Construction method and construction system of chronic lymphocytic leukemia prediction model based on machine learning, electronic equipment and storage medium

Abstract

The invention provides a construction method and a construction system of a chronic lymphocytic leukemia prediction model based on machine learning, electronic equipment and a storage medium, and relates to the field of chronic lymphocytic leukemia prognosis research. The construction method comprises the steps of obtaining original sample data, and performing preliminary screening; performing data cleaning on the screened sample data; determining a prediction factor from the plurality of features of the cleaned sample data; dividing the cleaned sample data corresponding to the prediction factor and the target variable into a training set and a test set; inputting the training set after unbalance processing into a LightGBM model for training to obtain a prediction model; inputting the test set into a prediction model, performing hyper-parameter optimization on the prediction model, and evaluating the performance of the prediction model; calibrating the prediction model to obtain a calibrated prediction model; the method has the beneficial effects that the method can be realized only by depending on conventionally available predictive factors, and the patient screening can be realized before the CLL clinical symptoms appear.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of chronic lymphocytic leukemia prognosis research, and in particular to a method and system for constructing a machine learning-based chronic lymphocytic leukemia prediction model, an electronic device, and a storage medium. BACKGROUND

[0002] Chronic lymphocytic leukemia (CLL) is the most common type of adult leukemia in Western countries, and its incidence increases with age, with an incidence of about 4.9 cases per 100,000 people per year. The median age of patients at diagnosis is about 70 years old. In the UK, more than 3800 new cases are diagnosed each year. This reflects the growing burden of CLL on public health, especially in the context of an aging population.

[0003] The clinical course of CLL exhibits diverse characteristics: abnormal lymphocytes begin to generate in the bone marrow, and due to the slow proliferation of leukemia cells, some patients may not develop symptoms for years or even longer. Others may experience rapid deterioration and be at risk of serious complications and death. Despite advances in targeted therapies and molecular diagnostic techniques, CLL remains incurable at present. These abnormal cells eventually spread to other parts of the body, including lymph nodes, liver, and spleen. The disease usually progresses slowly and may not require treatment or only conservative therapy.

[0004] Current research has found that there are many epidemiological and genetic risk factors for CLL. These factors include unchangeable risk factors such as older age, male gender, Caucasian race, and family history of hematological malignancies, as well as potential environmental and occupational exposure risks. Existing prognostic models are only suitable for post-diagnosis evaluation and have limited practicality in early screening or primary prevention. Therefore, there is still a key gap in clinical practice - how to identify individuals at high risk before the onset of clinical symptoms.

[0005] Machine learning (ML) algorithms can handle complex high-dimensional data and identify complex non-linear relationships that traditional analysis methods cannot capture. In recent years, machine learning has been mainly applied in oncology to predict prognostic indicators such as patient treatment response, disease progression, or survival. However, the use of machine learning to predict new-onset CLL in healthy populations is still a blank area. With the opening of large-scale prospective datasets such as the UK Biobank (UKB), we can develop robust individualized risk prediction models and provide a data-driven new approach to disease risk modeling. SUMMARY

[0006] To solve one of the above technical defects, the present application provides a method and system for constructing a machine learning-based chronic lymphocytic leukemia prediction model, an electronic device, and a storage medium.

[0007] According to a first aspect of the present application, a method for constructing a machine learning-based chronic lymphocytic leukemia prediction model is provided, comprising: obtaining original sample data in a biological sample library, and performing preliminary screening on the original sample data to obtain screened sample data; the original sample data includes multiple participants and multiple characteristics; performing data cleaning on the screened sample data to obtain cleaned sample data; determining a prediction factor from the multiple characteristics of the cleaned sample data; dividing the cleaned sample data corresponding to the prediction factor and the target variable into a training set and a test set, and performing unbalanced processing on the training set; inputting the unbalanced processed training set into a LightGBM model for training to obtain a prediction model; inputting the test set into the prediction model, optimizing the hyperparameters of the prediction model, and then evaluating the performance of the prediction model; calibrating the prediction model to obtain a calibrated prediction model.

[0008] Preferably, the prediction factor includes lymphocyte count, gender, sleep duration, smoking status, previous chemotherapy, hepatosplenomegaly, blood biochemical abnormalities, muscle and bone disease history, and secondary malignant tumor.

[0009] Preferably, the data cleaning on the screened sample data to obtain the cleaned sample data specifically includes: eliminating characteristics of participants with a data missing rate exceeding 50%; filling in mean values for numerical value type characteristics; filling in modes for classification type characteristics; splitting the answers to multiple-choice type characteristics in the questionnaire into independent binary variables; processing classification variables in non-numerical type characteristics using one-hot encoding.

[0010] More preferably, the determination of the prediction factor from the multiple characteristics of the cleaned sample data sequentially performs the following steps: training the multiple characteristics of the cleaned sample data using a Light Gradient Boosting Machine, and sorting the feature importance by gain value, and retaining the top 200 features; identifying and eliminating highly correlated features with r>0.9 in the 200 features using Pearson correlation coefficient; reducing the features to 50 using an iterative feature selector; retaining the top 20 features from the 50 features through SHAP analysis; Receiving selection instructions of experts, determining 9 predictors from 20 features.

[0011] Preferably, the performance of the evaluation prediction model specifically includes: evaluation by five-fold stratified cross-validation combined with gain importance analysis, SHAP value.

