Construction method of liver cancer risk prediction model based on non-virological indexes

Abstract

The invention discloses a construction method of a liver cancer risk prediction model based on non-virological indexes, and the method comprises the steps: collecting chronic HBV infected person data containing 12 parameters, carrying out desensitization, carrying out stratified random sampling to divide a data set, and carrying out two-stage screening to obtain TB, ALT, ALP, GGT, PLT and AFP core features; after Min-Max Scaling processing, L1 regularization logistic regression is combined with grid search and 5-fold cross validation to construct the model, and the model constructed by the method is obviously superior to a traditional CU-HCC model and a REACH-B model, especially has higher recognition capability for early HBV-PLC, and can effectively reduce the missed diagnosis rate.

Description

Technical Field

[0001] This invention relates to the field of medical informatics technology, specifically to a method for constructing a liver cancer risk prediction model based on non-virological indicators. Background Technology

[0002] The prognosis of early-stage and late-stage chronic hepatitis B virus (HBV)-related primary liver cancer (HBV-PLC) differs greatly. Therefore, early screening and diagnosis of HBV-PLC are very important in chronic disease management.

[0003] Ultrasound examination is a preliminary screening for HBV-PLC, but ordinary ultrasound examination cannot diagnose liver cancer. Contrast-enhanced liver ultrasound, CT, MRI, and PET-CT have better detection efficacy, but the latter have drawbacks such as contrast agent allergies, high cost, inability to repeat examinations multiple times in a short period, and limited accessibility in primary care settings. Although serum alpha-fetoprotein, abnormal prothrombin, CA199, microRNA, circulating tumor DNA, and liquid biopsy are also screening methods for HBV-PLC, these indicators suffer from insufficient sensitivity and / or specificity, low reimbursement rates under medical insurance, exclusion from medical insurance coverage, and inaccessibility in primary care hospitals, thus limiting their clinical application.

[0004] Therefore, numerous predictive models for liver cancer have emerged both domestically and internationally. However, existing liver cancer prediction models mostly consider age and gender, or incorporate virological indicators such as hepatitis B surface antigen (HBsAg), HBV DNA quantification, and gene mutations; or biochemical indicators such as alanine aminotransferase (ALT), albumin, and hematological indicators such as platelets; or the presence of conditions such as cirrhosis, alcohol consumption, and diabetes. These indicators and their combinations each have their advantages and disadvantages, and some indicators are difficult to obtain comprehensively in primary hospitals, thus limiting their widespread applicability. Therefore, how to further explore the intrinsic correlation between various clinical and laboratory indicators and the development and progression of HBV-related liver diseases and HBV-PLC, and construct HBV-PLC risk prediction models with high accessibility in primary hospitals, has become an urgent technical problem to be solved in this field. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0006] Therefore, the purpose of this invention is to provide a method for constructing a liver cancer risk prediction model based on cholestasis indicators. The constructed model has a stronger ability to identify early HBV-PLC and can effectively reduce the missed diagnosis rate.

[0007] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution:

[0008] A method for constructing a liver cancer risk prediction model based on non-virological indicators, comprising the following steps:

[0009] S1. Collect clinical test data of chronic HBV infected individuals. The clinical test data shall include at least alkaline phosphatase and gamma-glutamyl transferase. Valid data shall be screened according to preset inclusion and exclusion criteria. After desensitization of the valid data, it shall be split into training set and validation set by stratified random sampling.

[0010] S2. A two-stage feature screening process is used to screen the clinical test data in step S1 to obtain the core feature set;

[0011] S3. Based on the core feature set obtained in step S2, the HBV-PLC risk prediction model is constructed using machine learning algorithms, and the optimal parameters of the model are determined through hyperparameter tuning and cross-validation.

[0012] S4. The performance of the constructed prediction model is evaluated using the area under the ROC curve, sensitivity, and specificity. The qualified model is used for HBV-PLC risk prediction in chronic HBV infected individuals.

[0013] As a preferred embodiment of the method for constructing a liver cancer risk prediction model based on non-virological indicators as described in this invention, in step S1, the clinical test indicators are alkaline phosphatase, gamma-glutamyl transferase, alanine aminotransferase, total bilirubin, alpha-fetoprotein, and platelets.

