Prediction system for bone marrow suppression risk after radiotherapy of nephroblastoma patient

Abstract

The invention discloses a system for predicting the risk of myelosuppression after radiotherapy of a nephroblastoma patient. Belongs to the technical field of probability prediction, and particularly relates to the field of prediction of myelosuppression risks after radiotherapy of nephroblastoma patients. The technical problem that related technologies related to prediction of the myelosuppression risk after radiotherapy of a nephroblastoma patient are lacked at present is solved. The system comprises a data acquisition module, a data preprocessing module, a data set division module and a prediction module. The data preprocessing module is used for sequentially carrying out sample quality control processing, data missing value processing and feature conversion processing on the nephroblastoma patient data and then inputting the preprocessed nephroblastoma patient data into the data set division module; and the prediction module performs prediction model building and prediction model optimization in sequence, performs model training and testing by using the training set and the testing set, finally obtains a model capable of performing bone marrow suppression risk prediction after radiotherapy of a nephroblastoma patient, and applies the model to actual prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of probability prediction, and specifically relates to the field of prediction of the risk of bone marrow suppression in patients with nephroblastoma after radiotherapy. Background Art

[0002] Myelosuppression (CIM) is an important treatment-related toxicity in cancer patients, which occurs when actively dividing hematopoietic stem and progenitor cells (HSPCs) in the bone marrow (both of which give rise to blood cell lineages) are destroyed by cytotoxic chemotherapy. Chemotherapy-induced CIM usually presents as anemia, neutropenia, thrombocytopenia, and / or lymphopenia, and often leads to life-threatening infections, shortness of breath, fatigue, and increased risk of potential bleeding, which has a serious adverse effect on the patient's prognosis. Therefore, early identification of CIM and the adoption of appropriate preventive and therapeutic measures are very important for the clinical management of cancer patients.

[0003] The major factor in the development of CIM is the intensity of the cytotoxic drug administered to the patient, but other variables may influence the probability of such an event.

[0004] However, there is currently no simple and accurate method to identify patients at risk of CIM after chemotherapy. Existing patents, papers, etc. are also limited to the risk prediction of nephroblastoma itself, such as the Chinese invention patent application "A method for constructing a risk prediction model for nephroblastoma based on machine learning" (publication number: CN118155843A). For patients with nephroblastoma, the risk of bone marrow suppression is greater than 50%. There is an urgent need for an effective bone marrow suppression risk prediction method to provide timely prediction feedback during patient treatment to ensure that targeted adjustments can be made in a timely manner. Predicting the occurrence of bone marrow suppression in advance will greatly reduce the economic and disease burden of patients. Summary of the invention

[0005] In order to solve the problem that there is currently a lack of relevant technology related to the prediction of the risk of bone marrow suppression in patients with nephroblastoma after radiotherapy, the present invention provides a prediction system for the risk of bone marrow suppression in patients with nephroblastoma after radiotherapy, the system comprising a data acquisition module, a data preprocessing module, a data set division module and a prediction module;

[0006] The data acquisition module collects the data of patients with nephroblastoma and inputs it into the data preprocessing module;

[0007] The data preprocessing module sequentially performs sample quality control processing, data missing value processing and feature conversion processing on the nephroblastoma patient data, and then inputs the preprocessed nephroblastoma patient data into the data set division module;

[0008] The data set division module divides the preprocessed nephroblastoma patient data into a training set and a test set;

[0009] After the prediction module successively builds and optimizes the prediction model, it uses the training set and the test set to train and test the model, and finally obtains a model that can predict the risk of bone marrow suppression in patients with nephroblastoma after radiotherapy, and applies the model to actual predictions.

[0010] Furthermore, the inclusion criteria for the data of patients with nephroblastoma were: age ≤ 18 years; patients diagnosed with nephroblastoma; patients who had received at least 1 cycle of chemotherapy; patients who had received at least one routine test and laboratory biochemical index test before and after chemotherapy;

[0011] The collected data on patients with nephroblastoma include the first detection indicators and medication information of patients with nephroblastoma, specifically: demographic characteristics, physical examination, blood test, urine test, liver and kidney function, diagnostic information, treatment status and outcome indicators; demographic characteristics, physical examination, blood test, urine test, liver and kidney function, diagnostic information, treatment status and outcome indicators each contain several sub-features, each of which is a characteristic variable; the sub-features contained in the outcome indicators include the occurrence of bone marrow suppression and the absence of bone marrow suppression.

[0012] Furthermore, the sample quality control process is specifically as follows: statistically calculate the missing rate of characteristic variables of each sample; if there is a sample with a missing rate of characteristic variables greater than or equal to 50%, delete this sample; if there are 50% or more samples that are missing a certain characteristic variable, delete this characteristic variable; determine the sample size by the above method.

[0013] Furthermore, the data missing value processing is specifically as follows:

[0014] S41, deletion: After determining the sample size, check the missing rate of each characteristic variable. If 20% or more of the samples are missing a certain characteristic variable, delete this characteristic variable; count the frequency of use of chemotherapy drugs, and delete drugs with a frequency of use of 5% less;

[0015] S42. Interpolation: Use the mode to fill in missing categorical characteristic variables, use the median to fill in missing continuous characteristic variables, and use the corresponding first visit examination data to fill in the derived characteristic variables in the trend of laboratory indicators if they are missing.

