A method and system for constructing a machine learning model to assist in the diagnosis of hepatocellular carcinoma with steatohepatitis.

By screening 12 triglyceride biomarkers associated with SHHCC and constructing a lipidomics classification model, the quantitative diagnosis challenge of SHHCC was solved, achieving highly accurate and specific auxiliary diagnosis and supporting precision medicine in clinical practice.

CN123120345APending Publication Date: 2026-11-17NANJING DRUM TOWER HOSPITAL
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611208899.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-11
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Current technologies lack effective quantitative diagnostic methods to aid in the diagnosis of steatohepatitis-associated hepatocellular carcinoma (SHHCC), especially given the human differences in pathological diagnosis, and the lack of systematic application of lipidomics data to SHHCC identification.

Method used

By extracting lipidomics data from liver cancer tissue samples, 12 triglyceride biomarkers significantly associated with SHHCC were screened using LASSO and Boruta algorithms. A lipidomics classification model was constructed and combined with machine learning models such as XGBoost, support vector machine, logistic regression, Gaussian process classifier, K-nearest neighbor or random forest for auxiliary diagnosis.

Benefits of technology

It achieves highly accurate, specific, and sensitive SHHCC-assisted diagnosis, breaking through the technical bottleneck of traditional pathological diagnosis and providing reliable quantitative basis for accurate clinical diagnosis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123120345A_ABST
    Figure CN123120345A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of biomedical technology, specifically relating to a method and system for constructing a machine learning model for assisting in the diagnosis of steatohepatitis-type hepatocellular carcinoma (SHHCC). External validation in an independent patient cohort showed an AUC of up to 1.0, demonstrating high accuracy, high specificity, and high sensitivity. This invention overcomes the subjective limitations of traditional pathological diagnosis and can provide quantitative evidence for the clinical auxiliary identification and individualized decision-making in SHHCC.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of biomedical technology, specifically relating to a method and system for constructing a machine learning model to assist in the diagnosis of hepatocellular carcinoma of the steatohepatitis type. Background Technology

[0002] Steatohepatitic hepatocellular carcinoma (SHHCC) is a special subtype of liver cancer that has received widespread attention in recent years. Histopathologically, it is characterized by steatotic degeneration of tumor cells, accompanied by significant inflammatory cell infiltration, ballooning hepatocytes, Mallory-Denk body formation, and fibrosis, with some morphological features resembling non-alcoholic steatohepatitis (NASH). This subtype is closely associated with metabolic syndrome-related factors such as obesity, type 2 diabetes, and metabolic syndrome, and its incidence is gradually increasing among liver cancer patients, especially in non-viral liver cancer patients, making it of significant clinical and molecular biological research value.

[0003] Currently, the diagnosis of SHHCC primarily relies on postoperative histopathological examination. According to international consensus standards, SHHCC can only be diagnosed when more than 5% of the tumor cells in the pathological section exhibit fatty degeneration, accompanied by at least three of the following characteristics: inflammation, Mallory bodies, ballooning degeneration, and fibrosis. However, there is significant individual variation among pathologists in meeting these diagnostic criteria. A quantitative diagnostic method for SHHCC is currently lacking.

[0004] In recent years, multi-omics technologies have developed rapidly, especially lipidomics, which has shown great potential in molecular subtyping, subtype identification, risk assessment, and prognostic prediction of tumors. Lipomics can sensitively capture changes in lipids during metabolic reprogramming in tumor tissues, especially triglycerides (TAGs), which are specifically expressed in SHHCC, making it a relatively reliable and quantifiable method.

[0005] However, at present, no studies have systematically used this omics data for SHHCC identification, nor have any publicly available literature reported combinations of lipid biomarkers that can be used for the auxiliary diagnosis of SHHCC.

