Bile acid marker combination, model and application for t2dm-mafld liver fibrosis risk stratification

Abstract

The application discloses a kind of bile acid marker combination for assisting in assessing the risk of liver fibrosis of T2DM combined with MAFLD patient, kit, model and application.The bile acid marker includes lithocholic acid, ursodeoxycholic acid, glycine cholic acid, ursodeoxycholic acid-3-sulfate, glycine ursodeoxycholic acid-3-sulfate, glycine chenodeoxycholic acid-3-sulfate, glycine deoxycholic acid-3-sulfate, chenodeoxycholic acid-3-glucoside.The application only takes 8 kinds of bile acid concentration as input, and outputs risk level after preprocessing and calling gradient boosting classifier.The discrimination efficiency of the application is better than FIB-4 index and NFS score, does not need to rely on clinical biochemical or imaging index, has high specificity and accessibility, and is suitable for early risk stratification and follow-up management of liver fibrosis.

Description

Technical Field

[0001] This invention relates to a combination of bile acid biomarkers for assisting in the assessment of liver fibrosis risk in patients with type 2 diabetes mellitus (T2DM) and metabolic-associated fatty liver disease (MAFLD). It also relates to a corresponding liver fibrosis risk prediction model (hereinafter referred to as the prediction model), a test kit, a data processing system, and their applications, belonging to the field of biomarker detection technology. Background Technology

[0002] Type 2 diabetes mellitus (T2DM) combined with metabolic-associated fatty liver disease (MAFLD) is a common metabolic comorbidity in clinical practice, and this population has a significantly increased risk of developing and progressing liver fibrosis. Early identification and risk stratification of liver fibrosis are of great significance for delaying disease progression and guiding clinical follow-up and intervention management.

[0003] Currently, non-invasive methods used clinically to assess MAFLD-related liver fibrosis mainly include serological composite scoring systems (such as the FIB-4 index and NFS score) and imaging techniques (such as transient elastography of the liver, FibroScan). Among these, the FIB-4 index and NFS score rely on conventional biochemical indicators, are convenient to test, and have low cost, making them widely used primary screening tools in clinical practice. Transient elastography of the liver uses liver stiffness (LSM) as the gold standard and has high diagnostic accuracy, but it is highly dependent on equipment, expensive, and has a high failure rate in some patients due to factors such as obesity and ascites. Furthermore, it has poor accessibility in primary healthcare institutions, making it difficult to use for routine screening and long-term follow-up in large populations.

[0004] While existing non-invasive serological scoring tools (FIB-4 index, NFS score) are widely used, they suffer from core limitations such as insufficient discriminative power and limited stratification ability. Clinical studies have confirmed that these tools have low accuracy in identifying liver fibrosis in individuals with type 2 diabetes mellitus (T2DM) and metastatic liver disease (MAFLD), and are unable to effectively distinguish between individuals at high and low risk of liver fibrosis. Furthermore, these scores rely on multiple non-homologous clinical biochemical indicators, which are scattered and easily affected by confounding factors such as inflammation, renal function, and medication, resulting in poor stability. Moreover, they cannot directly reflect specific disorders of hepatic bile acid metabolism and have weak targeting of key pathophysiological mechanisms of liver fibrosis.

[0005] Bile acids are not only the end products of cholesterol metabolism but also important signaling molecules. Their composition, binding forms, and levels of derivatization modifications can specifically reflect changes in hepatic metabolic remodeling, bile acid transport, and detoxification pathways, and are closely related to the progression of liver fibrosis. However, current technologies have not yet established a combined risk assessment model for liver fibrosis in patients with type 2 diabetes mellitus (T2DM) and metastatic liver disease (MAFLD) using only bile acid metabolites as input, nor have they developed corresponding standardized testing kits and data processing systems. As a result, the clinical value of bile acid metabolomics in assessing the risk of liver fibrosis in this population has not been fully realized.

[0006] In summary, existing non-invasive assessment techniques suffer from insufficient accuracy, poor accessibility, complex indicator dependence, and low specificity, failing to meet the needs of individuals with type 2 diabetes mellitus (T2DM) and metastatic liver fibrosis (MAFLD) for early, accurate, and convenient risk stratification of liver fibrosis. Therefore, there is an urgent need to develop an auxiliary risk assessment scheme for liver fibrosis that relies solely on specific bile acid biomarkers, possesses high diagnostic efficacy, is easy to operate, and is readily applicable. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the primary technical problem this invention aims to solve is to provide a combination of bile acid biomarkers for risk stratification of liver fibrosis in patients with type 2 diabetes mellitus (T2DM) and metastatic liver disease (MAFLD). This combination exhibits high specificity and stability, and can accurately distinguish between high-risk and low-risk subgroups of liver fibrosis.

[0008] Another technical problem to be solved by the present invention is to provide a method for constructing a liver fibrosis risk prediction model, which is used to realize the scientific screening of biomarkers and the standardized construction of prediction models, so as to ensure the stability, repeatability and accuracy of prediction models.

[0009] Another technical problem to be solved by the present invention is to provide a data processing system for assisting in the assessment of the risk of liver fibrosis in patients with type 2 diabetes mellitus (T2DM) and metastatic liver disease (MAFLD).

