Method for predicting recurrence of CML patient after TKI treatment withdrawal based on machine learning
Through machine learning-based methods, a prediction model for recurrence after TKI treatment in CML patients was established, which solved the problem of difficult to accurately distinguish between patients with discontinuation of drugs and long-term treatment in the prior art, improved the accuracy of prediction and the choice of treatment plans, and reduced the financial burden and risk of recurrence.
Patent Information
- Application Number
- CN202411168244.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-23
- Publication Date
- 2025-01-10
AI Technical Summary
The prior art is difficult to accurately distinguish patients who are suitable for discontinuation of medication and require long-term continuous treatment based on the clinical characteristics and biometric characteristics of CML patients, resulting in a decrease in treatment compliance and an increase in economic burden.
Using a machine learning-based method, by collecting sample data from CML patients, building a data set and preprocessing, a recurrence prediction model with ensemble structure, including Randomforest and XGboost models, trained to predict the probability of recurrence after TKI treatment. .
Improve prediction accuracy and assist doctors in selecting appropriate treatment options to enable more patients to reach a treatment-free remission state while reducing the risk of recurrence in high-risk patients.
Smart Images
Figure CN119296717A_ABST
Abstract
Description
Technical Field
[0001] The present invention specifically relates to a method for predicting relapse of CML patients after discontinuation of TKI treatment based on machine learning. Background Art
[0002] Chronic myeloid leukemia (CML) is a myeloproliferative disease, which is biologically manifested by the abnormal expression of BCR-ABL protein in hematopoietic stem cells (HSC), which is responsible for regulating the phosphorylation, activation and dysregulation of intracellular signal transduction proteins. Tyrosine kinase inhibitors (TKIs) targeting BCR-ABL fusion protein have revolutionized the prognosis of patients with CML, especially for patients in the chronic stage of the disease. In clinical practice, a considerable number of patients have a survival period close to that of normal people after TKI treatment. The pursuit of treatment-free remission (TFR) has now become a practical clinical goal. The high cost of drugs required for continuous treatment has brought an increasing economic burden to the medical and health system and individual patients. Some CML patients have a reduced quality of life due to adverse reactions to TKIs. Therefore, how to achieve TFR has become an important clinical research direction in CML treatment with health economic significance and the ability to increase patient benefits. At the same time, studies have shown that some patients stop taking drugs due to financial difficulties, resulting in a decrease in treatment compliance and failure to achieve ideal treatment response. Therefore, how to accurately distinguish patients who are suitable for drug discontinuation and patients who require long-term continuous treatment to achieve satisfactory therapeutic effects based on clinical and biological characteristics has become an urgent but unmet clinical need. Summary of the invention
[0003] The present invention provides a method for predicting relapse of CML patients after TKI treatment cessation based on machine learning to solve the above-mentioned technical problems, and specifically adopts the following technical solutions:
[0004] A method for predicting relapse of CML patients after discontinuation of TKI treatment based on machine learning, comprising the following steps:
[0005] Collect sample data of CML patients to construct a data set, wherein the sample data includes input features and output quantities, wherein the input features include physiological indicators and medication regimens of CML patients, and the output quantity is the MMR maintenance within n months after TKI discontinuation, where n is one of 3, 6, and 12;
[0006] Establish a prediction model for relapse after CML drug withdrawal;
[0007] Training the CML relapse prediction model after drug withdrawal using the data set;
[0008] The physiological indexes and medication regimen of the CML patient to be predicted are input into the trained CML relapse prediction model after drug withdrawal to output the prediction result.
[0009] Furthermore, after collecting sample data of CML patients to construct a data set, the input features in the data set are also preprocessed.
[0010] Furthermore, the discrete feature data is subjected to hot encoding, and the data with irregular content or format is treated as null value;
[0011] Perform outlier detection on continuous feature data and treat data beyond the medical range as null values;
[0012] For missing continuous feature data, the mean and variance are calculated, and during training, missing values are filled by sampling from a Gaussian distribution with the same mean and variance.
[0013] For missing discrete feature data, the mean and variance are calculated, and during training, missing values are filled by sampling from a Gaussian distribution with the same mean and variance.
[0014] Furthermore, after the data set is preprocessed, the input features in the data set are screened.
