Multi-liver pathology assessment method and system based on machine learning

Abstract

The invention provides a method and a system for evaluating various liver pathologies based on machine learning. The method comprises the following steps: acquiring weight information, clinical examination information and evaluation results of liver tissue samples of a plurality of historical patients; performing data preprocessing based on the acquired information, and constructing a training set and a test set; training a Logistic regression model and a Ridge regression model by using the training set, determining the performance of the model in predicting balloon sample deformation, fatty degeneration, lobular inflammation, significant fibrosis and NASH, and determining final prediction index parameters and a prediction model; and obtaining weight information, clinical examination information and an evaluation result of the liver tissue sample of the patient, extracting related values of prediction index parameters in the evaluation result, and processing the prediction index parameters by the determined trained prediction model aiming at the corresponding prediction index parameters to obtain a final prediction result. According to the method, the simple model is constructed by obtaining the clinical indexes, and all aspects of the liver pathology of the chronic liver disease can be comprehensively evaluated.

Description

Technical Field

[0001] The present invention belongs to the field of pathological assessment, and specifically relates to a method and system for assessing multiple liver pathologies based on machine learning. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Studies have shown that chronic liver disease is closely related to the occurrence and development of a variety of extrahepatic diseases, including but not limited to obesity, diabetes, and cardiovascular disease (CVD). Large-scale cohort studies have shown that whether it is simple fatty liver, non-fibrotic metabolic dysfunctional steatohepatitis (MASH), or fibrosis that has not reached the stage of cirrhosis, the risk of all-cause mortality in patients increases significantly with the progression of metabolic-associated fatty liver disease (MAFLD). Therefore, for chronic liver disease, early diagnosis and intervention are of great significance for improving prognosis and reducing complications.

[0004] Currently, liver biopsy remains the gold standard for diagnosing chronic liver disease and evaluating histological changes. Early detection of liver histological changes can create opportunities to slow the progression of the disease and bring significant socioeconomic benefits. Although many current studies focus on blood-based omics testing as the main way of non-invasive testing, hoping to achieve early diagnosis of chronic liver disease, it still has limitations in accessibility and clinical translation. In addition, liver pathology diagnosis includes multiple indicators such as fatty degeneration, lobular inflammation, hepatocellular ballooning and fibrosis. For each liver pathology indicator, different models may need to be constructed to ensure the accuracy of the prediction, which also increases the difficulty of application. Therefore, there is an urgent need for a set of easy-to-use, low-cost, sensitive, accurate and comprehensive markers that can uniformly predict these characteristics and thus efficiently predict the occurrence of chronic liver disease.

[0005] Machine learning has emerged as a powerful tool for predicting and managing complex diseases. By training on large-scale clinical and multi-omics data, machine learning models are able to identify complex biomarkers that are critical for early diagnosis and risk stratification. However, these models need to learn from comprehensive datasets to identify relationships between different parameters and outcomes to accurately predict the onset and progression of chronic liver disease. Therefore, although machine learning has great potential to deepen our understanding and prediction of such diseases, the clinical translation of machine learning tools for the prediction and diagnosis of MAFLD is hampered by the lack of large-scale prospective studies integrating multiple types of clinical data. Summary of the invention

[0006] In order to solve the above problems, the present invention proposes a method and system for evaluating multiple liver pathologies based on machine learning. The present invention constructs a simple model by acquiring clinical indicators, which can achieve a comprehensive evaluation of various aspects of liver pathology in chronic liver diseases.

[0007] According to some embodiments, the present invention adopts the following technical solutions:

[0008] A method for assessing multiple liver pathologies based on machine learning, comprising the following steps:

[0009] Obtain weight information, clinical examination information, and pathological evaluation results of liver tissue samples from multiple historical patients;

[0010] Based on the information obtained, data preprocessing is performed to construct training and test sets, and key predictive features of each pathological indicator are determined;

[0011] Use the training set to train the machine learning model, determine the model's performance in predicting ballooning, fatty degeneration, lobular inflammation, significant fibrosis, NASH, NAS score, and risk of NASH, and determine the final predictive index parameters and prediction model;

[0012] The patient's weight information, clinical examination information and liver tissue sample evaluation results are obtained, and the relevant values ​​of the prediction index parameters are extracted. The corresponding prediction index parameters are processed with the determined trained prediction model to obtain the final prediction result.

