Liver stiffness-platelet ratio index and machine learning models for the noninvasive diagnosis of significant fibrosis in chronic hepatitis B

Abstract

BACKGROUND

Effective noninvasive tools for identifying significant liver fibrosis remain limited in untreated chronic hepatitis B (CHB).

AIM

To assess a liver stiffness-platelet ratio index (LPRI) derived from transient elastography and platelet count, and to evaluate its incremental value when incorporated into machine learning (ML) and deep learning (DL) models.

METHODS

We retrospectively enrolled 1098 therapy-naïve patients with CHB who underwent liver biopsy with Scheuer staging. Significant fibrosis was defined as S2-S4. We compared eight noninvasive tests and trained ML and DL models with and without the index (LPRI). Discrimination and clinical utility were assessed using the area under the receiver operating characteristic curve and decision curve analysis.

RESULTS

In feature selection analyses, liver stiffness measurement and platelet count ranked as the top predictors of significant fibrosis. LPRI outperformed conventional scores such as the aspartate aminotransferase-to-platelet ratio index and fibrosis-4, achieving an area under the receiver operating characteristic curve of approximately 0.84 in both the training and validation sets. Adding LPRI improved the performance of ML and DL models, and decision curve analysis suggested a net benefit across clinically relevant threshold probabilities.

CONCLUSION

A LPRI provides a simple, noninvasive approach for first-line screening of significant fibrosis in untreated CHB and may reduce unnecessary liver biopsies.

Key Words: Liver platelet ratio index; Significant fibrosis; Chronic hepatitis B; Machine learning; Noninvasive diagnosis

Core Tip: In this biopsy-based retrospective study of 1098 treatment-naïve patients with chronic hepatitis B, a liver stiffness-platelet ratio index, derived from transient elastography and platelet count, showed strong performance in identifying significant fibrosis (S2-S4) relative to conventional noninvasive scores. Liver stiffness-platelet ratio index also provided incremental value in selected machine learning and deep learning models and demonstrated a net benefit in decision curve analysis, supporting its use as a practical first-line triage tool when liver stiffness measurement and platelet count are routinely available.

INTRODUCTION

Chronic hepatitis B virus (HBV) infection remains a major global health burden, affecting approximately 296 million people[1]. According to World Health Organization data from 2022, approximately 820000 people die from hepatitis B annually, with a global mortality rate of 10 per 100000 people[2]. It is estimated that 20%-30% of individuals with chronic HBV infection will eventually develop cirrhosis. Chronic HBV infection is one of the leading causes of cirrhosis and hepatocellular carcinoma worldwide[3-5].

Liver fibrosis is a pivotal and potentially reversible stage in the progression of chronic liver disease[6]. Accurate assessment of significant liver fibrosis is essential for early intervention and the optimization of chronic hepatitis B (CHB) patient management. Clinical guidelines[7] designate significant fibrosis (Scheuer stage ≥ S2, i.e., S2-S4) as a key threshold for initiating critical therapeutic interventions, thereby directly influencing treatment decisions and prognostic evaluation.

Although liver biopsy remains the gold standard for diagnosing liver fibrosis, its invasive nature, cost, and risk of complications limit its clinical application. Therefore, noninvasive diagnostic methods have become increasingly important alternatives for screening and assessing liver fibrosis.

While existing noninvasive models perform well in identifying advanced fibrosis or cirrhosis, their effectiveness in detecting significant fibrosis (Scheuer ≥ S2) remains limited. As significant fibrosis has gained increasing attention in clinical practice[8], there is an urgent need for more effective noninvasive tools for its detection.

To address this issue, we collected the clinical characteristics of patients with CHB and applied feature selection methods to identify indicators most relevant to fibrosis severity. Liver stiffness measurement (LSM) and platelet (PLT) count consistently ranked among the top predictors. We evaluated a noninvasive test (NIT) model derived from LSM and PLT for assessing significant fibrosis in CHB.

Through a literature review, we found that combining PLT and LSM into a liver stiffness-platelet ratio index (LPRI) has shown predictive value in nonalcoholic steatohepatitis and metabolic dysfunction-associated steatotic liver disease populations, with a reported area under the receiver operating characteristic curve (AUROC) of 0.913[9]. However, this evidence was based on a small exploratory dataset, and the LPRI has not been validated in a large cohort of treatment-naïve CHB patients. Therefore, we validated the LPRI alongside seven other NITs in 1098 treatment-naïve CHB patients and assessed their performance in detecting significant fibrosis. In addition, we explored whether incorporating LPRI into machine learning (ML) and deep learning (DL) models could further improve diagnostic accuracy (Figure 1).