[0015] Furthermore, the specific method for screening the input features in the data set is:
[0016] For continuous characteristic data, one-way analysis of variance was used to analyze the correlation between the characteristic data and the maintenance of MMR within n months after TKI discontinuation, and the characteristic data that failed the significance test were deleted.
[0017] Furthermore, the characteristic data finally screened out included: gender, age at diagnosis, age at TKI discontinuation, duration of TKI treatment, DMR maintenance time before discontinuation, MMR duration before discontinuation, time to reach DMR, time to reach MMR, Sokal score, EUTOS long-term survival score, gene fluctuation within n months, degree of spleen enlargement, white blood cell count, hemoglobin level, proportion of primitive cells in peripheral blood, proportion of eosinophils in peripheral blood, proportion of basophils in peripheral blood, and platelet count in peripheral blood.
[0018] Furthermore, the CML relapse prediction model after drug withdrawal is an ensemble structure.
[0019] Furthermore, the CML relapse prediction model after drug withdrawal includes a Randomforest model and an XGboost model set.
[0020] Furthermore, during the training process of the CML relapse prediction model after drug withdrawal, each machine learning model of the CML relapse prediction model after drug withdrawal gives its own importance ranking result for the input feature data, and the feature importance order of all machine learning models is averaged to obtain the final feature data importance ranking.
[0021] Furthermore, the sample data of CML patients collected came from CML patients in multiple different centers.
[0022] The benefit of the present invention lies in the provided method for predicting relapse of CML patients after discontinuation of TKI treatment based on machine learning. By utilizing a machine learning algorithm, multiple feature data of multiple patients are integrated to train a prediction model that can be applied clinically. The prediction model can predict the probability of relapse of CML patients after discontinuation of TKI treatment, and assist doctors in selecting more appropriate treatment plans for patients, so that as many patients as possible can achieve TFR, while allowing high-risk patients to avoid the risk of relapse.
[0023] The present invention is also beneficial in that it provides a method for predicting relapse of CML patients after discontinuation of TKI treatment based on machine learning. The provided CML relapse prediction model after discontinuation of treatment combines the advantages of multiple models, reduces the degree of model overfitting, and improves prediction accuracy.
[0024] The present invention is also beneficial in that it provides a method for predicting relapse of CML patients after discontinuation of TKI treatment based on machine learning, which can also give an importance ranking of input features and provide suggestions on key focus objects when training the prediction model.
[0025] The present invention is also beneficial in that it provides a method for predicting relapse of CML patients after discontinuation of TKI treatment based on machine learning. When training the prediction model, it performs sampling to fill in the missing data, provides the first-order statistics of the features, and provides the second-order statistics of the features. The model processed by sampling improves the generalization accuracy of the model in the case of missing data. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0027] Figure 1 The present invention is a method for predicting relapse of CML patients after discontinuation of TKI treatment based on machine learning. DETAILED DESCRIPTION
[0028] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.
[0029] like Figure 1 The present invention shows a prediction method for relapse of CML patients after TKI treatment is discontinued based on machine learning, which includes the following steps: S1: Collect sample data of CML patients to construct a data set, the sample data includes input features and output quantities, the input features include physiological indicators and medication regimens of CML patients, and the output quantity is the MMR maintenance within n months after discontinuation of TKI. S2: Establish a prediction model for relapse after discontinuation of CML. S3: Train the prediction model for relapse after discontinuation of CML through the data set. S4: Input the physiological indicators and medication regimens of the CML patients to be predicted into the trained prediction model for relapse after discontinuation of CML and output the prediction results. The prediction method for relapse of CML patients after TKI treatment discontinuation based on machine learning of the present invention uses a machine learning algorithm to integrate multiple feature data of multiple patients to train a prediction model that can be applied to clinical practice, which can predict the probability of relapse of CML patients after TKI treatment is discontinued, and assist doctors in selecting more appropriate treatment plans for patients, so that as many patients as possible can achieve TFR, while allowing high-risk patients to avoid the risk of relapse. The above steps are described in detail below.
[0030] For step S1: sample data of CML patients are collected to construct a data set. The sample data includes input features and output quantities. The input features include physiological indicators and medication regimens of CML patients. The output quantity is the maintenance of MMR (Major Molecular Response) within n months after discontinuation of TKI.