[0010] Another technical problem that this invention aims to solve is to provide the application of the above-mentioned biomarker combination and prediction model in liver fibrosis risk assessment and stratified management, so as to provide technical support for early warning and precise intervention of liver fibrosis in patients with T2DM and MAFLD.

[0011] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution: A bile acid biomarker assembly comprising the following eight bile acid metabolites: lithocholic acid (LCA), ursodeoxycholic acid (UDCA), glycocholic acid (GCA), ursodeoxycholic acid-3-sulfate (UDCA-3S), glycoursodeoxycholic acid-3-sulfate (GUDCA-3S), glycochenodeoxycholic acid-3-sulfate (GCDCA-3S), glycodeoxycholic acid-3-sulfate (GDCA-3S), and chenodeoxycholic acid-3-glucoside (CDCA-3Gln).

[0012] Application of reagents for detecting the above-mentioned bile acid biomarker combination in the preparation of products to assist in the assessment of liver fibrosis risk in patients with type 2 diabetes mellitus (T2DM) and metastatic liver disease (MAFLD).

[0013] A method for constructing a liver fibrosis risk prediction model includes the following steps: (1) Obtain serum or EDTA-anticoagulated plasma samples from the subjects, and quantitatively detect bile acid metabolites in the samples to obtain concentration data; (2) Preprocess the concentration data; (3) The Mann-Whitney U test was used to screen for differences between groups, L1 regularized logistic regression was used to screen features, and gradient boosting classifier, random forest, extreme random forest, logistic regression and support vector machine were used to verify the stability of multiple models and determine the above 8 bile acids as the final biomarker combination. (4) Fit a machine learning classifier within the training set, use 5-fold cross-validation to evaluate the model’s generalization ability, and determine the optimal classification threshold based on the maximum value of the Youden exponent of the ROC curve in the training set to complete the model construction.

[0014] Preferably, in step (2), the preprocessing includes missing value processing, numerical transformation, and standardization; the missing value processing includes median imputation, minimum value imputation, detection lower limit imputation, half of the detection lower limit imputation, K-nearest neighbor imputation, or an interpolation method based on the training set distribution; the numerical transformation is preferably log1p transformation; and the standardization is preferably Z-score standardization based on the training reference data.

[0015] Preferably, the Z-score standardization is as follows: For the original concentration Ci of the i-th bile acid metabolite in any subject sample, the standardized value Zi before inputting into the pre-trained machine learning classifier is calculated using the following formula: Zi = [ln(1 + Ci) - μi] / σi In the formula, μi and σi are the reference mean and reference standard deviation of the log1p transformation value of the i-th bile acid metabolite in the training reference data, respectively.

[0016] Preferably, in step (4), the machine learning classifier is one or more of gradient boosting machine, random forest, extreme random forest, support vector machine, and logistic regression.

[0017] Preferably, the construction method further includes step (5): applying the locked combination of markers, preprocessing parameters and classifier parameters to an independent validation set to verify the discriminative ability of the model.

[0018] A liver fibrosis risk prediction model obtained by the above construction method takes the concentration data of the eight bile acid metabolites as input and outputs the risk probability value or risk level indication of liver fibrosis in patients with T2DM and MAFLD.

[0019] The above-mentioned liver fibrosis risk prediction model is used to assist in assessing the risk of liver fibrosis in patients with type 2 diabetes mellitus (T2DM) and metastatic liver fibrosis (MAFLD). The liver fibrosis risk prediction model uses the concentrations of the eight bile acid metabolites as inputs.

[0020] A data processing system for assisting in the assessment of liver fibrosis risk in patients with type 2 diabetes mellitus (T2DM) and metastatic liver disease (MAFLD) includes a processor and a memory. The memory stores program instructions, which, when executed by the processor, perform the following steps: receiving concentration data of the eight bile acid metabolites, performing preprocessing in the construction method, calling the liver fibrosis risk prediction model to calculate the risk probability, and outputting a risk level prompt.

[0021] Compared with existing technologies, the bile acid biomarker combination, liver fibrosis risk prediction model, and their applications provided by this invention have the following significant technical advantages and beneficial effects: 1. The biomarkers showed highly significant differences and had a strong ability to differentiate liver fibrosis risk. This invention screened eight core biomarkers from 58 bile acids, and their expression differed significantly between the high-risk group (LSM ≥ 7.0 kPa) and the low-risk group (LSM < 7.0 kPa) of liver fibrosis. Training set results showed that the levels of glycocholic acid (GCA), glycochenodeoxycholic acid-3-sulfate (GCDCA-3S), and glycodeoxycholic acid-3-sulfate (GDCA-3S) in the high-risk group were 1.87, 2.24, and 1.91 times higher than those in the low-risk group, respectively, with all differences reaching statistically significant levels (P values ​​were 1.86 × 10⁻⁶). -5 3.66×10 -6 3.09×10 -4 The overall distribution of the eight biomarkers can effectively distinguish between high- and low-risk subgroups. PCA analysis showed that the two groups of samples could be well separated, laying a reliable foundation for accurate stratification.