[0016] Further, the feature conversion process is specifically as follows:

[0017] S51. Convert units of some characteristic variables: for neutrophil percentage, monocyte percentage, basophil percentage and lymphocyte percentage, if the value is > 1, divide by 100 for unit conversion;

[0018] S52, adding a composite characteristic variable: adding composite characteristic variable 1, composite characteristic variable 2 and composite characteristic variable 3 to the characteristic variables;

[0019] Composite characteristic variable 1 = platelet count*absolute neutrophil count / absolute lymphocyte count;

[0020] Composite feature variable 2 = absolute value of neutrophils / absolute value of lymphocytes;

[0021] Composite characteristic variable 3 = platelet count / absolute lymphocyte value;

[0022] S52. Add derived characteristic variables: Add body surface area and combined use of highly toxic chemotherapy drugs to the characteristic variables:

[0023] Body surface area = 0.035*body weight + 0.1 (body weight ≤ 30 kg), body surface area = 1.05 + (body weight - 30) * 0.02 (body weight > 30 kg);

[0024] The method for determining the combination of highly toxic chemotherapy drugs is as follows: if one of cyclophosphamide, ifosfamide, doxorubicin, epirubicin, actinomycin D, carboplatin, etoposide, topotecan, and vindesine is used, it is a combination of highly toxic chemotherapy drugs;

[0025] S53. Use median values ​​to describe continuous characteristic variables, and use frequencies and percentages to describe categorical characteristic variables.

[0026] Furthermore, in the data set division module, the data were divided into two groups according to the outcome indicators: those with bone marrow suppression and those without bone marrow suppression. Then, the two groups of data were randomly divided into a training set and a test set in a ratio of 7:3.

[0027] Furthermore, the prediction model is specifically constructed as follows:

[0028] S71. Basic model selection: Select the prediction model based on the XGBoost model;

[0029] S72. Feature selection and weight enhancement:

[0030] S72.1. Use information value as the correlation index and set a threshold. Delete feature variables with information value less than the threshold. Increase the original weight of feature variables with information value greater than the threshold by 10%. Increase the original weight of feature variables with information value equal to the threshold by 15%.

[0031] S72.2. Calculate the correlation coefficient of the characteristic variables, take the absolute value of the correlation coefficient to obtain the collinear variables, and set a threshold. For the collinear variables exceeding the threshold, remove the characteristic variable with the smallest information value;

[0032] S73, cross-validation partitioning: Model hyperparameter search is performed based on cross-validation, with the goal of maximizing the average AUC on the cross-validation validation set, and the optimal hyperparameters of the model are obtained through the Bayesian optimization strategy.

[0033] Further, the prediction model optimization is specifically as follows:

[0034] S81, Adaptive hyperparameter tuning: The optimal hyperparameters are obtained through dynamic learning rate adjustment, tree depth and regularization adjustment, subsample ratio and column sampling ratio tuning, and adaptive hyperparameters are adjusted;

[0035] S82, multi-objective optimization mechanism: perform specificity optimization, sensitivity optimization and loss function optimization;

[0036] S83, weighted gradient boosting mechanism: increase the category weight for high-risk samples of myelosuppression, so that the model will incur a greater loss penalty for high-risk samples during training; assign a higher label distribution weight to high-risk samples of myelosuppression than to low-risk samples of myelosuppression, to avoid overfitting of the model to low-risk samples;

[0037] S84. Data augmentation and sample expansion: Use synthetic minority sample oversampling technology to additionally generate high-risk samples for bone marrow suppression; use stratified cross-validation method to ensure that each training set contains samples with and without bone marrow suppression in a consistent proportion.

[0038] Furthermore, the multi-objective optimization mechanism is specifically:

[0039] Specificity optimization: The specificity threshold was set at 85%. If the specificity was lower than 85%, the false positive rate was reduced by adjusting the weight of samples with low risk of myelosuppression.

[0040] Sensitivity optimization: Set the sensitivity threshold to 70%. If the sensitivity is lower than 70%, increase the penalty coefficient of high-risk samples of bone marrow suppression by weighting.

[0041] Loss function optimization: Based on the standard XGBoost binary classification loss function, specificity and sensitivity are added as weight factors.

[0042] The beneficial effects of the system of the present invention are:

[0043] Set up a multi-layer feature variable screening method to preprocess the data to prevent low-correlation feature variables from interfering with the prediction results.

[0044] By defining composite and derived characteristic variables based on basic characteristic variables, we seek to innovate types in traditional medical indicators and improve the accuracy of predicting bone marrow suppression risk.

[0045] According to the characteristics of the data to be processed, the standard XGBoost is adaptively improved to improve the prediction accuracy of the model.

[0046] The system of the present invention can be applied in the field of predicting the risk of bone marrow suppression in patients with nephroblastoma after radiotherapy. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 The ROC curve diagram of each model in the embodiment of the present invention;

[0048] Figure 2 Schematic diagram of the feature importance of each variable in the optimized XGBoost training set model in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of the present invention.