[0014] As a preferred embodiment of the method for constructing a liver cancer risk prediction model based on non-virological indicators as described in this invention, in step S1, the inclusion criteria are: chronic HBV infected individuals with serum HBsAg positivity duration > 6 months and age ≥ 18 years.

[0015] The exclusion criteria are: incomplete clinical data; coexisting hepatitis C, hepatitis D, autoimmune liver disease, alcoholic liver disease, Wilson's disease, or human immunodeficiency virus infection and having already received organ transplantation; or having been diagnosed with HBV-PLC or having a history of malignant tumors in organs other than the liver at the time of enrollment.

[0016] The stratified random sampling split ratio is 7:3 (training set:validation set), and the proportion of HBV-PLC patients in the training set is not less than 10%. The sampling process uses a fixed random seed to ensure that the results are reproducible.

[0017] As a preferred embodiment of the method for constructing a liver cancer risk prediction model based on non-virological indicators as described in this invention, the two-stage feature screening in step S2 specifically includes:

[0018] Phase 1: The Mann-Whitney U test was used to screen the clinical test data to identify features that showed significant differences in distribution between HBV-PLC patients and non-HBV-PLC patients, and to filter out noisy features that had no discriminative power.

[0019] The second stage: Based on the feature subset obtained from the initial screening in the first stage, the LassoCV algorithm is used for further screening. The feature space is compressed by L1 regularization to remove highly correlated or weakly contributing redundant features, and finally the core feature set containing TB, ALT, ALP, GGT, PLT and AFP is obtained.

[0020] As a preferred embodiment of the method for constructing a liver cancer risk prediction model based on non-virological indicators as described in this invention, step S3, the model construction process specifically includes:

[0021] Min-Max Scaling is used to linearly map each feature value in the core feature set to the [0,1] interval. The transformation formula is as follows:

[0022]

[0023] Where, x ij The original value of the feature to be scaled, min j The minimum value of this feature in the training set, max j To find the maximum value of this feature in the training set, x ij scaled These are the scaled feature values; and the min of the training set is used. j With max j Synchronously scale the validation set features;

[0024] L1 regularized logistic regression is used as the basic classifier for HBV-PLC risk prediction, and its objective function is:

[0025]

[0026] Where N is the number of samples, y i The true label for the i-th sample is 1 for HBV-PLC patients and 0 for non-HBV-PLC patients. iTo predict the probability that the i-th sample is HBV-PLC, β = (β1, β2, ..., β...). p ) represents the model parameter vector, and λ is the regularization strength hyperparameter, used to balance the model's fitting ability and complexity;

[0027] On the training set, a grid search is used to traverse the preset regularization strength parameter λ, and the performance of each parameter combination is evaluated by combining 5-fold hierarchical cross-validation. The parameter λ with the best average performance in the cross-validation is selected as the final parameter of the model.

[0028] As a preferred embodiment of the method for constructing a liver cancer risk prediction model based on non-virological indicators as described in this invention, the risk probability calculation formula for constructing the HBV-PLC risk prediction model in step S3 is as follows:

[0029] 1 / (1+exp(-(-0.399736-0.481501*((ALT-71.810948) / 193.273854)+0.48 7104*((ALP-88.136594) / 64.078400)+0.384761*((GGT-62.763811) / 111. 254706)-0.243420*((TB-27.410314) / 55.529364)+0.194164*((AFP-2560 .211518) / 21211.281471)-0.759320*((PLT-174.857831) / 74.673873)));

[0030] In this context, ALT represents alanine aminotransferase, ALP represents alkaline phosphatase, GGT represents gamma-glutamyl transferase, TB represents total bilirubin, AFP represents serum alpha-fetoprotein, and PLT represents platelets.

[0031] As a preferred embodiment of the method for constructing a liver cancer risk prediction model based on non-virological indicators as described in this invention, the cutoff value of the constructed HBV-PLC risk prediction model is 0.48; when the risk prediction value is ≥0.48, the subject is judged to be at high risk of HBV-PLC; when the risk prediction value is <0.48, the subject is judged to be at low risk of HBV-PLC.

[0032] A risk prediction device for HBV-related liver cancer, comprising:

[0033] Indicator acquisition module: used to acquire the detection values ​​of clinical test indicators of subjects, wherein the clinical test indicators include at least alkaline phosphatase and gamma-glutamyl transferase;

[0034] Feature scaling module: Used to perform feature scaling processing on the detection values ​​of the clinical laboratory indicators.

