Non-invasive prediction of significant hepatic fibrosis in individuals with chronic hepatitis C infection using fibrosis risk score and machine learning models

Abstract

BACKGROUND

Hepatitis C virus (HCV) poses a formidable global health concern. Non-invasive assessment of hepatic fibrosis has gained increasing importance as liver biopsy, the current gold standard, is less favored as a routine test due to its invasiveness, cost, and risk of complications. Although several clinical scoring systems exist, the potential contribution of lipid biomarkers has not been sufficiently explored.

AIM

To develop and validate a novel diagnostic tool for non-invasive detection of significant hepatic fibrosis in HCV patients using a composite of serum biomarkers, including lipid parameters. It also sought to establish and validate machine-learning (ML) models for predicting significant fibrosis in HCV.

METHODS

A retrospective of 316 chronic HCV patients was analyzed. Candidate predictors of significant hepatic fibrosis were screened by univariate analysis and selected using least absolute shrinkage and selection operator regression, followed by multivariate logistic regression to identify independent predictors. A novel fibrosis risk score (FRS) incorporating platelet count, liver function indices, and lipid biomarkers was developed and compared with three existing clinical scores. Six ML models were additionally developed, and diagnostic performance was assessed using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value, negative predictive value, and F1-score.

RESULTS

The newly developed FRS demonstrated excellent discriminatory performance for significant hepatic fibrosis, with scores ≥ 9 indicating higher risk and achieving an AUC of 0.83 (95% confidence interval: 0.75-0.88) in the training cohort. The model outperformed commonly used non-invasive indices, including aspartate aminotransferase-to-platelet ratio index (AUC = 0.59), fibrosis-4 (AUC = 0.65), and the gamma-glutamyl transpeptidase-to-platelet ratio (AUC = 0.67) at their respective optimal cutoff thresholds of 1.5, 1.45, and 0.4. Among the evaluated ML approaches, the random forest algorithm exhibited the highest diagnostic accuracy, yielding an AUC of 0.92 (95% confidence interval: 0.88-0.95) in the training cohort.

CONCLUSION

The FRS showed robust predictive accuracy for significant hepatic fibrosis in chronic HCV infection and offers a practical, reliable, and user-friendly non-invasive tool. Additionally, ML models substantially enhanced predictive performance. The random forest model demonstrated superior diagnostic potential among ML algorithms, supporting its applicability in fibrosis risk stratification and screening.

Key Words: Chronic hepatitis C; Non-invasive fibrosis assessment; Hepatic fibrosis; Lipid biomarkers; Platelet count; Machine learning; Random forest; AdaBoost

Core Tip: This study developed and validated a simple non-invasive fibrosis risk score for hepatitis C virus using platelet count, lipid markers, and liver function parameters. The score outperformed commonly used clinical tools for detecting significant hepatic fibrosis. Additionally, machine-learning models were evaluated, which outperformed the commonly used clinical risk scoring systems with random forest and AdaBoost demonstrating the highest diagnostic performance for detecting significant hepatic fibrosis in hepatitis C virus infection.

INTRODUCTION

Hepatitis C virus (HCV) poses a substantial global public health challenge. According to the World Health Organization (WHO)’s 2024 report, roughly 50 million people are living with chronic HCV infection globally. Each year, an estimated 1 million new cases are detected, and in 2022 alone, the disease accounted for approximately 240000 deaths, predominantly owing to its complications, including cirrhosis and hepatocellular carcinoma (HCC)[1].

HCV infection triggers a continuous spectrum of hepatic injury, ranging from fibrosis to cirrhosis and, in some cases, HCC. Hepatic fibrosis represents an aberrant reparative response to persistent hepatocellular injury induced by chronic HCV infection[2,3]. The fibrotic process is marked by the excessive accumulation of extracellular matrix components - particularly collagen - synthesized by hepatic stellate cells that have transitioned from a quiescent to an activated state[4,5]. Importantly, fibrosis is dynamic and potentially reversible when inflammation is controlled and the etiologic agent is eliminated[6,7]. Hence, fibrosis staging remains central to patient management and prognostication[7].

Liver biopsy remains the gold standard for assessing the degree of hepatic fibrosis[7,8]. However, it is fraught with certain limitations, including its invasive nature, associated procedure-related complications, high cost, and the risk of sampling error inherent in examining only a small portion of hepatic tissue[8]. Consequently, there has been a growing interest to develop alternative non-invasive tools that can accurately predict hepatic fibrosis without the need for biopsy[9]. Direct serum fibrosis markers, such as hyaluronic acid and procollagen peptides, remain costly and are not widely accessible for routine use[10,11]. In contrast, indirect biomarkers offer a more affordable and practical approach for fibrosis screening and form the basis of several commonly used clinical scoring systems - such as the aspartate aminotransferase-to-platelet ratio index (APRI), fibrosis-4 (FIB-4), and the gamma-glutamyl transpeptidase-to-platelet ratio (GPR) scores[12,13].