[0012] Preferably, the screening of the original sample data specifically includes: Eliminating participants and their corresponding feature data that meet the following conditions: withdrawal from the study, missing follow-up information, non-white ancestry, reporting gender difference with gene detection gender, lack of exome sequencing data, diagnosed with hematological malignancies, genetic association with other participants; Eliminating features that meet the following conditions: imaging data, online follow-up survey, algorithm-defined prognostic indicators, associated data other than diagnostic codes in cancer / death registration systems or general practitioners' cases, genomic comprehensive indicators, instrument processing or quality control indicators, and coded records recorded after the diagnosis of chronic lymphocytic leukemia.

[0013] Preferably, the unbalanced processing of the training set specifically includes: Increasing the number of positive samples in the training set to 2% of the number of negative samples; Then using a random undersampling method to reduce the number of negative samples by 50%.

[0014] According to a second aspect of the present application, a machine learning-based chronic lymphocytic leukemia prediction model construction system is provided, which includes a module for implementing the machine learning-based chronic lymphocytic leukemia prediction model construction method according to any one of the above.

[0015] According to a third aspect of the present application, an electronic device is provided, which includes: a memory; a processor; and a computer program; wherein the computer program is stored in the memory and configured to be executed by the processor to implement the machine learning-based chronic lymphocytic leukemia prediction model construction method according to any one of the above.

[0016] According to a fourth aspect of the present application, a computer readable storage medium is provided, which stores a computer program; the computer program is executed by a processor to implement the machine learning-based chronic lymphocytic leukemia prediction model construction method according to any one of the above.

[0017] The beneficial effects of the present application are: The prediction model of chronic lymphocytic leukemia based on machine learning is constructed in the application, the prediction model can more accurately predict the risk of CLL, and has excellent prediction performance and good calibration effect in multiple time dimensions. And has strong discrimination and reliable calibration, only needs to rely on 9 conventional available prediction factors. Can realize patient screening before the early stage of indolent CLL clinical symptoms appear, open up a new way for early intervention, and provide key basis for clinical diagnosis and treatment.

[0018] From the clinical point of view, patients marked as high risk by the prediction model can receive more frequent monitoring, early hematological evaluation or participate in preventive clinical trials. From the perspective of public health, the prediction model supports stratified prevention and control strategies, helps to shorten the diagnostic delay, reduce medical costs and alleviate the disease burden. With the popularization of gene typing technology and the increasing integration of data, integrating all this information into the electronic health record system can realize real-time risk warning and accurate monitoring, provide timely support for clinical decision-making, and develop personalized risk management programs. Lay the foundation for risk stratification at the group level, and provide guidance for early screening, targeted monitoring and prevention strategies in the future.

[0019] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent from the description, or can be learned by practice of the present application. The objects and other advantages of the present application can be achieved and obtained by what is written in the specification and the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0020] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation of the present application. In the drawings: Figure 1 A flowchart of a method for constructing a prediction model of chronic lymphocytic leukemia based on machine learning according to an embodiment of the present application is shown in FIG. 1; Figure 2 A flowchart of a method for constructing a prediction model of chronic lymphocytic leukemia based on machine learning according to an embodiment of the present application is shown in FIG. 1; Figure 1 A flowchart of a method for constructing a prediction model of chronic lymphocytic leukemia based on machine learning according to an embodiment of the present application is shown in FIG. 1; Figure 3 A model fitting effect diagram of a multi-gene risk score is shown in FIG. 4; Figure 4 A multi-gene risk score distribution diagram of positive and negative samples is shown in FIG. 5; Figure 5 An HLA evolutionary divergence analysis diagram of positive and negative samples is shown in FIG. 6; Figure 6 A flowchart of a method for constructing a prediction model of chronic lymphocytic leukemia based on machine learning according to an embodiment of the present application is shown in FIG. 1; Figure 1 A flowchart of a method for constructing a prediction model of chronic lymphocytic leukemia based on machine learning according to an embodiment of the present application is shown in FIG. 1; Figure 7For Figure 1 Flowchart of determining the predictor from the plurality of features of the cleaned sample data; Figure 8 Schematic diagram of feature importance of light gradient boosting machine; Figure 9 Feature selection diagram of iterative feature selector; Figure 10 Feature selection diagram of SHAP analysis; Figure 11 Overall contribution diagram of the predictor; Figure 12 ROC curve diagram of the embodiments, comparative example 1 to comparative example 3 provided in the present application; Figure 13 ROC curve diagram of the embodiments, comparative example 4 to comparative example 6 provided in the present application; Figure 14 ROC curve diagram of the embodiments provided in the present application on the 5-year, 10-year and full disease time axis in the training set; Figure 15 ROC curve diagram of the embodiments provided in the present application on the 5-year, 10-year and full disease time axis in the test set; Figure 16 Agreement analysis diagram of the embodiments provided in the present application on the 5-year, 10-year and full disease time axis; Figure 17 Structural schematic diagram of a machine learning-based chronic lymphocytic leukemia prediction model construction system provided in the embodiments of the present application; Figure 18 For Figure 17 Structural schematic diagram of the screening module; Figure 19 For Figure 17 Structural schematic diagram of the data cleaning module; Figure 20 For Figure 17 Structural schematic diagram of the predictor determination module; Figure 21 For Figure 17 Structural schematic diagram of the imbalance processing module; In the figure: 1 is a 5-year prediction curve, 2 is a 10-year prediction curve, and 3 is a full disease prediction curve; 10 is the acquisition module, 20 is the screening module, 30 is the data cleaning module, 40 is the predictor determination module, 50 is the partitioning module, 60 is the imbalance handling module, 70 is the predictive model training module, 80 is the optimization evaluation module, 90 is the calibration module, 201 is the first elimination unit, 202 is the second elimination unit, 301 is the missing data handling unit, 302 is the imputation unit, 303 is the splitting unit, 304 is the categorical variable handling unit, 401 is the sorting unit, 402 is the relevant feature elimination unit, 403 is the reduction unit, 404 is the analysis unit, 405 is the receiving unit, 406 is the determination unit, 601 is the positive sample enhancement unit, and 602 is the negative sample reduction unit. Detailed Implementation

[0021] To make the technical solutions and advantages of the embodiments of this application clearer, the exemplary embodiments of this application will be described in further detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0022] This application used data from the UK Biobank (UKB), hereinafter referred to as the Biobank, which is a population-based prospective cohort study. This application used data under UKKB-approved application number [Application ID: 95872]. Written informed consent was provided by all participants upon enrollment.