[0006] Therefore, there is an urgent need to discover a lipid biomarker that can be used for the auxiliary diagnosis of SHHCC, to break through the pathological dependence of existing SHHCC diagnosis, improve the ability of auxiliary judgment, and thus meet the urgent needs of the development of precision medicine for liver cancer. Summary of the Invention

[0007] This invention aims to solve the above-mentioned problems and provides an application of a triglyceride biomarker combination in the preparation of an auxiliary diagnostic kit for hepatocellular carcinoma of the steatohepatitis type.

[0008] This invention extracts lipidomics data from liver cancer tissue samples, uses the LASSO and Boruta algorithms to screen for triglyceride features significantly associated with SHHCC, constructs and validates a lipidomics classification model, and ultimately identifies 12 triglyceride biomarkers significantly associated with SHHCC. The biomarker combination described in this invention has advantages such as high accuracy and generalizability, and can be used for the auxiliary diagnosis and clinical decision support of SHHCC.

[0009] To achieve the above objectives, the present invention provides the following technical solution: The application of a reagent for detecting a combination of triglyceride biomarkers in the preparation of an auxiliary diagnostic kit for hepatocellular carcinoma of the steatohepatitis type, wherein the combination of triglyceride biomarkers includes the following 12 triglycerides: TAG 46:2-FA16:0, TAG 46:1-FA14:0, TAG 46:1-FA16:0, TAG 46:1-FA16:1, TAG 46:0-FA16:0, TAG 47:2-FA18:1, TAG48:2-FA16:0, TAG 50:4-FA18:1, TAG 50:3-FA18:1, TAG 53:1-FA18:1, TAG 54:5-FA18:1, and TAG 54:5-FA18:2.

[0010] Furthermore, the molecular formulas of the 12 triglyceride markers are as follows: TAG 46:2-FA16:0 is C 49 H 90 O6, TAG 46:1-FA14:0 is C 49 H 92 O6, TAG 46:1-FA16:0 are C 49 H 92 O6, TAG 46:1-FA16:1 are C 49 H 92 O6, TAG 46:0-FA16:0 are C 49 H 94 O6, TAG 47:2-FA18:1 is C 50 H 92 O6, TAG 48:2-FA16:0 is C 51 H 94 O6, TAG 50:4-FA18:1 is C 53 H 94 O6, TAG 50:3-FA18:1 is C 53 H 96 O6, TAG 53:1-FA18:1 is C 56 H 106O6, TAG 54:5-FA18:1 is C 57 H 100 O6, TAG 54:5-FA18:2 is C 57 H 100 O6.

[0011] Furthermore, the reagent is a reagent for detecting the triglyceride marker combination using liquid chromatography-mass spectrometry (LC-MS).

[0012] Furthermore, the triglyceride biomarker combination was obtained by screening as follows: lipidome data of liver cancer tissue samples were obtained, and feature screening was performed using LASSO regression and Boruta algorithm respectively. The intersection of the screening results of the two algorithms was taken to obtain the 12 triglyceride biomarkers.

[0013] Furthermore, the LASSO regression uses 10-fold cross-validation to determine the optimal penalty parameter λ, and the λ value corresponding to the minimum cross-validation bias is used as the optimal parameter to screen features with non-zero coefficients; the Boruta algorithm parameter maxRuns is set to 500.

[0014] Furthermore, the discrimination result of the auxiliary diagnostic kit is obtained based on a lipidomics classification model, which is selected from one or more of XGBoost, support vector machine, logistic regression, Gaussian process classifier, K-nearest neighbors or random forest.

[0015] Furthermore, the method for constructing the lipidomics classification model includes: using the quantitative detection results of the 12 triglyceride biomarkers as input features, constructing six machine learning classification models, namely XGBoost, support vector machine, logistic regression, Gaussian process classifier, K-nearest neighbor and random forest, to classify and distinguish the samples.