Figure 1 Study workflow and analytic pipeline. Treatment-naïve chronic hepatitis B patients were screened, eligible participants were enrolled, and fibrosis stages were recorded using Scheuer scoring (including intermediate half-stages). Nonsignificant fibrosis was defined as S0-S1.5 and significant fibrosis as S2-S4. The liver stiffness-platelet ratio index was derived from liver stiffness measurement and platelet count and evaluated alongside conventional noninvasive tests and machine learning/deep learning models using a 70%/30% training/validation split. DL: Deep learning; LPRI: Liver stiffness measurement to platelet ratio index; LSM: Liver stiffness measurement; ML: Machine learning; PLT: Platelet count; T_low/T_high: LPRI thresholds for triage.

MATERIALS AND METHODS

Reporting guideline compliance

This study is reported in accordance with the TRIPOD + artificial intelligence statement for prediction model studies. Where diagnostic accuracy against the liver biopsy reference standard is presented, reporting follows STARD 2015[10,11].

Study participants

The overall study workflow and analytic pipeline are summarized in Figure 1. This retrospective study included patients with untreated CHB from Xiamen Hospital of Traditional Chinese Medicine between January 2014 and May 2024. All patients completed laboratory tests within seven days of liver biopsy.

Inclusion criteria: (1) Hepatitis B surface antigen positivity for at least six consecutive months; and (2) HBV DNA level greater than 20 IU/mL.

Exclusion criteria: (1) Prior hepatitis B treatment; (2) Acute viral infections; (3) Coinfections with other chronic viruses (e.g., hepatitis C virus, hepatitis D virus, human immunodeficiency virus); (4) Hepatocellular carcinoma or other malignancies; (5) Other chronic liver diseases (autoimmune, genetic, drug-induced, or metabolic dysfunction-associated steatotic liver disease); (6) Long-term alcohol consumption (equivalent to ethanol intake > 20 g/day); and (7) End-stage systemic diseases or systemic inflammation.

Liver biopsy

All patients underwent ultrasound-guided percutaneous liver biopsy using a 16 G biopsy needle. Biopsy specimens were at least 1.5 cm in length or contained six or more portal tracts. Samples were fixed in 10% formalin, embedded in paraffin, and stained with hematoxylin-eosin, Masson’s trichrome, and reticulin stains. Scheuer’s scoring system was applied by the same pathologist to semi-quantify histologic necroinflammation (G0-G4) and fibrosis stage (S0-S4); half-stages (e.g., S0.5 and S1.5) were recorded where applicable[12]. In our cohort, Scheuer fibrosis staging included half-stages. For the primary endpoint, nonsignificant fibrosis was defined as S0-S1.5 and significant fibrosis as S2-S4. Representative liver biopsy images across Scheuer stages are provided in Supplementary Figures 1-4.

Serological and imaging variables

A total of 26 variables were included: (1) Baseline patient characteristics (e.g., sex, age, weight, and height); (2) Serological markers [e.g., HBV DNA, hepatitis B surface antigen, hepatitis B surface antibody, hepatitis B e antigen, hepatitis B e antibody, hepatitis B core antibody, white blood cell count, absolute neutrophil count, hemoglobin, PLT, prothrombin time, international normalized ratio, albumin, globulin, total bilirubin, direct bilirubin, alanine aminotransferase, aspartate aminotransferase (AST), gamma-glutamyl transferase, alkaline phosphatase, and alpha-fetoprotein]; and (3) Imaging parameters, including LSM (measured using FibroScan)[13].

Non-invasive fibrosis diagnostic models

Eight noninvasive diagnostic models were evaluated[14-20].

Statistical analysis

All data processing and statistical analyses were performed using Python 3.11 with the following primary libraries: (1) Pandas: Data reading, cleaning, and processing; (2) NumPy: Numerical computation and array operations; (3) Scikit-learn: ML (feature selection, data normalization, model construction, evaluation, and cross-validation); (4) Torch: DL model development; (5) Matplotlib: Data visualization; and (6) Statsmodels: Statistical modeling and regression analysis.