[0023] This application obtained data from 502,164 participants and their corresponding data (including multi-source and multi-dimensional data such as touch screen questionnaires, interview records, physical examination records, and biological tests) from the biobank as the original sample data. Among them, 53% were female; the age range was 39-72 years old, and the average age was 57 years old.

[0024] like Figure 1 As shown, to address the above problems, this application provides a method for constructing a machine learning-based predictive model for chronic lymphocytic leukemia, including: The raw sample data from the biobank was obtained and initially screened to obtain the screened sample data; the raw sample data included multiple participants and multiple characteristics. The filtered sample data is cleaned to obtain the cleaned sample data. Predictors are determined from multiple features of the cleaned sample data; The cleaned sample data corresponding to the predictor and the target variable (positive or negative samples) are divided into training and test sets with a ratio of 9:1. The training set is then imbalanced. The imbalance process includes: resampling the training set using the Synthetic Minority Oversampling (SMOTE) technique to increase the number of positive samples in the training set to 2% of the number of negative samples; and then reducing the number of negative samples by 50% using a random undersampling method.

[0025] The imbalanced training set is input into the gradient boosting decision tree model (LightGBM model) for training to obtain the prediction model; The test set is input into the prediction model, the hyperparameters of the prediction model are optimized, and then the performance of the prediction model is evaluated. The prediction model is calibrated to obtain the calibrated prediction model.

[0026] This application presents a machine learning-based predictive model for chronic lymphocytic leukemia (CLL). This model more accurately predicts the risk of CLL, demonstrating superior predictive performance and good calibration across multiple time dimensions. It possesses strong discriminative power and reliable calibration, requiring only nine commonly available predictive factors. This model enables patient screening before clinical symptoms appear in the early stages of indolent CLL, opening new avenues for early intervention and providing crucial evidence for clinical diagnosis and treatment.

[0027] From a clinical perspective, patients identified as high-risk can receive more frequent monitoring, earlier hematological assessments, or participate in preventative clinical trials. From a public health perspective, this predictive model supports a stratified prevention and control strategy, helping to reduce diagnostic delays, lower healthcare costs, and lessen the disease burden. With the increasing prevalence of genotyping technology and data integration, integrating this information into electronic health record systems enables precise monitoring and real-time risk warnings, providing timely support for clinical decision-making and the development of personalized risk management plans. This lays the foundation for population-level risk stratification and can guide future early screening, targeted surveillance, and prevention strategies.

[0028] like Figure 2 As shown, further, the filtering of the original sample data specifically includes: Participants meeting the following criteria and their corresponding characteristic data were excluded: mid-study withdrawal, missing follow-up information, non-Caucasian ancestry, discrepancy between reported gender and genetically determined gender, lack of exome sequencing data, prior diagnosis of hematologic malignancies, and genetic association with other participants. After excluding participants meeting the above criteria and their corresponding characteristic data, a total of 264,224 participants were identified. Among them, 579 were CLL-positive patients, identified through ICD-10 code C91.1, ICD-9 code 204.1, cancer registry data, self-reported disease and death registry records (hereinafter referred to as positive samples). The remaining 263,645 participants were classified as CLL-negative patients (hereinafter referred to as negative samples). Features that meet the following criteria will be excluded: imaging data, online follow-up questionnaires, algorithm-defined prognostic indicators, correlation data other than diagnostic codes in cancer / death registry systems or general practitioner cases, genomic comprehensive indicators, instrument handling or quality control indicators, and codes recorded after diagnosis of chronic lymphocytic leukemia.

[0029] In this application, the original sample data was screened, including screening of participants and features, making the filtered sample data cleaner and removing data that might affect the results or is useless to the model, thus facilitating subsequent analysis and processing. Specifically, the codes recorded after the diagnosis of chronic lymphocytic leukemia were removed to avoid reverse causality.

[0030] In this application, feature extraction was also performed on the screened sample data, resulting in the extraction of a first candidate feature and a second candidate feature, specifically: The selected sample data has multiple features, including genotyping features and exome sequencing features. The first candidate feature is obtained by performing multi-gene risk scoring on the genotyping feature data in the screened sample data; HLA evolutionary divergence analysis was performed on the exome sequencing feature data of the screened sample data to obtain the second candidate feature.

[0031] In this application, the first and second candidate features were customized, and complex genotyping and exome sequencing features were analyzed and visually presented using the first and second candidate features. However, as can be seen from the subsequent determination of predictive factors, the first and second candidate features did not appear as independent risk factors and were therefore not included in the predictive factors. Although the first candidate feature reflects lifelong genetic susceptibility, the development of chronic lymphocytic leukemia (CLL) may also be highly dependent on acquired somatic mutations, immune dysregulation, environmental exposure, and random events, factors that were not reflected in the first and second candidate features.

