Method for constructing early HCC prediction model based on machine learning and serum miRNA

Abstract

The invention belongs to the technical field of medical diagnosis, and particularly relates to a method for constructing an early HCC prediction model based on machine learning and serum miRNA. A large-scale serum miRNA chip data set and an analysis process driven by machine learning are utilized, feature selection and stacking integrated modeling are combined in the process, a public serum miRNA chip data set is utilized to identify a biomarker of the early hepatocellular carcinoma, and a high-performance early hepatocellular carcinoma diagnosis model is developed. As a non-invasive tool, the model has strong clinical transformation potential in early HCC screening. Subsequently, a stem-loop reverse transcription combined quantitative polymerase chain reaction technology is used for quantifying a serum miRNA spectrum type, an optimal HCC detection model XGBoostPLRglm is developed, and the provided serum miRNA characteristic is an effective tool for improving early hepatocellular carcinoma detection and possibly contributes to improvement of prognosis evaluation of hepatocellular carcinoma.

Description

Technical Field

[0001] This invention belongs to the technical field of medical diagnostics, specifically relating to a method for constructing an early HCC prediction model based on machine learning and serum miRNA. Background Technology

[0002] Primary liver cancer is the sixth most common cancer worldwide and the third leading cause of cancer-related deaths, with hepatocellular carcinoma (HCC) accounting for approximately 75% to 85% of primary liver cancers. Due to the lack of effective early screening methods, 70% to 80% of HCC patients are diagnosed at an intermediate or advanced stage. For these patients, comprehensive treatment methods, primarily surgery, are ineffective, leading to low survival rates. The poor prognosis of HCC is mainly attributed to the lack of reliable early detection biomarkers; therefore, developing highly specific non-invasive biomarkers for early detection of HCC is of great significance. Serum miRNAs, a class of small non-coding RNAs, are ideal candidates due to their unique characteristics. These miRNAs are secreted into the bloodstream by malignant cells and other cells via exosomes, reflecting disease-specific molecular characteristics and exhibiting extremely high stability during long-term storage, which facilitates clinical application.

[0003] Advances in artificial intelligence have brought revolutionary breakthroughs to biomarker discovery, enabling deep analysis of complex datasets. Machine learning algorithms have shown great potential in optimizing miRNA diagnostic models. For example, Poh et al. [Development of a miRNA-Based Model for Lung Cancer Detection. Cancers (Basel). 2025; 17(6). DOI:10.3390 / cancers17060942.] developed a lung cancer detection model by applying four machine learning algorithms to analyze six serum miRNA panels, achieving an AUC value of 0.78-0.86. Li et al. [A predictive system comprising serum microRNAs and radiomics for residual retroperitoneal masses in metastatic nonseminomatous germ cell tumors. CellRep Med. 2024; 5(12):101843. DOI:10.1016 / j.xcrm.2024.101843.] constructed a predictive machine learning algorithm that combines serum miRNAs and radiomics to predict residual retroperitoneal masses in metastatic nonseminomatous germ cell tumors, showing good accuracy. Wong et al. [Prognostic MicroRNA Fingerprints Predict Recurrence of Early-Stage Hepatocellular Carcinoma Following Hepatectomy. J Cancer. 2023; 14(3):480-489. DOI:10.7150 / jca.79593.] developed the HCCSeek-23 assay for the detection of early-stage hepatocellular carcinoma. Using the random forest algorithm, the AUC was 0.79, the sensitivity was 81%, and the specificity was 83%. The construction of cancer prognostic models depends on two key factors: the screening of effective tumor markers and the development of efficient diagnostic models, both of which are crucial for improving diagnostic accuracy. Given the incomplete understanding of tumor-associated miRNAs and the large accumulation of serum miRNA expression data, traditional literature review methods used in previous studies are still insufficient in screening serum miRNA biomarkers. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide a method for constructing an early HCC prediction model based on machine learning and serum miRNA.

[0005] The technical content of this invention is as follows:

[0006] This invention provides a method for constructing an early HCC prediction model based on machine learning and serum miRNA, comprising the following steps:

[0007] 1) Collect datasets

[0008] Serum miRNA expression data were collected from the Gene-Expression Omnibus (GEO) database.

[0009] The serum miRNA datasets include GSE106817, GSE112264, GSE113468, GSE113740, GSE122497, GSE124158, GSE137140, GSE139031, GSE164174, and GSE211692.

[0010] The datasets GSE106817, GSE112264, GSE113468, GSE113740, GSE122497, GSE124158, GSE137140, GSE139031, and GSE164174 are from platform GPL21263;

[0011] The GSE211692 dataset comes from platform GPL18941;

[0012] 2) Divided into training queue and validation queue

[0013] The serum miRNA dataset from platform GPL21263 was integrated and randomly divided into training and testing queues at a ratio of 2.5:1 using the createDataPartition function in the R package "caret" for model building and validation.

[0014] The dataset from platform GPL18941 is used as an external validation queue;

[0015] 3) Select characteristic serum miRNAs

[0016] Featured serum miRNAs were screened using RFE and LASSO machine learning functions;

[0017] To improve the performance and robustness of the model, the Z-score standardized data processing was performed using the preProcess function in the same R package;

[0018] Subsequently, 37 candidate serum biomarker miRNAs were screened from 2540 serum miRNAs using the recursive feature elimination (RFE) method of the “caret” package and the least absolute shrinkage and selection operator (LASSO) analysis of the “glmnet” package.