[0031] Specifically, n is one of 3, 6, and 12. In the present application, n is 12.
[0032] In the embodiments of the present application, the sample data of CML patients collected are from CML patients in 10 different centers, and samples from different regions are covered as much as possible to ensure clinical diversity, so that the constructed model can have a high consistency in performance on different patient groups. Specifically, the mark of maintaining MMR within 12 months after stopping TKI is 0, and the mark of losing MMR within 12 months is 1.
[0033] In an embodiment of the present application, after collecting sample data of CML patients to construct a data set, the input features in the data set are also preprocessed.
[0034] Specifically, discrete feature data are hot-coded, and data with irregular content or format are treated as null values. Outlier detection is performed on continuous feature data, and data beyond the medical range is treated as null values. For missing continuous feature data, the mean and variance are calculated, and during training, missing values are filled by sampling from a Gaussian distribution with the same mean and variance.
[0035] For missing discrete feature data, the mean and variance are calculated, and during training, missing values are filled by sampling from a Gaussian distribution with the same mean and variance.
[0036] In an embodiment of the present application, after the data set is preprocessed, the input features in the data set are also screened.
[0037] The specific method for screening the input features in the data set is as follows: for continuous feature data, one-way analysis of variance was used to analyze the correlation between the feature data and the maintenance of MMR within 12 months after discontinuation of TKI, and the feature data that failed the significance test was deleted.
[0038] In the embodiments of the present application, the characteristic data finally screened out include gender, age at diagnosis, age at TKI discontinuation, duration of TKI treatment, maintenance time of deep molecular response (DMR) before discontinuation, duration of MMR before discontinuation, time to reach DMR, time to reach MMR, Sokal score, EUTOS long-term survival score, gene fluctuation within n months (gene fluctuation within 12 months in the present application), degree of spleen enlargement, white blood cell count, hemoglobin level, proportion of primitive cells in peripheral blood, proportion of eosinophils in peripheral blood, proportion of basophils in peripheral blood and platelet count in peripheral blood.
[0039] For step S2: establish a prediction model for CML relapse after drug withdrawal.
[0040] In the embodiments of the present application, the CML relapse prediction model after drug withdrawal is an ensemble structure, which combines the advantages of multiple models to reduce the overfitting risk of the overall model and improve the prediction accuracy. Specifically, the CML relapse prediction model after drug withdrawal includes a Randomforest model and an XGboost model set.
[0041] For step S3: the CML relapse prediction model after drug discontinuation is trained using the dataset.
[0042] The data of CML patients from 10 different centers were divided into training sets and validation sets to train the CML relapse prediction model after drug withdrawal. In particular, the grid search optimization method was used to set the hyperparameters of each model. After training, 10 CML relapse prediction models after drug withdrawal were constructed, including 5 XGboost models and 5 randomforest models. The accuracy of the trained CML relapse prediction model after drug withdrawal is higher than the predictive factors given in the existing guidelines, and to a certain extent, it can assist doctors in judging the probability of molecular relapse after drug withdrawal.
[0043] In an embodiment of the present application, during the training of the CML relapse prediction model after drug withdrawal, each machine learning model of the CML relapse prediction model gives its own importance ranking result for the input feature data, and the feature importance order of all machine learning models is averaged to obtain the final feature data importance ranking.
[0044] It is understandable that during the training of the CML relapse prediction model after drug withdrawal, each model in the integrated model can rank the importance of features by calculating information entropy. The ranking results of all models are averaged to calculate the final importance ranking result of each feature.
[0045] Specifically, the feature importance ranking results obtained by the CML relapse prediction model after drug discontinuation are as follows: gene fluctuation within n months, platelet count in peripheral blood, DMR maintenance time before drug discontinuation, white blood cell count in peripheral blood, Sokal score, time to reach MMR, age at TKI discontinuation, time to reach DMR, duration of MMR before drug discontinuation, degree of spleen enlargement, proportion of eosinophils in peripheral blood, duration of TKI treatment, proportion of primitive cells in peripheral blood, gender, EUTOS long-term survival score, hemoglobin level in peripheral blood and proportion of basophils in peripheral blood.
[0046] The feature ranking results can suggest doctors to pay more attention to the indicators with higher rankings, thereby assisting doctors in designing targeted treatment plans for patients.