The dataset was stratified and randomly divided into a training set (70%) and a validation set (30%), preserving the proportional distribution of fibrosis stages (S0-S1.5 vs S2-S4). A complete-case analysis was performed; records with missing values in any model input were excluded. Data types were validated and converted to numeric format, with non-numeric data encoded or excluded as appropriate. Comparability between the training and validation sets was assessed using Student’s t-test or the Mann-Whitney U test for continuous variables, and the χ² test or Fisher’s exact test for categorical variables. Continuous variables were summarized as mean ± SD or median (interquartile range), and categorical variables as counts and percentages. A P value > 0.05 indicated no significant difference.

Categorical variables were encoded as numerical variables. Records with missing values were excluded, and normalization was performed using MinMaxScaler or StandardScaler.

Feature selection was conducted using SelectKBest, recursive feature elimination (RFE), and least absolute shrinkage and selection operator (LASSO) regression to identify the most relevant predictors of liver fibrosis stage[21]. The selected features were used to evaluate NIT model performance.

The performance of eight noninvasive models [AST to platelet ratio index (APRI), gamma-glutamyl transpeptidase to platelet ratio, fibrosis-4 (FIB-4), S-Index, Hui model, age-male-ALP-platelets risk score, AST to platelet and age-gender model, and LPRI] was compared using AUROC as the primary metric. Confidence intervals for AUROC were estimated using bootstrapping, and the DeLong test was used to assess differences between models.

To address class imbalance, probability thresholds were optimized in the training set by maximizing the Youden index. The selected threshold was then applied unchanged to the validation set for fixed-threshold evaluation. Sensitivity, specificity, positive predictive value, negative predictive value, accuracy, true positives, false positives, true negatives, false negatives, F1 score, precision-recall area under the curve, and Matthews correlation coefficient (MCC) were reported.

For ML and DL models, 12 ML classifiers (e.g., logistic regression, random forests, gradient boosting) and two DL models (neural network and Kolmogorov-Arnold Network) were used. Each classifier was trained with and without LPRI as a feature. AUROC, accuracy, and the Youden index were calculated, and grid search was used for hyperparameter optimization. Within the training set, fivefold cross-validation was used for hyperparameter tuning and model selection, and final performance was evaluated on the independent validation set.

To interpret feature contributions, SHapley Additive exPlanations (SHAP) analysis was performed for each ML model using the optimal NIT panel as input features, and SHAP summary bar plots were generated.

As a sensitivity analysis, propensity score matching was performed to generate a covariate-balanced dataset. Covariate balance diagnostics and model performance on the balanced dataset are reported in the Supplementary Tables 1 and 2.

Ethical statement

The study protocol was approved by the Medical Ethics Committee of Xiamen Hospital of TCM (No. 2024-k027-01). Given the retrospective design using de-identified data, the requirement for written informed consent was waived in accordance with the Declaration of Helsinki (2013 revision) and applicable local regulations.

RESULTS

Cohort characteristics and stratification

This study included 1098 eligible treatment-naïve CHB patients. Based on Scheuer staging with half-stages, 829 (75.5%) were classified as having nonsignificant fibrosis (S0-S1.5) and 269 (24.5%) as having significant fibrosis (S2-S4).

A stratified sampling method based on fibrosis stage was used to randomly divide the dataset into a training set (n = 768, 70%) and a validation set (n = 330, 30%), preserving the proportion of nonsignificant (S0-S1.5) and significant fibrosis (S2-S4) in both subsets. The training set included 580 (75.5%) nonsignificant and 188 (24.5%) significant cases, while the validation set included 249 (75.5%) nonsignificant and 81 (24.5%) significant cases. Baseline characteristics were broadly comparable between the training and validation sets (Table 1). Although PLT showed a small between-set difference (P = 0.02), the absolute difference was minimal, and PLT was retained as a predictor in all downstream models. Feature selection was performed on the 26 baseline variables using SelectKBest, RFE, and LASSO regression (Figure 2).

Figure 2 Comparison of feature selection methods. A: SelectKBest; B: Recursive feature elimination; C: Least absolute shrinkage and selection operator regression. AFP: Alpha-fetoprotein; ALP: Alkaline phosphatase; GGT: Gamma-glutamyl transferase; HBV: Hepatitis B virus; HBeAb: Hepatitis B e antibody; HBeAg: Hepatitis B e antigen; HBsAb: Hepatitis B surface antibody; HBsAg: Hepatitis B surface antigen; INR: International normalized ratio; WBC: White blood cell.