[0035] Model calculation module: used to substitute the scaled detection values ​​into the risk prediction model based on L1 regularized logistic regression to calculate the HBV-PLC risk prediction value;

[0036] Risk assessment module: used to determine the subject's HBV-PLC risk level based on the comparison between the predicted risk value and the preset threshold value.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] 1. This invention, through a two-stage feature screening process, integrates cholestasis-related indicators (ALP, GGT) with indicators such as TB, ALT, PLT, and AFP into a core feature set for the first time. The constructed A3GBP model achieves AUCs of 0.80 and 0.78 in the training and validation sets, respectively, with sensitivity and specificity of 0.71 and 0.74, respectively. This is significantly better than the traditional CU-HCC and REACH-B models, especially in identifying early HBV-PLC, which can effectively reduce the false negative rate.

[0039] 2. The risk prediction model constructed in this invention does not rely on special detection indicators such as HBV DNA. All indicators in the core feature set can be obtained through routine testing methods in primary healthcare institutions (including township health centers). No special equipment or professional testing personnel are required. It can accurately, conveniently, and cost-effectively predict the HBV-PLC risk of chronic HBV infected individuals, which helps to achieve early detection, early diagnosis, and early treatment of HBV-PLC, thereby reducing the incidence and mortality of HBV-PLC. It solves the problem that existing models are difficult to promote and apply in primary healthcare scenarios, and is conducive to large-scale screening of HBV-PLC risk.

[0040] 3. The predictive model of this invention has good predictive efficacy for both chronic HBV-infected individuals receiving antiviral treatment and those not receiving antiviral treatment. In particular, its predictive advantage is more obvious for those not receiving antiviral treatment, filling the application gap of existing models in this population. Moreover, this invention can directly output the HBV-PLC high and low risk determination results through a clear risk probability calculation formula and a cutoff value (0.52). Clinicians do not need to have machine learning expertise to quickly apply it, which is conducive to its promotion and use in medical institutions at all levels, especially primary hospitals. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0042] Figure 1 This is a flowchart illustrating a method for constructing a liver cancer risk prediction model based on non-virological indicators according to the present invention.

[0043] Figure 2 The area under the ROC curve of the liver cancer risk prediction model provided by this invention for predicting HBV-PLC in the training set and the test set is shown in the figure. Detailed Implementation

[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0045] like Figure 1 As shown, a method for constructing a liver cancer risk prediction model based on non-virological indicators is presented, and the specific steps are as follows:

[0046] S1. Collect clinical test data from chronic HBV-infected individuals. The clinical test data must include at least alkaline phosphatase and gamma-glutamyl transferase (GGT). Valid data are screened according to pre-defined inclusion and exclusion criteria. After desensitization, the valid data are split into training and validation sets using stratified random sampling. The clinical test indicators are alkaline phosphatase, GGT, alanine aminotransferase, total bilirubin, alpha-fetoprotein (AFP), and platelets. The inclusion criterion is: chronic HBV infection with serum HBsAg positivity lasting >6 months. Patients infected with BV and aged ≥18 years were excluded. The exclusion criteria were: incomplete clinical data; coexisting hepatitis C, hepatitis D, autoimmune liver disease, alcoholic liver disease, Wilson's disease, or human immunodeficiency virus infection and having undergone organ transplantation; or a history of HBV-PLC diagnosis or malignant tumors in organs other than the liver at the time of enrollment. The stratified random sampling split ratio was 7:3 (training set:validation set), and the proportion of HBV-PLC patients in the training set was not less than 10%. The sampling process used a fixed random seed to ensure reproducibility of results.

[0047] S2. A two-stage feature screening process is used to screen the clinical test data in step S1 to obtain a core feature set. The two-stage feature screening is as follows: First stage: The Mann-Whitney U test is used to initially screen the clinical test data to identify features that show significant differences in distribution between HBV-PLC patients and non-HBV-PLC patients, and to filter out noisy features that have no discriminative power; Second stage: On the feature subset obtained from the initial screening in the first stage, the LassoCV algorithm is used for further screening. The feature space is compressed by the L1 regularization term to remove highly correlated or weakly contributing redundant features, and finally a core feature set containing TB, ALT, ALP, GGT, PLT, and AFP is obtained.