[0016] Specifically, the parameter settings for each machine learning model are as follows: XGBoost: Data preprocessing is L1 norm regularization, n_estimators is 80, max_depth is 16, Eta is 0.4, Gamma is 0.1, and Objective is binary:logitraw; Logistic Regression: Data preprocessing included L2 norm regularization, C = 1.5, Class Weight = None, Penalty = l2, and Tol = 1e-05; K-Nearest Neighbors: Data preprocessing is Zscore, Algorithm is auto, Leaf Size is 30, N_Neighbors is 7, P is 1, and Weights is distance; Gaussian process classifier: Data preprocessing is Zscore, N Restart Optimizer is 0, and Length Scale is 1.0; Support Vector Machine: Data preprocessing is L2 norm regularization, C is 10, Class Weight is balanced, Gamma is 0.1, and Kernel is sigmoid; Random Forest: Data preprocessing is Z-score, Class Weight is balanced, Criterion is entropy, Max Depth is 2, Min Samples Leaf is 2, Min Samples Split is 2, and N Estimators is 5.

[0017] Furthermore, the lipidomics classification model undergoes cross-validation or external validation, with the external validation samples derived from independent patient cohorts.

[0018] The present invention also provides a kit for the auxiliary diagnosis of hepatocellular carcinoma of the steatohepatitis type. The kit contains reagents for detecting a combination of triglyceride biomarkers, the combination of triglyceride biomarkers including the following 12 triglycerides: TAG46:2-FA16:0, TAG 46:1-FA14:0, TAG 46:1-FA16:0, TAG 46:1-FA16:1, TAG 46:0-FA16:0, TAG 47:2-FA18:1, TAG 48:2-FA16:0, TAG 50:4-FA18:1, TAG 50:3-FA18:1, TAG 53:1-FA18:1, TAG 54:5-FA18:1, and TAG 54:5-FA18:2.

[0019] The beneficial effects of this invention are as follows: The triglyceride biomarker combination described in this invention has an AUC of up to 1.0, exhibiting high accuracy, high specificity, and high sensitivity. This invention successfully overcomes the human differences and technical bottlenecks of traditional pathological qualitative diagnosis. This invention provides a reliable quantitative basis for accurate postoperative diagnosis and individualized decision support in the clinical practice of hepatocellular carcinoma of the steatohepatitis type, and has significant clinical application value. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below, wherein: Figure 1 For the screening of features in lipidomics models.

[0021] Figure 2 The AUC curve is for a lipidomics classification model.

[0022] Figure 3 This is a machine learning evaluation metric for lipidomics classification models. Detailed Implementation

[0023] To make the above-mentioned objectives, 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 specific examples.

[0024] Example 1

[0025] Research cohort and sample: This example includes two independent cohorts of hepatocellular carcinoma (HCC) patients: Discovery and Modeling Cohort (Cohort 1): Contains 125 liver cancer tissue samples. These were randomly divided, with 100 samples used as the training set for classifier construction and parameter optimization, and 25 samples used as the internal test set for preliminary model performance evaluation.

[0026] Independent Validation Cohort (Cohort 2): This cohort contains liver cancer tissue samples from 24 independent patients and was used to rigorously evaluate the generalization ability and clinical applicability of the lipidomics classification model.

[0027] Inclusion criteria: ① Patients with hepatocellular carcinoma that has been surgically resected and pathologically confirmed; ② Patients with complete postoperative pathological classification records (clearly classified as SHHCC or conventional HCC); ③ Patients who have not received transarterial chemoembolization (TACE), ablation, or systemic antitumor therapy before surgery.

[0028] Example 2

[0029] Lipomics feature dimensionality reduction and core TAG screening: Raw lipidome data from hepatocellular carcinoma tissues were obtained using high-throughput liquid chromatography-mass spectrometry (LC-MS). Dual feature engineering screening was performed using the glmnet (v4.1.10) and Boruta (v9.0.0) packages in the Rv4.2.3 environment. Figure 1 ): LASSO regression dimensionality reduction: 10-fold cross-validation was used to determine the optimal penalty parameter λ. The λ value (lambda.min) corresponding to the minimum cross-validation deviation was taken as the optimal solution, and 38 lipid features with non-zero coefficients were initially screened.

[0030] Boruta Random Forest Full Feature Screening: With the parameter maxRuns=500, significance testing was performed by introducing shadow features, and a total of 32 important features were identified.