[0013] As an optional embodiment, the weight information includes BMI data; the anthropometric parameters include gender, age, body mass index (BMI), systolic blood pressure (SBP), diastolic blood pressure (DBP), heart rate (HR), waist circumference, hip circumference and waist-to-hip ratio (WHR); the complete blood cell count includes white blood cell count (WBC), neutrophil ratio (Neu ratio), lymphocyte ratio (Lym ratio), eosinophil ratio (EOS ratio), basophil ratio (BAS ratio), monocyte ratio (Mon ratio), neutrophil count (Neu), lymphocytes (Lym), eosinophils (EOS), basophils (BAS), monocytes (Mono), red blood cell count (RBC), hemoglobin (Hb), hematocrit (HCT), mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (MCHC), red blood cell distribution width (RDW), platelet count (PLT), platelet distribution width (PDW), mean platelet volume (MPV ) and platelet count (PCT); the biochemical indicators include prealbumin (PA), total protein (TP), albumin (ALB), globulin (GLB), albumin / globulin ratio (AGR), α1 antitrypsin (Alpha.1), α2 macroglobulin (Alpha.2), β1 globulin (Beta.1), β2 microglobulin (Beta.2), gamma globulin (Gamma), alanine aminotransferase (ALT), aspartate aminotransferase (AST), glutamate dehydrogenase (GDH), gamma glutamyl transaminase (GGT), transferase (GGT), alkaline phosphatase (AKP), adenosine deaminase (ADA), αL-fucosidase (AFU), total bilirubin (TBIL), direct bilirubin (DBIL), indirect bilirubin (IBIL), total bile acid (TBA), homocysteine ​​(Hcy), phospholipase A2 (PLA2), β-hydroxybutyric acid (HBUT), urea, creatinine (Cr), uric acid (UA), complement component 1q (C1q), creatine kinase (CK), creatine kinase MB isoenzyme (CK.MB), cardiac troponin I (CTNI ), lactate dehydrogenase (LDH), folic acid (FOL) and angiotensin II (AII); the blood lipid indexes include total cholesterol (TC), high-density lipoprotein cholesterol (HDL), low-density lipoprotein cholesterol (LDL), small dense low-density lipoprotein cholesterol (sdLDL), apolipoprotein A1 (ApoA1), apolipoprotein B (ApoB), triglycerides (TG), lipoprotein A (LP.a), non-esterified free fatty acids (NEFA); the blood glucose indexes include fasting blood glucose (GLU), glycated albumin index (GA.index), ischemia-modified albumin (IMA), sialic acid (SA), glycosylated hemoglobin (HbA1c), fasting insulin (FINS), fasting C-peptide (FCP), homeostasis model assessment of insulin resistance (HOMA-IR), adipose insulin resistance (Adipo-IR), triglyceride glycemic index (TyG), triglyceride glycemic body mass index (TyG-BMI) and triglyceride glycemic waist circumference index (TyG-WC); the coagulation indices include prothrombin time (PT-S), prothrombin time international normalized ratio (PT-INR), prothrombin activity (PT%), prothrombin time ratio (PT-R), activated partial thromboplastin time (APTT-S), activated The above mentioned indicators include partial thromboplastin time ratio (APTT-R), fibrinogen (FIB), thrombin time (TT-S), thrombin time ratio (TT-R) and D-dimer (DD.i); the above mentioned metal indicators include calcium (Ca), cadmium (Cd), cobalt (Co), chromium (Cr), copper (Cu), iron (Fe), lithium (Li), magnesium (Mg), manganese (Mn), nickel (Ni), lead (Pb) and zinc (Zn); the above mentioned endocrine hormone indicators include prolactin (PRL), testosterone (T), luteinizing hormone to follicle stimulating hormone ratio (LH / FSH), growth hormone (hGH), free triiodothyronine (FT3), free thyroxine (FT4) and thyroid stimulating hormone (TSH). .

[0014] As an optional embodiment, the pathological evaluation results of the liver tissue sample include ballooning, steatosis, lobular inflammation, significant fibrosis, NASH, NAS score and risk of NASH.

[0015] As an optional implementation, the process of data preprocessing includes:

[0016] In the training set, the original data is z-score transformed according to the mean and standard deviation of each indicator to achieve data standardization. The formula for z-score transformation is:

[0017] z=(x-μ) / σ;

[0018] Where z is the z-score value of each data, that is, the standardized value; x is the original value of the data; μ is the mean of each indicator in the original data; σ is the standard deviation of each indicator in the original data.

[0019] As an optional implementation, during the preprocessing of the test set data, each clinical indicator is standardized using the mean and standard deviation of the training set, that is, the mean of the training set is subtracted and divided by the standard deviation of the training set as the standardized value to achieve standardization of the test data.

[0020] As an optional implementation, during data preprocessing, samples with each pathological outcome are marked as 1, and samples without each outcome are marked as 0, so as to be input into the machine learning model.

[0021] As an optional embodiment, in the training set, the best predictive features for each liver pathological outcome are screened by recursive feature elimination.