Table 1 Baseline characteristics of the training and validation datasets, n (%)/median (interquartile range).

Variables	Training set (n = 768)	Validation set (n = 330)	P values
Gender (male/female)	Male 473 (61.6)	Male 211 (63.9)	0.50
	Female 295 (38.4%)	Female 119 (36.1%)
Age (years)	36.00 (30.00, 44.25)	37.00 (30.25, 45.75)	0.56
Weight (kg)	61.00 (53.00, 69.00)	61.80 (54.00, 69.00)	0.85
Height (cm)	167.00 (160.00, 171.00)	168.00 (160.00, 173.00)	0.27
Hepatitis B virus DNA (log10 IU/mL)	5.90 (3.78, 7.63)	6.14 (3.56, 7.62)	0.59
Hepatitis B surface antigen (log10 IU/mL)	3.59 (3.07, 4.22)	3.57 (3.04, 4.32)	0.76
Hepatitis B surface antibody (mIU/mL)	0.19 (0.00, 1.10)	0.10 (0.00, 1.00)	0.21
Hepatitis B e antigen (COI)	0.59 (0.11, 700.00)	0.50 (0.17, 700.00)	0.86
Hepatitis B e antibody (COI)	0.32 (0.01, 13.39)	0.22 (0.01, 7.99)	0.26
Hepatitis B core antibody (COI)	7.95 (0.01, 9.27)	7.66 (0.01, 9.35)	0.71
White blood cell (109/L)	5.30 (4.50, 6.20)	5.30 (4.40, 6.10)	0.84
Neutrophil Abs. count (109/L)	2.90 (2.40, 3.60)	3.00 (2.40, 3.70)	0.45
Hemoglobin (g/L)	145.00 (132.00, 155.25)	146.00 (132.25, 156.00)	0.67
Platelet (109/L)	202.00 (168.00, 239.00)	195.00 (165.00, 229.00)	0.02
Prothrombin time (seconds)	13.10 (12.62, 13.70)	13.10 (12.60, 13.70)	0.91
International normalized ratio	1.01 (0.97, 1.06)	1.02 (0.97, 1.06)	0.62
Albumin (g/L)	44.00 (42.00, 46.00)	44.00 (42.00, 46.00)	0.52
Globulin (g/L)	31.00 (28.00, 34.00)	30.00 (27.00, 34.00)	0.62
Total bilirubin (μmol/L)	16.05 (11.80, 21.50)	15.50 (12.10, 21.10)	0.88
Direct bilirubin (μmol/L)	4.85 (3.00, 7.20)	4.80 (2.90, 7.47)	0.92
Alanine aminotransferase (U/L)	68.00 (44.00, 116.00)	68.00 (47.00, 118.00)	0.65
Aspartate aminotransferase (U/L)	49.00 (27.75, 94.00)	44.00 (25.00, 89.75)	0.17
Gamma-glutamyl transferase (U/L)	33.00 (21.00, 62.00)	31.50 (21.00, 56.00)	0.44
Alkaline phosphatase (U/L)	63.00 (33.00, 83.00)	61.00 (31.25, 81.00)	0.64
Alpha-fetoprotein (ng/mL)	2.95 (1.75, 5.55)	2.79 (1.82, 4.64)	0.51
Liver stiffness measurement (kPa)	7.40 (5.60, 10.60)	7.30 (5.50, 10.20)	0.70

Performance of LPRI in fibrosis diagnosis associated with CHB

This study evaluated eight NITs, including LPRI, for diagnosing significant fibrosis in CHB (Table 2). Among the eight NITs, LPRI demonstrated the highest discriminative ability for significant fibrosis (S2-S4). In the training dataset, LPRI achieved an AUROC of 0.84 (95%CI: 0.80-0.87), exceeding comparator indices such as FIB-4 (AUROC of 0.65, 95%CI: 0.61-0.69) and APRI (AUROC of 0.66, 95%CI: 0.62-0.70) (Table 2, Figure 3A). DeLong’s tests supported the AUROC advantage of LPRI over conventional NITs (e.g., LPRI vs FIB-4, P < 0.001; LPRI vs APRI, P < 0.001) (Supplementary Table 1). In the validation dataset, LPRI retained superior performance, with an AUROC of 0.83 (95%CI: 0.78-0.88) (Table 2, Figure 3B). The overlapping confidence intervals between the training and validation sets suggest no evidence of overfitting; the small difference in AUROC is plausibly attributable to sampling variability and case-mix differences. At the Youden-optimal cutoff, LPRI achieved balanced operating characteristics, with a Youden index of 49.65%, sensitivity of 82.72%, specificity of 66.94%, positive predictive value of 44.97%, and negative predictive value of 92.22% in the validation set (Table 2). The MCC further supported overall agreement (MCC = 0.43). Compared with conventional indices (e.g., APRI and FIB-4), LPRI provided a more favorable sensitivity–specificity trade-off, supporting its potential role as a first-line triage test. Decision curve analysis indicated potential clinical utility of LPRI by demonstrating a favorable net benefit across clinically relevant threshold probabilities (Supplementary Figure 5).