[0031] By taking the intersection of the results from the LASSO and Boruta algorithms, 12 core TAG features were finally determined. Figure 1 The names and molecular formulas of the 12 core TAG features are as follows: TAG 46:2-FA16:0, C 49 H 90 O6; TAG 46:1-FA14:0, C 49 H 92 O6; TAG 46:1-FA16:0, C 49 H 92 O6; TAG 46:1-FA16:1, C 49 H 92 O6; TAG 46:0-FA16:0, C 49 H 94 O6; TAG 47:2-FA18:1, C 50 H 92 O6; TAG 48:2-FA16:0, C 51 H 94 O6; TAG 50:4-FA18:1, C 53 H 94 O6; TAG 50:3-FA18:1, C 53 H 96 O6; TAG 53:1-FA18:1, C 56 H 106 O6; TAG 54:5-FA18:1, C 57 H 100 O6; TAG 54:5-FA18:2, C 57 H 100 O6.

[0032] Example 3

[0033] Machine learning model building: Using the aforementioned 12 TAG features, this invention constructs six machine learning classification models: XGBoost, Support Vector Machine (SVM), Logistic Regression, Gaussian Process, K-Nearest Neighbors (KNN), and Random Forest. Figure 2 , Figure 3 ).

[0034] The specific model parameters are as follows: Training / test set partitioning method: five-fold cross-validation.

[0035] The parameter settings for each model are as follows: XGBoost: Data preprocessing - L1 norm regularization, n_estimators: 80, max_depth: 16, Eta: 0.4, Gamma: 0.1, Objective: binary: logitraw.

[0036] Logistic: Data preprocessing - L2 norm regularization, C: 1.5, Class Weight: None, Penalty: l2, Tol: 1e-05.

[0037] KNN: Data preprocessing - Zscore, Algorithm: auto, Leaf Size: 30, N_Neighbors: 7, P: 1, Weights: distance.

[0038] Gaussian Process: Data preprocessing - Zscore, N Restart Optimizer: 0, Length Scale: 1.0.

[0039] SVM (Support Vector Machine): Data preprocessing - L2 norm regularization, C: 10, Class Weight: balanced, Gamma: 0.1, Kernel: sigmoid.

[0040] Random Forest: Data preprocessing - Z-score, Class Weight: balanced, Criterion: entropy, Max Depth: 2, Min Samples Leaf: 2, Min Samples Split: 2, NEstimators: 5.

[0041] Example 4

[0042] Model performance evaluation: The specific quantization performance of each model on different datasets is described below: 4.1 Training Set Performance: During the modeling and training phase (n=100), due to the integration of specific lipid features, each model exhibited extremely high fitting and classification capabilities. Figure 2 , Figure 3 ): XGBoost and KNN models achieved perfect sample differentiation, with their area under the receiver operating characteristic curve (AUC) reaching as high as 1.0; their accuracy, F1-score, precision, sensitivity / recall, and specificity all reached the perfect score of 1.00.

[0043] The AUCs of the Logistic and SVM models were 0.940 and 0.942, respectively; the SVM had a sensitivity (Recall) of 0.82 and a specificity as high as 0.94; the Logistic model had a sensitivity of 0.73 and a specificity as high as 0.98.

[0044] The Gaussian process and random forest models also achieved high AUC values ​​of 0.945 and 0.925, respectively.

[0045] 4.2 Internal Test Set Validation: On the test set (n=25) with internal five-fold cross-validation, the model showed good robustness and no serious overfitting occurred. Figure 2 , Figure 3 ): The Logistic model has an AUC of 0.933, an accuracy of 0.82, a precision of 0.92, and a specificity of 0.93.

[0046] The AUCs of XGBoost and Gaussian process models remained high at 0.921 and 0.915, respectively.

[0047] Both SVM and KNN models achieved an accuracy of 0.82 and an F1 score of 0.81, respectively, with test set AUCs of 0.904 and 0.911.