[0032] This application uses pooled data from the FinnGen 11th Genome-wide Association Study of Chronic Lymphoblastic Leukemia (CLL GWAS), hereinafter referred to as the pooled data of the Genome-wide Association Study of Chronic Lymphoblastic Leukemia, which includes 6,000 patients and 180,000 controls.

[0033] Furthermore, multi-gene risk scoring is performed on the genotyping characteristic data in the screened sample data, specifically including: Acquire the aggregated data from the genome-wide association study of chronic lymphocytic leukemia and input the aggregated data into the constructed multigene risk scoring model; The multi-gene risk scoring model was optimized, specifically including: Quality control of the aggregated data was performed using a multigene risk scoring model, specifically including the following criteria for exclusion: minor allele frequency (MAF) < 0.01, minor allele count (MAC) < 100, genotype loss rate (geno) > 0.1, Hardy-Weinberg equilibrium p-value (<1×10) > 0.1; individuals with heterozygosity or outliers, sex chromosome aneuploidy, or ancestral outliers as shown by principal component analysis were excluded. -15 Sample-level missing rate (mind) The remaining genotyping data were further screened using linkage disequilibrium (r < 0.1, 250 kb window); 2 The optimized multi-gene risk scoring model was obtained; Genotyping feature data from the screened sample data were input into the optimized polygenic risk scoring model. The PRSice-2 software was used to calculate the polygenic risk score by taking p-values ​​at different thresholds. The optimal threshold PT was selected, with PT = 5 × 10⁻⁶. -8 Population stratification effects were corrected using covariates (age, sex, and the top ten principal genetic components). The multi-gene risk score corresponding to the optimal threshold of p-value is used as the first candidate feature.

[0034] like Figure 3 As shown in (a), the optimal threshold PT = 5 × 10 -8 This explains approximately 1% of the variance (R, p = 4.1 × 10⁻⁶). -23 ).like Figure 3 As shown in (b), the high-resolution curves confirm that PT = 5 × 10 is the optimal threshold. At this optimal threshold, the final polygenic risk scoring model includes 16 single nucleotide polymorphisms (SNPs). The polygenic risk scoring model, plus covariates, explains approximately 4% of the variation (R), of which covariates explain approximately 3%. 2= 0.01 -82 = 0.04 Table 1: Polygenic Risk Scores at Optimal Thresholds

[0035] like Figure 4 As shown, the polygenic risk scores of positive samples (Cases) were generally slightly higher than those of negative samples (Controls), and this trend continued. Figure 4 (c), (d), and (e) are all consistent.

[0036] This application incorporates a polygenic risk score, a quantitative indicator that assesses an individual's genetic susceptibility to disease by integrating the effects of multiple genetic variations across the entire genome.

[0037] Furthermore, HLA evolutionary divergence analysis was performed on the exome sequencing characteristics of the screened sample data, specifically including: Obtain HLA reference databases, including the IMGT / HLA database and the Ensembl database; HLA typing was performed on the exome sequencing data of the screened samples based on the HLA reference database to obtain two alleles for each participant at each HLA gene locus; the allele sequences and exons could be obtained from the IMGT / HLA database and the Ensembl database; HLA gene loci were divided into class I HLA gene loci (HLA-A, -B, -C) and class II HLA gene loci (HLA-DPB1, -DQB1, -DRB1). The second and third exons of the two alleles at each class I HLA gene locus were compared codon-by-codon; the second exon set of the two alleles at each class II HLA gene locus was compared codon-by-codon. The Grantham distance metric was used to calculate the Grantham distance between two alleles at HLA gene loci based on the HLA reference database, and the distance values ​​of each HLA gene locus were obtained; the amino acid differences between alleles were quantified. The Type I and Type II evolutionary divergence scores for each participant were calculated. Specifically, this involved summing and standardizing the distance values ​​of the Type I and Type II HLA loci for each participant to obtain the Type I and Type II HLA loci evolutionary divergence scores. The average of the Type I and Type II HLA loci evolutionary divergence scores for each participant was then calculated to obtain the participant's Type I and Type II evolutionary divergence scores. (When two alleles of an HLA locus are completely identical, the HLA locus is considered homozygous, and its evolutionary divergence score is assigned as 0.) Both Type I and Type II evolutionary differentiation values ​​with a calculated confidence score ≥ 0.5 were considered as second candidate features.

[0038] like Figure 5 As shown in (f) and (h), the distribution of HLA evolutionary differences in type I is as follows: there is no significant difference between positive and negative samples (the median and interquartile range are almost identical; t-test p = 0.3650).

[0039] like Figure 5 As shown in (g) and (i), the distribution characteristics of HLA evolutionary differences in type II are shown, with the median of positive samples being significantly higher than that of negative samples (t-test p = 0.0451).

[0040] Figure 5 (f) and (g) are histograms, which can intuitively show the distribution pattern of evolutionary differentiation values. Figure 5 (h) and (i) are box plots that can show in detail the median, interquartile range and outlier distribution of each group.

[0041] This application includes an HLA evolutionary divergence analysis and an analysis of immunogenetic characteristics.

[0042] like Figure 6 As shown, further, the step of cleaning the screened sample data to obtain cleaned sample data specifically includes: Participants with a data missing rate exceeding 50% were excluded; responses related to "unwilling to answer" or "don't know" in the sample data were marked as missing data. The mean is used to impute numerical features; The mode is used to impute categorical features; The responses to the multiple-choice feature in the questionnaire were split into independent binary variables; One-hot encoding is used to process categorical variables in non-numerical features (for example, when the non-numerical feature is gender, the categorical variables include "female" and "male").