[0050] Embodiment 1,

[0051] This embodiment provides a prediction system for the risk of myelosuppression after radiotherapy in patients with nephroblastoma, the system comprising a data acquisition module, a data preprocessing module, a data set division module and a prediction module;

[0052] The data acquisition module collects the data of patients with nephroblastoma and inputs it into the data preprocessing module;

[0053] The data preprocessing module sequentially performs sample quality control processing, data missing value processing and feature conversion processing on the nephroblastoma patient data, and then inputs the preprocessed nephroblastoma patient data into the data set division module;

[0054] The data set division module divides the preprocessed nephroblastoma patient data into a training set and a test set;

[0055] After the prediction module successively builds and optimizes the prediction model, it uses the training set and the test set to train and test the model, and finally obtains a model that can predict the risk of bone marrow suppression in patients with nephroblastoma after radiotherapy, and applies the model to actual predictions.

[0056] The system of the present invention is used to evaluate the risk factors for CIM in patients with nephroblastoma after chemotherapy, and to construct an optimized model based on XGBoost to form a CIM risk prediction model, thereby assisting clinicians in identifying patients with a high risk of CIM after chemotherapy.

[0057] Embodiment 2,

[0058] This embodiment further limits the embodiment 1 and further illustrates the data acquisition module.

[0059] (1) Collect data on patients with Wilms tumor.

[0060] The inclusion criteria for the Wilms tumor population were: 1) age ≤ 18 years; 2) patients diagnosed with Wilms tumor; 3) all patients received at least one cycle of chemotherapy; 4) received at least one routine test and laboratory biochemical index test before and after chemotherapy.

[0061] Exclusion criteria: 1) patients with severe infection; 2) incomplete medical records (variables used for analysis are missing more than 50% at the same time); 3) patients with treatment interruption. Patients were divided into two groups according to whether bone marrow suppression occurred. The risk factors for bone marrow suppression in patients were analyzed and a risk prediction model was constructed.

[0062] Collect variables such as the patient's first test index and medication information after admission, including:

[0063] 1) Demographic characteristics: age, gender;

[0064] 2) Physical examination: height, weight, body surface area;

[0065] 3) Blood test: such as: neutrophil percentage, absolute neutrophil count, monocyte percentage, absolute monocyte count, eosinophil percentage, absolute eosinophil count, large platelet ratio, mean corpuscular volume, mean hemoglobin concentration, mean hemoglobin amount, lymphocyte percentage, absolute lymphocyte count, white blood cells, red blood cells, red blood cell distribution width, absolute red blood cell distribution, hematocrit, platelet distribution width, hematocrit, mean platelet volume, platelet count, hemoglobin, reactive protein, variant lymphocytes, basophil percentage, absolute basophil count, eosinophil count, specimen status, white blood cell morphology, red blood cell morphology, reticulocyte, serum amyloid a, etc.;

[0066] 4) Urine examination: such as pH, nitrite, urobilinogen, urine transparency, urine color, qualitative analysis of microalbumin, unclassified crystals, specific gravity, pathological casts, white blood cells, leukocyte enzymes, red blood cells, vitamin C, bilirubin, pus cell clusters, glucose, protein, transparent casts, ketone body yeast, non-squamous epithelium, squamous epithelium, mucus filaments, etc.;

[0067] 5) Liver and kidney function: such as lactate dehydrogenase, prealbumin, uric acid, total bilirubin, total protein, globulin, albumin, alkaline phosphatase, creatinine, cholinesterase, alanine aminotransferase, aspartate aminotransferase, urea, total calcium, unconjugated bilirubin, chloride ion, direct bilirubin, phosphorus, conjugated bilirubin, cystine protease inhibitor, glucose, glutamyl transpeptidase, sodium ion, potassium ion, magnesium, indirect bilirubin, etc.;

[0068] 4) Diagnostic information: tumor stage, COG grade;

[0069] 7) Treatment status: chemotherapy drugs, dosage, chemotherapy course, etc.;

[0070] 8) Outcome indicators: whether bone marrow suppression occurs.

[0071] (2) Definition of myelosuppression outcome

[0072] Definition of myelosuppression: According to the U.S. Department of Health and Human Services Common Criteria for Adverse Events (CTCAE) version 5.0, myelosuppression of grade 2 or above requires treatment, so myelosuppression is defined as:

[0073] (1) White blood cells <3.0 x 10 9 / L;

[0074] (2) Neutrophils <1.5 x 10 9 / L;

[0075] (3) Hemoglobin <100 g / L;

[0076] (4) Platelets <75 x 10 9 / L;

[0077] If any one of the above four points is met at any time point during or after chemotherapy, it can be defined as post-chemotherapy bone marrow suppression. Bone marrow suppression mainly includes anemia, lymphopenia, neutropenia and thrombocytopenia.

[0078] Embodiment 3,

[0079] This embodiment further limits the embodiment 1 and further illustrates the data preprocessing module.

[0080] (1) Sample quality control

[0081] In this example, 1478 medical records of patients with Wilms tumor were initially extracted.