[0019] The 37 candidate serum biomarker miRNAs are hsa-miR-4708-5p, hsa-miR-6825-5p, hsa-miR-1268b, hsa-miR-3663-3p, hsa-miR-4327, hsa-miR-3620-5p, hsa-miR-1307-5p, hsa-miR-6781-5p, hsa-miR-4707-5p, hsa-miR-4260, hsa-miR-195-5p, hsa-miR-211-3p, hsa-miR-658, hsa-miR-6794-3p, hsa-miR-663b, hsa-miR-5100, hsa-miR-6784-5p, and hsa-miR-3195. hsa-miR-6789-5p, hsa-miR-4534, hsa-miR-6875-5p, hsa-miR-6785-5p, hsa-miR-36 21. hsa-miR-6076, hsa-miR-3162-3p, hsa-miR-6845-5p, hsa-miR-221-3p, hsa-miR-6 765-5p, hsa-miR-125a-3p, hsa-miR-516a-5p, hsa-miR-6877-5p, hsa-miR-187-5p, hs a-miR-3619-3p, hsa-miR-4649-5p, hsa-miR-3122, hsa-miR-1228-5p, hsa-miR-4463;

[0020] 4) Construct HCC-related serum miRNA features using machine learning algorithms.

[0021] Thirty-seven candidate serum biomarker miRNAs were analyzed using 13 different machine learning algorithms, including Support Vector Machine (SVM), Support Vector Regression (SGBT), Multilayer Perceptron (MLP), Classification Tree (CART), C5, AdaBoost, Multivariate Analysis Regression (MARS), Adaptive Neural Network (avNNET), Neural Network (NNET), Naive Bayes (NB), Random Forest (RF), K Nearest Neighbors (KNN), and Probabilistic Logistic Regression (PLR). Thirteen basic models were constructed, and the models were evaluated based on accuracy, AUC value, sensitivity, and specificity using metrics such as residual analysis, cumulative gain, boosting plot, precision-recall curves, and ROC analysis. The optimal model (SVM) was then determined.

[0022] Furthermore, by using the generalized linear model (GLM) or weighted regression model (GBM) method, the paired combinations of 13 models were integrated to generate 156 hybrid models;

[0023] ROC evaluation was performed on 13 basic models and 156 hybrid models to obtain the best model (SVM_SGBT_glm), which showed the best cancer detection capability. This model was designated as "HCC-related serum miRNA signature".

[0024] The above model development process used the "caret" R package and employed 10-fold cross-validation to build a binary classifier, ensuring performance optimization and preventing overfitting; the model evaluation used the "DALEX" package.

[0025] 5) Clinical translation and validation of HCC diagnostic models

[0026] Through multiple rounds of primer design, a set of PCR amplification primers for 37 HCC characteristic serum miRNAs and internal reference miRNAs was established. The expression profiles of 37 HCC characteristic serum miRNAs in clinical serum samples were detected by stem-loop reverse transcription-qPCR, and a retrospective study dataset was constructed.

[0027] In this dataset, a machine learning-driven analysis process was used to build an HCC diagnostic model. The best model was selected based on AUC value, sensitivity, and specificity. The results showed that the XGBoost_PLR_glm model was the best HCC diagnostic model.

[0028] The internal reference miRNAs were selected from three internal reference miRNAs in the microarray dataset, including hsa-miR-4463, hsa-miR-2861 and hsa-miR-149-3p;

[0029] In the stem-loop reverse transcription method, the nucleotide sequences of the reverse transcription primers for the 37 HCC characteristic serum miRNAs and internal reference miRNAs are shown in the sequence listing SEQ NO.1 to 39, respectively.

[0030] In the qPCR, the nucleotide sequences of the F primers for the 37 HCC characteristic serum miRNAs and internal reference miRNAs are shown in SEQ NO. 40-78, respectively, and the nucleotide sequences of the universal primers are shown in SEQ NO. 79-81, respectively.

[0031] The universal primers for hsa-miR-4708-5p, hsa-miR-6825-5p, hsa-miR-1268b, hsa-miR-3663-3p, hsa-miR-4327, hsa-miR-3620-5p, hsa-miR-1307-5p, hsa-miR-6781-5p, hsa-miR-4707-5p, hsa-miR-4260, hsa-miR-195-5p, hsa-miR-211-3p, hsa-miR-658, hsa-miR-6794-3p, hsa-miR-663b, hsa-miR-5100, hsa-miR-6784-5p, and hsa-miR-3195 are listed in the sequence listing SEQ. The sequences shown in NO.81; universal primers for hsa-miR-6789-5p, hsa-miR-4534, hsa-miR-6875-5p, hsa-miR-6785-5p, hsa-miR-3621, hsa-miR-6076, hsa-miR-3162-3p, hsa-miR-6845-5p, hsa-miR-221-3p, hsa-miR-6765-5p, hsa-miR-125a-3p, hsa-miR-516a-5p, hsa-miR-6877-5p, hsa-miR-187-5p, hsa-miR-3619-3p, hsa-miR-4649-5p, and hsa-miR-3122 are listed in SEQ. The sequence shown in NO.82; the universal primers for hsa-miR-1228-5p, hsa-miR-4463, hsa-miR-2861, and hsa-miR-149-3p are shown in the sequence listing SEQ NO.83.

[0032] This invention also provides the application of the above-mentioned 37 candidate serum biomarker miRNAs in the preparation of early HCC prediction and diagnostic products, which include prediction and diagnostic models, the same below.

[0033] This invention also provides the application of the above-mentioned HCC-related serum miRNA characteristics in the preparation of products that differentiate early HCC from chronic hepatitis B / cirrhosis.

[0034] This invention also provides the application of the above-mentioned optimal HCC diagnostic model in the preparation of early HCC prediction and diagnostic products.

[0035] This invention also provides the application of the above method in the preparation of early HCC prediction and diagnostic models.