[0022] As an optional implementation, 23 common machine learning algorithms were used to build models to determine the performance of each model in predicting ballooning, fatty degeneration, lobular inflammation, significant fibrosis, NASH, NAS score, and risk NASH. On the basis of balancing model accuracy and clinical practicality, we screened the optimal algorithm and number of features for each liver pathology result to determine the final predictive indicator parameters and prediction model for each pathological outcome. At the same time, the model performance was tested in internal and external test cohorts;

[0023] As an optional implementation, for the Logistic regression model, a regularization term is added to prevent overfitting, and the loss function of the model is set to:

[0024] J(w)=-(1 / N)Σ[y i *log(p i )+(1-y i )*log(1-p i )]+λ*R(w);

[0025] Among them, J(w) is the loss function; N is the total number of samples; y i is the actual label of the i-th sample (0 or 1); p i is the probability that the i-th sample is predicted as category 1; λ is the regularization strength, which is automatically selected by cross-validation; R(w) is the regularization term, using L1 regularization;

[0026] The prediction formula of the model is:

[0027] p(y=1|x)=σ(w0+Σ(w j *x j ))

[0028] Among them: w0 is the bias term, output by the model; w j is the weight, that is, the contribution of each feature to the model, output by the model; x j is the jth component of the input feature vector, which is determined by the specific value of the indicator of each sample in the training set;

[0029] If p(y=1|x)≥0.5, the sample is predicted to be of category 1;

[0030] If p(y=1|x)<0.5, the sample is predicted to be of category 0.

[0031] As an optional implementation, for the Ridge regression model, its loss function is defined as:

[0032] L(w)=Σ(y i -(w0+Σ(w j *x ij ))) 2 +α*Σ(w j 2 )

[0033] Among them, y i is the actual label of the i-th sample; w0 is the bias term of the model; w j is the weight parameter of the jth feature; x ij is the jth eigenvalue of the i-th sample; α is the regularization strength;

[0034] The prediction formula of the model is:

[0035]

[0036] Among them, w j is the weight coefficient of the jth feature, output by the model; xi j is the jth eigenvalue of the ith sample, which is determined by the specific value of the indicator of each sample in the training set.

[0037] if Predict the category of this sample to be 1;

[0038] if The predicted class of this sample is 0.

[0039] The machine learning models of ballooning, fatty degeneration, lobular inflammation, significant fibrosis, NASH, NAS score, and risk NASH were established in sequence, and the model's predicted category for each sample was output, i.e., 0 or 1;

[0040] Under repeated stratified 5-fold cross validation, the balanced accuracy, precision, recall, F1 score and AUROC of each model were calculated;

[0041] The calculation method of each indicator is:

[0042] Balanced accuracy = (1 / 2)*[(TP / (TP+FN))+(TN / (TN+FP))];

[0043] Precision = TP / (TP+FP);

[0044] Recall = TP / (TP+FN);

[0045] F1 score = 2*(precision*recall) / (precision+recall);

[0046] AUROC = ∫(TPR versus FPR curve);

[0047] Among them, TP is a true positive example, TN is a true negative example, FN is a false negative example, and FP is a false positive example.

[0048] As an optional implementation, the final selected prediction indicators include, for example Figure 3 As shown:

[0049] For ballooning, alanine aminotransferase (ALT), total bile acid (TBA), and homeostasis model assessment of insulin resistance (HOMA-IR) were selected as predictors, and the Ridge regression model was selected;

[0050] For steatosis, ALT and glutamate dehydrogenase (GDH) were selected as predictors, and the Ridge regression model was selected;

[0051] For lobular inflammation, ALT, prothrombin activity (PT%), and cadmium (Cd) were selected as predictors, and the logistic regression model was selected;

[0052] For significant fibrosis, peripheral blood mononuclear cell ratio (Mono.ratio), aspartate aminotransferase (AST), FPG, glycated albumin index (GA.index), and β2-microglobulin (Beta2) were selected as predictive indicators, and the Logistic regression model was selected;

[0053] For NASH, phospholipase A2 (PLA2), Cd, magnesium (Mg), fasting insulin (FINS), HOMA-IR, and adipose tissue insulin resistance index (Adipo-IR) were selected as predictors, and the logistic regression model was selected;

[0054] For NAS score, ALT, cobalt (Co), and triglyceride-glucose body mass index (TyG-BMI) were selected as predictors, and the logistic regression model was selected;

[0055] For risk NASH, ALT, GA.index, and Beta2 were selected as predictors, and the Logistic regression model was selected.

[0056] Figure 3 The left side of the figure shows the ROC-AUC of each model in the training set and the test set, and the right side shows the recall of each model. The table on the right lists the features used by each model.

[0057] Furthermore, ALT, HOMA-IR, and Mono.ratio were selected as key clinical indicators from the 7 independent models for each pathological outcome, and a Logistic regression model was constructed to simultaneously predict multiple liver pathological outcomes as the prediction target.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] The present invention constructs a machine learning model by acquiring simple and easily available clinical indicators, which can achieve a comprehensive assessment of various aspects of liver pathology in chronic liver diseases and improve the efficiency and performance of pathology prediction.

[0060] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0062] Figure 1 A schematic diagram of a method flow chart of an embodiment;

[0063] Figure 2 The present invention is a schematic diagram of a method flow chart of another embodiment.

[0064] Figure 3 Performance of different prediction models for various liver pathology types.

[0065] Figure 4 A simplified general model for predicting outcomes in multiple liver pathologies.

[0066] Figure 5 To simplify the model's performance in the Chinese external validation cohort.