[0022] 2. The predictive model exhibits excellent discriminative power, significantly outperforming existing non-invasive clinical scoring methods. This invention presents a liver fibrosis risk prediction model based on eight bile acids. The model achieved an AUC of 0.9488 on the training set and 0.7310 ± 0.1180 on the 5-fold cross-validation set. On the independent validation set, the AUC reached 0.8403, with a common control subset AUC of 0.837, significantly higher than FIB-4 (AUC = 0.605, P = 0.0312) and NFS score (AUC = 0.579, P = 0.0127). The model's optimal operating point sensitivity reached 84.6% and specificity reached 75.5%, while FIB-4's optimal sensitivity was only 61.5% and specificity 65.3%, and the NFS score's specificity was only 22.4%, indicating that this prediction model has a stronger and more balanced ability to discriminate liver fibrosis risk.

[0023] 3. Precise risk stratification, highly correlated with the severity of liver fibrosis. The stratified results of this predictive model are in high agreement with the gold standard for liver fibrosis (LSM). Stratified validation set analysis showed that the mean LSM value for the high-risk group was 7.99 ± 3.38 kPa, significantly higher than the 5.45 ± 2.04 kPa value for the low-risk group (P = 2.15 × 10⁻⁶). -4 The mean ALT level in the high-risk group was 30.74±15.56 U / L, significantly higher than that in the low-risk group (21.11±8.45 U / L, P=0.0012). Stratification results accurately reflect the risk of liver fibrosis progression, providing a reliable basis for clinical intervention.

[0024] 4. The biomarker selection process is rigorous, and the model exhibits strong stability and repeatability. This invention employs a four-level hierarchical screening strategy. After quality filtering, inter-group difference screening, LASSO regularization dimensionality reduction, and multi-model stability verification of the original 58 bile acids, eight core biomarkers were finally determined. The entire process was completed solely on the training set, with the validation set remaining completely independent. Multi-model validation results show that the AUC of Random Forest and Extreme Random Forest on the training set both reached 1.0000, the AUC of Gradient Boosting was 0.9488, the AUC of Support Vector Machine was 0.8988, and the AUC of Logistic Regression was 0.8012. High consistency in feature importance across algorithms indicates that the biomarker combination is stable, reliable, and reproducible, avoiding the risk of overfitting by a single model.

[0025] 5. The test is convenient and non-invasive, suitable for large-scale screening and long-term follow-up scenarios. This invention requires only 20 μL of serum or plasma sample and can accurately quantify eight bile acids using targeted LC-MS / MS, without relying on expensive equipment such as transient liver elastography or integrating multiple clinical biochemical indicators. The accompanying kit provides standardized pretreatment reagents, standards, internal standards, and interpretation instructions, making it simple to operate, minimally invasive, cost-effective, and with a low failure rate. It can be used for early screening, stratified management, and dynamic monitoring of liver fibrosis risk in individuals with type 2 diabetes mellitus (T2DM) and metastatic liver disease (MAFLD), and is particularly suitable for primary healthcare institutions and large-scale population follow-ups, significantly improving clinical accessibility and practicality. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the overall technical solution implementation of the present invention; Figure 2 A flowchart for the graded screening of bile acid markers provided by the present invention; Figure 3 PCA scatter plot of a T2DM-MAFLD population based on eight bile acid characteristics; Figure 4 Box plots of eight bile acids for high- and low-risk groups stratified according to a liver fibrosis risk prediction model; Figure 5 Median heatmaps of eight bile acids in the high-risk and low-risk groups; Figure 6 ROC curve for a liver fibrosis risk prediction model; Figure 7 To validate the findings, a comparison chart of the ROC curves of the liver fibrosis risk prediction model, FIB-4 index, and NFS score provided in this invention with the gold standard for liver fibrosis is presented. Detailed Implementation

[0027] The embodiments of this invention are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Experimental methods in the following embodiments, unless otherwise specified, are generally performed under conventional conditions or as recommended by the manufacturer. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of skill in the art. Furthermore, any methods and materials similar or equivalent to those described herein may be applied to the embodiments of this invention. The preferred embodiments and materials described herein are for illustrative purposes only.

[0028] Figure 1This is a flowchart illustrating the overall technical solution implementation of the present invention. The implementation process includes, in sequence: screening and grouping of the study population, collection of serum samples, quantitative detection of bile acid-targeted LCMS / MS, data preprocessing, screening of bile acid biomarkers, construction of a liver fibrosis risk prediction model, internal cross-validation of the model, independent validation set validation, risk stratification, and result output. The overall process design is rigorous and the steps are coherent, ensuring the standardization and reproducibility of biomarker screening and model construction.

[0029] Example 1: Screening and Model Construction of Bile Acids 1. Clinical cohort, subject inclusion / exclusion criteria, and ethical approval The original data for this invention comes from the Mentougou MAFLD cross-sectional survey database. This database was established by Beijing You'an Hospital, affiliated with Capital Medical University, and has been approved by the hospital's ethics committee (ethics approval number: Jingyoukelunzi

[2024] No. 104). All subjects have signed written informed consent forms.