[0036] The beneficial effects of the present invention are as follows:

[0037] This invention presents a method for constructing an early HCC prediction model based on machine learning and serum miRNA. Utilizing a large-scale serum miRNA microarray dataset and a machine learning-driven analysis workflow that combines feature selection and stacked ensemble modeling, the method identifies biomarkers for early hepatocellular carcinoma (HCC) using public serum miRNA microarray datasets. This leads to the development of a high-performance early HCC diagnostic model, namely, HCC-related serum miRNA features, which performed excellently in a test cohort (n=384). Its reliability was further validated in an independent external cohort (n=5991), achieving an AUC of 0.984, with a sensitivity of 98.28% and a specificity of 99.2%. Notably, the model demonstrated excellent performance in early HCC detection, with a sensitivity of 99.52% and a specificity of 99.49%. These results indicate that this model has strong clinical translational potential as a non-invasive tool for early HCC screening. Subsequently, serum miRNA profiles were quantified using stem-loop reverse transcription combined with quantitative polymerase chain reaction (PCR) to establish a retrospective clinical cohort (n=100; half HCC patients and half non-cancer patients) for translational validation. The optimal HCC detection model developed using the XGBoost_PLR_glm algorithm achieved an AUC of 0.889 in the clinical test cohort, with both sensitivity and specificity at 86.7%. TNM stage I / II HCC samples accounted for more than 70% of all HCC cases in this cohort. Based on the modeling characteristics of machine learning algorithms, increasing the amount of training data in the future is expected to further improve the diagnostic accuracy of this model in clinical practice, bringing it closer to the validation results in the microarray dataset. This not only demonstrates the clinical effectiveness of the proposed early HCC detection model, but also shows that the proposed serum miRNA features are an effective tool for improving early hepatocellular carcinoma detection, potentially contributing to improved prognostic assessment of hepatocellular carcinoma and the development of personalized prevention and screening programs. It also validates the practical application value of machine learning-driven workflows. Attached Figure Description

[0038] Figure 1 This is a schematic diagram illustrating the early construction, clinical translation, and validation process of the HCC diagnostic model of this invention.

[0039] Figure 2 The results of candidate miRNA screening using LASSO and RFE algorithms are shown in the figure (A. Using the LASSO algorithm in the training queue to screen corresponding serum miRNAs according to the minimum criteria; B. Using the RFE algorithm in the training queue).

[0040] Figure 3 Characteristic description and functional annotation of serum microRNAs (miRNAs) associated with hepatocellular carcinoma (HCC);

[0041] Figure 4 A graph showing the hyperparameter settings for thirteen machine learning algorithms selected based on the optimal ROC curve;

[0042] Figure 5 A schematic diagram illustrating the development of HCC-related miRNA characteristics;

[0043] Figure 6 A graph showing the diagnostic efficacy assessment results of HCC-related serum miRNA features in different cohorts;

[0044] Figure 7 Diagnostic performance of HCC-related serum miRNA features in early HCC detection;

[0045] Figure 8 Diagnostic performance graph of a single serum miRNA biomarker in a retrospective clinical cohort;

[0046] Figure 9 This is a diagram showing the reconstruction results of the HCC detection model in a retrospective clinical cohort.

[0047] Figure 10 HCC diagnostic models were built using a machine learning-driven workflow in a retrospective clinical cohort. Detailed Implementation

[0048] The present invention will be further described in detail below through specific implementation examples and accompanying drawings. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of protection of the present invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art fall within the scope of the appended claims.

[0049] Unless otherwise specified, all raw materials and reagents used in this invention are from the conventional market.

[0050] In this invention, serum miRNA expression data from 10 datasets (GSE106817, GSE112264, GSE113468, GSE113740, GSE122497, GSE124158, GSE137140, GSE139031, GSE164174, and GSE211692) were collected from the Gene-Expression Omnibus (GEO) database. The first nine serum miRNA datasets from the same platform (GPL21263) were integrated and randomly divided into training and testing queues at a ratio of 2.5:1 for model building and validation. The dataset from platform GPL18941 (GSE211692) was used as an external validation queue. Subsequently, a machine learning-based workflow was developed to screen candidate features from 2540 serum miRNAs, and an HCC diagnostic model was built using the training queue. The testing and external validation queues were used to validate the effectiveness of the model. To further validate the model's diagnostic capability for HCC, candidate serum miRNA expression profiles were identified in clinical patients, and a retrospective clinical cohort was constructed. In this cohort, a machine learning-based workflow was again used to validate and translate the HCC detection model. The study design is as follows: Figure 1 As shown, it is divided into two stages: the first stage is the early construction of the HCC diagnostic model, and the second stage is the clinical translation and validation of the HCC diagnostic model.

[0051] The following data processing used R software (version 4.1.3) and its Bioconductor package. To assess the diagnostic potential of serum miRNA signatures, receiver operating characteristic (ROC) curves were generated, and performance metrics including AUC, sensitivity, specificity, and accuracy were calculated. Normality and homogeneity of variance were assessed using the Shapiro-Wilk and Bartlett tests, respectively. Between-group differences were analyzed using parametric methods (e.g., t-tests or one-way ANOVA) or non-parametric methods (e.g., Wilcoxon or Kruskal-Wallis tests) depending on the data distribution. Chi-square tests were used for comparisons of categorical variables. A two-sided significance threshold of P < 0.05 was applied in this study.

[0052] Example 1

[0053] A method for building early HCC prediction and diagnostic models based on machine learning

[0054] 1. Data Preprocessing and Feature Selection

[0055] The standardized matrix dataset and corresponding medical records were obtained directly from the source database, including patients clinically diagnosed with HCC and participants without cancer. Individuals who had received any cancer treatment (including surgery, radiotherapy, chemotherapy, interventional therapy, or biologics such as targeted / immunomodulatory agents) before serum sample collection were systematically excluded from the analysis. To avoid the impact of imbalanced data on the accuracy of machine learning models in the ensemble cohort, a random undersampling method was used to eliminate redundant non-tumor samples until the sample sizes of the non-cancer group and the hepatocellular carcinoma group were similar. The 'ComBat' function from the 'sva' package was used to correct for batch-to-batch variability in serum miRNA profiles.