[0043] In this application, the sample data corresponding to the candidate feature set was cleaned. The cleaned sample data includes 2932 features. The cleaned sample data is high-quality data suitable for analysis, which improves data credibility and work efficiency, and avoids analysis bias caused by problems such as large data missing rate.

[0044] like Figure 7 As shown, further, the step of determining the predictor from multiple features of the cleaned sample data involves sequentially performing the following steps: A light gradient boosting machine was used to train multiple features on the cleaned sample data, and the importance of the features was ranked by the gain value, retaining the top 200 features. Pearson correlation coefficient was used to identify and remove highly correlated features with r > 0.9 from 200 features; An iterative feature selector (SFS) is used to continuously add features to maximize the cross-validation AUROC metric, reducing the number of features to 50. The first 20 features out of 50 were retained through SHAP analysis; Receive selection instructions from experts and determine 9 predictors from 20 features.

[0045] This application employs tools such as Light Gradient Boosting Machine, Pearson correlation coefficient, Iterative Feature Selector, and SHAP analysis for feature ranking and selection. Combined with expert selection instructions, nine predictive factors were identified. These factors were used to train the predictive model, resulting in more accurate predictions. This approach enables CLL risk stratification before clinical diagnosis using a minimal number of readily available predictive factors. The nine predictive factors provide the best predictive value for the model, and all are information readily available or obtainable in routine outpatient settings.

[0046] like Figure 8 As shown, feature importance was ranked by gain value, revealing the relative contribution of features during tree splitting. Figure 9 As shown, this illustrates the incremental contribution of each selected feature to the model performance in successive forward selection iterations. Figure 10 As shown, the SHAP value illustrates the impact of each feature on the output of the prediction model. Figure 11 As shown, the average SHAP values ​​are ranked according to the overall contribution of features to the model's predictions. A multi-stage strategy ensures that the most predictive and interpretable predictors are selected.

[0047] The nine predictors are: Lymphocyte count: An elevated lymphocyte count is the leading predictor of CLL development. CLL is essentially a lymphocyte disorder. Patients with monoclonal B-cell lymphoblastic disease (MBL) (a subclinical clonal B-cell proliferation often accompanied by mild lymphocytosis) have a significantly increased risk of developing overt CLL.

[0048] Gender: The higher incidence of CLL in men is consistent with the findings of multiple epidemiological studies. Data show that there are an average of 1.5 male patients and only 1 female patient (male-to-female ratio of 1.5:1). Possible pathogenic factors include hormone-mediated immune regulation, differences in occupational / environmental exposure, and genetic susceptibility loci enriched in men.

[0049] Sleep duration: Short or irregular sleep patterns are the third most important factor. Studies have shown that circadian rhythm disruption and chronic sleep deprivation can trigger immune dysregulation, systemic inflammatory responses, and impaired DNA repair function. These factors have all been shown to increase the risk of hematologic malignancies.

[0050] Smoking: Smoking is clearly closely linked to the development of lung cancer. It also significantly increases the risk of other malignant tumors, including oral cancer, throat cancer, esophageal cancer, stomach cancer, bladder cancer, kidney cancer, liver cancer, pancreatic cancer, cervical cancer, colon cancer, rectal cancer, and hematologic malignancies (such as acute myeloid leukemia (AML)). Tobacco smoke contains thousands of chemicals, at least 70 of which have been identified as carcinogens. These substances directly damage DNA and increase the risk of cancer.

[0051] Previous chemotherapy: Chemotherapy kills malignant cells in various ways, but it can also induce gene mutations in normal healthy cells. CLL survivors have a significantly increased risk of developing a second primary malignancy (SPM). Possible mechanisms include: underlying genetic susceptibility to malignancy; chemotherapy-induced immune dysregulation and changes in the bone marrow microenvironment may promote B-cell clonal expansion; enhanced medical monitoring of cancer survivors may lead to early detection of asymptomatic CLL; studies have also shown that although cytotoxic drugs are more strongly associated with myeloid tumors, they may also cause genomic instability in lymphoprogenitor cells.

[0052] Hepatomegaly and splenomegaly: Enlargement of the liver and / or spleen, while not specific, may reflect early lymphoid tissue infiltration or immune activation. These organs are often involved in CLL and related lymphoproliferative disorders, even before a clear hematological diagnosis.

[0053] Secondary malignancies (respiratory / digestive organs): These reflect shared biological predispositions, such as genomic instability, abnormal immune regulation, or clonal hematopoiesis, factors that may collectively contribute to the development of solid tumors and lymphatic system malignancies. Furthermore, previously received cancer treatments (including chemotherapy or radiotherapy) may create favorable conditions for B-cell clonal proliferation by altering the characteristics of bone marrow progenitor cells. This finding is also consistent with the aforementioned phenomenon of "previous chemotherapy."

[0054] Blood biochemistry abnormalities: "Abnormal blood chemistry markers" is a broad label defined by the International Classification of Diseases (ICD) to record non-specific laboratory abnormalities observed in routine medical care (such as claims / electronic health record systems), rather than research-level biomarkers. This differs from the lymphocyte count mentioned earlier, which is a continuous value obtained through analysis at baseline in the UK Biobank. In practice, when blood biochemistry tests (such as coagulation parameters, blood lipids, and inflammatory markers) show clinically significant but unexplained abnormalities, clinicians or coders will label this code, which is then recorded by the attending physician (usually as an auxiliary diagnosis to support subsequent testing or further investigation). Once a specific cause is identified (such as hepatitis, kidney failure, or specific electrolyte disturbances), a more specific diagnostic code is usually used instead of or accompanying this non-specific category. Therefore, any interpretation related to "abnormal blood chemistry markers" should be considered a marker of early or concurrent disease activity, potentially indicating reverse causality or testing bias. It may reflect an underlying disease burden that other covariates have not fully captured. Non-specific abnormalities in routine biochemical tests may reflect subclinical hematopoietic dysfunction, early metabolic or inflammatory changes, or unrecognized disease activity related to the pathogenesis of CLL.

