intervention measures for clinicians.

[0031] Furthermore, the data acquisition module includes a missing value processing module, a standardization coding module, and a balancing processing module; The missing value processing module is configured to perform median filling and mode filling on the clinical data of children; among them, for continuous numerical data, the median is used to fill in the missing values to avoid the influence of extreme values on the data distribution; for categorical data, the mode (the most frequently occurring value) is used to fill in the missing values to retain the representativeness and consistency of the data. The missing value processing module effectively reduces the bias and inaccuracy caused by data loss, ensuring that subsequent data processing and analysis can be carried out.

[0032] The standardization and encoding module is configured to perform standardization and dummy variable encoding on the clinical data of children; among them, standardization converts continuous numerical data to a unified scale, such as zero mean and unit variance, which can accelerate the convergence speed of the model and improve the accuracy. Dummy variable encoding converts categorical data into binary dummy variables to facilitate the model to process and interpret the influence of categorical features. Standardization and dummy variable encoding are important steps in data preprocessing, which can effectively eliminate the dimension difference and category bias, enabling the model to better understand and utilize data features.

[0033] The balancing processing module is configured to identify the minority class samples in the clinical data of children, synthesize new samples based on the analysis results of the minority class samples, and fuse them into the clinical data of children. The balancing processing module can be based on the Synthetic Minority Over-sampling Technique (SMOTE) algorithm, which can improve the performance of the model on imbalanced data, making the model more accurate and reliable when predicting minority class situations, such as rare but severe delayed nausea and vomiting reactions.

[0034] Further, the balancing processing module includes a sample identification module, a nearest neighbor matching module, and a sample synthesis module; The sample identification module is used to identify minority class samples from the clinical data of children; minority class samples refer to the situation where the samples belong to a specific category in the target classification task. Specifically, the sample category refers to the category of the target variable (label) in the dataset. In the case of chemotherapy-induced nausea and vomiting, the target variable is whether the patient has delayed nausea and vomiting.

[0035] The nearest neighbor matching module is used to match the nearest neighbor sample set of each minority class sample by calculating the Euclidean distance; the number of samples in the nearest neighbor sample set is preset to k, and k takes the value of 3 in one embodiment, that is, each minority class sample matches 3 nearest neighbor samples; The sample synthesis module is used to synthesize new samples according to each minority class sample and its nearest neighbor sample set, and fuse the new samples into the clinical data of children; The new sample is represented as: ; Wherein, represents the new sample, denote the i-th minority class sample, is a random number within the range of [0, 1], is the z-th nearest neighbor sample of the i-th minority class sample.

[0036] In the target classification task, class imbalance will cause the model to be biased towards the majority class and ignore the minority class. In this embodiment, by identifying minority class samples and synthesizing more samples, the problem of class imbalance can be effectively reduced, and the representativeness of the minority class can be improved; since more minority class samples are generated, the model can better learn the features of the minority class during the training process; the balancing process makes the model more stable and more robust in dealing with the minority class in practical applications.

[0037] In this embodiment, the sample construction module simulates the biological evolution process (such as selection, crossover, mutation) through genetic programming to automatically generate the optimal feature combination to expand the clinical dataset of children, thereby improving the training effect of the model.

[0038] Please refer to Figure 2 , further, the method of processing the clinical data of children through the genetic programming algorithm to generate extended feature data and generating a model sample dataset according to the extended feature data and the clinical data of children includes the steps: S11. Initialize the population: Randomly generate a plurality of different individuals based on the clinical data of children, and each individual represents a feature combination formula; Based on the foregoing, the clinical data of children contains 29 original features. The population size is set to 100 - 300 individuals, and each individual is a feature combination formula composed of basic operators (such as addition, subtraction, multiplication, division, logarithm, exponent) and original features.

[0039] S12. Evaluate the fitness of individuals: Use each individual to generate new features; add the generated features to the clinical data of children to form an extended feature dataset; use the extended feature dataset to train a logistic regression model, evaluate the prediction accuracy of the logistic regression model corresponding to each individual, and calculate the fitness value of each individual; Using each individual to generate new features means generating new features according to the feature combination formula of each individual. For example, the clinical data of children includes the original features acute CINV and chemotherapy days, and the new feature = acute CINV × log(chemotherapy days).

[0040] Use the extended feature dataset to train a logistic regression model, and the prediction target is the binary classification label of delayed nausea and vomiting (such as occurrence = 1, non-occurrence = 0), and then calculate the individual fitness value based on the classification performance index of the logistic regression model. In one embodiment, the fitness value is the AUC (area under the curve) or F1 score of the logistic regression model.