[0056] To develop serum diagnostic biomarkers for the detection of hepatocellular carcinoma (HCC), the `createDataPartition` function from the R package `caret` was used to randomly divide the synthesis cohort into training and testing cohorts at a ratio of 2.5:1. To improve model performance and robustness, Z-score normalization was performed using the `preProcess` function from the same package. Subsequently, recursive feature elimination (RFE) from the `caret` package and least absolute shrinkage and selection operator (LASSO) analysis from the `glmnet` package were performed, such as... Figure 2 , Figure 3 As shown, among 2540 serum miRNAs, the Venn diagram reveals the overlap between 81 LASSO-recognized miRNAs and 146 RFE-selected miRNAs, of which 37 miRNAs exhibited different expression patterns in HCC and non-tumor samples (e.g., Figure 3 (As shown in A). Furthermore, a heatmap of unsupervised hierarchical clustering analysis shows the differential expression patterns of 37 candidate miRNAs between the HCC group (red) and the non-cancer group (blue). Figure 3 Principal component analysis also confirmed that these miRNAs could effectively distinguish HCC samples from control samples (B). Figure 3 (C) The study revealed distinct clustering between HCC samples (red) and control groups (blue) based on 37 miRNA features. These findings collectively demonstrate that these miRNAs are suitable for constructing predictive models, and miRNAs that showed consistency in RFE and LASSO analyses were identified as candidate serum biomarkers.

[0057] The 37 candidate serum biomarker miRNAs are hsa-miR-4708-5p, hsa-miR-6825-5p, hsa-miR-1268b, hsa-miR-3663-3p, hsa-miR-4327, hsa-miR-3620-5p, hsa-miR-1307-5p, hsa-miR-6781-5p, hsa-miR-4707-5p, hsa-miR-4260, hsa-miR-195-5p, hsa-miR-211-3p, hsa-miR-658, hsa-miR-6794-3p, hsa-miR-663b, hsa-miR-5100, hsa-miR-6784-5p, and hsa-miR-3195. hsa-miR-6789-5p, hsa-miR-4534, hsa-miR-6875-5p, hsa-miR-6785-5p, hsa-miR-36 21. hsa-miR-6076, hsa-miR-3162-3p, hsa-miR-6845-5p, hsa-miR-221-3p, hsa-miR-6 765-5p, hsa-miR-125a-3p, hsa-miR-516a-5p, hsa-miR-6877-5p, hsa-miR-187-5p, hs a-miR-3619-3p, hsa-miR-4649-5p, hsa-miR-3122, hsa-miR-1228-5p, hsa-miR-4463.

[0058] To investigate the biological significance of identified miRNAs in the pathogenesis of hepatocellular carcinoma (HCC), functional enrichment analysis was performed using mirPath v.3. Kyoto Encyclopedia of Genetics and Genomes (KEGG) pathway analysis showed that these miRNAs were significantly enriched in multiple cancer-related pathways, including signaling pathways regulating stem cell pluripotency, the Hippo signaling pathway, proteoglycans in cancer, cell adhesion molecules, cancer-related pathways, and other pathways closely related to tumorigenesis and development. Figure 3 (D in the original text). Supplemental gene ontology (GO) analysis revealed that these miRNAs have a wide range of involvement in organelles, ion binding, cellular nitrogen metabolism, and biological processes. Figure 3 (E in the text). These meta-analyses indicate that these 37 serum miRNAs are involved in multiple cellular processes that are crucial for tumorigenesis and cancer progression.

[0059] 2. Establish diagnostic features for HCC

[0060] 2.1 Constructing HCC-related serum miRNA features using machine learning algorithms

[0061] The above results confirm the diagnostic potential of candidate serum miRNAs in HCC detection. To improve detection accuracy, a machine learning framework integrating multiple algorithms was developed. After feature selection, 13 different machine learning algorithms were constructed using 37 miRNAs. According to a standardized protocol, the serum miRNA expression dataset (HCC samples and non-cancer control groups) was divided into a training set (n=906) and a test set (n=384) at a ratio of 2.5:1. Thirteen established machine learning algorithms were independently applied, including Support Vector Machine (SVM), Support Vector Regression (SGBT), Multilayer Perceptron (MLP), Classification Tree (CART), C5, AdaBoost, Multivariate Analysis Regression (MARS), Adaptive Neural Network (avNNET), Neural Network (NNET), Naive Bayes (NB), Random Forest (RF), K-Nearest Neighbors (KNN), and Probabilistic Logistic Regression (PLR), to construct a diagnostic model (results are shown in the figure). Figure 4 As shown). Through comprehensive evaluation using five machine learning metrics (residual analysis, cumulative gain, boost plot, precision-recall curve, and ROC analysis), the SVM-based model demonstrated the best performance, almost perfectly distinguishing between HCC and non-cancer samples (median AUC of 0.9997; sensitivity of 98.60%; specificity of 99.64%). Figure 5 A in the middle.

[0062] During model development, the "caret" R package was used, and a binary classification framework was established through 10-fold cross-validation to ensure performance optimization and prevent overfitting. The "DALEX" package was used for model evaluation. To improve the diagnostic accuracy and computational robustness of the models, the caretStack function was configured to integrate pairs of 13 models using either the Generalized Linear Model (GLM) or Weighted Regression Model (GBM) methods, generating 156 hybrid models. ROC evaluation of all 169 candidate models (13 base models + 156 ensemble models) revealed that the SVM_SGBT_glm ensemble model performed best, achieving unprecedented accuracy (median AUC of 0.99995). Figure 5 Among the models (B and C), the one exhibiting the best cancer detection capability was designated as "HCC-related serum miRNA profiling." Decision curve analysis showed that it was comparable to single miRNA biomarkers (AUC range 0.527–0.899). Figure 5 The comparison of D in the data shows that the integrated model provides significantly higher net clinical benefit at different decision thresholds. Figure 5The results (E) validated the effectiveness of the proposed machine learning-driven workflow in improving HCC detection capabilities and revealed the superiority of integrating miRNA features in clinical applications, with net benefits at different threshold probabilities exceeding those of individual biomarkers.