[0067] Figure 6 Performance of the simplified model in the US NHANES cohort. DETAILED DESCRIPTION

[0068] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0069] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0070] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0071] In the absence of conflict, the embodiments in this application and the features in the embodiments may be combined with each other.

[0072] Embodiment 1

[0073] A machine learning-based method for assessing multiple liver pathologies, e.g. Figure 1 As shown, the following steps are included:

[0074] A data acquisition module, which acquires weight information, clinical examination information, and evaluation results of liver tissue samples of multiple historical patients;

[0075] The preprocessing module performs data preprocessing based on the acquired information and constructs training sets and test sets;

[0076] Model training and validation module, using the training set to train 23 different machine learning models, to determine the performance of each model in predicting ballooning, fatty degeneration, lobular inflammation, significant fibrosis, NASH, NAS score, and risk NASH. On the basis of balancing model accuracy and clinical practicality, we screened the optimal algorithm and number of features for each liver pathology result to determine the final predictive indicator parameters and prediction model for each pathological outcome. At the same time, the model performance was tested in internal and external test cohorts;

[0077] The prediction module obtains the patient's weight information, clinical examination information and liver tissue sample evaluation results, extracts the relevant values ​​of the prediction index parameters, processes the corresponding prediction index parameters with the determined trained prediction model, and obtains the final prediction result.

[0078] The following is an introduction in steps.

[0079] The data acquisition module aims to obtain weight information, clinical examination information and evaluation results of liver tissue samples of multiple historical patients.

[0080] In this example, the training cohort included 909 individuals who underwent sleeve gastrectomy for morbid obesity at Qilu Hospital of Shandong University between May 2018 and October 2023.

[0081] The test cohort consisted of 186 individuals who underwent sleeve gastrectomy at the same hospital between December 2023 and June 2024. Liver tissue samples were collected from the patients during surgery to assess liver damage and liver pathological changes.

[0082] Physical examination, clinical and laboratory examinations are performed 1-2 days before surgery. The formula for calculating body mass index (BMI) is weight (kg) / height (m) 2 The calculation formula of insulin resistance index (HOMA-IR) is fasting glucose (mmol / L) × Fasting insulin (μIU / mL) / 22.5.

[0083] Diabetes is diagnosed based on a fasting blood glucose (FPG) level of ≥7.0mmol / L or a glycosylated hemoglobin (HbA1c) level of ≥6.5% (48mmol / mol). Metabolic syndrome is diagnosed based on the CDS2004 criteria, and individuals must meet at least three of the following diagnostic criteria: body mass index (BMI) ≥25kg / m 2 ; Blood pressure ≥140 / 90mmHg or taking antihypertensive drugs; fasting blood glucose ≥6.1mmol / L and / or oral glucose tolerance test (OGTT) 2-hour blood glucose ≥7.8mmol / L, or taking hypoglycemic drugs; triglycerides (TG) ≥1.7mmol / L and / or high-density lipoprotein cholesterol (HDL-c) <0.9mmol / L for men and <1.0mmol / L for women.

[0084] Liver biopsy samples were systematically evaluated by experienced liver pathologists, and the analysis of the training and test sets was performed by different experts. Liver fibrosis was graded using a 5-point scale (0 to 4 points). Histological changes—steatosis, lobular inflammation (0 to 3 points), and ballooning (0 to 2 points)—were evaluated according to the NASH CRN guidelines and the NAFLD Activity Scoring System (NAS). The NAS score is the sum of the scores for steatosis, lobular inflammation, and ballooning, ranging from 0 to 8 points.

[0085] Preprocessing module, the process of data preprocessing includes:

[0086] In the training set, the original data is z-score transformed according to the mean and standard deviation of each indicator to achieve data standardization. The formula for z-score transformation is:

[0087] z=(x-μ) / σ;

[0088] Where z is the z-score value of each data, that is, the standardized value; x is the original value of the data; μ is the mean of each indicator in the original data; σ is the standard deviation of each indicator in the original data.

[0089] During the data preprocessing process, for the test cohort, each clinical indicator was standardized using the mean and standard deviation of the training set, that is, the mean of the training set was subtracted and divided by the standard deviation of the training set as the standardized value to achieve standardization of the test data.

[0090] During the data preprocessing process, samples with each pathological outcome were marked as 1, and samples without each outcome were marked as 0 for input into the machine learning model.

[0091] The baseline information of the training set and the test set is as follows: the average age of the cohort is 31.5 years old, and the average body mass index (BMI) is 41.85 kg / m 2 . Of the subjects, 29.6% were male, 35.7% had type 2 diabetes (T2D), and 48% had metabolic syndrome (MetS). Liver pathology data showed that 89.8% showed ballooning (ballooning score>0), 73.9% had steatosis (steatosis score>0), 60.9% had lobular inflammation (inflammation score>0), and 22.2% had significant fibrosis (fibrosis score≥2). A total of 8.4% were diagnosed with NASH, 27.5% had a NAS score≥5, and 14.1% had NASH with fibrosis, referred to as "risk NASH" (NASH with fibrosis score≥2).