Figure 3 Receiver operating characteristic curves for eight noninvasive tests in the training set and validation set. A: Training set; B: Validation set. Pairwise DeLong tests are reported in Supplementary Table 1. aMAP: Age-male-alkaline phosphatase-platelets risk score; APAG: Aspartate aminotransferase to platelet and age-gender model; APRI: Aspartate aminotransferase to platelet ratio index; AUC: Area under the curve; FIB-4: Fibrosis-4; GPR: Gamma-glutamyl transpeptidase to platelet ratio; LPRI: Liver stiffness-platelet ratio index.

Table 2 Detailed model performance metrics for liver fibrosis diagnostic models.

Metric	Liver stiffness-platelet ratio index	Fibrosis-4	AST to platelet ratio index	Gamma-glutamyl transpeptidase to platelet ratio	S-Index	Hui model	Age-male-ALP-platelets risk score	AST to platelet and age-gender model
Training data
AUROC	0.84	0.65	0.66	0.70	0.71	0.75	0.69	0.71
95%CI lower	0.80	0.61	0.62	0.66	0.67	0.71	0.65	0.66
95%CI upper	0.87	0.69	0.70	0.74	0.75	0.78	0.74	0.75
UA	0.06	0.09	0.08	0.08	0.08	0.08	0.09	0.09
Matthews correlation coefficient	0.46	0.21	0.20	0.26	0.26	0.33	0.27	0.31
Youden index (%)	52.96	20.38	22.87	28.58	30.00	37.87	31.13	33.25
PPV (%)	47.55	40.56	32.08	40.39	36.54	40.29	38.04	44.83
NPV (%)	92.53	80.44	85.04	83.43	85.78	88.08	85.52	84.33
Sensitivity (%)	82.45	38.83	72.87	54.79	68.62	72.87	65.96	55.32
Specificity (%)	70.52	81.55	50.00	73.79	61.38	65.00	65.17	77.93
Accuracy (%)	73.44	71.09	55.60	69.14	63.15	66.93	65.36	72.40
TP (%)	20.18	9.51	17.84	13.41	16.80	17.84	16.15	13.54
Validation data
AUROC	0.83	0.65	0.61	0.71	0.71	0.74	0.70	0.73
95%CI lower	0.78	0.58	0.54	0.64	0.64	0.67	0.63	0.67
95%CI upper	0.88	0.72	0.68	0.78	0.78	0.79	0.76	0.80
UA	0.10	0.14	0.13	0.13	0.13	0.12	0.13	0.13
Matthews correlation coefficient	0.43	0.23	0.08	0.30	0.26	0.29	0.26	0.30
Youden index (%)	49.65	23.45	9.26	31.74	30.02	33.75	30.02	31.72
PPV (%)	44.97	41.67	27.91	44.00	36.60	37.50	36.60	44.79
NPV (%)	92.22	81.22	78.98	83.84	85.80	87.57	85.80	83.69
Sensitivity (%)	82.72	43.21	59.26	54.32	69.14	74.07	69.14	53.09
Specificity (%)	66.94	80.24	50.00	77.42	60.89	59.68	60.89	78.63
accuracy (%)	70.82	71.12	52.28	71.73	62.92	63.22	62.92	72.34
TP (%)	20.36	10.64	14.59	13.37	17.02	18.24	17.02	13.07

AST: Aspartate aminotransferase; AUROC: Area under the receiver operating characteristic curve; NPV: Negative predictive value; PPV: Positive predictive value; TP: True positives; UA: Uncertainty area.