[0030] Subject selection process: A total of 1,425 subjects were registered in the original database (1,263 valid data). The subjects were selected step by step according to the following criteria: ① Meeting the diagnosis of type 2 diabetes mellitus (T2DM) (DM=Yes in the database field, referring to the Chinese guidelines for the prevention and treatment of type 2 diabetes mellitus); ② Meeting the diagnosis of metabolic-associated fatty liver disease (MAFLD) (MAFLD Diagnosis=MAFLD, referring to the 2020 international expert consensus of Eslam et al., i.e., abdominal imaging showing hepatic steatosis and at least one of overweight / obesity, T2DM or metabolic syndrome); ③ Completing FibroScan transient liver elastography and obtaining a valid LSM value; ④ Archived serum or plasma samples can be used for bile acid targeted LC-MS / MS quantitative detection.

[0031] Exclusion criteria: (1) Hepatitis B (HBV, n=26); (2) Hepatitis C (HCV, n=1); (3) Primary cholestatic cirrhosis or other known organic liver diseases that cause progression of liver fibrosis (n=1); (4) Insufficient sample size or missing key detection indicators, making it impossible to complete all necessary analyses. After the above screening, 289 subjects were finally included (aged 64–85 years, mean approximately 69.8 years; 97 males and 192 females).

[0032] Gold Standard Explanation: The liver stiffness measurement (LSM, unit: kPa) obtained by FibroScan (EchoSens, France) transient elastography (TE) is used as the reference gold standard for the degree of liver fibrosis. In this embodiment of the invention, LSM ≥ 7.0 kPa is defined as the high-risk group (indicating a clinically significant risk of liver fibrosis, approximately corresponding to ≥ F2 level), and LSM < 7.0 kPa is defined as the low-risk group. This cutoff value is based on the recommended criteria for diagnosing MAFLD-related liver fibrosis in the current international clinical guidelines (Castera et al., J Hepatol, 2019). It should be noted that LSM ≥ 7.0 kPa is not used as the pathological diagnostic criterion for liver fibrosis in this embodiment of the invention, but as a reference threshold for auxiliary risk stratification and clinical follow-up management; this threshold can be adjusted according to different testing equipment, testing populations, or etiological backgrounds.

[0033] The 289 subjects were stratified and randomly assigned according to the LSM gold standard to a training set of 216 subjects (49 high-risk and 167 low-risk) and a validation set of 73 subjects (17 high-risk and 56 low-risk). The randomization was performed using a fixed random seed (random_state=42) to ensure reproducibility.

[0034] 2. Sample Acquisition and Bile Acid Detection Serum or EDTA-anticoagulated plasma samples were obtained from the subjects, and bile acid metabolites in the samples were quantitatively detected using an internal standard method combined with targeted liquid chromatography-tandem mass spectrometry (LC-MS / MS). The instrument platform consisted of an AB Sciex ExionLC AD ultra-high performance liquid chromatograph connected in series with an AB Sciex QTRAP 6500+ triple quadrupole / linear ion trap mass spectrometer (ABSciex, USA). Targeted detection covered 58 bile acid metabolites, and six stable isotope-labeled internal standards were used to correct for matrix effects and instrument response fluctuations.

[0035] Sample pretreatment steps: Accurately measure 20 µL of serum or plasma sample into a 1.5 mL microcentrifuge tube, add 60 µL of protein precipitant containing mixed isotope internal standards (acetonitrile:methanol = 8:2, v / v), and vortex thoroughly to mix; place in an ultrasonic water bath and sonicate for 10 min; incubate at -20℃ for 60 min to allow the protein to precipitate fully; centrifuge at 12000 rpm and 4℃ for 10 min, and transfer all supernatant to a vial for instrumental analysis.

[0036] Reagents and standards: 58 bile acid standards and 6 stable isotope internal standards were purchased from Shanghai Zhenzhun Biotechnology Co., Ltd.; LC-MS grade methanol, acetonitrile, and formic acid were purchased from Thermo Fisher Scientific (USA); ammonium acetate (analytical grade). Standards were accurately weighed and dissolved in methanol to prepare standard stock solutions (approximately 1 mg / mL). Before use, a series of standard working solutions were prepared by serial dilution. Isotope internal standard solutions were prepared using acetonitrile:methanol (8:2, v / v). All standard and internal standard solutions were stored at -20°C to avoid repeated freeze-thaw cycles.

[0037] Chromatographic conditions can be method-validated targeted LC-MS / MS bile acid detection conditions. For example, the column can be a Waters ACQUITY UPLC BEH C18 column (size: 2.1 mm × 100 mm, particle size 1.7 µm, Waters, USA); column temperature 50 °C; injection volume 2 µL; flow rate 0.35 mL / min; mobile phase A can be 0.1% formic acid + 5 mM ammonium acetate aqueous solution, and mobile phase B can be acetonitrile (LC-MS grade). The gradient elution program can be optimized and determined based on the instrument platform, column, mobile phase system, and target analyte retention behavior.

[0038] Mass spectrometry conditions can be achieved using electrospray ionization (ESI) negative ion mode, with data acquired via multiple reaction monitoring (MRM) scanning. Example parameters include: ion source temperature 550℃; ion spray voltage (IS) -4500 V; curtain gas (CUR) 30 psi; nebulizer gas (GS1) 60 psi; auxiliary heating gas (GS2) 60 psi. MRM parameters such as retention time, precursor / daughter ion pairs, declustering voltage, and collision energy for the eight target bile acid metabolites can be optimized using conventional methods based on the instrument platform, column, mobile phase system, and corresponding standards.