[0092]

[0093] Model training and validation modules

[0094] In this embodiment, when performing feature screening, we use recursive feature elimination in the training set to screen the best predictive features for each liver pathology outcome. Specifically, Logistic regression is used as the basic model for fitting data, and one feature is removed in each recursion to obtain the optimal predictive features for each liver pathology outcome.

[0095] In this example, we used 23 different algorithms to train the machine learning model, and used repeated stratified 5-fold cross-validation to evaluate the performance of the model by calculating the balanced accuracy, precision, recall, F1 score, and area under the receiver operating characteristic curve (AUROC). On the basis of balancing model accuracy and clinical practicality, we screened out the optimal algorithm and number of features for each liver pathology result.

[0096] For all seven liver pathological outcomes, 23 models were established in turn, including ballooning, fatty degeneration, lobular inflammation, significant fibrosis, NASH, NAS score, and risk NASH, and the model's predicted category for each sample was output, i.e., 0 or 1;

[0097] Under repeated stratified 5-fold cross validation, the balanced accuracy, precision, recall, F1 score and AUROC of each model were calculated;

[0098] The calculation method of each indicator is:

[0099] Balanced accuracy = (1 / 2)*[(TP / (TP+FN))+(TN / (TN+FP))];

[0100] Precision = TP / (TP+FP);

[0101] Recall = TP / (TP+FN);

[0102] F1 score = 2*(precision*recall) / (precision+recall);

[0103] AUROC = ∫(TPR versus FPR curve);

[0104] Among them, TP is a true positive example, TN is a true negative example, FN is a false negative example, and FP is a false positive example.

[0105] On the basis of balancing model accuracy and clinical practicality, the optimal number of features and the optimal model were screened for each liver pathology result.

[0106] For ballooning, we selected alanine aminotransferase (ALT), total bile acid (TBA), and homeostasis model assessment of insulin resistance (HOMA-IR), and established a model by Ridge regression, which had an AUROC of 0.72 (95% CI: 0.67-0.76) and a recall rate of 0.52 (95% CI: 0.48-0.56) on the training set;

[0107] For steatosis, we selected ALT and glutamate dehydrogenase (GDH) and established a model by Ridge regression, which had an AUROC of 0.74 (95% CI: 0.72-0.77) and a recall rate of 0.55 (95% CI: 0.52-0.59) on the training set;

[0108] For lobular inflammation, we selected ALT, prothrombin activity (PT%), and cadmium (Cd), and established a model by logistic regression, which had an AUROC of 0.63 (95% CI: 0.58-0.68) and a recall rate of 0.59 (95% CI: 0.54-0.65) on the training set;

[0109] For significant fibrosis, we selected peripheral blood mononuclear cell ratio (Mono.ratio), aspartate aminotransferase (AST), FPG, glycosylated albumin index (GA.index), and β2-microglobulin (Beta2), and established a model through Logistic regression. The AUROC on the training set was 0.63 (95% CI: 0.58-0.67), and the recall rate was 0.48 (95% CI: 0.43-0.53);

[0110] For NASH, we selected phospholipase A2 (PLA2), Cd, magnesium (Mg), fasting insulin (FINS), HOMA-IR, and adipose tissue insulin resistance index (Adipo-IR), and established a model through logistic regression, with an AUROC of 0.66 (95% CI: 0.59-0.63) and a recall rate of 0.51 (95% CI: 0.44-0.59) on the training set;

[0111] For NAS score, we selected ALT, cobalt (Co), triglyceride glucose body mass index (TyG-BMI), and established the model by logistic regression, which had an AUROC of 0.73 (95% CI: 0.70-0.76) and a recall rate of 0.62 (95% CI: 0.55-0.70) on the training set;

[0112] For risk NASH, we selected ALT, GA.index, and Beta2, and established a model through Logistic regression, with an AUROC of 0.69 (95% CI: 0.66-0.72) and a recall rate of 0.53 (95% CI: 0.46-0.60) on the training set.

[0113] The Logistic regression and Ridge regression models involved are constructed as follows:

[0114] For the Logistic regression model, a regularization term is added to prevent overfitting, and the loss function of the model is set as:

[0115] J(w)=-(1 / N)Σ[y i *log(p i )+(1-y i )*log(1-p i )]+λ*R(w);

[0116] Among them, J(w) is the loss function; N is the total number of samples; y i is the actual label of the i-th sample (0 or 1); p iis the probability that the i-th sample is predicted as category 1; λ is the regularization strength, which is automatically selected by cross-validation; R(w) is the regularization term, using L1 regularization;

[0117] The prediction formula of the model is:

[0118] p(y=1|x)=σ(w0+Σ(w j *x j ))

[0119] Among them: w0 is the bias term, output by the model; w j is the weight, that is, the contribution of each feature to the model, output by the model; x j is the jth component of the input feature vector, which is determined by the specific value of the indicator of each sample in the training set;

[0120] If p(y=1|x)≥0.5, the sample is predicted to be of category 1;

[0121] If p(y=1|x)<0.5, the sample is predicted to be of category 0.