[0047] For step S4: input the physiological indexes and medication regimen of the CML patient to be predicted into the trained CML relapse prediction model after drug withdrawal to output the prediction result.
[0048] The physiological indicators and medication regimens of the CML patients to be predicted are obtained and used as input features to enter the trained CML relapse prediction model after drug withdrawal. The model outputs the probability of molecular relapse after drug withdrawal in CML patients. Specifically, the data of the CML patients to be predicted are calculated 10 times, and 5 Random Forest models and 5 XGboost models are run each time to output 10 values. The average of the corresponding category probability values output by the 10 models is used as the final probability value.
[0049] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the present invention in any form, and any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.
Claims
1. A method for predicting relapse of CML patients after discontinuation of TKI treatment based on machine learning, characterized in that: The following steps are involved: Collect sample data of CML patients to construct a data set, wherein the sample data includes input features and output quantities, wherein the input features include physiological indicators and medication regimens of CML patients, and the output quantity is the MMR maintenance within n months after TKI discontinuation, where n is one of 3, 6, and 12; Establish a prediction model for relapse after CML drug withdrawal; Training the CML relapse prediction model after drug withdrawal using the data set; The physiological indexes and medication regimen of the CML patient to be predicted are input into the trained CML relapse prediction model after drug withdrawal to output the prediction result.
2. The method for predicting relapse of CML patients after discontinuation of TKI treatment based on machine learning according to claim 1, characterized in that: After collecting sample data of CML patients to construct a data set, the input features in the data set are also preprocessed.
3. The method for predicting relapse of CML patients after discontinuation of TKI treatment based on machine learning according to claim 2, characterized in that: Perform hot encoding on discrete feature data and treat data with null values if the content or format is not standardized; Perform outlier detection on continuous feature data and treat data beyond the medical range as null values; For missing continuous feature data, the mean and variance are calculated, and during training, missing values are filled by sampling from a Gaussian distribution with the same mean and variance. For missing discrete feature data, the mean and variance are calculated, and during training, missing values are filled by sampling from a Gaussian distribution with the same mean and variance.
4. The method for predicting relapse of CML patients after discontinuation of TKI treatment based on machine learning according to claim 2, characterized in that: After the data set is preprocessed, the input features in the data set are also screened.
5. The method for predicting relapse of CML patients after discontinuation of TKI treatment based on machine learning according to claim 4, characterized in that: The specific method for screening the input features in the data set is: For continuous characteristic data, one-way analysis of variance was used to analyze the correlation between the characteristic data and the maintenance of MMR within n months after TKI discontinuation, and the characteristic data that failed the significance test were deleted.
6. The method for predicting relapse of CML patients after discontinuation of TKI treatment based on machine learning according to claim 5, characterized in that: The characteristic data finally screened out included: gender, age at diagnosis, age at TKI discontinuation, duration of TKI treatment, DMR maintenance time before discontinuation, MMR duration before discontinuation, time to reach DMR, time to reach MMR, Sokal score, EUTOS long-term survival score, gene fluctuation within n months, degree of spleen enlargement, white blood cell count, hemoglobin level, proportion of primitive cells in peripheral blood, proportion of eosinophils in peripheral blood, proportion of basophils in peripheral blood, and platelet count in peripheral blood.
7. The method for predicting relapse of CML patients after discontinuation of TKI treatment based on machine learning according to claim 1, characterized in that: The CML relapse prediction model after drug withdrawal is an ensemble structure.
8. The method for predicting relapse of CML patients after discontinuation of TKI treatment based on machine learning according to claim 7, characterized in that: The CML relapse prediction model after drug withdrawal includes a Randomforest model and an XGboost model set.
9. The method for predicting relapse of CML patients after discontinuation of TKI treatment based on machine learning according to claim 8, characterized in that: During the training process of the CML relapse prediction model after drug withdrawal, each machine learning model of the CML relapse prediction model after drug withdrawal gives its own importance ranking result for the input feature data, and the feature importance order of all machine learning models is averaged to obtain the final feature data importance ranking.
10. The method for predicting relapse of CML patients after discontinuation of TKI treatment based on machine learning according to claim 1, characterized in that: The sample data collected from CML patients came from CML patients in multiple different centers.