[0063] 2.2 Diagnostic performance of HCC-related serum miRNA characteristics

[0064] The diagnostic metric was quantified using the predicted probability output of a dual-classifier ensemble. Feature scores were calculated in both the training and testing queues using a "prediction" function from data preprocessing. In the training queue, the predicted scores of HCC-related serum miRNA features were evaluated. Figure 6 As shown in Figure A, the diagnostic scores of HCC patients were significantly higher than those of the non-cancer control group, with relatively small differences in distribution, highlighting the model's strong discriminative ability. Subsequently, the model was validated in three independent cohorts: a training cohort and a testing cohort (internal validation), and an external validation dataset (GSE211692). In the training cohort, the miRNA features achieved near-perfect diagnostic accuracy, with an AUC of 0.99995 (95% confidence interval [CI], 0.99990–1.000), a sensitivity of 99.15%, and a specificity of 99.64%. Figure 6 (B in the original text). Similar performance was observed in the internal test cohort, with an AUC of 0.998 (95% confidence interval [CI], 0.996–1.000), a sensitivity of 99.14%, and a specificity of 98.01%. Figure 6 (C in the original text). Notably, the model maintained strong diagnostic stability on a large external validation set (GSE211692; n=5991), achieving an AUC of 0.984 (95% CI, 0.972–0.993), with a sensitivity of 98.28% and a specificity of 99.20%. Figure 6 D). For comparative evaluation, this study used seven miRNA combinations (miR-16, miR-122, miR-21, miR-223, miR-25, miR-375, and let-7f) identified by Fan et al. [Zhou J, Yu L, Gao X, et al. Plasma microRNA panel to diagnose hepatitis B virus-related hepatocellular carcinoma. J Clin Oncol. 2011; 29(36):4781-4788. doi:10.1200 / JCO.2011.38.2697], and applied the SGBT algorithm to generate a diagnostic model on the dataset. This method yielded an AUC of 0.908 on the training set ( Figure 6 The value of E in the equation yielded an AUC of 0.901 on the external validation set. Figure 6 The sensitivity and specificity of the optimized miRNAs were 80.23% and 88.82%, respectively, similar to the results of the original study. Importantly, the optimized miRNA profiles demonstrated superior diagnostic performance compared to established miRNA combinations.

[0065] To assess potential confounding effects, this study systematically analyzed age- and sex-related differences. Correlation analysis showed no significant association between diagnostic scores and age (Pearson's r = 0.04, p = 0.52). Figure 6 (G in the text). Although gender stratification showed a slight difference in feature scores between men and women (p = 0.03, ... Figure 6 The H in the text is missing, but this observation may reflect more the statistical power of the large sample size than the biological relevance. ROC analysis confirmed the consistency of diagnostic efficacy between the sexes, with comparable sensitivity (99.37% vs. 99.47%) and specificity (99.43% vs. 99.44%) in the male and female subgroups. Figure 6 (I in 6 and J in 6). These findings collectively indicate that demographic variables do not affect the diagnostic reliability of the model.

[0066] 2.3 HCC-related serum miRNA characteristics demonstrated a strong diagnostic ability to differentiate early HCC from chronic hepatitis B / cirrhosis.

[0067] Improving the efficiency of early HCC detection is crucial for improving patient survival outcomes. This study stratified the comprehensive dataset into early (I / II) and late (III / IV) HCC subgroups for analysis. ROC curve analysis showed that serum miRNA signatures demonstrated excellent diagnostic performance in both subgroups, achieving a sensitivity of 99.52% and a specificity of 99.49% for early HCC detection. Figure 7 (A and B in the text).

[0068] To evaluate the clinical value of HCC screening in high-risk populations (such as patients with chronic hepatitis B [CHB] or cirrhosis), a dedicated cohort was established, including only patients with CHB / cirrhosis and HCC-positive individuals. Density analysis showed a clear separation in the intensity distribution of HCC-related serum miRNA signatures between the two groups, with only minimal overlap. Figure 7 (C in the original text). Subsequent ROC analysis further confirmed the model's excellent diagnostic accuracy, with an AUC of 0.999, sensitivity of 98.81%, and specificity of 95.86%, effectively distinguishing HCC from CHB / cirrhosis. Figure 7 (D in the middle).

[0069] The reliability of this feature was further confirmed through phased verification across multiple queues, with results as follows: Figure 7 As shown in E, in the training cohort, 748 out of 751 samples were correctly classified, with only two non-tumor samples and one early-stage hepatocellular carcinoma (HCC) misclassified; in the test cohort, 325 out of 327 samples were correctly predicted, including two non-tumor samples that were incorrectly classified. External validation showed ( Figure 7 The accuracy rate for identifying non-tumor HCC (F) was 99.2% (5598 / 5643), for stage I HCC it was 96.1% (148 / 154), and for advanced HCC it was 100%. These results systematically demonstrate the consistent diagnostic performance of miRNA signatures across different disease stages.

[0070] 3. Clinical serum sample collection

[0071] Serum samples were collected from patients with informed consent. The study protocol was approved by the Ethics Committee of Foshan First People's Hospital (Approval No.: 54

[2025] ). Between March 2024 and March 2025, 100 blood samples were collected from the Department of Hepatobiliary Surgery of Foshan First People's Hospital, along with corresponding clinical information from eligible HCC patients and non-cancer patients. Each participant provided 2 mL of fresh blood for routine examinations, and any remaining blood samples were then rapidly sent to the laboratory and stored at 4°C. Blood samples were centrifuged twice at 1900×g for 10 min at 4°C to separate serum. The separated serum samples were stored at -80°C for further analysis.

[0072] 4. Clinical serum miRNA qRT-PCR and hepatocellular carcinoma detection model construction. In this study, candidate serum miRNAs and three internal reference miRNAs (hsa-miR-4463, hsa-miR-2861, and hsa-miR-149-3p) were measured, selected from the above.