[0082] After matching the patients' first visit examination indicators, a total of 46 clinically relevant characteristic variables were extracted, including age, sex, height, weight, BMI, body surface area, tumor stage, COG grade, chemotherapy course, whether bone marrow suppression occurred, neutrophil percentage, absolute neutrophil value, monocyte percentage, absolute monocyte value, basophil percentage, absolute basophil value, large platelet ratio, mean corpuscular volume, mean hemoglobin concentration, mean hemoglobin amount, lymphocyte percentage, absolute lymphocyte value, white blood cells, red blood cells, red blood cell distribution width, absolute red blood cell distribution, hematocrit, platelet distribution width, platelet hematocrit, mean platelet volume, platelet count, hemoglobin, pH, lactate dehydrogenase, prealbumin, uric acid, total bilirubin, total protein, globulin, albumin, alkaline phosphatase, creatinine, bile acid, cholinesterase, alanine aminotransferase and aspartate aminotransferase. The missing rate of characteristic variables of each sample was calculated, and 50% was selected as the threshold according to the distribution of characteristic variables of each sample and the requirements for model entry. If 50% or more of the 46 characteristic variables were missing at the same time, it was considered that the characteristic variables of the sample were seriously missing and were included in the exclusion criteria.

[0083] Secondly, the chemotherapy drugs used by patients were considered and new variables were included: chemotherapy course, bleomycin, fluorouracil, topotecan, vindesine, ifosfamide, cisplatin, doxorubicin, epirubicin, carboplatin, etoposide, actinomycin D, cyclophosphamide, and vincristine. Among them, the use frequency of bleomycin, fluorouracil, topotecan, vindesine, and ifosfamide was less than 5%, and the difference in use frequency with other drugs was particularly large, so they were deleted. Cisplatin, doxorubicin, epirubicin, carboplatin, etoposide, actinomycin D, cyclophosphamide, and vincristine were included in the model variables.

[0084] According to the above criteria, there were 55 patient baseline characteristic variables in total, and the final sample size was 1433 cases.

[0085] Based on the patient's baseline characteristics, some characteristic variables need to be processed to form composite characteristic variables and derived characteristic variables, with a total of 48 variables. For details, see the description in "Feature Conversion".

[0086] (2) Processing of missing data values

[0087] (2.1) Deletion method:

[0088] Clinical characteristics: After determining the sample size, check the missing rate of each characteristic variable, and select 20% as the threshold according to the model entry requirements. If the missing rate of a characteristic variable exceeds 20%, the variable will be deleted and not included in the model construction.

[0089] Chemotherapy drugs: The frequency of use of chemotherapy drugs was counted, and drugs with a usage frequency of 5% lower were deleted and not included in the model construction.

[0090] Frequency of use = number of times used / total sample size

[0091] (2.2) Interpolation method:

[0092] For the patient's first visit examination indicators, missing categorical variables are filled with the mode, and missing continuous variables are filled with the median. For the derived variables in the laboratory indicator change trend, if they are missing, they are filled with the corresponding first visit examination data, such as mean corpuscular volume, mean hemoglobin concentration, etc., which need to be calculated based on multiple values. The average is called a derived variable in the laboratory indicator change trend. If the mean cannot be calculated, the laboratory test value of the first visit is used.

[0093] (3) Feature conversion

[0094] (3.1) Unit conversion

[0095] For the percentage of neutrophils, monocytes, basophils, and lymphocytes, if the value is > 1, divide by 100 for unit conversion.

[0096] (3.2) Calculation of composite characteristic variables

[0097] Add composite characteristic variable 1, composite characteristic variable 2, composite characteristic variable 3, body surface area and combined use of highly toxic chemotherapy drugs to the characteristic variables;

[0098] Composite characteristic variable 1 = platelet count*neutrophil absolute value / lymphocyte absolute value;

[0099] Composite feature variable 2 = absolute value of neutrophils / absolute value of lymphocytes;

[0100] Composite characteristic variable 3 = platelet count / absolute lymphocyte value;

[0101] Body surface area = 0.035*weight + 0.1 (weight ≤ 30 kg);

[0102] Body surface area = 1.05 + (body weight - 30) * 0.02 (body weight > 30 kg);

[0103] The method for determining the combination of highly toxic chemotherapy drugs is as follows: if one of cyclophosphamide, ifosfamide, doxorubicin, epirubicin, actinomycin D, carboplatin, etoposide, topotecan, and vindesine is used, it is a combination of highly toxic chemotherapy drugs;

[0104] Classification of hematologic toxicity of chemotherapy drugs:

[0105] High: cyclophosphamide, ifosfamide, doxorubicin, epirubicin, actinomycin D, carboplatin, etoposide, topotecan, and vindesine;

[0106] Medium / low: cisplatin, vincristine, bleomycin, fluorouracil.

[0107] (3.3) Statistical methods

[0108] Continuous variables were described using median values ​​(1 / 4-3 / 4 quantiles), and categorical variables were described using frequencies and percentages. The Wilcoxon rank sum test was used for comparisons between groups of continuous variables, and the chi-square test (χ 2 ).