Performance of ML and DL models in identifying fibrosis stages in CHB

We trained and evaluated multiple ML and DL models to distinguish significant fibrosis (S2-S4) from nonsignificant fibrosis (S0-S1.5). Model performance on the imbalanced dataset is summarized in Table 3 and illustrated in Figure 4. Incorporating LPRI improved performance in several models, although the magnitude of improvement varied across algorithms. For example, the ExtraTree model showed an AUROC increase from 0.63 to 0.66 and an increase in the Youden index from 0.27 to 0.32 after incorporating LPRI (Table 3). In contrast, the DecisionTree model showed reduced performance after adding LPRI (AUROC of 0.70-0.67; Youden index of 0.40-0.35) (Table 3).

Figure 4 Receiver operating characteristic curves comparing machine learning/deep learning models trained on 26 baseline features and 26 baseline features plus liver stiffness-platelet ratio index on the imbalanced dataset. A: 26 baseline features; B: 26 baseline features plus liver stiffness-platelet ratio index on the imbalanced dataset. AUC: Area under the curve; KAN: Kolmogorov-Arnold Network; NuSVC: Nu-support vector classification; SVM: Support vector machine.

Table 3 Performance of machine learning and deep learning models with 26 features and liver stiffness-platelet ratio index + 26 features on imbalanced dataset.

Models	Area under the receiver operating characteristic curve	Youden	Positive predictive value (%)	Negative predictive value (%)	Accuracy (%)	True positives (n)	False positive (n)	True negative (n)	False negative (n)	Sensitivity (%)	Specificity (%)
Using 26 features
LogisticRegression	0.85	0.58	0.57	0.92	0.80	41.40	31.60	134.00	12.40	0.77	0.81
SVM	0.83	0.55	0.55	0.91	0.79	40.40	33.40	132.20	13.40	0.75	0.80
NuSVC	0.81	0.52	0.51	0.91	0.76	41.40	40.80	124.80	12.40	0.77	0.75
DecisionTree	0.70	0.40	0.54	0.85	0.77	29.80	25.60	140.00	24.00	0.55	0.85
ExtraTree	0.63	0.27	0.44	0.82	0.72	24.80	32.00	133.60	29.00	0.46	0.81
GaussianNB	0.80	0.49	0.52	0.89	0.77	38.20	35.80	129.80	15.60	0.71	0.78
GradientBoosting	0.88	0.63	0.60	0.93	0.82	43.20	29.00	136.60	10.60	0.80	0.82
HistGradientBoosting	0.87	0.62	0.55	0.94	0.79	45.20	37.00	128.60	8.60	0.84	0.78
AdaBoost	0.86	0.61	0.54	0.95	0.78	46.20	41.60	124.00	7.60	0.86	0.75
RandomForest	0.87	0.62	0.58	0.93	0.81	44.00	32.00	133.60	9.80	0.82	0.81
KNeighbors	0.77	0.43	0.47	0.89	0.72	38.00	46.40	119.20	15.80	0.71	0.72
KAN	0.83	0.55	0.56	0.91	0.79	40.80	34.20	131.40	13.00	0.76	0.79
NeuralNetwork	0.84	0.57	0.55	0.92	0.78	42.80	36.80	128.80	11.00	0.80	0.78
Using liver stiffness-platelet ratio index + 26 features
LogisticRegression	0.85	0.58	0.58	0.91	0.81	40.80	29.00	136.60	13.00	0.76	0.82
SVM	0.83	0.55	0.54	0.91	0.78	41.00	34.80	130.80	12.80	0.76	0.79
NuSVC	0.81	0.53	0.54	0.91	0.77	40.40	37.20	128.40	13.40	0.75	0.78
DecisionTree	0.67	0.35	0.51	0.84	0.76	26.80	25.00	140.60	27.00	0.50	0.85
ExtraTree	0.66	0.32	0.48	0.83	0.75	26.40	28.40	137.20	27.40	0.49	0.83
GaussianNB	0.81	0.52	0.58	0.89	0.80	37.20	27.80	137.80	16.60	0.69	0.83
GradientBoosting	0.87	0.62	0.56	0.94	0.80	44.80	34.80	130.80	9.00	0.83	0.79
HistGradientBoosting	0.87	0.61	0.53	0.94	0.78	46.00	41.40	124.20	7.80	0.86	0.75
AdaBoost	0.85	0.58	0.59	0.92	0.80	42.00	33.00	132.60	11.80	0.78	0.80
RandomForest	0.86	0.64	0.62	0.93	0.82	43.60	28.80	136.80	10.20	0.81	0.83
KNeighbors	0.78	0.42	0.47	0.88	0.72	37.00	44.20	121.40	16.80	0.69	0.73
KAN	0.84	0.58	0.55	0.92	0.79	42.60	35.80	129.80	11.20	0.79	0.78
NeuralNetwork	0.85	0.59	0.59	0.92	0.81	41.00	28.80	136.80	12.80	0.76	0.83

SVM: Support vector machine; NuSVC: Nu-support vector classification; KAN: Kolmogorov-Arnold Network.