[0073] Expression levels of the internal reference miRNA in the microarray dataset were determined. Total miRNA was extracted from 200 μL of serum using the miRNeasy serum / plasma kit (217184) from QIAGEN (Shanghai, China), following the manufacturer's instructions. Concentrations were determined using a NanoDrop spectrophotometer from Thermo Fisher Technology (Beijing, China). Total miRNA was transcribed into corresponding cDNAs using the miRNA first-strand cDNA synthesis kit (MR101-02; Vazyme Biotechnology, Nanjing, China) combined with the stem-loop method from Vazyme Biotechnology (Nanjing, China). The concentration of cDNA was then determined by real-time PCR on a Roche LightCycler 480 from Roche (Shanghai, China) using the miRNA universal SYBR qPCR master mix (MQ101-02) from Vazyme Biotechnology (Nanjing, China). The stem-loop and specific qPCR primers used in this study were designed and tested multiple times to ensure the specificity and amplification efficiency of the amplified products. The expression of candidate serum miRNAs was analyzed using the average CT value of three internal reference miRNAs and the 2-ΔΔCT method. Using the expression profiles of candidate miRNAs in clinical samples, an HCC detection model was reconstructed using the machine learning-driven method described above.

[0074] 4.1 Patient characteristics in retrospective clinical cohorts

[0075] The above analysis confirms the feasibility of the machine learning-driven workflow and demonstrates the diagnostic efficacy of serum miRNA signatures associated with HCC. To facilitate the clinical application of these findings, stem-loop reverse transcription-qPCR was used to systematically quantify the expression profiles of 37 candidate serum miRNAs. A retrospective clinical cohort containing comprehensive clinical data was constructed with 100 participants (half HCC patients and half non-cancer controls). This cohort was prospectively divided into a clinical training group and a test group at a 7:3 ratio, while maintaining a 1:1 case-control ratio.

[0076] In the stem-loop reverse transcription method, the nucleotide sequences of the reverse transcription primers for 37 HCC characteristic serum miRNAs and internal reference miRNAs are shown in the sequence listing SEQ NO.1 to 39, respectively.

[0077] In the qPCR, the nucleotide sequences of the F primers for 37 HCC characteristic serum miRNAs and internal reference miRNAs are shown in SEQ NO. 40-78, respectively, and the nucleotide sequences of the universal primers are shown in SEQ NO. 79-81, respectively.

[0078] The universal primers for hsa-miR-4708-5p, hsa-miR-6825-5p, hsa-miR-1268b, hsa-miR-3663-3p, hsa-miR-4327, hsa-miR-3620-5p, hsa-miR-1307-5p, hsa-miR-6781-5p, hsa-miR-4707-5p, hsa-miR-4260, hsa-miR-195-5p, hsa-miR-211-3p, hsa-miR-658, hsa-miR-6794-3p, hsa-miR-663b, hsa-miR-5100, hsa-miR-6784-5p, and hsa-miR-3195 are listed in the sequence listing SEQ. The sequences shown in NO.81; universal primers for hsa-miR-6789-5p, hsa-miR-4534, hsa-miR-6875-5p, hsa-miR-6785-5p, hsa-miR-3621, hsa-miR-6076, hsa-miR-3162-3p, hsa-miR-6845-5p, hsa-miR-221-3p, hsa-miR-6765-5p, hsa-miR-125a-3p, hsa-miR-516a-5p, hsa-miR-6877-5p, hsa-miR-187-5p, hsa-miR-3619-3p, hsa-miR-4649-5p, and hsa-miR-3122 are listed in SEQ. The sequence shown in NO.82; the universal primers for hsa-miR-1228-5p, hsa-miR-4463, hsa-miR-2861, and hsa-miR-149-3p are shown in the sequence listing SEQ NO.83.

[0079] Table 1. Characteristics of participants in the centralized study data of the clinical training and testing groups.

[0080]

[0081]

[0082] Note: Due to rounding, some percentages total > 100.

[0083] As shown in Table 1, the two groups exhibited similar distributions in baseline characteristics, including age (P = 0.680), sex (P = 0.410), serum alanine aminotransferase level (P = 0.261), and alpha-fetoprotein concentration. The incidence of cirrhosis was similar in both groups (P > 0.999). For patients with hepatocellular carcinoma, key disease parameters such as tumor size (P = 0.643), multifocal presentation (P = 0.325), TNM stage (P = 0.653), Edmondson histological grade (P = 0.204), and vascular invasion status (P > 0.999) did not differ significantly between the two groups. Notably, this cohort primarily consisted of patients with early-stage hepatocellular carcinoma, with TNM stage I patients accounting for 65.7% (23 / 35) of the clinical training cohort and 66.7% (10 / 15) of the clinical testing cohort.

[0084] 4.2 Predictive performance of individual serum miRNAs in a retrospective clinical cohort

[0085] This study evaluated the diagnostic performance of 37 candidate serum miRNAs in HCC detection. The AUC values ​​of these miRNAs in the retrospective cohort ranged from 0.508 to 0.847. Figure 8 (A in the text). The results showed that, compared with the non-cancer control group, the first three miRNAs—hsa-miR-3663-3p, hsa-miR-6784-5p, and hsa-miR-3620-5p—were significantly upregulated in HCC patients. Figure 8 Among them, hsa-miR-3663-3p showed the strongest discriminative ability, with an AUC of 0.847 (95% CI, 0.772–0.925). At the optimal cutoff, its sensitivity was 76.0% and its specificity was 84.0%. The second-ranked biomarker, hsa-miR-6784-5p, also performed well, with an AUC of 0.838 (95% CI, 0.817–0.946), a sensitivity of 68.0%, and a specificity of 88.0%. The third-ranked biomarker, hsa-miR-3620-5p, had an AUC of 0.828 (95% CI, 0.795–0.935). At the optimal threshold, its sensitivity was 740% and its specificity was 84.0%. Figure 8 (C in the middle).