[0041] S13. Generate a new population: Determine the candidate population for genetic operations based on fitness values and selection algorithms; perform genetic operations on the candidate population to obtain new individuals, and update the population based on the new individuals. The selection algorithm can adopt strategies such as roulette wheel selection or tournament selection. Among them, roulette wheel selection selects individuals proportionally according to fitness values. The higher the fitness value, the greater the probability of being selected; tournament selection randomly selects several individuals and selects the individual with the highest fitness value from them.

[0042] The genetic operations include replication, mutation, transposition, and recombination. Replication directly replicates individuals with high fitness values to the next generation; crossover exchanges subtrees or formula parts of two parent individuals to generate new individuals; mutation randomly modifies subtrees or formula parts of individuals (such as replacing operators, adjusting coefficients); transposition moves subtrees or formula parts of individuals to other positions.

[0043] S14. Iterative optimization: Repeat the evaluation of individual fitness and the generation of a new population until the termination condition is reached. Repeat steps S12 and S13 until the termination condition is reached; the termination conditions include the maximum number of generations and the convergence of the population fitness value. The maximum number of generations can be set to 100, and the population fitness value converges (for example, the change in fitness value is less than 1% for 10 consecutive generations).

[0044] S15. Output of feature combination: Generate extended feature data according to the feature combination of the individual with the highest fitness value in the last generation population; generate a model sample data set based on the extended feature data and the clinical data of the child patients.

[0045] In this step, extended feature data is generated according to the feature combination of the individual with the highest fitness value in the last generation population, and the generated new features are merged with the original data set to form the final model sample data set.

[0046] In this embodiment, the clinical data of child patients is processed by a genetic programming algorithm to generate an extended feature data set, and a logistic regression model is used as a fitness evaluation tool for the extended feature data set; considering that the genetic programming algorithm needs to repeatedly generate and evaluate a large number of feature combinations, using a logistic regression model can significantly reduce the calculation time and resource consumption. This embodiment takes into account both the computational efficiency and evaluation accuracy of generating the extended feature data set, and the generated extended feature data set has a clear mathematical meaning, which is convenient for clinicians to understand and apply.

[0047] Based on the foregoing, this solution aims to use the model sample dataset generated by genetic programming (GP) to train a machine learning model for predicting the risk of chemotherapy-induced nausea and vomiting (CINV) in children. In the specific implementation process, this solution trained 7 different machine learning models, including random forest, SVM, XGBoost, ANN, LightGBM, CatBoost, and Lasso + Logistic regression, and evaluated them through the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 score, and Brier score. Since the artificial neural network (ANN) model performed the best (AUC 0.815, 95% CI 0.695 - 0.903), the artificial neural network (ANN) was finally selected for visualization and deployment on the web page.

[0048] Furthermore, the training of the artificial neural network model by the model sample dataset includes the steps of: S21. Divide the model sample dataset into a training set and a validation set; S22. Set the architecture of the artificial neural network model; the artificial neural network model includes an input layer, a hidden layer, and an output layer; Input layer: The number of nodes is equal to the number of features in the model sample dataset. For example, the model sample dataset includes original features and new features generated by the genetic programming algorithm.

[0049] Hidden layer: One or more hidden layers can be selected. Each hidden layer can contain several nodes. In one embodiment, it is designed as a 2-layer fully connected network, with 64 neurons in each layer, and the activation function is ReLU (rectified linear unit).

[0050] Output layer: One node, used to predict the risk of delayed CINV with continuous values.

[0051] S23. Use the training set and the validation set to optimize the weights and biases of the neural network; During the training process, binary cross-entropy is used as the loss function; the Adam optimizer is used, combined with the backpropagation algorithm to update the weights and biases; L2 regularization (weight decay) is used to prevent overfitting, and the validation set loss is monitored based on early stopping. If the loss does not decrease for 10 consecutive epochs, the training is stopped in advance.

[0052] The validation set is used for hyperparameter tuning and performance evaluation of the model, and is used during training to monitor the model's performance on unseen data. It helps prevent overfitting during training, i.e., the model performs well on the training data but poorly on new data. In this experiment, validation was performed through a 10-fold cross-validation method, where a portion of the data was retained as the validation set each time, and the remaining data was used for training. This method ensures the robustness and generalization ability of the model on the entire training set, making the tuned model parameters more representative and universal.