Feature importance for Random Forest, Gradient Boosting, and HistGradientBoosting models was examined using SHAP bar plots (Figure 5). After incorporating LPRI, the attribution profile shifted: LPRI ranked among the top predictors, becoming the leading contributor in the boosting models and the second-largest contributor (after LSM) in the Random Forest model.

Figure 5 Global feature importance assessed by SHapley Additive exPlanations. A-C: Models without liver stiffness-platelet ratio index – random forest, gradient boosting, and histogram-based gradient boosting; D-F: Corresponding models with liver stiffness-platelet ratio index. AFP: Alpha-fetoprotein; ALB: Albumin; ALP: Alkaline phosphatase; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; DBIL: Direct bilirubin; GLB: Globulin; HBV: Hepatitis B virus; HBeAb: Hepatitis B e antibody; HBeAg: Hepatitis B e antigen; HBsAb: Hepatitis B surface antibody; HBsAg: Hepatitis B surface antigen; HGB: Hemoglobin; INR: International normalized ratio; LPRI: Liver stiffness-platelet ratio index; LSM: Liver stiffness measurement; NE: Neutrophil; PLT: Platelet; PT: Prothrombin time; TBIL: Total bilirubin; WBC: White blood cell; γGGT: Gamma-glutamyl transferase.

DISCUSSION

The accurate identification of significant (S2-S4) fibrosis in patients with CHB is crucial for timely intervention, effective disease management, and improved long-term outcomes. While existing NITs adequately detect advanced fibrosis or cirrhosis, their performance often wanes in earlier yet clinically critical stages. Our study addresses this gap by focusing on S2-S4 fibrosis and introducing LPRI as a simplified, robust indicator that could reshape early diagnostic strategies.

LSM and PLT were consistently identified as the most potent predictors of significant fibrosis severity using multiple feature selection methods (SelectKBest, RFE, and LASSO). This finding aligns with existing literature that underscores LSM’s reliability as a noninvasive surrogate of liver stiffness and fibrosis burden[22-24], as well as the established association between thrombocytopenia and progressive hepatic fibrosis[25-27]. Pathophysiologically, advancing fibrosis increases liver stiffness while also predisposing patients to portal hypertension and hypersplenism, which promote splenic sequestration of platelets. In parallel, impaired hepatic synthetic function may reduce thrombopoietin production, further lowering PLTs. By coupling these two biologically linked processes into a single ratio, LPRI provides a parsimonious yet mechanistically coherent index that is intuitive for clinicians and readily integrated into routine workflows.

Notably, previous investigations of LPRI have been limited to smaller cohorts and have focused primarily on metabolic dysfunction-associated steatotic liver disease populations, with limited validation in untreated CHB. To our knowledge, this study represents one of the earliest and largest validations of LPRI in a substantial cohort of antiviral therapy-naïve CHB patients. By demonstrating robust performance in a previously underexamined clinical setting, our findings broaden the evidence base for LPRI and enhance its potential clinical utility and translational impact across diverse patient populations.

Compared with traditional NITs (e.g., APRI and FIB-4), LPRI demonstrated consistently superior diagnostic performance for significant fibrosis. By offering both high sensitivity and specificity, it reduces the likelihood of missed cases while minimizing unnecessary follow-up in patients without advanced disease, thereby strengthening its potential as a frontline screening tool. Although other NITs incorporate additional features, many have not demonstrated equivalent predictive strength or consistency across diverse cohorts. External factors – such as inflammation, metabolic comorbidities, or assay variability – may influence their performance, rendering them less universally reliable than LSM and PLT. In contrast, LPRI’s reliance on widely recognized, high-value predictors fosters stable and reproducible results, further supported by the alignment of our findings with previously reported AUROCs for APRI[14], FIB-4[16], the Hui model[18], and age-male-ALP-platelets risk score[20].