[0039] The above chromatographic and mass spectrometric conditions are only optional implementation methods and are not intended to limit the scope of protection of the present invention; those skilled in the art can perform routine optimization of detection parameters such as chromatographic gradient, ion pair and collision energy, provided that the eight target bile acid metabolites are accurately quantified.

[0040] Quality Control Standards: A standard calibration curve should be prepared using a blank matrix (mixed serum treated with activated charcoal) that does not contain the target bile acids. At least six concentration gradient points (including the lower limit of quantitation (LLOQ) point) should be set, and quantification should be performed using the internal standard method (IS method). Quality control samples at three concentration levels (Low, Mid, and High) should be included with each analytical batch. Intra-batch and inter-batch accuracy bias and precision (CV) should both be ≤15%. Analytical results can only be reported as quantifiable when the signal-to-noise ratio (S / N) is ≥10. Measurements below the LLOQ should be recorded as missing and imputed according to the above method.

[0041] 3. Data Preprocessing The bile acid detection data are numerically processed. Missing values ​​may be caused by reasons such as being below the detection limit, peak identification failure, or signal interference; data imputation algorithms can be used for processing, including but not limited to median imputation, minimum imputation, lower detection limit imputation, half of the lower detection limit imputation, K-nearest neighbor imputation, or imputation methods based on the training set distribution. In one embodiment of the present invention, bile acid variables with a missing ratio exceeding 20% ​​are removed, and bile acid variables with a missing ratio not exceeding 20% ​​are imputed by the lower detection limit (i.e., missing measurements below the lower limit of quantitation are imputed by the lower detection limit of the metabolite). Subsequently, the concentration of each bile acid is transformed by log1p and then Z-score standardized. Specifically, for the original concentration Ci of the i-th bile acid metabolite in any subject sample, the standardized value Zi before inputting into the pre-trained machine learning classifier is calculated by the following formula (1).

[0042] Zi=[ln(1+Ci)-μi] / σi (1) In the formula, μi and σi are the reference mean and reference standard deviation of the log1p transformation value of the i-th bile acid metabolite in the training reference data, respectively. Their specific values ​​are shown in Table 2.

[0043] 4. Screening and Model Construction of Bile Acid Biomarkers In one embodiment of the present invention, the liver stiffness value (LSM) measured by transient elastography of the liver is used as the gold standard for risk reference, where LSM ≥ 7.0 kPa is the high-value group and LSM < 7.0 kPa is the low-value group. The 289 subjects included in the analysis were stratified and randomly divided into a training set of 216 cases and a validation set of 73 cases according to the gold standard. In the training set, there were 49 positive and 167 negative cases, and in the validation set, there were 17 positive and 56 negative cases. The aforementioned eight bile acids were used as fixed feature combinations. Missing value imputation parameters, log1p-Z standardization parameters, gradient boosting classifier fitting, and ROC-Youden threshold were determined only in the training set, and internal validation was performed using 5-fold cross-validation in the training set. Subsequently, the locked parameters were directly applied to the validation set for validation.

[0044] In one embodiment of the invention, the "training set" (n=216) is a sample set used for bile acid biomarker screening, model training, and threshold determination; the "validation set" (n=73) is an independent sample set that does not participate in any biomarker screening, parameter determination, or threshold adjustment, and is used only to validate the efficacy of a fixed biomarker combination. The validation set is derived from a stratified random partition of the same study population and is considered an internal validation set.

[0045] Table 1. Grading and Screening Process and Technical Role of Bile Acid Biomarkers This invention employs a rigorous four-level screening strategy to determine the final eight core bile acid biomarkers from the initial targeted metabolites. All screening steps are performed solely on the training set (n=216), with the validation set remaining completely independent and not participating in any screening or parameter determination steps. The specific screening process and technical functions are shown in Table 1 and... Figure 2 As shown.

[0046] In one embodiment of the present invention, the selected bile acid feature subset is: LCA, UDCA, GCA, UDCA-3S, GUDCA-3S, GCDCA-3S, GDCA-3S, CDCA-3Gln.

[0047] Table 2. Example preprocessing parameters and model-related indices for 8 bile acid metabolites in the training set. Note: The L1 coefficients in Table 2 are multivariate regression coefficients obtained by jointly fitting the eight markers using L1 regularized logistic regression (solver = liblinear, C = 1.0), representing the independent discriminative contribution of each marker after controlling for the remaining markers. Their signs may not be completely consistent with the correlation between the markers and the univariate distribution trend (see Table 4); feature importance is the relative importance of the gradient boost classifier output.

[0048] The screening results are as follows Figure 3 As shown, this is a PCA scatter plot of the T2DM combined with MAFLD population based on eight bile acid characteristics. Hollow circles represent the low-risk group (n=167), and solid triangles represent the high-risk group (n=49). Principal component 1 (PC1) explains 34.5%, and principal component 2 (PC2) explains 18.7%. The results show a clear separation trend between the high- and low-risk groups in the PCA space, indicating that the eight bile acid characteristics can effectively distinguish between high- and low-risk individuals with liver fibrosis and have good stratification potential.