[0055] History of musculoskeletal disease: This is based on ICD coding of routine clinical records (claims / electronic medical record systems) and reflects diagnoses classified as musculoskeletal / connective tissue diseases. Many musculoskeletal diseases coded in medical records are essentially autoimmune diseases but are categorized under "musculoskeletal / connective tissue" (e.g., rheumatoid arthritis, polymyalgia, lupus), so the observed risk may primarily reflect its autoimmune component. A history of musculoskeletal or autoimmune connective tissue disease may indicate persistent immune stimulation and altered inflammatory microenvironment that can impair immune surveillance and create favorable conditions for the formation of malignant lymphocyte clones.

[0056] Table 2: Predictors and their corresponding data

[0057] In this application, logistic regression was used to calculate the 95% confidence intervals and p-values ​​of the selected predictive factors in the prediction model. As shown in Table 2, the 95% confidence intervals and p-values ​​demonstrate that the predictive factors provided in this application are of practical significance and can accurately predict CLL patients.

[0058] To demonstrate the accuracy of the prediction model obtained using the nine predictive factors provided in this application, comparative examples one through three are also provided in this application.

[0059] Comparative Example 1: Training the LightGBM model using cleaned sample data (multiple features); Comparative Example 2: The LightGBM model was trained using sample data corresponding to the first 50 features; Comparative Example 3: The LightGBM model was trained using sample data corresponding to the first 20 features.

[0060] Precision, recall, accuracy, and F1 score were calculated based on regression prediction probabilities for the embodiments and comparative examples 1 to 3 provided in this application. The AUROC values ​​in the operating characteristic (ROC) curves of the LightGBM models trained in the embodiments and comparative examples 1 to 3 provided in this application were analyzed, and the optimal cutoff value was determined using the maximum Oden index (sensitivity + specificity - 1). The calculation results are shown in Table 3 below.

[0061] Table 3: Calculation results of the examples and comparative examples 1 to 3

[0062] As shown in Table 3 and Figure 12 As shown, the AUROC value of the embodiment is 0.890; the AUROC value of Comparative Example 1 is 0.881; the AUROC value of Comparative Example 2 is 0.864; and the AUROC value of Comparative Example 3 is 0.824. Therefore, the prediction model provided by the embodiments of this application has high prediction accuracy.

[0063] To demonstrate the beneficial effects of the LightGBM model used in the embodiments of this application, this application also provides Comparative Examples Four to Six.

[0064] Comparative Example 4: The model was trained using a logistic regression model (LogReg model), with an AUROC value of 0.879; Comparative Example 5: The SGD-SVM model was trained, and the AUROC value was 0.881; Comparative Example 6: The model was trained using a Random Forest (RF) model, with an AUROC value of 0.778.

[0065] Precision, recall, accuracy, and F1 score were calculated based on regression prediction probabilities for the embodiments and comparative examples four to six provided in this application. The AUROC values ​​in the ROC curves of the models trained in the embodiments and comparative examples four to six provided in this application were analyzed, and the optimal cutoff value was determined using the maximum Yangen index (sensitivity + specificity - 1). The calculation results are shown in Table 4 below.

[0066] Table 4: Calculation results of Examples 4 to 6

[0067] As shown in Table 4, Figure 13 As shown, the LightGBM model provided in this application embodiment has the best overall discrimination ability. On this basis, the LightGBM model also maintains a high recall rate and better prediction performance.

[0068] Furthermore, the test set is input into the prediction model, the hyperparameters of the prediction model are optimized, and then the performance of the prediction model is evaluated. The hyperparameter optimization of the prediction model specifically includes: selecting 100 different configuration parameters of the prediction model through GridSearchCV, and selecting the configuration parameter combination with the highest average AUROC value and acceptable calibration effect, specifically: learning rate = 0.05, number of trees (n_estimators) = 200, number of leaf nodes (num_leaves) = 6, maximum depth (max_depth) = 15, row sampling ratio (subsample) = 0.7, column sampling ratio (colsample_bytree) = 0.64, type (boosting) = gbdt, maximum binary threshold = 225.

[0069] Table 5: Hyperparameter search range and final selected values ​​of the LightGBM classifier

[0070] The search space described above covers a reasonable range of values ​​for each parameter to ensure that the model performance reaches its optimal level.

[0071] The evaluation of the predictive model's performance specifically includes: using a five-fold stratified cross-validation method combined with gain importance analysis and SHAP value. Evaluation metrics include log loss, area under the ROC curve (AUROC), sensitivity, specificity, precision, recall, F1 score, gain importance analysis, and SHAP value.

[0072] This application quantifies the contribution of features using SHAP values, enabling interpretation of prediction results at both the global and individual levels. SHAP analysis revealed that lymphocyte count is the most influential predictor, contributing the most to the model output (odds ratio = 2.26, p < 0.001). Gender is the second most important feature, with male participants showing a higher predictive risk. Previous chemotherapy for cancer (International Classification of Diseases Z512) is also one of the strongest predictive factors.

[0073] Specifically, the prediction model is calibrated to obtain a calibrated prediction model. This includes calibrating the prediction probability of the prediction model through isotonic regression, obtaining a calibrated prediction model, and then verifying it through the Hosmer-Lemeshow test (HL test).