[0122] For the Ridge regression model, the loss function is defined as:

[0123] L(w)=Σ(y i -(w0+Σ(w j *x ij ))) 2 +α*Σ(w j 2 )

[0124] Among them, y i is the actual label of the i-th sample; w0 is the bias term of the model; w j is the weight parameter of the jth feature; x ij is the jth eigenvalue of the i-th sample; α is the regularization strength;

[0125] The prediction formula of the model is:

[0126]

[0127] Among them, w j is the weight coefficient of the jth feature, output by the model; xi j is the jth eigenvalue of the ith sample, which is determined by the specific value of the indicator of each sample in the training set.

[0128] if Predict the category of this sample to be 1;

[0129] if The predicted class of this sample is 0.

[0130] In the test set, each clinical indicator was first standardized using the mean and standard deviation of the training set. Subsequently, we calculated the recall and AUROC of the model to evaluate the performance of the model on the test set. Specifically, the AUROC of the model in predicting ballooning, predicting ballooning, fatty degeneration, lobular inflammation, significant fibrosis, NASH, NAS score, and risk of NASH were 0.66, 0.78, 0.56, 0.73, 0.64, 0.73, and 0.76, respectively.

[0131] Comparison with existing models

[0132] The proposed model was compared with existing clinical models, including AAR, APRI, FCI, FI, FIB4, Fibroindex, Forns, GUCI, HGM1, HGM2, MODEL3, AP index, and CDS score (for predicting significant fibrosis), as well as LFS, HSI, and FLI (for predicting steatosis). For each participant, we first calculated each person's score using these models and trained them through a logistic regression model. We evaluated the model performance by comparing the AUROC in the test cohort to assess whether our model outperformed the existing models. The AUROC results showed that our model consistently outperformed the existing prediction tools in the test cohort. For example, in predicting steatosis and significant fibrosis, the maximum AUROC of the existing models was 0.61 and 0.66, respectively, while our machine learning model achieved an AUROC of 0.71 in both predictions.

[0133] Embodiment 2

[0134] A machine learning-based system for assessing multiple liver pathologies, e.g. Figure 2 As shown, including:

[0135] The data acquisition module is used to obtain the ALT, HOMA-IR, Mono.ratio test values ​​of multiple historical patients, as well as the evaluation results of liver tissue samples;

[0136] The preprocessing module is used to preprocess data based on the acquired information and construct training sets and test sets;

[0137] Model training and validation module, used to establish a Logistic regression model using the training set data and three clinical indicators: ALT, HOMA-IR, and Mono.ratio. Determine the performance of the model in predicting ballooning, fatty degeneration, lobular inflammation, significant fibrosis, and NASH; and test the model performance in internal and external test sets;

[0138] The prediction module is used to obtain the specific values ​​of the three clinical indicators ALT, HOMA-IR, and Mono.ratio, input them into the trained model, and obtain the final prediction results.

[0139] Specifically, in order to improve the clinical practicality of the model, we manually screened out three key clinical indicators, ALT, HOMA-IR, and Mono.ratio, from the seven independent models originally designed for each pathological result in Example 1, and tried to build a simplified model that can simultaneously predict multiple pathological results.

[0140] like Figure 4 A. Machine learning workflow for training and testing a logistic regression model to predict multiple liver pathological outcomes. Training ROC, test ROC, and confusion matrix for bloating (B), steatosis (C), lobular inflammation (D), significant fibrosis (E), and NASH (F). TPR: true positive rate; FPR: false positive rate.

[0141] The training set and internal test set in the second embodiment are the same as the data used in the first embodiment.

[0142] In addition, an external Chinese testing cohort was established with data from Shandong Qianfoshan Hospital and Jiangsu Provincial People's Hospital. Both cohorts also consisted of obese individuals who underwent sleeve gastrectomy for morbid obesity. Informed consent was obtained from each patient.

[0143] Data for the NHANES cohort were obtained from examinations from 2017 to March 2020. Participants with missing data were excluded from the analysis. Liver fibrosis and steatosis were diagnosed by transient elastography of the liver ultrasound, with steatosis defined as a controlled attenuation parameter (CAP) ≥ 274 and significant fibrosis as a liver stiffness measurement (LSM) ≥ 8 kPa.

[0144] To avoid the problem of collinearity, among the pathological results related to NASH, we selected NASH as the prediction target and excluded risk NASH and NAS score. Subsequently, we constructed logistic regression models for the other five pathological results based on these three indicators.

[0145] For the Logistic regression model, a regularization term is added to prevent overfitting, and the loss function of the model is set as:

[0146] J(w)=-(1 / N)Σ[y i *log(p i )+(1-y i )*log(1-p i )]+λ*R(w);

[0147] Among them, J(w) is the loss function; N is the total number of samples; y i is the actual label of the i-th sample (0 or 1); p i is the probability that the i-th sample is predicted as category 1; λ is the regularization strength, which is automatically selected by cross-validation; R(w) is the regularization term, using L1 regularization;

[0148] The prediction formula of the model is:

[0149] p(y=1|x)=σ(w0+Σ(w j *x j ))

[0150] Among them: w0 is the bias term, output by the model; w j is the weight, that is, the contribution of each feature to the model, output by the model; x j is the jth component of the input feature vector, which is determined by the specific value of the indicator of each sample in the training set;

[0151] If p(y=1|x)≥0.5, the sample is predicted to be of category 1;

[0152] If p(y=1|x)<0.5, the sample is predicted to be of category 0.