[0049] Figure 4Box plots showing the levels of eight bile acids in high- and low-risk groups stratified according to a liver fibrosis risk prediction model (vertical axis represents the log-transformed concentration). Results showed that the expression levels of all eight bile acids were higher in the high-risk group than in the low-risk group; among them, GCA, GCDCA-3S, and GDCA-3S showed the most significant differences, consistent with the intergroup statistical test results, further validating the core value of these biomarkers in liver fibrosis risk assessment.

[0050] Figure 5 The heatmap shows the median expression levels of eight bile acids in the high- and low-risk groups, with darker colors indicating higher expression levels. Results showed that the overall expression levels of the eight bile acids were significantly higher in the high-risk group than in the low-risk group. GCA, GCDCA-3S, and GDCA-3S exhibited high expression characteristics, consistent with the box plot and statistical analysis results, directly reflecting the specific disordered pattern of bile acid metabolism in high-risk individuals with liver fibrosis.

[0051] 5. Example Classifier and Output Rules In one embodiment of the present invention, the pre-trained machine learning classifier is a gradient boosting classifier, with example hyperparameters of: n_estimators = 50, max_depth = 2, learning_rate = 0.08, min_samples_split = 8, subsample = 0.6, and min_samples_leaf = 3. These hyperparameters are merely specific implementations of this embodiment and are not intended to limit the scope of protection of the present invention.

[0052] In one embodiment of the present invention, the threshold for the risk probability is T = 0.249. When the risk probability p reaches or exceeds this threshold, a high-risk warning is output; when p is below the threshold, a low-risk warning is output. The threshold is determined by the ROC-Youden index in the training set and is directly used in the validation set.

[0053] The Youden index (J) is calculated as follows: J = Sensitivity + Specificity - 1.

[0054] In the ROC curve of the training set, the model prediction probability values ​​of all training samples are used as candidate probability thresholds. The sensitivity (TP / (TP+FN)) and specificity (TN / (TN+FP)) corresponding to each candidate threshold are calculated one by one, and the probability value that maximizes the J value is taken as the optimal risk classification threshold T. In one embodiment of the present invention, the Youden index reaches its maximum value at T=0.249, corresponding to a training set sensitivity of 91.8% (45 / 49) and a specificity of 90.4% (151 / 167). This threshold is then directly applied to the validation set for risk stratification without any reassessment or adjustment to ensure the independence of the evaluation.

[0055] Software Environment and Data Format: The software environment for data processing and model building is Python 3.10 (Python Software Foundation). Major dependencies include: scikit-learn 1.3.x (Gradient BoostingClassifier, Logistic Regression LASSO, Stratified K-Fold, and roc_auc_score ROC analysis), pandas 2.0.x (data frame processing and file reading / writing), numpy 1.24.x (numerical matrix operations), and scipy 1.11.x (Mann-Whitney U test and statistical analysis). The model input format is a two-dimensional numerical matrix, with each row representing one subject and columns corresponding to the original concentration values ​​(unit: nmol / L) of LCA, UDCA, GCA, UDCA-3S, GUDCA-3S, GCDCA-3S, GDCA-3S, and CDCA-3Gln. CSV (.csv) or Excel (.xlsx) formats are supported. Model output format: After reading the original concentration, the system sequentially performs missing value imputation → log1p transformation → Z-score standardization (parameters are shown in Table 2) → gradient boosting classifier inference, and outputs the risk probability value p (0 to 1 real number) and risk level indication for each subject (p≥T outputs "high risk", p<T outputs "low risk"); the results are saved in text report, CSV or structured JSON format.

[0056] Table 3 Performance evaluation of the prediction models on the training and validation sets. Figure 6The figure shows the ROC curves of the liver fibrosis risk prediction model on the training and validation sets. The solid line represents the ROC curve on the training set (AUC = 0.9488), and the dashed line represents the ROC curve on the validation set (AUC = 0.8403). The results indicate that the liver fibrosis risk prediction model provided by this invention exhibits excellent fitting performance on the training set and maintains high discriminative power on the independent validation set, demonstrating good generalization ability and stability.

[0057] 6. Examples of differences in bile acid distribution For the eight bile acids, a distribution comparison was performed between the high and low value groups of the LSM gold standard in the training set. This comparison is used to illustrate the consistency between bile acid characteristics and the direction of the liver fibrosis risk reference gold standard, and does not indicate that the model input includes non-bile acid indicators.

[0058] Table 4. Distribution differences of 8 bile acids among the LSM gold standard groups in the training set. Note: The total number of cases in the two groups was 167 in the low-risk group and 49 in the high-risk group. Due to missing values ​​for some bile acids, the median, IQR, and P-value in each row were calculated based on non-missing observations. The actual number of cases was -LCA 163 / 48, UDCA 155 / 47, GCA 163 / 49, UDCA-3S 163 / 49, GUDCA-3S 159 / 48, GCDCA-3S 167 / 49, GDCA-3S 139 / 42, and CDCA-3Gln 157 / 46 (low-risk group / high-risk group). The Mann-Whitney U test (two-tailed) was used for comparisons between groups.

[0059] To illustrate the clinical characteristics of the population after risk stratification in the validation set, the primary clinical baseline was described for the low-risk and high-risk groups obtained by stratifying the 73 subjects in the validation set according to the training set threshold, without changing the training set parameters, fixing the 8 bile acid combinations, fixing the example classifier hyperparameters and fixing the threshold.