A substantial body of evidence indicates that serum lipid levels are significantly reduced in patients with chronic viral hepatitis, particularly HCV[14,15]. The mechanisms underlying this serum lipid reduction are multifactorial and include impaired hepatic synthetic function, activation of pro-inflammatory and immune-mediated pathways, and direct viral interference with lipid metabolism[16,17]. Despite this biological plausibility, lipid parameters remain underutilized in fibrosis assessment, and their adjunctive role as non-invasive biomarkers is inadequately explored[18,19].

Non-invasive imaging modalities, including shear-wave elastography, transient elastography (FibroScan), and magnetic resonance elastography demonstrate high diagnostic performance[20-22]. Among these, transient elastography has emerged as the most widely used and well-validated tool for evaluating hepatic fibrosis. Transient elastography is a non-invasive, simple, and easy-to-perform imaging technique[23,24]. It has demonstrated high diagnostic accuracy in viral hepatitis, with meta-analysis studies reporting a high accuracy for detecting significant fibrosis (Meta-analysis of Histological Data in Viral Hepatitis ≥ F2) showing an area under the receiver operating characteristic curve (AUC) of 0.89. Its performance is even stronger for diagnosing cirrhosis (Meta-analysis of Histological Data in Viral Hepatitis F3-F4), with an AUC of around 0.94[25,26].

In this context, the present study aimed to integrate demographic data, hematological indices, liver function parameters and lipid profile indices to develop a reliable non-invasive tool for predicting significant hepatic fibrosis in chronic HCV patients. Additionally, the study assessed the added diagnostic value of machine-learning (ML) models in identifying significant hepatic fibrosis in chronic HCV patients.

MATERIALS AND METHODS

This retrospective study comprised 316 patients with confirmed chronic HCV infection, defined by the presence of anti-HCV antibodies and a detectable HCV RNA assay for ≥ 6 months. Adults aged 18 years or older with chronic HCV infection with comprehensive clinical and laboratory records - including complete blood counts, liver function tests (LFT), lipid profiles, and documented transient elastography (FibroScan) assessments - were eligible for inclusion. Patients were excluded if they had hepatitis B co-infection or other etiologies of chronic liver disease, such as autoimmune hepatitis, metabolic-associated fatty liver disease, primary biliary cholangitis, primary sclerosing cholangitis, drug-induced liver injury, alcoholic liver disease, decompensated cirrhosis, or HCC. Patients with conditions that alter lipid metabolism, including those undergoing treatment with hypolipidemic drugs or pregnant individuals, were excluded from the study.

The study protocol was approved by the Institutional Review Board, and the requirement for written informed consent was waived owing to the retrospective nature of the study.

Transient elastography (FibroScan) was performed in strict accordance with established protocols, ensuring optimal measurement accuracy and methodological rigor[27,28]. Only scans with ≥ 10 valid measurements and an interquartile-range/median ratio ≤ 30% were considered valid and reliable. Liver fibrosis assessed by transient elastography was classified according to liver stiffness measurements expressed in kilopascals (kPa). Fibrosis was categorized as F0-F1 for stiffness values < 7.0 kPa, F2 for values between 7.0 kPa and 8.99 kPa, F3 for values ranging from 9.0 kPa to 12.49 kPa, and F4 for measurements ≥ 12.5 kPa. These cut-off points were selected based on widely accepted thresholds in the literature, which identify these values as optimal for detecting significant hepatic fibrosis in chronic viral hepatitis, including HCV. For analytical purposes, transient elastography scores of F0-F1 were considered non-significant fibrosis, whereas scores ≥ F2 (≥ 7.0 kPa) were classified as significant hepatic fibrosis[29,30].

Data collection

Patient data were retrieved from the medical records system and included demographic variables such as age, gender, presence of diabetes mellitus, and body mass index (BMI). Laboratory parameters comprised complete blood count, LFT [serum bilirubin, total protein, serum albumin, aspartate aminotransferase (AST), alanine aminotransferase (ALT), alkaline phosphatase (ALP), γ-glutamyl transferase (GGT)], lipid profiles [total cholesterol, triglycerides, low-density lipoprotein (LDL), high-density lipoprotein (HDL), and very low density lipoprotein (VLDL)]. Fasting venous blood samples were used for LFT and lipid profile using standardized laboratory protocols. Transient elastography measurements were also recorded. To ensure temporal consistency between the measurements, only cases in which the blood tests and transient elastography were conducted within seven days of each other were included. These variables were selected for analysis because they represent routinely available clinical indicators and are objective, easy to obtain, and closely associated with the severity of HCV infection. Serum biomarker–based fibrosis scores were calculated using standard formulas: The APRI was determined as [(AST/upper limit of normal) × 100]/platelet count (PLT); the FIB-4 index as (age × AST)/(PLT × √ALT); and the GPR as (GGT/upper limit of normal)/(PLT × 100)[11,12].