[0153] In the training set, the AUROC of the simplified model for predicting ballooning, steatosis, lobular inflammation, significant fibrosis, and NASH were 0.71, 0.75, 0.56, 0.60, and 0.63, respectively.

[0154] In both the internal and external test sets, each clinical indicator was first normalized using the mean and standard deviation of the training set.

[0155] Subsequently, the recall and AUROC of the model were calculated to evaluate the performance of the model on the test set. In the internal test set, the AUROC of the simplified model in predicting ballooning, steatosis, lobular inflammation, significant fibrosis, and NASH were 0.70, 0.66, 0.62, 0.65, and 0.71, respectively.

[0156] In the external test set, the model also shows similar performance to the training set and the internal test set. Figure 5As shown, the ROC curves and confusion matrix of the Qianfoshan cohort in China (sample number = 217) show the diagnostic performance of the simplified model in identifying steatosis (A), lobular inflammation (B), significant fibrosis (C), and NASH (D). The ROC-AUCs and confusion matrix of the Jiangsu cohort (sample number = 103) show the performance of the model in identifying ballooning (E), steatosis (F), lobular inflammation (G), significant fibrosis (H), and NASH (I).

[0157] For the test cohort of Shandong Qianfoshan Hospital, the AUROC of the simplified model in predicting steatosis, lobular inflammation, significant fibrosis, and NASH were 0.86, 0.67, 0.76, and 0.66, respectively.

[0158] For the test cohort of Jiangsu Provincial People's Hospital, the AUROC of the simplified model in predicting ballooning, fatty degeneration, lobular inflammation, significant fibrosis, and NASH were 0.84, 0.79, 0.54, 0.76, and 0.64, respectively. The above results verify the effectiveness and wide applicability of the machine learning model we established based on ALT, HOMA-IR, and Mono.ratio.

[0159] like Figure 6 Shown, A Schematic representation of the criteria used to identify individuals with steatosis (controlled attenuation parameter (CAP) ≥ 274) and marked fibrosis (liver stiffness measurement (LSM) ≥ 8 kPa) in the NHANES cohort. ROC curves and confusion matrices show the diagnostic performance of the combined model for steatosis (B) and marked fibrosis (C) in 1794 individuals serologically confirmed to be free of hepatitis virus infection. ROC curves and confusion matrices show the diagnostic performance of the combined model for steatosis (D) and marked fibrosis (E) in the NHANES cohort of 4299 individuals with no missing data. TPR: true positive rate; FPR: false positive rate.

[0160] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principle of the present invention without creative labor shall be included in the protection scope of the present invention.

Claims

1. A method for assessing multiple liver pathologies based on machine learning, characterized in that: The following steps are involved: Obtain weight information, clinical examination information, and pathological evaluation results of liver tissue samples from multiple historical patients; Based on the information obtained, data preprocessing is performed to construct training and test sets, and key predictive features of each pathological indicator are determined; Use the training set to train the machine learning model, determine the model's performance in predicting ballooning, fatty degeneration, lobular inflammation, significant fibrosis, NASH, NAS score, and risk of NASH, and determine the final predictive index parameters and prediction model; The patient's weight information, clinical examination information and liver tissue sample evaluation results are obtained, and the relevant values ​​of the prediction index parameters are extracted. The corresponding prediction index parameters are processed with the determined trained prediction model to obtain the final prediction result.

2. The method for assessing multiple liver pathologies based on machine learning as claimed in claim 1, characterized in that: The weight information includes BMI data; the clinical examination information includes insulin resistance index, fasting blood sugar or glycosylated hemoglobin and triglycerides; Pathological evaluation results of the liver tissue samples included ballooning, steatosis, lobular inflammation, significant fibrosis, NASH, NAS score, and risk of NASH.

3. The method for assessing multiple liver pathologies based on machine learning as claimed in claim 1, characterized in that: The process of data preprocessing includes: In the training set, the original data is z-score transformed according to the mean and standard deviation of each indicator to achieve data standardization. The formula for z-score transformation is: z=(x-μ) / σ; Where z is the z-score value of each data, that is, the standardized value; x is the original value of the data; μ is the mean of each indicator in the original data; σ is the standard deviation of each indicator in the original data.

4. The method for assessing multiple liver pathologies based on machine learning as claimed in claim 1, characterized in that: During the preprocessing of the test set data, each clinical indicator was standardized using the mean and standard deviation of the training set, that is, the mean of the training set was subtracted and divided by the standard deviation of the training set as the standardized value to achieve standardization of the test data; During data preprocessing, samples with each pathological outcome were marked as 1, and samples without each outcome were marked as 0 to be input into the machine learning model; In the training set, the best predictive features for each liver pathological outcome were screened by recursive feature elimination.