[0053] Please refer to Figure 3 , further, obtaining the prediction contribution of each input feature through the current model sample dataset and risk prediction results includes the steps of: S31. Set a benchmark reference value for each input feature of the current model sample dataset; input the benchmark reference values corresponding to all input features into the artificial neural network model to obtain a benchmark prediction result; The benchmark reference value adopts the statistics (such as mean, median, or typical state) of the training dataset of the model construction module. For example, the mean is used for continuous features, the mode is used for categorical features, and the high-frequency scheme is used as the benchmark for combined features such as chemotherapy regimens. Then, all features are replaced with the benchmark values and input into the artificial neural network model to obtain a benchmark prediction result, which serves as a comparison baseline for feature contributions.

[0054] S32. Select the prediction sample instance to be analyzed, and generate all feature subsets of the prediction sample instance through mathematical combination; input each feature subset into the artificial neural network model to obtain a prediction result; This step includes two parts: feature subset enumeration and perturbation prediction calculation. Feature subset enumeration specifically means: for the prediction instance to be analyzed, generate all possible feature subsets through mathematical combination. Each subset has some features maintaining the original values of the instance, and the remaining features are replaced with the benchmark values. For example, for the feature set {age, chemotherapy days}, the subsets may include combinations such as only retaining the original value of age, only retaining the original value of chemotherapy days, and retaining the original values of both.

[0055] Perturbation prediction calculation specifically means: input each feature subset into the artificial neural network model and record its corresponding prediction result. By comparing the difference between the subset prediction value and the benchmark prediction value, quantify the impact of this subset on the prediction result.

[0056] S33. Calculate the marginal contribution of each feature in the feature subset among all feature subsets according to the prediction result; dynamically configure weights according to the size of the feature subset; In this step, for each feature in the feature subset, calculate its marginal contribution among all possible subset combinations. Specifically, for the feature subset S containing this feature, its marginal contribution is expressed as: ; Among them, represents the marginal contribution; represents the feature subset; i represents the feature number. The marginal contribution is the change in the prediction result after adding this feature to the subset S.

[0057] Dynamically configure the weights according to the size of the feature subset. Specifically, for the subset S containing k features, its weight is: ; Among them, M is the total number of features; represents the feature subset; represents the number of combinations of selecting k features from M - 1 features. For the subset S of size k, the weight can ensure that the contributions of all subsets are fairly measured.

[0058] S34. Weighted sum the marginal contributions of each feature in all subset environments and perform normalization processing to obtain the predicted contribution of the feature; The predicted contribution is expressed as: ; Among them, i represents the feature number; represents the predicted contribution of feature i; represents all feature subsets that do not contain feature i.

[0059] It should be noted that for the predicted contribution obtained through the above formula, it needs to be proportionally normalized to the range of the difference between the prediction result and the baseline prediction to ensure the interpretability of the contribution value.

[0060] In this step, the marginal contributions of each subset are weighted and summed to obtain the predicted contribution of feature i in each feature subset environment. The value of the predicted contribution represents the net impact of this feature on the current prediction result. A positive value indicates a positive promotion of risk prediction, and a negative value indicates an inhibition of risk.

[0061] In this embodiment, the exhaustive feature subset combination and dynamic weight allocation are performed through the risk analysis module, satisfying the fairness of contribution allocation; and for the artificial neural network model, it can effectively handle multi - variable and non - linear relationships, enabling the capture of complex interaction effects inside the model; by comprehensively evaluating the individual and interaction contributions of features, the model can provide more reliable predictions and decision - making support.

[0062] As a preferred embodiment, the risk prediction system for delayed nausea and vomiting caused by pediatric chemotherapy of the present invention further includes a visualization module; The visualization module is used to construct a contribution bar chart based on the prediction contributions of all input features of any sample in the current model sample dataset; it is also used to statistically calculate the absolute mean or frequency proportion of the prediction contributions of all input features of all samples in the current model sample dataset, and generate a global feature importance ranking.

[0063] The visualization module provided in this embodiment supports from single-case analysis (explaining the reasons for the high risk of a certain child) to population pattern mining (identifying the global association rules between chemotherapy cycles and age). For a single prediction instance, the absolute value and sign of the contribution value of each feature are output to form a visualization chart (such as a SHAP chart), and the contribution value of each feature is displayed through the SHAP chart, so as to intuitively reflect the influence degree of each feature in this sample. Also, the feature contribution values of all samples are statistically calculated, the absolute mean or frequency proportion of the contributions of each feature is calculated, and a global feature importance ranking is generated (such as the influence weight of the chemotherapy regimen type on the overall prediction task). This function helps users understand which features have a greater impact on the model prediction results in general.