[0048] S3. Based on the core feature set obtained in step S2, an HBV-PLC risk prediction model is constructed using machine learning algorithms. The optimal parameters of the model are determined through hyperparameter tuning and cross-validation. In this step, the model construction process specifically includes:

[0049] Min-Max Scaling is used to linearly map each feature value in the core feature set to the [0,1] interval. The transformation formula is as follows:

[0050]

[0051] Where, x ij The original value of the feature to be scaled, min j The minimum value of this feature in the training set, max j To find the maximum value of this feature in the training set, x ij scaled These are the scaled feature values; and the min of the training set is used. j With max j Synchronously scale the validation set features;

[0052] L1 regularized logistic regression is used as the basic classifier for HBV-PLC risk prediction, and its objective function is:

[0053]

[0054] Where N is the number of samples, y i The true label for the i-th sample is 1 for HBV-PLC patients and 0 for non-HBV-PLC patients. i To predict the probability that the i-th sample is HBV-PLC, β = (β1, β2, ..., β...). p ) represents the model parameter vector, and λ is the regularization strength hyperparameter, used to balance the model's fitting ability and complexity;

[0055] On the training set, a grid search is used to traverse the preset regularization strength parameter λ, and the performance of each parameter combination is evaluated by combining 5-fold hierarchical cross-validation. The parameter λ with the best average performance in the cross-validation is selected as the final parameter of the model.

[0056] The formula for calculating the risk probability in the HBV-PLC risk prediction model is as follows:

[0057] 1 / (1+exp(-(-0.399736-0.481501*((ALT-71.810948) / 193.273854)+0.48 7104*((ALP-88.136594) / 64.078400)+0.384761*((GGT-62.763811) / 111. 254706)-0.243420*((TB-27.410314) / 55.529364)+0.194164*((AFP-2560 .211518) / 21211.281471)-0.759320*((PLT-174.857831) / 74.673873)));

[0058] Among them, ALT represents alanine aminotransferase, ALP represents alkaline phosphatase, GGT represents gamma-glutamyl transferase, TB represents total bilirubin, AFP represents serum alpha-fetoprotein, and PLT represents platelets.

[0059] The cutoff value of the constructed HBV-PLC risk prediction model is 0.48; when the risk prediction value is ≥0.48, the subject is judged to be at high risk of HBV-PLC; when the risk prediction value is <0.48, the subject is judged to be at low risk of HBV-PLC.

[0060] S4. The performance of the constructed prediction model is evaluated using the area under the ROC curve, sensitivity, and specificity. The qualified model is used for HBV-PLC risk prediction in chronic HBV infected individuals.

[0061] To verify the effectiveness of the method for constructing a liver cancer risk prediction model based on non-virological indicators and the performance of the obtained liver cancer risk prediction model, the following specific embodiment is provided:

[0062] I. Experimental Preparation

[0063] 1. Selection of research subjects

[0064] 1.1 The clinical data in this embodiment comes from the clinical data of 5,872 patients with chronic HBV infection (1,833 received antiviral therapy and 4,039 did not) who visited the Eastern Theater Command General Hospital from January 1, 2013 to December 31, 2023 and whose information was relatively complete.

[0065] 1.2 Inclusion criteria:

[0066] Serum HBsAg positivity duration > 6 months;

[0067] Age ≥ 18 years old;

[0068] The clinical data is complete and includes all the detection indicators required for this invention.

[0069] 1.3 Exclusion criteria:

[0070] Combined with hepatitis C, hepatitis D, autoimmune liver disease, alcoholic liver disease, or Wilson's disease;

[0071] Human immunodeficiency virus (HIV) infection;

[0072] He has already received a liver transplant or other organ transplant;

[0073] Patients were diagnosed with HBV-PLC or had a history of other malignant tumors at the time of enrollment.

[0074] Clinical test data contain significant outliers or missing values ​​(missing percentage > 5%).

[0075] 2. Indicator Selection and Detection

[0076] Twelve all-factor parameters were included as initial candidate indicators. All indicators were obtained through routine clinical testing methods, as detailed below:

[0077]

[0078]

[0079] The IQR method is used to identify outliers, and outliers that exceed the reasonable range are professionally judged and processed to ensure data reliability.