Decision curve analysis further confirmed that LPRI provides a meaningful net clinical benefit across relevant risk thresholds, consistently outperforming established NITs. This highlights the practical advantages of LPRI for real-world risk stratification and patient management.

To facilitate clinical implementation, we propose an LPRI-based triage workflow that stratifies patients into rule-out, indeterminate, and rule-in zones using optimized lower and upper thresholds (Figure 6), reserving more complex ML/DL models for selected scenarios in which incremental benefit justifies their use.

Figure 6 Proposed clinical workflow for liver stiffness-platelet ratio index-based triage of significant fibrosis in treatment-naïve chronic hepatitis B. Patients are stratified into rule-out, indeterminate, and rule-in zones using lower and upper liver stiffness-platelet ratio index thresholds; complex machine learning/deep learning models may provide incremental accuracy in selected settings. aMAP: Age-male-alkaline phosphatase-platelets risk score; APAG: Aspartate aminotransferase to platelet and age-gender model; APRI: Aspartate aminotransferase to platelet ratio index; FIB-4: Fibrosis-4; GPR: Gamma-glutamyl transpeptidase to platelet ratio; KAN: Kolmogorov-Arnold Network; LPRI: Liver stiffness-platelet ratio index; NIT: Noninvasive test; PF: Random forest; S Index: S-Index; SVM: Support vector machine.

Class imbalance was present in the study cohort (75.5% nonsignificant vs 24.5% significant fibrosis), reflecting real-world clinical distributions. Accordingly, the primary analyses were conducted on the original imbalanced dataset, with resampling-based sensitivity analyses provided. Specifically, SMOTE-balanced model performance was reported as a sensitivity analysis (Supplementary Table 2, Supplementary Figures 6 and 7), and PSM[28] was included as a supplementary covariate balance diagnostic (Supplementary Table 3, Supplementary Figure 8). These analyses were intended to assess the robustness of model performance under alternative sampling schemes rather than to redefine the primary evaluation. Given that the most clinically challenging decision point lies near the S1-S2 boundary, we additionally performed a borderline-case analysis (Scheuer S1 vs S2) and summarized operating characteristics under different LPRI threshold strategies (Supplementary Figure 9, Supplementary Table 4).

We further integrated LPRI into multiple ML/DL frameworks to assess its incremental value for fibrosis prediction[29-31]. Overall, incorporating LPRI improved diagnostic performance in several models, although the magnitude of improvement varied across algorithms and metrics. Notably, it enhanced the identification of patients with significant fibrosis in selected classifiers (Supplementary Figure 10). However, a minority of classifiers exhibited reduced accuracy, likely due to feature redundancy or complex interactions arising from LPRI’s derivation from pre-existing variables. This finding underscores the need for careful feature engineering and regularization in future studies.

SHAP analysis consistently identified LPRI as a major contributor across the ensemble tree-based models examined (Figure 6). In the augmented models, LPRI ranked first in Gradient Boosting and HistGradientBoosting and second (after LSM) in Random Forest. The observed reduction in the apparent importance of LSM and PLT in some augmented models likely reflects shared information captured by LPRI rather than diminished clinical relevance. Thus, LPRI can consolidate key indicators while improving interpretability and, in many settings, predictive performance in ML/DL-based liver fibrosis classification in CHB.

Our study benefits from a large sample size, rigorous statistical analysis, and robust validation procedures (including cross-validation), which reduce bias and enhance the credibility of the findings. However, several limitations should be acknowledged. The retrospective, single-center design may limit generalizability. Prospective, multicenter studies are needed to confirm the predictive value of LPRI across diverse patient populations, healthcare systems, and liver disease etiologies. Additionally, evaluating LPRI in combination with emerging biomarkers or advanced imaging techniques may further improve diagnostic performance.

CONCLUSION

In conclusion, this large retrospective, biopsy-based study provides robust large-scale validation of LPRI in treatment-naïve CHB and supports its use as a simplified first-line tool for triaging significant fibrosis (S2-S4). Incorporating LPRI into ML/DL classifiers yielded modest but meaningful improvements in selected models, while highlighting the need for careful regularization when derived features are introduced. Prospective, multicenter studies are warranted to confirm generalizability across healthcare systems and disease etiologies and to determine whether combining LPRI with emerging biomarkers or imaging modalities can further improve risk stratification.