[0109] Embodiment 4,

[0110] This embodiment further limits the embodiment 1 and further illustrates the data set division module.

[0111] (1) According to the outcome stratification, the data set was divided into two groups: those with bone marrow suppression and those without bone marrow suppression, as shown in Table 1:

[0112] Table 1:

[0113]

[0114]

[0115]

[0116]

[0117]

[0118] M(Q1-Q3): median value (1 / 4-3 / 4 quantile value);

[0119] COG: Children's Oncology Group;

[0120] (2) According to the ratio of 7:3, the data were randomly divided into two groups: a training set (N = 1003) and an independent test set (N = 430), as shown in Table 2:

[0121] Table 2:

[0122]

[0123]

[0124]

[0125]

[0126]

[0127] Embodiment 5,

[0128] This embodiment further limits the embodiment 1 and further illustrates the prediction module.

[0129] The model building is specifically a prediction model based on the XGBoost model. XGBoost (Extreme Gradient Boosting, XGB) is a boosting algorithm, which is mainly used for regression, classification, sorting and other problems. XGBoost gradually adds new models by optimizing the loss function, and each step attempts to correct the prediction error of the previous step. Compared with the traditional GBM (Gradient Boosting Machine), XGBoost adds a regularization term to the objective function and expands the loss function Taylor to the second order, which improves the efficiency and accuracy of the calculation. The XGBoost algorithm is the most flexible and popular algorithm in the boosting family, but XGBoost has the problem of feature engineering dependence and sensitivity to missing values ​​and outliers. Based on the problems of XGBoost, in the present invention, XGBoost is improved to make it more suitable for the application of the present invention. The specific improvements are:

[0130] (1) Feature selection and weight enhancement:

[0131] Feature screening: Use information value (IV) as a correlation indicator. Information value IV is a statistical indicator of information theory, which is used to measure the difference in the distribution of a characteristic variable between two groups of samples with different outcomes (myelosuppression and no myelosuppression), and then characterize the predictive ability of the characteristic variable for the outcome. Set the threshold of the IV value to 0.2, and delete the characteristic variables with a value less than 0.2. Perform a preliminary screening of the characteristic variables based on the information value, and assign higher initial weights to features that are strongly correlated with the risk of myelosuppression (such as hemoglobin, white blood cells, alkaline phosphatase, and albumin). This can be achieved by constructing a new feature importance matrix so that the model is more inclined to focus on these high-weight features during training.

[0132] (2) Multicollinearity calculation

[0133] For the characteristic variables screened out and related to the outcome, the absolute value of the correlation coefficient is calculated to examine collinearity, and the threshold is set to 0.8. For collinear variables that exceed the threshold, it means that two or more independent variables are highly correlated, and it is difficult to distinguish the impact of an independent variable on the dependent variable. Therefore, the variables with smaller IV values ​​are removed from the variables, and the variables with the smallest IV values ​​can also be removed. For example, the correlation coefficient between weight and body surface area is greater than 0.8, and the IV value of weight is lower than that of body surface area. Therefore, weight is removed and body surface area is retained.

[0134] (3) Hierarchical weight allocation

[0135] For feature variables with higher IV values ​​(IV>0.2), the weight is increased by 10% on the basis of the original weight. For feature variables with IV values ​​close to 0.2 (such as chemotherapy courses and red blood cell distribution width), the weight is also considered to be increased by 15% to ensure that the model can effectively use important features when distinguishing patient risks.

[0136] (4) Feature interaction enhancement

[0137] The XGBoost-based quadratic interaction feature generation method is used to mine the nonlinear relationship between features. For example, the interaction features of "hemoglobin" and "leukocytes" are added to the model to capture their potential compound effects on bone marrow suppression.

[0138] Among the 48 clinical characteristics, 6 characteristic variables had a missing rate of more than 20% (absolute value of basophils, percentage of basophils, cholinesterase, prealbumin, and bile pH), so they were excluded. Finally, 42 clinical characteristic variables were selected as model variables.

[0139] Based on the IV value threshold of 0.2 and the collinearity threshold of 0.8, we screened out variables that were strongly correlated with the outcome. Since the chemotherapy course variable has very important clinical significance, although the IV value is <0.2, its weight is increased through "stratified weight allocation" and put into the model. Therefore, a total of 11 characteristic variables were finally included in the model, as shown in Table 3:

[0140] Table 3:

[0141] variable IV Value Hemoglobin 1.77 Red blood cells 0.708 Alkaline phosphatase 0.422 Red blood cell distribution width 0.392 Combination with highly toxic chemotherapy drugs 0.388 leukocyte 0.372 Neutrophil absolute value 0.369 albumin 0.328 Mean hemoglobin concentration 0.243 Platelet count 0.213 Chemotherapy course 0.082

[0142] (2) Cross-validation division: divide the training set into 5-fold training sets and validation sets:

[0143] The model hyperparameter search is performed based on cross-validation (CV), with the goal of maximizing the average AUC (area under the curve) on the CV validation set, and the optimal hyperparameters of the model are obtained through the Bayesian optimization strategy.