[0080] II. Model Building Steps

[0081] 1. Dataset Splitting: A stratified random sampling method was used to divide the clinical data of 5872 patients into a training set and a validation set in a 7:3 ratio. The training set contained 3946 patients, and the validation set contained 1926 patients. During the splitting process, the proportion of HBV-PLC patients in the training set was strictly controlled to be no less than 10% to ensure the model's ability to learn key subgroups of HBV-PLC. At the same time, a fixed random seed (seed=42) was used to ensure the reproducibility of the dataset splitting results. The training set was used for model training and hyperparameter tuning, while the validation set was only used for model generalization performance evaluation and did not participate in the training process.

[0082] 2. Feature Filtering:

[0083] 2.1 First Stage: Mann-Whitney U Test

[0084] The Mann-Whitney U test was used to screen 12 initial candidate indicators. This method does not require the assumption that the data follows a normal distribution and is suitable for scenarios with skewed distributions, small samples, and outliers common in clinical biological data. The screening logic was as follows: the distribution difference of each indicator in the HBV-PLC group (positive class) and the non-HBV-PLC group (negative class) was compared, and indicators with statistically significant differences (P < 0.05) were retained, while "noise features" without classification ability were filtered out, resulting in the feature subset after the first stage of screening.

[0085] 2.2 Second Stage: LassoCV Screening

[0086] Based on the feature subset selected by the Mann-Whitney U test, L1 regularized logistic regression combined with cross-validation (LassoCV) is used to further select core features. LassoCV achieves feature coefficient compression and selection by adding an L1 regularization term to the logistic regression objective function;

[0087] By using LassoCV filtering, when the regularization strength lambda reaches its optimal value, the regression coefficients of unimportant features are compressed to 0, thereby automatically removing highly correlated or weakly contributing redundant features, ultimately yielding 6 core features: TB, ALT, ALP, GGT, PLT, and AFP.

[0088] 3. Machine learning modeling

[0089] Step 1: Data Preprocessing – Feature Scaling (Min-Max Scaling)

[0090] The six core features selected are then subjected to Min-Max scaling (normalization), which linearly maps the feature values ​​to the [0,1] interval, eliminating the impact of differences in the units of measurement of different indicators on model training. The specific steps are as follows:

[0091] Calculate the minimum value of each core feature based on the training set data. and maximum value Where x ij This represents the j-th feature value of the i-th sample in the training set;

[0092] The feature values ​​of the training and validation sets are scaled separately using the following formulas:

[0093]

[0094] Furthermore, the scaling of the validation set strictly follows the minimum value calculated from the training set. j and maxj To avoid data leaks.

[0095] Step 2: Model Selection and Training

[0096] L1 regularized logistic regression was chosen as the base classifier for the model. This classifier operates on the same principle as LassoCV in the feature selection stage, further compressing redundant feature coefficients and outputting clear predicted probabilities and feature coefficients, facilitating model interpretation. The training process used the "whether it is HBV-PLC" label of the samples in the training set as the label, and the scaled six core features as input to construct a preliminary prediction model.

[0097] Step 3: Hyperparameter tuning

[0098] The model hyperparameters (regularization strength parameter λ) were optimized using GridSearchCV combined with 5-fold hierarchical cross-validation. The specific steps are as follows:

[0099] The candidate value range for the predefined regularization strength parameter λ is set (e.g., 0.001, 0.01, 0.1, 1, 10, 100).

[0100] On the training set, each candidate λ value is traversed using a 5-fold stratified cross-validation method. The average AUC corresponding to different λ values ​​is calculated using the area under the receiver operating characteristic curve (AUC) as the performance scoring standard.

[0101] The λ value with the largest average AUC is selected as the final hyperparameter of the model to determine the optimal L1 regularized logistic regression model.

[0102] Step 4: Model Performance Evaluation

[0103] Using the entire training set data, refit the model with the optimal hyperparameter λ, and calculate model performance metrics on both the training and validation sets, including: sensitivity, specificity, accuracy, and AUC. Figure 2 The evaluation results show that the model has an AUC of 0.8 (95% confidence interval: 0.79-0.83) on the training set and an AUC of 0.78 (95% confidence interval: 0.77-0.81) on the validation set; the sensitivity is 0.71, the specificity is 0.74, and the high-risk / low-risk cutoff value is 0.48 (i.e., a prediction probability ≥0.48 is considered high risk for HBV-PLC, and <0.48 is considered low risk).