Statistical analysis

The statistical analyses were performed using Python 3 and IBM SPSS Statistics version 24 (IBM Corporation, Armonk, NY, United States). Data following a normal distribution are summarized as mean ± SD, while non-normal data are summarized as median (interquartile range). Differences between categorical variables were examined using the χ² test. For continuous variables, comparisons were made using the independent Student’s t-test when data were normally distributed and the Mann-Whitney U test when data were non-normally distributed. Data normality was assessed with the Shapiro-Wilk test, and correlations with fibrosis stage were analyzed using Spearman rank correlation test. Variables demonstrating statistically significant differences between the two groups - non-significant fibrosis (F0-F1) and significant fibrosis (F2+) - were identified using univariate analysis. To minimize potential multicollinearity among predictors, feature selection was performed using least absolute shrinkage and selection operator regression with 10-fold cross-validation. Variables with non-zero coefficients at the optimal λ value were subsequently entered into a multivariate logistic regression analysis to determine their independent significance, expressed as adjusted odds ratios (ORs) with corresponding 95% confidence intervals (CIs). The regression coefficients were derived from the final multivariate analysis. Statistical significance was defined as P < 0.05.

A total of 316 patients with HCV were included in the study. The study was randomly split into a training cohort (n = 214; 68%) and a validation cohort (n = 102; 32%) using a simple random allocation method. Randomization was performed by assigning a random number to each patient using a computer-based random number generator. No stratification was employed during the randomization process, as the overall sample was deemed adequate and representative to ensure a balanced distribution of baseline characteristics between the two cohorts. The allocation ratio of approximately 2:1 was intentionally chosen to allocate a sufficient number of patients for model development in the training cohort while retaining an adequate number of patients in the validation cohort to allow for robust internal validation. The 2:1 train-validation partition ensures a larger training cohort for feature selection, optimal model fitting and model stability, whereas a sufficient number of patients in the validation cohort ensures meaningful evaluation of model performance and reduces optimism bias and overfitting.

The fibrosis risk score (FRS) was developed in the training cohort by using the regression coefficients obtained in multivariate model. Each variable was assigned a weighted score, and the cumulative score indicated the probability of an individual of having significant hepatic fibrosis. Cut-off values for variables were determined using the Youden Index. The FRS was subsequently applied to the validation cohort to assess its diagnostic performance by calculating the AUC, sensitivity, specificity, positive predictive value, and negative predictive value in both the training and validation cohorts. Diagnostic performance of FRS was compared with commonly used fibrosis scores (APRI, FIB-4, GPR).

Development and performance assessment of the ML algorithms

Subsequently, several ML algorithms were developed to predict significant hepatic fibrosis using demographic and biochemical variables. The models evaluated included random forest (RF), AdaBoost, logistic regression, support vector machine (SVM), K-nearest neighbors (KNN), and Naive Bayes. These models were chosen because they represent diverse ML paradigms with unique strengths and weaknesses when applied in healthcare settings. To identify the optimal algorithm, a comparative evaluation was performed in both training and validation cohorts to identify the algorithm which is best suited to the demands of this clinical dataset.

RF and AdaBoost are robust ensemble models capable of handling complex clinical data and non-linear interactions. Logistic Regression is fast and easily interpretable but unable to capture non-linear relationships in complex clinical data. SVM performs well in handling high-dimensional data effectively through kernel functions but can be less efficient with larger datasets. KNN provides instance-based learning that adapts to local data structures; and naive Bayes can be efficient in limited sample sizes and provide probabilistic modeling. Evaluating multiple models together allows a balanced comparison between accuracy, interpretability, and generalizability in healthcare domain.

Five-fold cross-validation was used to optimize performance and mitigate class imbalance. Model performance was assessed using receiver operating characteristic curves, accuracy, precision, recall, and F1 score. The use of a validation set simulated real-world clinical conditions, allowing for internal validation and aids in assessment of model generalizability. The methodological workflow of the study is presented in Figure 1.

Figure 1 Methodological workflow of the study. HCV: Hepatitis C virus; HBV: Hepatitis B virus; HIV: Human immunodeficiency virus; LASSO: Least absolute shrinkage and selection operator; RF: Random forest; SVM: Support vector machine; KNN: K-nearest neighbors; NB: Naive Bayes; APRI: Aspartate aminotransferase to platelet ratio index; FIB-4: Fibrosis-4 index; GPR: Gamma-glutamyl transferase to platelet ratio.

RESULTS

Baseline characteristics

A total of 316 patients with confirmed HCV infection were enrolled in the study, of whom 182 (57.6%) had non-significant hepatic fibrosis (F0-F1) and 134 (42.4%) had significant fibrosis (F2-F4). There were no significant differences between the two groups with respect to some baseline characteristics including age, sex distribution, body mass index, some biochemical parameters (total protein, triglycerides, VLDL) (all P values > 0.05) (Table 1). However, the univariate analysis identified platelet count, ALT, AST, GGT, ALP, serum albumin, total bilirubin, total cholesterol, HDL, and LDL as parameters significantly associated with significant hepatic fibrosis (all P values < 0.05). A detailed comparison of variables between the non-significant hepatic fibrosis and significant hepatic fibrosis groups is provided in Table 1.