[0048] 4.3 Performance of Independent External Validating Sets: To confirm the clinical generalization and scalability of the lipidomics model of this invention, blinded validation was performed in an independent validation cohort (cohort 2, n=24) where no one participated in the modeling. Figure 2 , Figure 3 Multiple mainstream machine learning models demonstrated consistently strong diagnostic performance in external validation. The Gaussian process classification model performed best, with an AUC of 0.944 for independent external validation, demonstrating a very strong combination of specificity and sensitivity.

[0049] The XGBoost classification model followed closely behind, with an external validation AUC of 0.923. In terms of various quantitative metrics, its accuracy reached 0.83, F1-score was 0.82, precision was 0.82, recall was 0.82, and specificity performed exceptionally well, reaching 0.85.

[0050] The SVM (Support Vector Machine) model exhibited excellent specificity (high diagnostic exclusion ability), with an external validation AUC of 0.909, an accuracy of 0.83, and an F1-score of 0.83. Notably, its specificity significantly increased to 0.91 and its recall was 0.77, making it highly valuable for clinical auxiliary diagnosis.

[0051] The Logistic and KNN models performed identically on the external validation set, with an AUC of 0.909. Their accuracy (0.83), F1-score (0.82), precision (0.82), sensitivity (0.82), and specificity (0.85) were all very balanced and stable.

[0052] Example 5

[0053] Preparation and application of the reagent kit: Based on the aforementioned 12 triglyceride biomarkers, a kit for the auxiliary diagnosis of SHHCC was prepared. This kit contains reagents for detecting the combination of the aforementioned triglyceride biomarkers using liquid chromatography-mass spectrometry (LC-MS / MS). The kit also includes instructions or a label specifying a discrimination threshold, which is determined based on the quantitative detection results of the triglyceride biomarker combination using the aforementioned lipidomics classification model.

[0054] The application method is as follows: 1) Obtain liver cancer tissue samples for testing; 2) Using the detection reagents in the kit, the content of the 12 triglyceride markers in the sample was detected by liquid chromatography-mass spectrometry (LC-MS). 3) Input the detection results into a pre-established lipidomics classification model, wherein the model is selected from one or more of XGBoost, support vector machine, logistic regression, Gaussian process classifier, K-nearest neighbors or random forest; 4) Based on the classification results output by the model, help determine whether the sample is SHHCC.

[0055] In summary, this invention, through lipidomics data from a large-scale cohort of liver cancer patients and combined with the LASSO and Boruta dual-feature screening algorithms, has for the first time discovered 12 core triglyceride biomarkers significantly associated with SHHCC. The lipidomics classification model constructed based on these 12 biomarkers consistently outputs highly accurate, specific, and sensitive classification results, both in internal multi-round cross-validation and in independent external patient cohorts. This successfully overcomes the human error and technical bottlenecks inherent in traditional pathological qualitative diagnosis, providing a reliable quantitative basis for accurate postoperative clinical diagnosis and individualized decision support in steatohepatitis-type hepatocellular carcinoma (SHHCC). The triglyceride biomarker combination and its detection reagents described in this invention have broad application prospects in the preparation of SHHCC auxiliary diagnostic kits.

[0056] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. The application of a reagent for detecting a combination of triglyceride markers in the preparation of an auxiliary diagnostic kit for hepatocellular carcinoma of the steatohepatitis type, characterized in that, The triglyceride biomarker combination includes 12 triglycerides: TAG 46:2-FA16:0, TAG 46:1-FA14:0, TAG 46:1-FA16:0, TAG 46:1-FA16:1, TAG 46:0-FA16:0, TAG 47:2-FA18:1, TAG 48:2-FA16:0, TAG 50:4-FA18:1, TAG 50:3-FA18:1, TAG 53:1-FA18:1, TAG54:5-FA18:1, and TAG 54:5-FA18:

2.