[0104] 4. Final Model (A3GBP) Determined

[0105] Through the above steps, an HBV-PLC risk prediction model based on six core features is finally constructed, named the A3GBP model, and its prediction formula is as follows:

[0106]

[0107] III. Examples of Clinical Application of the Model

[0108] 1. Sample collection: Collect 4 tubes of fasting venous blood from the patient, totaling 8 ml. Centrifuge to separate serum (3000 r / min, centrifuge for 10 min) for the detection of biochemical indicators, virological indicators and tumor markers; at the same time, retain a portion of whole blood for platelet count detection.

[0109] Indicator detection: Following the method in "Indicator Selection and Detection" of this implementation method, the patient's TB, ALT, ALP, GGT, PLT and AFP, a total of 6 core indicators, were detected and the results were recorded;

[0110] Risk calculation: Substitute the test values ​​of the 6 indicators into the A3GBP model formula to calculate the patient's HBV-PLC risk value (the value range is 0-1).

[0111] Risk assessment and clinical recommendations:

[0112] If the risk value is ≥0.48 (the high-low risk cutoff value), the individual is considered to be at high risk for HBV-PLC. It is recommended to have an abdominal ultrasound examination and AFP test every 3-6 months. If necessary, further examinations such as liver ultrasound contrast imaging, CT or MRI can be performed to achieve early screening and diagnosis of HBV-PLC.

[0113] If the risk value is <0.48, the individual is considered to be at low risk for HBV-PLC and is advised to undergo an abdominal ultrasound examination and AFP test every 6-12 months, with routine follow-up management.

[0114] (III) Model Comparison and Validation

[0115] The performance of the A3GBP model was compared with that of existing main HBV-PL prediction models (such as the CU-HCC model and the REACH-B model) on the validation set. The results showed that in the subgroup of HBV-infected individuals not receiving antiviral treatment, the AUC of the A3GBP model [0.832 (0.796, 0.868)] was significantly higher than that of the CU-HCC model [0.681 (0.633, 0.730)] and the REACH-B model [0.712 (0.674, 0.750)]. Furthermore, the A3GBP model included six indicators in the study: ALP, GGT, AST, PLT, TB, and AFP. These six indicators can be detected in primary hospitals and even township health centers, requiring minimal laboratory and technician expertise, making them highly accessible. The area under the curve (AUC) for predicting HBV-PLC in the untreated population was superior to the existing CU-HCC and REACH-B models.

[0116] The present invention also provides a risk prediction device for HBV-related liver cancer, comprising:

[0117] Indicator acquisition module: used to acquire the detection values ​​of clinical test indicators of subjects, wherein the clinical test indicators include at least alkaline phosphatase and gamma-glutamyl transferase;

[0118] Feature scaling module: Used to perform feature scaling processing on the detection values ​​of the clinical laboratory indicators.

[0119] Model calculation module: used to substitute the scaled detection values ​​into the risk prediction model based on L1 regularized logistic regression to calculate the HBV-PLC risk prediction value;

[0120] Risk assessment module: used to determine the subject's HBV-PLC risk level based on the comparison between the predicted risk value and the preset threshold value.

[0121] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for constructing a liver cancer risk prediction model based on non-virological indicators, characterized in that, The steps are as follows: S1. Collect clinical test data of chronic HBV infected individuals. The clinical test data shall include at least alkaline phosphatase and gamma-glutamyl transferase. Valid data shall be screened according to preset inclusion and exclusion criteria. After desensitization of the valid data, it shall be split into training set and validation set by stratified random sampling. S2. A two-stage feature screening process is used to screen the clinical test data in step S1 to obtain the core feature set; S3. Based on the core feature set obtained in step S2, the HBV-PLC risk prediction model is constructed using machine learning algorithms, and the optimal parameters of the model are determined through hyperparameter tuning and cross-validation. S4. The performance of the constructed prediction model is evaluated using the area under the ROC curve, sensitivity, and specificity. The qualified model is used for HBV-PLC risk prediction in chronic HBV infected individuals.

2. The method for constructing a liver cancer risk prediction model based on non-virological indicators according to claim 1, characterized in that, In step S1, the clinical test indicators are alkaline phosphatase, gamma-glutamyl transferase, alanine aminotransferase, total bilirubin, alpha-fetoprotein, and platelets.