[0074] Given the low prevalence of CLL (approximately 0.2%), the predictive model tends to overestimate the risk in the test set, thus requiring calibration of the predicted probabilities. Isotonic regression was used to correct for the prevalence discrepancies, improving the consistency between the predicted and observed risks. After calibration, the original and calibrated probabilities were obtained. The calibrated probabilities were validated against the true probabilities using the HL test, further demonstrating the accuracy of the calibration results.

[0075] To evaluate the generalization ability of the prediction model over time, the prediction results were validated at 5-year, 10-year, and all-onset timelines (119 cases followed up for 5 years, 245 cases followed up for 10 years, and 579 cases with all-onset disease). Due to differences in event incidence rates across time groups, isotonic regression analysis was performed within each group to dynamically calibrate the predicted probabilities. The analysis showed a high degree of consistency between the predicted probabilities and actual observations. These results indicate that the prediction model provided in this application has good predictive performance and can provide reliable disease risk assessment across different time spans.

[0076] Table 6: Time-follow-up results

[0077] As shown in Table 6, Figure 14 As shown, in the training set, the area under the ROC curve (AUROC) for predicting the overall incidence rate reaches 0.916, the 5-year AUROC is 0.997, and the 10-year AUROC is 0.973. This indicates that the prediction model has strong discriminative performance. Figure 15 As shown in the test set, the AUROC value for predicting the overall incidence rate was 0.884, the 5-year AUROC value was 0.876, and the 10-year AUROC value was 0.873. These results indicate the accuracy of the predictive model, its robustness across different time dimensions, and its good temporal generalization ability.

[0078] like Figure 16 As shown, the 5-year, 10-year, and full onset timelines (corresponding to...) Figure 16 (j) Figure 16 (k) Figure 16 (l) The agreement between the predicted values ​​and the actual observed values ​​of the prediction model was analyzed. In the HL test for 5-year, 10-year and all-cause prediction, the p-values ​​were 0.18, 0.30 and 0.08, respectively, indicating that the prediction effect after calibration was good. The corresponding Brier scores were 0.02, 0.05 and 0.10, respectively, indicating that the prediction results were relatively accurate.

[0079] It should be understood that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order constraint on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the diagram may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0080] This application also provides a prediction method for chronic lymphocytic leukemia based on a machine learning-based prediction model, specifically including: Obtain the data corresponding to the predictive factors for the individuals to be tested; Input the data corresponding to the predictor into the calibrated prediction model; The prediction model is used to perform calculations and output the prediction results.

[0081] This application also provides a system for constructing a machine learning-based chronic lymphocytic leukemia prediction model, including modules for implementing the machine learning-based chronic lymphocytic leukemia prediction model construction method as described in any of the above claims.

[0082] like Figure 17 As shown, specifically, the system construction includes: The acquisition module 10 is used to acquire raw sample data from the biobank, which includes multiple participants and multiple features; The filtering module 20 is used to perform preliminary filtering on the original sample data to obtain filtered sample data; The data cleaning module 30 is used to clean the filtered sample data to obtain cleaned sample data. The predictor determination module 40 is used to determine predictors from multiple features of the cleaned sample data; The partitioning module 50 is used to divide the cleaned sample data corresponding to the predictor and the target variable into training set and test set; Imbalance processing module 60 is used to process the training set to be imbalanced; The prediction model training module 70 is used to input the imbalanced training set into the LightGBM model for training to obtain the prediction model. The optimization and evaluation module 80 is used to input the test set into the prediction model, optimize the hyperparameters of the prediction model, and evaluate the performance of the prediction model. The calibration module 90 is used to calibrate the prediction model to obtain the calibrated prediction model.

[0083] like Figure 18 As shown, specifically, the filtering module 20 includes: The first elimination unit 201 is used to eliminate participants and their corresponding characteristic data who meet the following conditions: withdrawal from the study, lack of follow-up information, non-Caucasian ancestry, discrepancy between reported gender and genetic test gender, lack of exome sequencing data, previous diagnosis of hematologic malignancies, and genetic association with other participants. The second elimination unit 202 is used to eliminate features that meet the following conditions: imaging data, online random access questionnaires, algorithm-defined prognostic indicators, correlation data other than diagnostic codes in cancer / death registration systems or general practitioner cases, genomic comprehensive indicators, instrument handling or quality control indicators, and codes recorded after diagnosis of chronic lymphocytic leukemia.

[0084] like Figure 19 As shown, more specifically, the data cleaning module 30 includes: Missing data processing unit 301 is used to remove features of participants with a missing data rate of more than 50%. Imputation unit 302 is used to impute numerical features using the mean and to impute categorical features using the mode; Splitting unit 303 is used to split the responses to multiple-choice features in the questionnaire into independent binary variables; The categorical variable processing unit 304 is used to process categorical variables in non-numerical features using one-hot encoding.

[0085] like Figure 20 As shown, more specifically, the predictor determination module 40 includes: The sorting unit 401 is used to train multiple features of the cleaned sample data using a light gradient boosting machine, and sort the feature importance by gain value, retaining the top 200 features. The relevant feature removal unit 402 is used to identify and remove highly correlated features (r > 0.9) from 200 features using the Pearson correlation coefficient; The reduction unit 403 is used to reduce the features to 50 using an iterative feature selector; Analysis unit 404 is used to retain the first 20 features out of 50 features through SHAP analysis; The receiving unit 405 is used to receive the expert's selection instructions; The determination unit 406 is used to determine 9 predictors from 20 features.