Table 1 Results of univariate and multivariate logistic regression analysis in the study cohort, n (%).

Parameter	Univariate analysis				Multivariate analysis
	Non-significant fibrosis (F0-F1) (n = 182; 57.6%)	Significant fibrosis (F2-F4) (n = 134; 42.4%)	Spearman ρ	P value	OR (95%CI)	P value
Gender			0.049	0.38	-	-
Males	96 (52.7)	64 (47.7)
Females	86 (47.3)	70 (52.3)
Age (year), mean ± SD	46.23 ± 14.5	48.10 ± 10.93	0.070	0.192	-	-
Body mass index, mean ± SD	24.03 ± 3.7	23.96 ± 3.2	0.0500	0.659	-	-
Diabetes			0.082	0.146	-	-
Yes	22 (12)	24 (17.91)
No	160 (88)	110 (82.09)
Platelet count (× 109/L)	157.7 ± 30.93	127.15 ± 28.00	-0.4421	< 0.001	0.72 (0.60-0.86)	< 0.001
Bilirubin (mg/dL)	0.726 ± 0.29	0.818 ± 0.31	0.1056	0.007	-	-
AST (U/L)	33.2 (21, 50)	49 (30, 66)	0.4768	< 0.001	1.18 (1.05-1.33)	0.003
ALT (U/L)	36 (24, 49)	47 (30, 64)	0.33	< 0.001	-	-
GGT (U/L)	26 (16, 33)	46 (25, 79)	0.50	< 0.001	1.07 (1.02-1.13)	0.01
ALP (U/L)	66 (52, 77)	81 (64, 97)	0.4298	< 0.001	1.10 (1.02-1.13)	0.01
Serum albumin (g/dL)	3.83 ± 0.29	3.16 ± 0.36	-0.6278	< 0.001	0.81 (0.68-0.96)	0.02
Total protein (g/dL)	7.082 ± 1.23	6.912 ± 1.12	-0.07	0.2	-	-
Total cholesterol (mg/dL)	160.4 ± 21.1	129.1 ± 23.9	-0.6156	< 0.001	0.84 (0.76-1.00)	0.01
Triglycerides (mg/dL)	131 (120, 170)	126 (115, 167)	-0.108	0.25	-	-
HDL (mg/dL)	37 (30, 54)	30 (26, 44)	-0.4414	< 0.001	-	-
LDL (mg/dL)	90.0 ± 4.3	65.4 ± 5.2	-0.5992	< 0.001	0.84 (0.76-1.00)	0.01
VLDL (mg/dL)	24.97 ± 12.11	23.45 ± 18.50	-0.1656	0.1424	-	-

OR: Odds ratio; CI: Confidence interval; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; GGT: γ-glutamyl transferase; ALP: Alkaline phosphatase; LDL: Low-density lipoprotein; HDL: High-density lipoprotein; VLDL: Very low-density lipoprotein.

Identifying predictors of significant hepatic fibrosis

Ten biochemical and hematological parameters - platelet count, AST, ALT, GGT, ALP, serum albumin, total bilirubin, total cholesterol, LDL, and HDL - were significantly different between the non-significant and significant hepatic fibrosis groups on univariate analysis. To mitigate multicollinearity and to identify the most relevant predictors these ten significant variables were subjected to least absolute shrinkage and selection operator regression with 10-fold cross-validation. The variables that demonstrated non-zero coefficients were used for multivariate modeling. The seven variables identified by multivariate logistic regression as independent predictors of significant hepatic fibrosis were used to develop the FRS. The predictors included: Platelet count (OR = 0.72; 95%CI: 0.60-0.86; P < 0.001), AST (OR = 1.18; 95%CI: 1.05-1.33; P = 0.003), GGT (OR = 1.07; 95%CI: 1.02-1.13; P = 0.01), ALP (OR = 1.10; 95%CI: 1.02-1.13; P = 0.01), serum albumin (OR = 0.81; 95%CI: 0.68-0.96; P = 0.02), total cholesterol (OR = 0.84; 95%CI: 0.76-1.00; P = 0.01), and LDL (OR = 0.84; 95%CI: 0.76-1.00; P = 0.01).

Development of FRS and its comparison with existing scoring systems

To develop a simple, practical and accurate FRS based on demographic and biochemical parameters, patients were allocated into a training set (n = 214; 68%) and a validation set (n = 102; 32%) using a random sampling method. There were no significant differences between the two groups with respect to demographic characteristics, biochemical parameters, or the proportion of patients with significant and non-significant hepatic fibrosis, indicating that both groups were comparable and suitable for subsequent model development and validation (Table 2).

Table 2 Comparison of baseline characteristics and biochemical parameters between the training and validation cohorts, n (%).