[0144] The specific model optimization is as follows:

[0145] (1) Adaptive hyperparameter tuning: Based on the optimal hyperparameters, the algorithm model is trained again on the entire training set:

[0146] Dynamic learning rate adjustment:

[0147] The learning rate in deep learning is a very important hyperparameter, which determines the step size of parameter updates during the optimization process. Use a higher learning rate (0.003-0.005) in the early stages of model training to accelerate convergence, and gradually reduce the learning rate (about 0.001) when the model approaches the optimum to ensure that the model can accurately fit the data during the final optimization process.

[0148] Tree depth and regularization adjustment:

[0149] The depth of the tree is the maximum value of each node in the tree. According to the Bayesian optimization strategy, the depth of the tree and the L1 and L2 regularization parameters are dynamically adjusted to avoid overfitting. Deeper trees can capture more complex patterns, but the depth should be controlled within a moderate range of 8 in the prediction of bone marrow suppression risk to retain generalization ability.

[0150] Subsample ratio and column sampling ratio tuning:

[0151] In order to improve the robustness of the model and reduce the variance, Bayesian optimization is used to dynamically adjust the subsample ratio (subsample) and column sampling ratio (colsample_bytree), so that each tree uses different subsamples and feature sets to enhance the generalization performance of the model.

[0152] (2) Multi-objective optimization mechanism

[0153] Optimize the balance of AUC, specificity and sensitivity: By introducing multi-objective optimization, AUC, specificity (TNR) and sensitivity (TPR) are used as optimization objectives. Combined with the weighted mechanism, specificity and sensitivity loss terms are added to the objective function to achieve multi-objective balance.

[0154] The specific method is:

[0155] Sensitivity and specificity thresholds were set, and weights were dynamically adjusted in the feature space to improve the performance of the model in patients at high risk of myelosuppression by focusing on the identification of high-risk groups.

[0156] Sensitivity threshold setting:

[0157] Goal: To ensure that the detection rate of high-risk patients is high enough to reduce the risk of missed diagnosis.

[0158] Range:70%.

[0159] Actual adjustment: If the sensitivity of the current model is lower than 70%, the penalty coefficient for high-risk groups will be increased through weighting.

[0160] Specificity threshold setting:

[0161] Goal: Avoid excessive false positives and improve prediction accuracy for low-risk patients.

[0162] Range:85%.

[0163] Practical adjustment: If the specificity is lower than 85%, the false positive rate is reduced by adjusting the weights of low-risk samples.

[0164] In myelosuppression risk prediction, the definition of “high-risk patients” and “low-risk patients” can be based on the following factors:

[0165] (1) Based on the label variables in the data set (such as whether myelosuppression occurs), the patients are divided into two categories:

[0166] High-risk patients: patients who actually develop myelosuppression (labeled 1).

[0167] Low-risk patients: patients who do not develop myelosuppression (label 0).

[0168] (2) Clinical threshold criteria:

[0169] According to the definition of bone marrow suppression (such as the indicators of white blood cells, neutrophils, hemoglobin and platelets in the CTCAE 5.0 standard), if any one of the conditions is met, it can be judged as high risk.

[0170] In machine learning, the loss function can be used to define the error between the predicted value and the true value of a single training sample. For continuous variables, the present invention uses regression loss; for discrete variables, the present invention uses classification loss. Custom optimization of loss function: Based on the standard XGBoost binary classification loss function, specificity and sensitivity are added as weight factors, so that when the model is dealing with patients with different risks, it can improve specificity for high-risk patients without losing sensitivity.

[0171] The present invention modifies the binary classification loss function of the standard XGBoost and introduces a weighting parameter:

[0172]

[0173] N represents the total number of samples in the data set, that is, the sum of all samples evaluated by the loss function. For example, if the training set contains 1003 data, then N = 1003.

[0174] ω i Represents the weighted parameter, which is the weight of the i-th sample and is used to adjust the contribution of different samples in the loss function.

[0175] When the weights of high-risk and low-risk samples are unbalanced, high-risk samples are usually given a larger ω i , for example: high risk sample weight ω i=9, indicating that the prediction errors of high-risk samples will be punished more severely. Low-risk sample weight ω i =1, indicating that the prediction error penalty for low-risk samples is lighter.

[0176] y i represents the true label of the sample, y i is the true classification label of the i-th sample, indicating whether the sample is a high-risk patient:

[0177] y i =1: high-risk patients (bone marrow suppression occurs).

[0178] y i =0: low-risk patients (no bone marrow suppression).

[0179] (3) Weighted Gradient Boosting Mechanism

[0180] Category weight adjustment:

[0181] In view of the fact that there are few high-risk samples of bone marrow suppression, the category weights are increased for high-risk samples, so that the model will incur greater loss penalties for high-risk samples during training. The specific method is to set the "scale_pos_weight" parameter in the XGBoost model, and adjust it adaptively according to the ratio of high-risk samples to low-risk samples, thereby enhancing the model's ability to identify high-risk patients.

[0182] Soft label mechanism introduction:

[0183] By assigning higher label distribution weights (softlabels) to high-risk samples of myelosuppression, the model can avoid overfitting of low-risk samples. Adding label weights to the loss function allows the model to impose higher penalties on high-risk samples when calculating gradients, thereby favoring high-risk samples of myelosuppression in classification decisions.