[0060] It should be noted that the clinical variables listed in Table 5 are only used to describe the clinical characteristics of the high-risk and low-risk groups after stratification according to the output of the prediction model in the validation set. They are not used as input variables for the liver fibrosis risk prediction model provided by this invention, nor are they involved in feature screening, model training or threshold determination.

[0061] Table 5. Clinical baseline characteristics of the validation set divided into high-risk and low-risk groups according to the prediction model. Note: Percentages for each count variable are calculated with the number of non-missing cases as the denominator. In the low-risk group, there was 1 case of missing data each for hypertension, coronary heart disease, and lipid-lowering treatment (effective n = 53); and 2 cases of missing data for antihypertensive treatment (effective n = 52). In the high-risk group, there were no missing data for any variable (n = 19).

[0062] As shown in Table 5, in the validation set, the high-risk group, as determined by the prediction model, generally exhibited higher levels of liver fibrosis-related indicators such as LSM and ALT; specifically, the LSM levels in the low-risk group and the high-risk group were 5.45±2.04 and 7.99±3.38, respectively. The values ​​were 21.11±8.45 and 30.74±15.56, respectively, P=0.0012, suggesting that this stratification is consistent with the direction of more unfavorable clinical phenotypes.

[0063] 7. Head-to-head comparison of the validation set with existing non-invasive liver fibrosis scoring tools, FIB-4 index and NFS score. Based on the model construction and 5-fold cross-validation completed on the aforementioned training set, in order to further illustrate the discriminative advantage of the bile acid risk probability p output by the data processing system provided by this invention in real-world validation scenarios compared to existing clinically commonly used non-invasive liver fibrosis scoring tools, this embodiment of the invention conducted a supplementary head-to-head comparison between the liver fibrosis risk prediction model provided by this invention and the FIB-4 index and NFS score on the same validation set.

[0064] It should be noted that the liver fibrosis risk prediction model provided by this invention was only trained and threshold determined once on the training set (n=216), and no retraining, rescreening or threshold re-evaluation was performed on the validation set; the AUC of the training set was 0.9488, the mean AUC of the 5-fold cross-validation of the training set was 0.7310±0.1180, and the AUC of the validation set was 0.8403 (see Table 3).

[0065] The specific evaluation subjects were a subset of 62 cases in the validation set where the liver fibrosis risk prediction model, FIB-4 index, and NFS score provided by this invention were all non-missing. LSM≥7.0kPa was used as the gold standard for liver fibrosis risk reference, with 13 positive cases and 49 negative cases. The DeLong test was used to compare the differences in AUC, and the Youden index was used to determine the optimal operating point of each tool in this subset. The sensitivity and specificity of the optimal point were reported.

[0066] Table 6. Comparison of the ROC discriminant ability of the liver fibrosis risk prediction model provided by this invention with the FIB-4 index and NFS score on the validation set for the gold standard of liver fibrosis. Note: The "Youden optimal threshold" for each tool in Table 6 (0.136 for the liver fibrosis risk prediction model provided by this invention) was recalculated within a common control subset (n=62) and is only for horizontal comparison of the optimal operating points of each tool. It does not represent the actual application threshold of the liver fibrosis risk prediction model provided by this invention. The application threshold of the liver fibrosis risk prediction model provided by this invention is T=0.249, determined by the ROC-Youden index of the training set (n=216) and fixed in the validation set without any re-estimation.

[0067] As shown in Table 6, the AUC (0.837) of the liver fibrosis risk prediction model provided by this invention on the validation subset is higher than that of the FIB-4 index (0.605, DeLong P = 0.0312) and the NFS score (0.579, DeLong P = 0.0127). Its optimal operating point has both good sensitivity and specificity, indicating that the prediction model still maintains the ability to assist in the discrimination of liver fibrosis risk on the validation set.

[0068] like Figure 7 As shown, the ROC curve of the liver fibrosis risk prediction model provided by this invention is generally above the FIB-4 index and NFS score; the corresponding AUC differences are 0.232 and 0.258, respectively.

[0069] 8. Combined application of bile acid markers and systemic protection This invention provides a bile acid detection kit for assisting in the assessment of liver fibrosis risk in patients with type 2 diabetes mellitus (T2DM) and metastatic liver disease (MAFLD), comprising at least the following steps: (1) Detection reagents, standards, internal standards and / or quality control products for the quantitative detection of eight bile acid metabolites, including LCA, UDCA, GCA, UDCA-3S, GUDCA-3S, GCDCA-3S, GDCA-3S and CDCA-3Gln; (2) Sample pretreatment reagents for processing serum or plasma samples; (3) Record the instruction manual, parameter file or data processing module for the above 8 bile acid data preprocessing parameters, pre-trained classifier parameters, risk thresholds and interpretation rules.

[0070] This invention also provides the application of the above-mentioned eight bile acid metabolites in the preparation of products for the auxiliary assessment of liver fibrosis risk in patients with type 2 diabetes mellitus (T2DM) and metastatic liver disease (MAFLD). The products include diagnostic reagents, diagnostic kits, diagnostic systems, data processing systems, computer-readable storage media, local terminal software, cloud analysis modules, or combinations thereof.