Parameters	Training dataset (n = 214)	Validation dataset (n = 102)	P value
Group			0.815
F0-F1	106 (49.5)	49 (48)
F2	36 (16.8)	21 (20.5)
F3	38 (17.8)	18 (17.6)
F4	34 (15.9)	14 (13.7)
Gender			0.994
Males	109 (50.9)	52 (51)
Females	105 (49.1)	50 (49)
Age (years)	46.88 ± 13.5	47.12 ± 11.6	0.871
Body mass index	23.86 ± 3.44	24.01 ± 3.39	0.715
Diabetes			0.70
Yes	30 (14)	16 (15.6)
No	184 (86)	86 (84.3)
Platelet count (× 109/L)	145.44 ± 30.83	144.1 ± 29.2	0.708
Bilirubin (mg/dL)	0.75 ± 0.26	0.756 ± 0.27	0.852
AST (U/L)	36 (25, 60)	40 (25, 63)	0.55
ALT (U/L)	42 (23, 47)	39 (25, 51)	0.66
GGT (U/L)	34 (21, 55)	36 (20, 53)	0.90
ALP (U/L)	59 (48, 77)	63.7 (54, 80)	0.07
Serum albumin (g/dL)	3.53 ± 0.26	3.57 ± 0.33	0.284
Total protein (g/dL)	7.02 ± 1.21	7.01 ± 1.13	0.943
Total cholesterol (mg/dL)	148.33 ± 19.12	146.6 ± 22.61	0.505
Triglycerides (mg/dL)	128 (111, 167)	133.2 (112, 174)	0.38
HDL (mg/dL)	35 (27, 43)	31.9 (25, 48)	0.98
LDL (mg/dL)	80.11 ± 14.3	78.33 ± 15.2	0.323
VLDL (mg/dL)	24.32 ± 11.16	23.89 ± 17.68	0.822

ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; GGT: γ-glutamyl transferase; ALP: Alkaline phosphatase; LDL: Low-density lipoprotein; HDL: High-density lipoprotein; VLDL: Very low-density lipoprotein.

Based on the results of the multivariate logistic regression, a simple and practical scoring model, FRS was developed using seven clinically relevant variables that demonstrated significant association with significant hepatic fibrosis. The components and scoring scheme of the FRS score are shown in Table 3. By summing the scores assigned to each variable, a total score is obtained that reflects the risk of having significant hepatic fibrosis.

Table 3 Newly developed fibrosis risk score system and its scoring scheme.

Parameter		Points
Platelet count < 150 × 109/L		5
AST > 45 U/L		3
Serum albumin < 3.5 g/dL		3
Total cholesterol < 140 mg/dL		3
LDL < 80 mg/dL		3
ALP > 75 U/L		2
GGT > 40 U/L		1
Interpretation of the score
Total score	Risk category	Interpretation
< 5	Low risk	Minimal risk of significant fibrosis
≥ 5 and < 9	Intermediate risk	Borderline risk → further evaluation suggested
9-20	High risk	Strongly predictive of F2-F4 fibrosis

AST: Aspartate aminotransferase; GGT: γ-glutamyl transferase; ALP: Alkaline phosphatase; LDL: Low-density lipoprotein.

Following the development of FRS the data from both the training and validation sets were analyzed to assess its predictive performance by comparing it against the established non-invasive fibrosis scores, including APRI, FIB-4, and GPR (Table 4). In the training set, the FRS demonstrated the highest predictive performance, yielding an AUC of 0.83 (95%CI: 0.75-0.88), which was significantly superior to APRI (AUC: 0.59, 95%CI: 0.49-0.66), FIB-4 (AUC: 0.65, 95%CI: 0.57-0.75), and GPR (AUC: 0.67, 95%CI: 0.55-0.76) for the same training set. A similar trend was observed in the validation set, where the FRS model developed in this study consistently outperformed these three clinical scoring systems, confirming its superior diagnostic accuracy for predicting significant hepatic fibrosis in HCV (Figure 2).

Figure 2 Receiver operating characteristic curves for the training and validation datasets. Figure illustrating the comparative performance of the fibrosis risk score against other established clinical scoring systems (aspartate aminotransferase to platelet ratio index, fibrosis-4 index, gamma-glutamyl transferase to platelet ratio) for predicting significant hepatic fibrosis. AUC: Area under the receiver operating characteristic curve; ROC: Receiver operating characteristic; FRS: Fibrosis risk score; APRI: Aspartate aminotransferase to platelet ratio index; FIB-4: Fibrosis-4 index; GPR: Gamma-glutamyl transferase to platelet ratio.

Table 4 Comparison of the predictive performance of the newly developed fibrosis risk score and established clinical scoring systems.

Models	Cut-off	Training set					Validation set
		AUC (95%CI)	Sensitivity	Specificity	NPV	PPV	AUC (95%CI)	Sensitivity	Specificity	NPV	PPV
FRS	9	0.83 (0.75-0.88)	0.71	0.85	0.79	0.71	0.82 (0.76-0.90)	0.72	0.84	0.79	0.72
APRI	1.5	0.59 (0.49-0.66)	0.31	0.87	0.67	0.61	0.61 (0.48-0.73)	0.31	0.83	0.65	0.62
FIB-4	1.45	0.65 (0.57-0.75)	0.50	0.78	0.72	0.59	0.69 (0.59-0.78)	0.48	0.81	0.65	0.63
GPR	0.40	0.67 (0.55-0.76)	0.49	0.83	0.71	0.68	0.70 (0.59-0.80)	0.48	0.83	0.69	0.70

CI: Confidence interval; FRS: Fibrosis risk score; APRI: Aspartate-to-platelet ratio index; FIB-4: Fibrosis-4 score; GPR: γ-glutamyl transferase-to-platelet ratio; AUC: Area under the curve; NPV: Negative predictive value; PPV: Positive predictive value.