[0086] 4.3 Reconstructing HCC Detection Models in Retrospective Clinical Cohorts Using Machine Learning-Driven Analysis

[0087] Given the structural differences in retrospective clinical and public HCC cohort data (based on microarray detection platforms) due to methodological variations, this study utilized an established machine learning-driven workflow to reconstruct an HCC diagnostic model within the clinical cohort. This process was applied to the expression profiles of 37 candidate serum miRNAs identified in previous analyses.

[0088] According to the established plan, 13 different machine learning algorithms were independently applied to develop diagnostic features in a clinical training cohort. Three base models with the best performance were constructed using the SGBT, PLR, and NNET algorithms. Figure 9 A in the middle.

[0089] By integrating these 13 basic models in pairs using generalized linear model (GLM) or generalized Bayesian model (GBM) functions, 156 composite diagnostic features for hepatocellular carcinoma were generated. Figure 10 The three most effective ensemble models come from a combination of the XGBoost_PLR_glm, XGBoost_PLR_gbm, and XGBoost_NB_glm algorithms. Figure 9 (B) Evaluation results showed that these models demonstrated satisfactory performance in both cohorts. In the training cohort, the six leading models achieved AUC values ​​of 0.844–0.922, with sensitivity and specificity of 73.3–85.7% and 77.1–91.4%, respectively. The corresponding test cohort showed AUC values ​​of 0.827–0.893, with sensitivity (80.0–86.7%) and specificity (73.3–84.7%) demonstrating diagnostic effectiveness. Figure 9 (CH in the text). Overall, the XGBoost_PLR_glm model with the integrated algorithm exhibits the best performance, with an AUC of 0.915 in the training queue (sensitivity 82.9%, specificity 91.4%) and an AUC of 0.889 in the test queue (sensitivity 86.7%, specificity 86.7%). Figure 9 (F in the original text). These results indicate that the diagnostic capability of the machine learning model is superior to that of a single miRNA biomarker. Decision curve analysis further confirms the clinical applicability of the XGBoost_PLR_glm model, showing that its net benefit is greater than that of the single miRNA method (F in the original text). Figure 9 (I in the middle).

[0090] To validate this study, the effectiveness of serum alpha-fetoprotein (AFP) as a biomarker was evaluated in the same clinical cohort. The results showed an AUC of only 0.644. Figure 9The J in the data is significantly lower than that of machine learning models. This comprehensive analysis validates the diagnostic potential of the panel of 37 serum miRNAs and the effectiveness of the machine learning framework in developing robust HCC detection models for clinical applications.

[0091] In summary, a machine learning-based model for the accurate early detection of hepatocellular carcinoma (HCC) was developed to screen serum miRNA biomarkers and construct a diagnostic model. From an initial pool of 2540 serum miRNAs, 37 candidate miRNAs were identified in a training cohort of 906 cases using LASSO and RFE algorithms. Subsequently, 13 different machine learning algorithms were used to construct diagnostic features within the same cohort. These 13 models were then paired and integrated using generalized linear model (GLM) and generalized regression model (GBM) functions to generate 169 HCC diagnostic features. This innovative machine learning framework significantly improves the efficiency of HCC biomarker screening and algorithm application. The optimal model, representing HCC-related serum miRNA features, performed excellently in the test cohort (n=384). Its reliability was further validated in an independent external cohort (n=5991), achieving an AUC of 0.984, with a sensitivity of 98.28% and a specificity of 99.2%. Notably, the model performed exceptionally well in early HCC detection, achieving a sensitivity of 99.52% and a specificity of 99.49%. These results demonstrate its strong clinical translational potential as a non-invasive tool for early HCC screening.

[0092] Cost reduction remains a key objective in promoting the clinical and widespread application of early-stage hepatocellular carcinoma (HCC) diagnostic models. Although the proposed model is developed based on miRNA microarray data, this method has a high cost per sample, limiting its scalability for clinical application. This study aims to address this issue. To improve cost-effectiveness and further validate the model's performance in real-world clinical settings, a dataset containing the expression of 37 serum miRNAs was constructed from clinical patient samples using qRT-PCR, a cost-effective method for miRNA detection. Following the same model building workflow described previously, this method was applied to a retrospective clinical cohort. Results showed that the optimal HCC detection model developed using the XGBoost_PLR_glm algorithm achieved an AUC of 0.889 in the clinical test cohort, with both sensitivity and specificity at 86.7%. Notably, TNM stage I HCC samples accounted for 66.7% of all HCC cases in this cohort. These findings not only demonstrate the clinical effectiveness of the proposed early-stage HCC detection model but also validate the practical value of the machine learning-driven workflow implemented in this study.

[0093] Thirty-seven candidate serum miRNAs, selected from 2540 miRNAs, significantly improved the diagnostic accuracy of this early HCC detection model using LASSO and RFE machine learning algorithms. KEGG and GO enrichment analyses showed that these miRNAs were enriched in biological pathways closely related to HCC initiation and progression. Among these candidates, the three serum miRNAs with the highest diagnostic AUC values ​​in the retrospective clinical cohort were hsa-miR-3663-3p, hsa-miR-6784-5p, and hsa-miR-3620-5p. Previous studies have shown that hsa-miR-3663-3p can regulate HCC cell proliferation and serves as a biomarker for malignant pleural effusion. Furthermore, hsa-miR-6784-5p has been identified as a potential biomarker for gastric cancer, while hsa-miR-3620-5p is associated with lymph node metastasis in intramucosal gastric cancer and participates in the development of hepatocellular carcinoma as a p53-related miRNA. These findings collectively highlight the close link between identified miRNAs and tumor progression. Further investigation into the mechanisms of action of these miRNAs in hepatocellular carcinoma (HCC) may provide new insights into the molecular mechanisms of this disease. This study shows that the optimal machine learning algorithm for constructing the best HCC diagnostic model differs between microarray-based and retrospective clinical cohorts, with the microarray-based cohort significantly outperforming the retrospective clinical cohort in diagnostic performance. The effectiveness of machine learning algorithms is influenced by various factors, including the size and structural composition of the dataset, which explains why most studies employ a multi-algorithm approach to construct and select the optimal diagnostic model. The sample size of the retrospective clinical cohort is much smaller than that of the microarray-based cohort, further exacerbating the differences in dataset structure due to different detection methods. These differences directly influence the divergence in optimal algorithm selection between the two cohorts. Furthermore, this machine learning-driven workflow employs a stacked ensemble approach to enhance model robustness and predictive accuracy, a well-established strategy that improves the generalization signal by averaging random variations across different base models.