[0184] (4) Data enhancement and sample expansion

[0185] Data synthesis method (SMOTE) combined with XGBoost:

[0186] The synthetic minority oversampling technique (SMOTE) is used to generate more high-risk sample data, combined with XGBoost training to alleviate the problem of sample imbalance. By generating new high-risk samples, the model's recognition accuracy for patients at high risk of bone marrow suppression can be further improved.

[0187] Stratified cross-validation: The stratified cross-validation method was used to ensure that each training set contained a consistent proportion of samples with and without myelosuppression, thereby improving the stability and generalization ability of the model.

[0188] (5) Feature interpretation and enhanced interpretability

[0189] Improved model interpretability: The contribution of each feature in the XGBoost model is analyzed based on the SHAP (SHapley Additive ex Planations) value to help clinicians understand the impact of the feature on the risk of bone marrow suppression.

[0190] Feature Importance Visualization: Provides a feature importance graph of the variables in the XGBoost model, and visualizes the variables that contribute most to the model, such as hemoglobin, alkaline phosphatase, and albumin, to intuitively present the impact of features on risk.

[0191] (6) Evaluate the performance of the model on the training set and independent test set.

[0192] The specific implementation effects are:

[0193] The area under the receiver operating characteristic curve (ROC), sensitivity (TPR), specificity (TNR), accuracy (ACC) and precision (PPV) are used to characterize the predictive ability of the model. The ROC curves of each model are shown in Figure 1 As shown, Specificity means specificity, Sensiticity means sensitivity, from Figure 1 It can be seen that the optimized XGBoost model of the present invention is significantly better than other algorithm models, and is also significantly improved over the original XGBoost model.

[0194] The AUC value range is 0 to 1. The closer it is to 1, the better the prediction effect of the model. The model stability index (PSI) is used to measure the stability of the model in the training set samples and the validation set samples. PSI < 0.1, the model is stable; PSI: 0.1 to 0.25, the model is slightly unstable; PSI > 0.25, the model is unstable. The fitting effects of each model are shown in Table 4, and the authenticity evaluation of each model is shown in Table 5, where Bestcut-off represents the best cutoff point.

[0195] Table 4:

[0196]

[0197] Table 5:

[0198]

[0199]

[0200] The results show that the optimized XGBoost model has the largest AUC, the highest model sensitivity and specificity, and good model stability. The AUC of the original XGBoost is lower than that of the optimized XGBoost.

[0201] The feature importance of each variable in the optimized XGBoost training set model is as follows Figure 2 As shown, the five variables that contribute most to the model are: hemoglobin, leukocytes, alkaline phosphatase, whether or not highly toxic chemotherapy drugs are used in combination, and albumin.

[0202] After implementing the above improvements, the optimized XGBoost model can achieve:

[0203] Significantly improved AUC values:

[0204] An AUC of 0.896 was achieved in the test set, and the sensitivity and specificity were 76.2% and 93.3%, respectively.

[0205] Enhanced feature interpretability:

[0206] Through SHAP value analysis, the model is able to provide an explanatory weight for each feature, making clinical applications more transparent.

[0207] Stable generalization performance:

[0208] Through stratified cross-validation and soft labeling mechanism, the model can maintain stable performance in high-risk predictions, not only in the training set, but also in the test set, showing good robustness.

[0209] The system of the present invention can be used to predict the risk of myelosuppression in patients with nephroblastoma at an early stage, provide reminders to clinicians, and help them take corresponding preventive and therapeutic measures in a timely manner, thereby improving the prognosis of patients.

Claims

1. A prediction system for the risk of bone marrow suppression in patients with nephroblastoma after radiotherapy, characterized in that: The system includes a data acquisition module, a data preprocessing module, a data set division module and a prediction module; The data acquisition module collects the data of patients with nephroblastoma and inputs it into the data preprocessing module; The data preprocessing module sequentially performs sample quality control processing, data missing value processing and feature conversion processing on the nephroblastoma patient data, and then inputs the preprocessed nephroblastoma patient data into the data set division module; The data set division module divides the preprocessed nephroblastoma patient data into a training set and a test set; After the prediction module successively builds and optimizes the prediction model, it uses the training set and the test set to train and test the model, and finally obtains a model that can predict the risk of bone marrow suppression in patients with nephroblastoma after radiotherapy, and applies the model to actual predictions.

2. The prediction system according to claim 1, characterized in that The inclusion criteria for the data of patients with nephroblastoma were: age ≤ 18 years; patients diagnosed with nephroblastoma; patients who received at least 1 cycle of chemotherapy; patients who received at least one routine test and laboratory biochemical index test before and after chemotherapy; The collected data on patients with nephroblastoma include the first detection indicators and medication information of patients with nephroblastoma, specifically: demographic characteristics, physical examination, blood test, urine test, liver and kidney function, diagnostic information, treatment status and outcome indicators; demographic characteristics, physical examination, blood test, urine test, liver and kidney function, diagnostic information, treatment status and outcome indicators each contain several sub-features, each of which is a characteristic variable; the sub-features contained in the outcome indicators include the occurrence of bone marrow suppression and the absence of bone marrow suppression.