2. The application according to claim 1, characterized in that, The molecular formulas of the 12 triglyceride markers are as follows: TAG 46:2-FA16:0 (C 49 H 90 O6, TAG 46:1-FA14:0 is C 49 H 92 O6, TAG 46:1-FA16:0 are C 49 H 92 O6, TAG 46:1-FA16:1 are C 49 H 92 O6, TAG 46:0-FA16:0 are C 49 H 94 O6, TAG 47:2-FA18:1 is C 50 H 92 O6, TAG48:2-FA16:0 is C 51 H 94 O6, TAG 50:4-FA18:1 is C 53 H 94 O6, TAG 50:3-FA18:1 is C 53 H 96 O6, TAG 53:1-FA18:1 is C 56 H 106 O6, TAG 54:5-FA18:1 is C 57 H 100 O6, TAG 54:5-FA18:2 is C 57 H 100 O6.

3. The application according to claim 1 or 2, characterized in that, The reagent is used to detect the triglyceride marker combination by liquid chromatography-mass spectrometry.

4. The application according to claim 1 or 2, characterized in that, The triglyceride biomarker combination was obtained through screening using the following method: Lipidome data of liver cancer tissue samples were obtained, and feature screening was performed using LASSO regression and Boruta algorithm respectively. The intersection of the screening results of the two algorithms was taken to obtain the 12 triglyceride biomarkers.

5. The application according to claim 4, characterized in that, The LASSO regression uses 10-fold cross-validation to determine the optimal penalty parameter λ, and the λ value corresponding to the minimum cross-validation bias is used as the optimal parameter to screen features with non-zero coefficients; the parameter maxRuns of the Boruta algorithm is set to 500.

6. The application according to claim 1 or 2, characterized in that, The discrimination results of the auxiliary diagnostic kit are obtained based on a lipidomics classification model, which is selected from one or more of XGBoost, support vector machine, logistic regression, Gaussian process classifier, K-nearest neighbors or random forest.

7. The application according to claim 6, characterized in that, The method for constructing the lipidomics classification model includes: Using the quantitative detection results of the 12 triglyceride biomarkers as input features, six machine learning classification models were constructed, namely XGBoost, support vector machine, logistic regression, Gaussian process classifier, K-nearest neighbor and random forest, to classify and distinguish the samples.

8. The application according to claim 7, characterized in that, The parameter settings for each machine learning model are as follows: XGBoost: Data preprocessing is L1 norm regularization, n_estimators is 80, max_depth is 16, Eta is 0.4, Gamma is 0.1, and Objective is binary:logitraw; Logistic Regression: Data preprocessing included L2 norm regularization, C = 1.5, Class Weight = None, Penalty = l2, and Tol = 1e-05; K-Nearest Neighbors: Data preprocessing is Zscore, Algorithm is auto, Leaf Size is 30, N_Neighbors is 7, P is 1, and Weights is distance; Gaussian process classifier: Data preprocessing is Zscore, N Restart Optimizer is 0, and Length Scale is 1.0; Support Vector Machine: Data preprocessing is L2 norm regularization, C is 10, Class Weight is balanced, Gamma is 0.1, and Kernel is sigmoid; Random Forest: Data preprocessing is Z-score, Class Weight is balanced, Criterion is entropy, MaxDepth is 2, Min Samples Leaf is 2, Min Samples Split is 2, and N Estimators is 5.

9. The application according to claim 6, characterized in that, The lipidomics classification model has undergone cross-validation or external validation.

10. A reagent kit for the auxiliary diagnosis of hepatocellular carcinoma of the steatohepatitis type, characterized in that, The kit contains reagents for detecting a combination of triglyceride biomarkers, which includes the following 12 triglycerides: TAG46:2-FA16:0, TAG 46:1-FA14:0, TAG 46:1-FA16:0, TAG 46:1-FA16:1, TAG 46:0-FA16:0, TAG 47:2-FA18:1, TAG 48:2-FA16:0, TAG 50:4-FA18:1, TAG 50:3-FA18:1, TAG 53:1-FA18:1, TAG 54:5-FA18:1, and TAG 54:5-FA18:2.