Claims

1. A method for constructing an early HCC prediction model based on machine learning and serum miRNA, characterized in that, The steps include: 1) Collect datasets Serum miRNA expression data were collected from the Gene Expression Comprehensive Database (GEO) dataset. 2) Divided into training queue and validation queue The serum miRNA dataset from platform GPL21263 was integrated and randomly divided into training and testing queues at a ratio of 2.5:1 using the createDataPartition function in the R package "caret" for model building and validation. The dataset from platform GPL18941 is used as an external validation queue; 3) Select characteristic serum miRNAs By using the recursive feature elimination RFE method of the "caret" package and the minimum absolute contraction and selection operator LASSO analysis of the "glmnet" package, 37 candidate serum biomarker miRNAs were screened from 2540 serum miRNAs. 4) Construct HCC-related serum miRNA features using machine learning algorithms. The basic models of 37 candidate serum biomarker miRNAs were obtained through 13 different machine learning algorithms; Furthermore, by using the generalized linear model (GLM) or the weighted regression model (GBM) method, the 13 basic models were integrated in pairs to generate 156 hybrid models. ROC evaluation was performed on 13 basic models and 156 mixed models to obtain the best model, which was designated as the HCC-related serum miRNA feature. 5) Clinical translation and validation of HCC diagnostic models Through multiple rounds of primer design, a set of PCR amplification primers for 37 HCC characteristic serum miRNAs and internal reference miRNAs was established. The expression profiles of the 37 HCC characteristic serum miRNAs in clinical serum samples were detected using the stem-loop reverse transcription-qPCR method, and a retrospective study dataset was constructed. This dataset uses a machine learning-driven analysis process to build HCC diagnostic models, and selects the best HCC diagnostic model based on AUC value, sensitivity, and specificity.

2. The method for constructing an early HCC prediction model based on machine learning and serum miRNA according to claim 1, characterized in that, The serum miRNA datasets include GSE106817, GSE112264, GSE113468, GSE113740, GSE122497, GSE124158, GSE137140, GSE139031, GSE164174, and GSE211692. The datasets GSE106817, GSE112264, GSE113468, GSE113740, GSE122497, GSE124158, GSE137140, GSE139031, and GSE164174 are from platform GPL21263; The GSE211692 dataset comes from platform GPL18941.

3. The method for constructing an early HCC prediction model based on machine learning and serum miRNA according to claim 1, characterized in that, The 13 different machine learning algorithms include Support Vector Machine (SVM), Support Vector Regression (SGBT), Multilayer Perceptron (MLP), Classification Tree (CART), C5, AdaBoost, Multivariate Analysis Regression (MARS), Adaptive Neural Network (avNNET), Neural Network (NNET), Naive Bayes (NB), Random Forest (RF), K-Nearest Neighbors (KNN), and Probabilistic Logistic Regression (PLR).

4. The method for constructing an early HCC prediction model based on machine learning and serum miRNA according to claim 1, characterized in that, The 37 candidate serum biomarker miRNAs are hsa-miR-4708-5p, hsa-miR-6825-5p, hsa-miR-1268b, hsa-miR-3663-3p, hsa-miR-4327, hsa-miR-3620-5p, hsa-miR-1307-5p, hsa-miR-6781-5p, hsa-miR-4707-5p, hsa-miR-4260, hsa-miR-195-5p, hsa-miR-211-3p, hsa-miR-658, hsa-miR-6794-3p, hsa-miR-663b, hsa-miR-5100, hsa-miR-6784-5p, and hsa-miR-3195. hsa-miR-6789-5p, hsa-miR-4534, hsa-miR-6875-5p, hsa-miR-6785-5p, hsa-miR-36 21. hsa-miR-6076, hsa-miR-3162-3p, hsa-miR-6845-5p, hsa-miR-221-3p, hsa-miR-6 765-5p, hsa-miR-125a-3p, hsa-miR-516a-5p, hsa-miR-6877-5p, hsa-miR-187-5p, hs a-miR-3619-3p, hsa-miR-4649-5p, hsa-miR-3122, hsa-miR-1228-5p, hsa-miR-4463.

5. The method for constructing an early HCC prediction model based on machine learning and serum miRNA according to claim 1, characterized in that, In the stem-loop reverse transcription method, the nucleotide sequences of the reverse transcription primers for 37 HCC characteristic serum miRNAs and internal reference miRNAs are shown in the sequence listing SEQ NO.1 to 39, respectively.

6. The method for constructing an early HCC prediction model based on machine learning and serum miRNA according to claim 1, characterized in that, In the qPCR, the nucleotide sequences of the F primers for 37 HCC characteristic serum miRNAs and internal reference miRNAs are shown in SEQ NO.40-78, respectively, and the nucleotide sequences of the universal primers are shown in SEQ NO.79-81, respectively.

7. The use of the 37 candidate serum biomarker miRNAs of claim 4 in the preparation of early HCC prediction and diagnostic products.

8. The use of the HCC-related serum miRNA characteristics described in claim 1 in the preparation of a product that distinguishes early HCC from chronic hepatitis B / cirrhosis.

9. The application of the optimal HCC diagnostic model of claim 1 in the preparation of early HCC prediction and diagnostic products.

10. The application of the method according to any one of claims 1-6 in the preparation of an early HCC prediction and diagnostic model.

Images