3. The method for constructing a liver cancer risk prediction model based on non-virological indicators according to claim 1, characterized in that, In step S1, the inclusion criteria are: chronic HBV infected individuals with a serum HBsAg positivity duration of >6 months and an age ≥18 years; The exclusion criteria are: incomplete clinical data; coexisting hepatitis C, hepatitis D, autoimmune liver disease, alcoholic liver disease, Wilson's disease, or human immunodeficiency virus infection and having already received organ transplantation; or having been diagnosed with HBV-PLC or having a history of malignant tumors in organs other than the liver at the time of enrollment. The stratified random sampling split ratio is 7:3 (training set:validation set), and the proportion of HBV-PLC patients in the training set is not less than 10%. The sampling process uses a fixed random seed to ensure that the results are reproducible.

4. The method for constructing a liver cancer risk prediction model based on non-virological indicators according to claim 1, characterized in that, In step S2, the two-stage feature selection is specifically as follows: Phase 1: The Mann-Whitney U test was used to screen the clinical test data to identify features that showed significant differences in distribution between HBV-PLC patients and non-HBV-PLC patients, and to filter out noisy features that had no discriminative power. The second stage: Based on the feature subset obtained from the initial screening in the first stage, the LassoCV algorithm is used for further screening. The feature space is compressed by L1 regularization to remove highly correlated or weakly contributing redundant features, and finally the core feature set containing TB, ALT, ALP, GGT, PLT and AFP is obtained.

5. The method for constructing a liver cancer risk prediction model based on non-virological indicators according to claim 1, characterized in that, In step S3, the model building process specifically includes: Min-Max Scaling is used to linearly map each feature value in the core feature set to the [0,1] interval. The transformation formula is as follows: Where, x ij The original value of the feature to be scaled, min j The minimum value of this feature in the training set, max j To find the maximum value of this feature in the training set, x ij scaled These are the scaled feature values; and the min of the training set is used. j With max j Synchronously scale the validation set features; L1 regularized logistic regression is used as the basic classifier for HBV-PLC risk prediction, and its objective function is: Where N is the number of samples, y i The true label for the i-th sample is 1 for HBV-PLC patients and 0 for non-HBV-PLC patients. i To predict the probability that the i-th sample is HBV-PLC, β = (β1, β2, ..., β...). p ) represents the model parameter vector, and λ is the regularization strength hyperparameter, used to balance the model's fitting ability and complexity; On the training set, a grid search is used to traverse the preset regularization strength parameter λ, and the performance of each parameter combination is evaluated by combining 5-fold hierarchical cross-validation. The parameter λ with the best average performance in the cross-validation is selected as the final parameter of the model.

6. The method for constructing a liver cancer risk prediction model based on non-virological indicators according to claim 1, characterized in that, The risk probability calculation formula for the HBV-PLC risk prediction model constructed in step S3 is as follows: 1 / (1+exp(-(-0.399736-0.481501*((ALT-71.810948) / 193.273854)+0.48 7104*((ALP-88.136594) / 64.078400)+0.384761*((GGT-62.763811) / 111. 254706)-0.243420*((TB-27.410314) / 55.529364)+0.194164*((AFP-2560 .211518) / 21211.281471)-0.759320*((PLT-174.857831) / 74.673873))); In this context, ALT represents alanine aminotransferase, ALP represents alkaline phosphatase, GGT represents gamma-glutamyl transferase, TB represents total bilirubin, AFP represents serum alpha-fetoprotein, and PLT represents platelets.

7. The method for constructing a liver cancer risk prediction model based on non-virological indicators according to claim 6, characterized in that, The cutoff value of the constructed HBV-PLC risk prediction model is 0.48; when the risk prediction value is ≥0.48, the subject is judged to be at high risk of HBV-PLC; when the risk prediction value is <0.48, the subject is judged to be at low risk of HBV-PLC.

8. A risk prediction device for HBV-related liver cancer, characterized in that, include: Indicator acquisition module: used to acquire the detection values ​​of clinical test indicators of subjects, wherein the clinical test indicators include at least alkaline phosphatase and gamma-glutamyl transferase; Feature scaling module: Used to perform feature scaling processing on the detection values ​​of the clinical laboratory indicators. Model calculation module: used to substitute the scaled detection values ​​into the risk prediction model based on L1 regularized logistic regression to calculate the HBV-PLC risk prediction value; Risk assessment module: used to determine the subject's HBV-PLC risk level based on the comparison between the predicted risk value and the preset threshold value.

Images