Assessment of ML models for predicting significant hepatic fibrosis

Finally, the predictive performance of multiple ML algorithms, including RF, AdaBoost, logistic regression, SVM, KNN and Naive Bayes was assessed to determine if they could perform better in predicting the significant hepatic fibrosis. The performance metrics for all models were evaluated in terms of AUC, sensitivity, specificity, positive predictive value, negative predictive value, F1 score, and Brier score, as summarized in Table 5. In the training dataset, among all models, RF demonstrated the best predictive performance with the highest AUC of 0.921, along with a sensitivity of 0.825, specificity of 0.90, positive predictive value of 0.89, negative predictive value of 0.83, F1 score of 0.85, and a Brier score of 0.14. AdaBoost was the second-best performing model, yielding an AUC of 0.881, a sensitivity of 0.80, specificity of 0.77, positive predictive value of 0.78, negative predictive value of 0.79, an F1 score of 0.79, and a Brier score of 0.15. Other ML algorithms showed relatively lower predictive efficiency in this dataset (Figure 3). The superior AUC value of the RF model highlights its strong discriminative ability. The relative importance of variables contributing to the ML prediction models is illustrated in Figure 4, with platelet count, total cholesterol, LDL, GGT, AST, ALP, and serum albumin emerging as the most influential predictors.

Figure 3 Receiver operating characteristic curves comparing the diagnostic performance of various machine-learning algorithms in predicting significant hepatic fibrosis in training cohort. ROC: Receiver operating characteristic; AUC: Area under the receiver operating characteristic curve; SVM: Support vector machine; KNN: K-nearest neighbors.

Figure 4 Feature importance ranking of clinical and biochemical variables in the random forest model. LDL: Low-density lipoprotein; AST: Aspartate aminotransferase; GGT: Gamma-glutamyl transferase; ALP: Alkaline phosphatase; HDL: High-density lipoprotein; VLDL: Very-low-density lipoprotein; ALT: Alanine transaminase.

Table 5 Comparison of machine learning algorithms for predicting significant hepatic fibrosis.

Parameter	Model	AUC (95%CI)	Sensitivity	Specificity	PPV	NPV	F1 score	Brier score
Training	Random forest	0.921 (0.889-0.953)	0.825	0.900	0.892	0.837	0.857	0.141
	AdaBoost	0.881 (0.842-0.921)	0.800	0.775	0.780	0.795	0.790	0.157
	SVM	0.791 (0.742-0.842)	0.675	0.850	0.818	0.723	0.740	0.188
	Logistic regression	0.750 (0.698-0.803)	0.625	0.825	0.781	0.688	0.694	0.198
	Naive Bayes	0.751 (0.700-0.806)	0.925	0.525	0.661	0.875	0.771	0.217
	KNN	0.658 (0.602-0.716)	0.850	0.450	0.607	0.750	0.708	0.227
Validation	Random forest	0.905 (0.870-0.940	0.800	0.880	0.870	0.820	0.835	0.152
	AdaBoost	0.860 (0.820-0.903)	0.770	0.750	0.760	0.765	0.765	0.169
	SVM	0.760 (0.710-0.812)	0.650	0.820	0.790	0.700	0.710	0.205
	Logistic regression	0.735 (0.680-0.788)	0.600	0.800	0.760	0.660	0.675	0.210
	Naive Bayes	0.740 (0.690-0.795)	0.900	0.500	0.640	0.840	0.745	0.230
	KNN	0.630 (0.575-0.690)	0.820	0.420	0.580	0.720	0.685	0.245

AUC: Area under the curve; CI: Confidence interval; SVM: Support vector machine; KNN: K-nearest neighbors; NPV: Negative predictive value; PPV: Positive predictive value.

DISCUSSION

Liver biopsy remains the gold standard for diagnosing liver disease and grading hepatic fibrosis; however, its use has declined due to several limitations, including high cost, invasive nature, patient discomfort, potential procedural complications, and the requirement for specialized histopathological expertise. Consequently, a slew of non-invasive modalities has gained traction for the evaluation and staging of hepatic fibrosis and are advocated by the WHO. These methods incorporate either serum-based biomarkers - such as APRI, FIB-4, and GPR - or ultrasound-derived techniques, most notably transient elastography (FibroScan)[29,30]. The adoption of these approaches in clinical practice has substantially reduced the reliance on liver biopsy, particularly in individuals with an already established etiology of chronic liver disease[31,32].