3. The prediction system according to claim 2, characterized in that The sample quality control process is specifically as follows: The missing rate of characteristic variables of each sample is counted. If there is a sample with a missing rate of characteristic variables greater than or equal to 50%, this sample will be deleted; if 50% or more of the samples are missing a certain characteristic variable, this characteristic variable will be deleted; the sample size is determined by the above method.

4. The prediction system according to claim 3, characterized in that The data missing value processing is specifically as follows: S41, deletion: After determining the sample size, check the missing rate of each characteristic variable. If 20% or more of the samples are missing a certain characteristic variable, delete this characteristic variable; count the frequency of use of chemotherapy drugs, and delete drugs with a frequency of use of 5% less; S42. Interpolation: Use the mode to fill in missing categorical characteristic variables, use the median to fill in missing continuous characteristic variables, and use the corresponding first visit examination data to fill in the derived characteristic variables in the trend of laboratory indicators if they are missing.

5. The prediction system according to claim 4, characterized in that The feature conversion process is specifically as follows: S51. Convert units of some characteristic variables: for neutrophil percentage, monocyte percentage, basophil percentage and lymphocyte percentage, if the value is > 1, divide by 100 for unit conversion; S52, adding a composite characteristic variable: adding composite characteristic variable 1, composite characteristic variable 2 and composite characteristic variable 3 to the characteristic variables; Composite characteristic variable 1 = platelet count*absolute neutrophil count / absolute lymphocyte count; Composite feature variable 2 = absolute value of neutrophils / absolute value of lymphocytes; Composite characteristic variable 3 = platelet count / absolute lymphocyte value; S52. Add derived characteristic variables: Add body surface area and combined use of highly toxic chemotherapy drugs to the characteristic variables: Body surface area = 0.035*body weight + 0.1 (body weight ≤ 30 kg), body surface area = 1.05 + (body weight - 30) * 0.02 (body weight > 30 kg); The method for determining the combination of highly toxic chemotherapy drugs is as follows: if one of cyclophosphamide, ifosfamide, doxorubicin, epirubicin, actinomycin D, carboplatin, etoposide, topotecan, and vindesine is used, it is a combination of highly toxic chemotherapy drugs; S53. Use median values ​​to describe continuous characteristic variables, and use frequencies and percentages to describe categorical characteristic variables.

6. The prediction system according to claim 5, characterized in that In the data set division module, the data were divided into two groups according to the outcome indicators: one with bone marrow suppression and one without bone marrow suppression. Then the two groups of data were randomly divided into a training set and a test set in a ratio of 7:

3.

7. The prediction system according to claim 6, characterized in that The prediction model is specifically constructed as follows: S71. Basic model selection: Select the prediction model based on the XGBoost model; S72. Feature selection and weight enhancement: S72.

1. Use information value as the correlation index and set a threshold. Delete feature variables with information value less than the threshold. Increase the original weight of feature variables with information value greater than the threshold by 10%. Increase the original weight of feature variables with information value equal to the threshold by 15%. S72.

2. Calculate the correlation coefficient of the characteristic variables, take the absolute value of the correlation coefficient to obtain the collinear variables, and set a threshold. For the collinear variables exceeding the threshold, remove the characteristic variable with the smallest information value; S73, cross-validation partitioning: Model hyperparameter search is performed based on cross-validation, with the goal of maximizing the average AUC on the cross-validation validation set, and the optimal hyperparameters of the model are obtained through the Bayesian optimization strategy.

8. The prediction system according to claim 7, characterized in that The prediction model optimization is specifically as follows: S81, Adaptive hyperparameter tuning: The optimal hyperparameters are obtained through dynamic learning rate adjustment, tree depth and regularization adjustment, subsample ratio and column sampling ratio tuning, and adaptive hyperparameters are adjusted; S82, multi-objective optimization mechanism: perform specificity optimization, sensitivity optimization and loss function optimization; S83, weighted gradient boosting mechanism: increase the category weight for high-risk samples of bone marrow suppression, so that the model will produce a greater loss penalty for high-risk samples of bone marrow suppression during training; Assign higher label distribution weights to samples with high risk of myelosuppression than to samples with low risk of myelosuppression to avoid overfitting of the model to low-risk samples; S84, Data augmentation and sample expansion: Use synthetic minority sample oversampling technology to generate additional high-risk samples for bone marrow suppression; A stratified cross-validation method was used to ensure that each training set contained samples with and without myelosuppression in a consistent proportion.

9. The prediction system according to claim 8, characterized in that The multi-objective optimization mechanism is specifically: Specificity optimization: The specificity threshold was set at 85%. If the specificity was lower than 85%, the false positive rate was reduced by adjusting the weight of samples with low risk of myelosuppression. Sensitivity optimization: Set the sensitivity threshold to 70%. If the sensitivity is lower than 70%, increase the penalty coefficient of high-risk samples of bone marrow suppression by weighting. Loss function optimization: Based on the standard XGBoost binary classification loss function, specificity and sensitivity are added as weight factors.

Images