[0064] During the implementation process, each visualization chart can be accompanied by a detailed text explanation to illustrate the specific contribution, positive and negative impacts, and importance of each feature to the prediction result of the current sample. Example explanation: "The contribution value of Feature A to the current sample is +0.35, indicating that Feature A positively promotes the prediction of the risk of delayed nausea and vomiting."; An explanation is provided beside the global feature importance chart to illustrate why some features have higher importance in the global model and need to be focused on. Example explanation: "Feature B has the highest absolute mean of contributions among all samples, indicating that it is one of the key factors for predicting the risk of delayed nausea and vomiting."

[0065] Based on the global feature importance ranking obtained from the above embodiment, in one embodiment, the visualization module is further used to screen a key feature set according to the global feature importance ranking, and design a user input form according to the key feature set; the risk analysis module obtains the clinical data of the child according to the user input form.

[0066] In this embodiment, the visualization module removes redundant or low-contribution features based on the results of the aforementioned global feature importance ranking to obtain a key feature set. Based on the aforementioned embodiment, the clinical data of the child is divided into 29 original input features. After removing redundant or low-contribution features, the remaining key feature set has 11 features, including age, gender, height, number of chemotherapy days, whether the chemotherapy regimen includes cisplatin, anticipatory CINV, acute CINV, previous CINV control effect, number of entertainment items, static activities, and the expectation of CINV. Then, a user input form is designed according to the key feature set, and an artificial neural network model is deployed using a Web framework (such as Flask). The visualization result of the user input form after deployment is asFigure 4 As shown, according to the user input received from the user input form, the visualization module transmits the clinical data of the child to the risk analysis module to obtain the risk prediction result.

[0067] In this embodiment, 29 features are used in the training stage, aiming to include all potential relevant features in the initial training stage to avoid missing important information; while reducing the feature input in the final deployment stage can speed up the prediction speed, reduce the data acquisition cost, and better meet the actual clinical needs. Through feature reduction, dynamic interpretation, and global analysis, the visualization module realizes the transformation from a complex model to a clinically practical tool. The 29 features in the training stage are used to fully explore the data potential, while the 11 features in the deployment stage balance efficiency and accuracy. Finally, through web interaction and visualization charts, it helps clinicians quickly identify high-risk children and develop personalized intervention strategies.

[0068] In several embodiments provided by the present application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of the modules is only a logical function division, and there can be other division methods in actual implementation. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces, and the indirect couplings or communication connections of the modules can be in electrical, mechanical or other forms.

[0069] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, that is, they can be located in one place or distributed to multiple network modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0070] In addition, in each embodiment of the present application, the functional modules can be integrated into one processing module, or each module can exist physically alone, or two or more modules can be integrated into one module. The above integrated modules can be implemented in the form of hardware or in the form of software functional modules.

[0071] When the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of this application. The foregoing storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM, read-only memory), random access memories (RAM, random access memory), magnetic disks, or optical discs.

Claims

1. A risk prediction system for delayed nausea and vomiting caused by chemotherapy in children, characterized by: It includes data collection module, sample construction module, model construction module and risk analysis module; The data acquisition module is used to collect clinical data of children in a standardized manner and perform data preprocessing; The sample construction module is used to process the clinical data of the patient through a genetic programming algorithm to generate extended feature data, and to generate a model sample data set based on the extended feature data and the clinical data of the patient; The model building module is used to train the artificial neural network model through the model sample data set; The risk analysis module is used to input the clinical data of the child obtained in real time into the sample construction module to obtain the current model sample data set, input the current model sample data set into the artificial neural network model to obtain the risk prediction result, and obtain the prediction contribution of each input feature through the current model sample data set and the risk prediction result.

2. The risk prediction system for delayed nausea and vomiting caused by chemotherapy in children according to claim 1, characterized in that: The clinical data of the child include at least age, gender, height, number of chemotherapy days, whether the chemotherapy regimen includes cisplatin, expected CINV, acute CINV, previous CINV control effect, number of recreational activities, static activities and expectations for CINV.