The 2022 WHO guidelines for HCV emphasize on universal screening and widespread access to direct-acting antiviral therapy, in order to broaden treatment coverage irrespective of the stage of disease. These guidelines advocate the use of non-invasive tools (APRI, FIB-4, and FibroScan) to evaluate for significant hepatic fibrosis and cirrhosis, however, treatment eligibility is not dependent on degree of hepatic fibrosis, as treatment is recommended for all individuals with HCV[33,34].

In contrast, the WHO 2024 guidelines for chronic hepatitis B also advocate use of non-invasive tests such as APRI and FibroScan, particularly in resource-limited settings for the detection of significant hepatic fibrosis (≥ F2) or cirrhosis (F4), as it serves as an eligibility criterion to ensure that antiviral therapy is directed toward those at greatest risk of liver-related morbidity and mortality[35].

Numerous studies have demonstrated that serum lipid levels decline significantly in chronic viral hepatitis, particularly HCV. The hypolipidemic trend observed in HCV has a multifactorial etiology[36,37]. Despite this recognized association, lipid profiles are not routinely performed in such patients. As a consequence, the potential utility of lipid biomarkers, either alone or in combination with other biochemical parameters for evaluating hepatic fibrosis has not been adequately investigated. To date, most validated clinical scoring systems rely primarily on demographic factors (e.g., age), LFT and hematological parameters (platelet count), while the contribution of lipid markers in predicting hepatic fibrosis in HCV and other chronic liver diseases remains insufficiently explored[37,38]. To our knowledge, the Forns index is the only available non-invasive fibrosis score that incorporates a lipid profile parameter, namely serum cholesterol for estimation of risk of hepatic fibrosis. Developed for patients with HCV, it combines age, platelet count, GGT, and serum cholesterol to estimate the probability of significant hepatic fibrosis[39]. A Forns score < 4.2 indicates a low likelihood of significant fibrosis, whereas a score > 6.9 suggests a high probability of fibrosis. Values between 4.2 and 6.9 are considered indeterminate, as they cannot reliably rule in or rule out significant fibrosis[39]. However, the Forns score is based on a complex algorithmic formula in which serum cholesterol is inversely related to risk of fibrosis. Consequently, as the concentration of cholesterol reduces there is a corresponding increase in the Forns score, thereby indicating a greater likelihood of significant hepatic fibrosis.

The present study evaluated the predictive value of lipid biomarkers in conjunction with liver function parameters and platelet count in predicting significant hepatic fibrosis in patients with chronic HCV. The initial part of this study used conventional logistic regression to identify variables associated with hepatic fibrosis and subsequently derive a practical, reliable, and user-friendly FRS.

The FRS was developed by combining multiple parameters (platelet count, AST, ALP, GGT, serum albumin and serum lipid concentrations of total cholesterol and LDL), and its diagnostic performance for significant hepatic fibrosis was evaluated in the training cohort and internally validated against the validation cohort. The diagnostic performance of the FRS was subsequently compared with that of commonly used noninvasive fibrosis scores, including APRI, FIB-4 and GPR. In our study, the newly developed FRS demonstrated superior diagnostic performance, with an AUC of 0.83, compared with APRI (AUC = 0.59), FIB-4 (AUC = 0.65), and GPR (AUC = 0.67) at their respective optimal cutoff thresholds of 9, 1.5, 1.45, and 0.4.

The FIB-4 and APRI scores have demonstrated similar diagnostic performance for predicting significant hepatic fibrosis in patients with HCV in previous studies as they did in the current study. In a large retrospective cohort study comprising of 19510 chronic HCV patients, FIB-4 demonstrated the highest diagnostic performance (AUC = 0.70), followed by the King’s score (AUC = 0.70), while APRI showed a lower diagnostic performance (AUC = 0.68)[40].

This enhanced diagnostic performance of FRS may, in part, be due to integration of multiple complementary parameters in it, whereas existing scores primarily rely on age, platelet count and liver enzymes. Furthermore, differences in patient characteristics and population heterogeneity may have potentially contributed to the higher diagnostic performance of FRS in our cohort. The findings of this study suggest that the routine integration of lipid profile parameters with LFT and hematological indices may enhance diagnostic performance for the assessment of significant hepatic fibrosis in HCV. The FRS may offer improved diagnostic performance for significant hepatic fibrosis in HCV compared to the existing noninvasive scores. However, external validation in larger independent cohorts is warranted to confirm its generalizability.

At a threshold of ≥ 9 points, the FRS yielded an optimal balance of sensitivity (71%) and specificity (85%). Clinically, this cut-off allows easy and meaningful risk stratification: Patients scoring ≥ 9 are more likely to have significant fibrosis and may warrant invasive assessment or close monitoring, whereas those below this value can be managed with routine follow-up. Although statistically derived, this threshold offers clear practical value for preliminary screening and prioritizing patients for further diagnostic workup.

In the second step, to further enhance diagnostic efficiency, we developed predictive models using ML algorithms. Given the potential concerns regarding overfitting and the potential optimism of ML approaches, 5-fold cross-validation was employed to improve model stability and robustness. Following multiple attempts, algorithms with consistently high diagnostic accuracy for significant hepatic fibrosis were identified. These models offer a concrete theoretical ground for subsequent development of practical, accurate, accessible, and clinically deployable fibrosis assessment tools, with promising implications for wider clinical application.