5. The method for assessing multiple liver pathologies based on machine learning as claimed in claim 1, characterized in that: For the Logistic regression model, a regularization term is added to prevent overfitting, and the loss function of the model is set as: J(w)=-(1 / N)Σ[y i *log(p i )+(1-y i )*log(1-p i )]+λ*R(w); Among them, J(w) is the loss function; N is the total number of samples; y i is the actual label of the i-th sample (0 or 1); p i is the probability that the i-th sample is predicted as category 1; λ is the regularization strength, which is automatically selected by cross-validation; R(w) is the regularization term, using L1 regularization; The prediction formula of the model is: p(y=1|x)=σ(w0+Σ(w j *x j )) Among them: w0 is the bias term, output by the model; w j is the weight, that is, the contribution of each feature to the model, output by the model; x j is the jth component of the input feature vector, which is determined by the specific value of the indicator of each sample in the training set; If p(y=1|x)≥0.5, the sample is predicted to be of category 1; If p(y=1|x)<0.5, the sample is predicted to be of category 0.

6. The method for assessing multiple liver pathologies based on machine learning as claimed in claim 1, characterized in that: For the Ridge regression model, the loss function is defined as: L(in)=Σ(y i -(w0+Σ(w j *x ij ))) 2 +α*Σ(in j 2 ) Among them, y i is the actual label of the i-th sample; w0 is the bias term of the model; w j is the weight parameter of the jth feature; x ij is the jth eigenvalue of the i-th sample; α is the regularization strength; The prediction formula of the model is: Among them, w j is the weight coefficient of the jth feature, output by the model; x ij is the jth eigenvalue of the ith sample, which is determined by the specific value of the indicator of each sample in the training set. if Predict the category of this sample to be 1; if The predicted class of this sample is 0.

7. The method for assessing multiple liver pathologies based on machine learning as claimed in claim 1, characterized in that: The machine learning models of ballooning, fatty degeneration, lobular inflammation, significant fibrosis, NASH, NAS score, and risk NASH were established in sequence, and the model's predicted category for each sample was output, i.e., 0 or 1; Under repeated stratified 5-fold cross validation, the balanced accuracy, precision, recall, F1 score and AUROC of each model were calculated; The calculation method of each indicator is: Balanced accuracy = (1 / 2)*[(TP / (TP+FN))+(TN / (TN+FP))]; Precision = TP / (TP+FP); Recall = TP / (TP+FN); F1 score = 2*(precision*recall) / (precision+recall); AUROC = ∫(TPR versus FPR curve); Among them, TP is a true positive example, TN is a true negative example, FN is a false negative example, and FP is a false positive example.

8. The method for assessing multiple liver pathologies based on machine learning as claimed in claim 1, characterized in that: Predictors of the final choice included: For ballooning, alanine aminotransferase, total bile acid, and homeostasis model assessment of insulin resistance were selected as predictors, and the Ridge regression model was selected; For steatosis, alanine aminotransferase and glutamate dehydrogenase were selected as predictors, and the Ridge regression model was selected; For lobular inflammation, alanine aminotransferase, prothrombin activity, and cadmium were selected as predictors, and the logistic regression model was selected; For significant fibrosis, peripheral blood mononuclear cell ratio, aspartate aminotransferase, FPG, glycated albumin index, and β2-microglobulin were selected as predictive indicators, and the logistic regression model was selected; For NASH, phospholipase A2, Cd, magnesium, fasting insulin, HOMA-IR, and adipose tissue insulin resistance index were selected as predictors, and the logistic regression model was selected; For NAS score, ALT, cobalt, triglyceride glucose body mass index were selected as predictors, and the logistic regression model was selected; For risk NASH, ALT, GA.index, and Beta2 were selected as predictors, and the Logistic regression model was selected.

9. The method for assessing multiple liver pathologies based on machine learning as claimed in claim 8, characterized in that: From the 7 independent models for each pathological outcome, ALT, HOMA-IR and Mono.ratio were screened as key clinical indicators, and a Logistic regression model was constructed to simultaneously predict multiple liver pathological outcomes as the prediction target.

10. A system for assessing multiple liver pathologies based on machine learning, characterized in that: include: A data acquisition module, used to obtain weight information, clinical examination information and evaluation results of liver tissue samples of multiple historical patients; The preprocessing module is used to preprocess data based on the acquired information and construct training sets and test sets; Model training module, used to train Logistic regression and Ridge regression models using training sets, determine the performance of the models in predicting ballooning, fatty degeneration, lobular inflammation, significant fibrosis, and NASH, and determine the final prediction index parameters and prediction model; The prediction module is used to obtain the patient's weight information, clinical examination information and liver tissue sample evaluation results, extract the relevant values ​​of the prediction index parameters, and process the corresponding prediction index parameters with the determined trained prediction model to obtain the final prediction result.

Images