[0086] like Figure 21 As shown, more specifically, the imbalance processing module 60 includes: Positive sample boosting unit 601 is used to increase the number of positive samples in the training set to 2% of the number of negative samples; Negative sample reduction unit 602 is used to reduce the number of negative samples by 50% using a random undersampling method.

[0087] This application also provides an electronic device, including: Memory; Processor; and Computer programs; The computer program is stored in the memory and configured to be executed by the processor to implement the method for constructing a machine learning-based predictive model for chronic lymphocytic leukemia as described in any of the above claims.

[0088] This application also provides a computer-readable storage medium having a computer program stored thereon; the computer program is executed by a processor to implement the method for constructing a machine learning-based predictive model for chronic lymphocytic leukemia as described in any of the above claims.

[0089] All analyses in this application were performed using Python 3.9, and the packages and versions used included: NumPy (v1.26.4), SciPy (v1.14.1), pandas (v2.2.3), scikit-learn (v1.2.2), imbalanced-learn (v0.12.4), matplotlib (v3.9.3), LightGBM (v3.3.2), and SHAP (v0.46.0).

[0090] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as C, VHDL, Verilog, the object-oriented programming language Java, and the interpreted scripting language JavaScript.

[0091] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure One One or more processes and / or boxes Figure One A device that provides the functions specified in one or more boxes.

[0092] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure One One or more processes and / or boxes Figure One The function specified in one or more boxes.

[0093] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure One One or more processes and / or boxes Figure One The steps of the function specified in one or more boxes.

[0094] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0095] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0096] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for constructing a machine learning-based predictive model for chronic lymphocytic leukemia, characterized in that, include: Obtain raw sample data from the biobank and perform preliminary screening of the raw sample data to obtain the screened sample data; The original sample data includes multiple participants and multiple features; The filtered sample data is cleaned to obtain the cleaned sample data. Predictors are determined from multiple features of the cleaned sample data; The cleaned sample data corresponding to the predictor and the target variable are divided into training set and test set, and the training set is imbalanced. The imbalanced training set is input into the LightGBM model for training to obtain the prediction model. The test set is input into the prediction model, the hyperparameters of the prediction model are optimized, and then the performance of the prediction model is evaluated. The prediction model is calibrated to obtain the calibrated prediction model.

2. The method for constructing a machine learning-based predictive model for chronic lymphocytic leukemia according to claim 1, characterized in that, The predictive factors include: lymphocyte count, sex, sleep duration, smoking status, previous chemotherapy, hepatosplenomegaly, abnormal blood biochemistry, history of musculoskeletal diseases, and secondary malignant tumors.

3. The method for constructing a machine learning-based predictive model for chronic lymphocytic leukemia according to claim 1, characterized in that, The step of cleaning the screened sample data to obtain cleaned sample data specifically includes: Features of participants with a data missing rate exceeding 50% were excluded; The mean is used to impute numerical features; The mode is used to impute categorical features; The responses to the multiple-choice feature in the questionnaire were split into independent binary variables; One-hot encoding is used to process categorical variables in non-numerical features.

4. The method for constructing a machine learning-based predictive model for chronic lymphocytic leukemia according to claim 1 or 3, characterized in that, The predictor is determined from multiple features of the cleaned sample data by performing the following steps in sequence: A light gradient boosting machine was used to train multiple features on the cleaned sample data, and the importance of the features was ranked by the gain value, retaining the top 200 features. Pearson correlation coefficient was used to identify and remove highly correlated features with r > 0.9 from 200 features; An iterative feature selector was used to reduce the number of features to 50. The first 20 features out of 50 were retained through SHAP analysis; Receive selection instructions from experts and determine 9 predictors from 20 features.

5. The method for constructing a machine learning-based predictive model for chronic lymphocytic leukemia according to claim 1, characterized in that, The evaluation of the predictive model's performance specifically includes: evaluation using a five-fold stratified cross-validation method combined with gain importance analysis and SHAP value.

6. The method for constructing a machine learning-based predictive model for chronic lymphocytic leukemia according to claim 1, characterized in that, The filtering of the original sample data specifically includes: Participants and their corresponding characteristic data who meet the following criteria will be excluded: withdrawal from the study, lack of follow-up information, non-Caucasian ancestry, discrepancy between reported gender and genetic test gender, lack of exome sequencing data, previous diagnosis of hematologic malignancy, or genetic association with other participants. Features that meet the following criteria will be excluded: imaging data, online follow-up questionnaires, algorithm-defined prognostic indicators, correlation data other than diagnostic codes in cancer / death registry systems or general practitioner cases, genomic comprehensive indicators, instrument handling or quality control indicators, and codes recorded after diagnosis of chronic lymphocytic leukemia.

7. The method for constructing a machine learning-based predictive model for chronic lymphocytic leukemia according to claim 1, characterized in that, The imbalance processing of the training set specifically includes: Increase the number of positive samples in the training set to 2% of the number of negative samples; Then, a random undersampling method was used to reduce the number of negative samples by 50%.

8. A system for constructing a machine learning-based predictive model for chronic lymphocytic leukemia, characterized in that, Includes a module that implements the method for constructing a machine learning-based chronic lymphocytic leukemia prediction model as described in any one of claims 1 to 7.

9. An electronic device, characterized in that, include: Memory; processor; as well as Computer programs; The computer program is stored in the memory and configured to be executed by the processor to implement the method for constructing a machine learning-based predictive model for chronic lymphocytic leukemia as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program; the computer program is executed by a processor to implement the method for constructing a machine learning-based predictive model for chronic lymphocytic leukemia as described in any one of claims 1 to 7.

Images