In comparison to existing conventional noninvasive fibrosis scores such as APRI, FIB-4, and GPR, both the logistic regression-derived FRS and the ML-based RF and AdaBoost models demonstrated superior diagnostic performance for detecting significant hepatic fibrosis. Moreover, the reliance of FRS on routine laboratory parameters enhances feasibility, accessibility and affordability particularly in low-resource settings, by reducing dependence on specialized and costly tests. These models may obviate the need for advanced diagnostic tests in low-risk patients, improving cost-effectiveness and accessibility, especially in resource-limited regions or economically disadvantaged populations.

Both the FRS and RF model yielded higher specificity than sensitivity, indicating strong utility in ruling out patients without significant fibrosis and minimizing misclassification. However, their moderately lower sensitivity suggests that mild cases of fibrosis may be under detected. The positive and negative predictive values of the models were comparable, indicating balanced diagnostic efficiency. Although high specificity is advantageous and desirable, sensitivity remains an area for improvement, potentially achievable through integration of emerging biomarkers or optimization and refinement of the algorithms in future investigations. It is plausible that owing to the complex interactions among liver function parameters, lipid profile indices, and platelet count, the non-linear ML models - RF and AdaBoost - outperformed linear models in predictive performance for fibrosis.

The results reflect that, within the context of this study, the RF model may be potentially more efficient, reliable, and precise method for diagnosing significant hepatic fibrosis in patients with HCV infection, compared with conventional scoring systems including FRS, APRI, FIB-4, and GPR.

Overall, the ML models in this study offered considerable promise as non-invasive tools for the assessment of significant hepatic fibrosis in HCV-infected individuals, particularly for screening high-risk populations and steering targeted diagnostic evaluation.

Wei et al[41] employed ML techniques to develop a non-invasive model for predicting hepatic fibrosis in patients with hepatitis B and C using same indicators as in FIB-4. They compared the diagnostic performance of multiple ML models and developed a novel fibrosis prediction tool, LiveBoost, which was evaluated against the conventional FIB-4 index. The authors reported that the gradient boosting ML model significantly outperformed FIB-4 in predicting advanced hepatic fibrosis and cirrhosis, achieving an AUC of 0.91, compared with 0.84 for FIB-4, thereby demonstrating higher discriminatory accuracy[41].

Song et al[42] investigated the application of ML models for the non-invasive prediction of advanced liver fibrosis in patients with chronic hepatitis B. Their findings showed that the RF algorithm achieved the highest predictive performance, with an AUC of 0.87, outperforming XGBoost, logistic regression, and SVM algorithms. Moreover, when compared with traditional fibrosis indices such as APRI (AUC of 0.58) and FIB-4 (AUC of 0.63), the RF model exhibited markedly superior diagnostic accuracy[42].

Some studies have also evaluated the role of ML-based algorithms for the non-invasive detection of liver fibrosis in non-alcoholic steatohepatitis, reporting that these models achieve significantly higher diagnostic accuracy than conventional clinical indices, thereby supporting their potential clinical utility in detection of hepatic fibrosis[43,44].

CONCLUSION

This study developed a novel FRS by integrating liver function indices, lipid profile markers, and platelet count, and validated its performance against three widely adopted non-invasive tools including APRI, FIB-4, and GPR. The FRS demonstrated superior diagnostic accuracy in comparison to these commonly used clinical scores. Furthermore, the use of ML approaches boosted diagnostic discrimination, with the RF algorithm yielding the highest diagnostic performance for detecting significant hepatic fibrosis among chronic HCV patients. Collectively, these findings suggest that the FRS offers a simple, practical, and clinically valuable alternative for estimating the likelihood of significant hepatic fibrosis, particularly within HCV-affected individuals.

The main limitation of the present study is the lack of liver biopsy as a reference standard to directly compare the diagnostic or predictive performance of the proposed score. Transient elastography is vulnerable to technical alterations and possesses inherent limitations that may impact fibrosis classification. Nevertheless, a growing shift in clinical practice over recent years has led to decline in the routine use of liver biopsy for detecting significant fibrosis, resulting in limited availability of biopsy as reference standard. Another limitation of the study is the absence of an external validation cohort, although internal validation was conducted by dividing the sample into training and validation sets. Another shortcoming of the study is retrospective design of the study. Additionally, the variations in laboratory testing protocols and the inherent heterogeneity of the disease may influence the model’s robustness and efficiency, potentially resulting in an overestimation of its generalizability.

Despite these limitations, the integration of lipid biomarkers with standard liver function indices and platelet count shows promise and can provide a practical and biologically plausible alternative tool for non-invasive fibrosis evaluation.

Future research should focus on multicenter external validation, include diverse patient populations to assess the real-world applicability of the model, and incorporate multidimensional factors - such as emerging molecular markers and imaging data - to develop a more comprehensive and robust predictive system.