3. The risk prediction system for delayed nausea and vomiting caused by chemotherapy in children according to claim 1, characterized in that: The data acquisition module includes a missing value processing module, a standardized coding module and a balanced processing module; the missing value processing module is configured to perform median filling and mode filling processing on the clinical data of the children; the standardized coding module is configured to perform standardization and dummy variable coding processing on the clinical data of the children; the balanced processing module is configured to identify minority samples in the clinical data of the children, synthesize new samples based on the analysis results of the minority samples and merge them into the clinical data of the children.

4. The risk prediction system for delayed nausea and vomiting caused by chemotherapy in children according to claim 3, characterized in that: The balancing processing module includes a sample identification module, a nearest neighbor matching module and a sample synthesis module; the sample identification module is used to identify minority samples from the clinical data of the child; the nearest neighbor matching module is used to match the nearest neighbor sample set of each minority sample by calculating the Euclidean distance; the sample synthesis module is used to synthesize a new sample based on each minority sample and its nearest neighbor sample set, and merge the new sample into the clinical data of the child.

5. The risk prediction system for delayed nausea and vomiting caused by chemotherapy in children according to claim 1, characterized in that: The method of processing the clinical data of the child with a genetic programming algorithm to generate extended feature data, and generating a model sample data set according to the extended feature data and the clinical data of the child with a genetic programming algorithm, comprises the following steps: Initialize the population: randomly generate multiple different individuals based on the clinical data of the children, each of which represents a feature combination formula; Evaluate individual fitness: Use each individual to generate new features; add the generated features to the clinical data of the child to form an extended feature data set; Use the extended feature dataset to train the logistic regression model, evaluate the prediction accuracy of the logistic regression model corresponding to each individual, and calculate the fitness value of each individual; Generate a new population: determine the candidate population for genetic operations based on the fitness value and selection algorithm; perform genetic operations on the candidate population to obtain new individuals, and update the population based on the new individuals; Iterative optimization: Repeatedly evaluate individual fitness and generate new populations until the termination condition is reached; Feature combination output: Generate extended feature data based on the feature combination of the individual with the highest fitness in the last generation of the population; Generate a model sample data set based on the extended feature data and the clinical data of the children.

6. The risk prediction system for delayed nausea and vomiting caused by chemotherapy in children according to claim 1, characterized in that: The method of training the artificial neural network model by using the model sample data set comprises the following steps: Divide the model sample data set into a training set and a validation set; The architecture of the artificial neural network model is set; the artificial neural network model includes an input layer, a hidden layer and an output layer; the number of nodes in the input layer is equal to the number of features of the model sample data set; the output layer includes a node for outputting delayed CINV risk; Optimize the weights and biases of the neural network using the training and validation sets.

7. The risk prediction system for delayed nausea and vomiting caused by chemotherapy in children according to claim 1, characterized in that: The method of obtaining the prediction contribution of each input feature through the current model sample data set and the risk prediction result includes the following steps: Set a benchmark reference value for each input feature of the current model sample data set; The benchmark reference values ​​corresponding to all input features are input into the artificial neural network model to obtain the benchmark prediction results; Select the prediction sample instance to be analyzed, generate all feature subsets of the prediction sample instance through mathematical combination; input each feature subset into the artificial neural network model to obtain the prediction result; Calculate the marginal contribution of each feature in the feature subset in all feature subsets based on the prediction results; dynamically configure the weight according to the size of the feature subset; The marginal contribution of each feature in all subset environments is weighted and normalized to obtain the predicted contribution of the feature.

8. The risk prediction system for delayed nausea and vomiting caused by chemotherapy in children according to claim 7, characterized in that: The marginal contribution is expressed as: ； in, represents marginal contribution; represents a feature subset; i represents a feature number.

9. The risk prediction system for delayed nausea and vomiting caused by chemotherapy in children according to claim 8, characterized in that: The weight is expressed as: ； Among them, M is the total number of features; Represents a feature subset; Indicates the number of combinations of selecting k features from M-1 features; The predicted contribution is expressed as: ； Where i represents the feature number; represents the prediction contribution of feature i; represents all feature subsets that do not contain feature i.

10. The risk prediction system for delayed nausea and vomiting caused by chemotherapy in children according to claim 1, characterized in that: It also includes a visualization module; the visualization module is used to construct a contribution bar chart based on the predicted contribution of all input features of any sample in the current model sample data set; it is also used to count the absolute mean or frequency ratio of the predicted contribution of all input features of all samples in the current model sample data set, and generate a global feature importance ranking.

Images