[0071] In addition, embodiments of the present invention provide a method for using a liver fibrosis risk prediction model based on the above-mentioned eight bile acid combined biomarkers, comprising at least the following steps: (1) Obtain the concentration data of the above 8 bile acid metabolites in the serum or plasma of the subjects; (2) Perform data preprocessing on the concentration data, including missing value imputation, log1p transformation and Z-score standardization; (3) Input the standardized 8-dimensional feature vector into the pre-trained machine learning classifier, select the gradient boosting classifier GBM model for inference, and output the risk probability value p; (4) Generate risk level prompts based on preset threshold T; (5) The model parameters are determined only in the training set and are applied directly without adjustment in the independent validation set.

[0072] The present invention also provides a data processing system for assisting in the assessment of liver fibrosis risk in patients with type 2 diabetes mellitus (T2DM) and metastatic liver disease (MAFLD), comprising a processor and a memory. The memory stores program instructions, which, when executed by the processor, perform the following steps: receiving concentration data of the above-mentioned eight bile acid metabolites, performing preprocessing in the above-mentioned construction method, calling the model to calculate the risk probability, and outputting a risk level prompt.

[0073] It should be noted that the output results of the above-mentioned test kits, products, model construction methods, and data processing systems are used to assist in risk stratification and follow-up management, and do not directly form disease diagnosis conclusions.

Claims

1. A combination of bile acid markers, characterized in that... It consists of the following eight bile acid metabolites: lithocholic acid (LCA), ursodeoxycholic acid (UDCA), glycocholic acid (GCA), ursodeoxycholic acid-3-sulfate (UDCA-3S), glycoursodeoxycholic acid-3-sulfate (GUDCA-3S), glycochenodeoxycholic acid-3-sulfate (GCDCA-3S), glycodeoxycholic acid-3-sulfate (GDCA-3S), and chenodeoxycholic acid-3-glucoside (CDCA-3Gln).

2. The use of the reagent for detecting the bile acid biomarker combination of claim 1 in the preparation of a product to aid in the assessment of liver fibrosis risk in patients with type 2 diabetes mellitus (T2DM) and metastatic liver disease (MAFLD).

3. A method for constructing a liver fibrosis risk prediction model, characterized in that... Includes the following steps: (1) Obtain serum or EDTA-anticoagulated plasma samples from the subjects, and quantitatively detect bile acid metabolites in the samples to obtain concentration data; (2) Preprocess the concentration data; (3) The Mann-Whitney U test was used to screen for differences between groups, L1 regularized logistic regression was used to screen for features, and multi-model stability verification was used to determine the eight bile acids described in claim 1 as the final biomarker combination. (4) Fit a machine learning classifier within the training set, use 5-fold cross-validation to evaluate the model’s generalization ability, and determine the optimal classification threshold based on the maximum value of the Youden exponent of the ROC curve in the training set to complete the model construction.

4. The construction method as described in claim 3, characterized in that... In step (2), the preprocessing includes missing value processing, numerical transformation, and standardization; the missing value processing includes median imputation, minimum value imputation, detection lower limit imputation, half of the detection lower limit imputation, K-nearest neighbor imputation, or imputation method based on the training set distribution; the numerical transformation is log1p transformation; the standardization is Z-score standardization based on the training reference data.

5. The construction method as described in claim 4, characterized in that: The Z-score standardization is as follows: For any subject sample, the original concentration Ci of the i-th bile acid metabolite, before being input into the pre-trained machine learning classifier, is calculated using the following formula: Zi = [ln(1 + Ci) - μi] / σi In the formula, μi and σi are the reference mean and reference standard deviation of the log1p transformation value of the i-th bile acid metabolite in the training reference data, respectively.

6. The construction method as described in claim 3, characterized in that... In step (4), the machine learning classifier is one or more of gradient boosting machine, random forest, extreme random forest, support vector machine, and logistic regression.

7. The construction method as described in claim 3, characterized in that... It also includes step (5): The locked combination of markers, preprocessing parameters, and classifier parameters are applied to an independent validation set to verify the model's discriminative ability.

8. A liver fibrosis risk prediction model obtained by the construction method according to any one of claims 3 to 7, characterized in that: The input is the concentration data of the eight bile acid metabolites described in claim 1, and the output is the risk probability value or risk level indication of liver fibrosis in patients with T2DM and MAFLD.

9. The application of the liver fibrosis risk prediction model according to claim 8 in assisting in the assessment of liver fibrosis risk in patients with type 2 diabetes mellitus (T2DM) and metastatic liver disease (MAFLD), characterized in that: The liver fibrosis risk prediction model uses the concentrations of the eight bile acid metabolites described in claim 1 as input.

10. A data processing system for assisting in the assessment of liver fibrosis risk in patients with type 2 diabetes mellitus (T2DM) and metastatic liver disease (MAFLD), characterized in that... It includes a processor and a memory, wherein the memory stores program instructions, and when the program instructions are executed by the processor, the following steps are performed: receiving the concentration data of the eight bile acid metabolites of claim 1, performing the preprocessing in the construction method of any one of claims 3 to 5, calling the liver fibrosis risk prediction model of claim 8 to calculate the risk probability, and outputting a risk level prompt.

Images