Liver stiffness measurement-based risk score for predicting liver decompensation risk: a single-center retrospective Chinese study

Abstract

Liver decompensation represents a critical milestone in compensated advanced chronic liver disease (cACLD). Liver stiffness measurement (LSM) has emerged as a valuable non-invasive marker. This study aimed to develop an LSM-based liver decompensation risk prediction. We retrospectively recruited 1064 cACLD patients (LSM ≥ 10 kPa), divided into derivation (n = 745) and validation (n = 319) groups. Fine-Gray competing risk regression identified independent risk factors. Optimal cut-off values for risk stratification were determined using X-tile software. Model performance was evaluated using C-index and calibration curves. The main etiology was hepatitis B virus infection (69.8%). During follow-up, 328 patients (30.8%) developed liver decompensation with median decompensation time of 33 (16–50) months. Six independent predictors were identified: age, LSM, spleen diameter, hemoglobin, platelet, and international normalized ratio. The model demonstrated good discrimination [C-index: 0.779 (0.714–0.845)], calibration and overall performance (Brier Score 0.139). LSM contributed significantly (likelihood ratio test = 47.99, P < 0.001) with hazard ratio increasing substantially when LSM > 20 kPa. Patients were stratified using optimal cut-offs into low-risk (≤ 147.7 points), medium-risk (147.7–206.6 points), and high-risk (≥ 206.6 points) groups, with decompensation rates of 13.4%, 58.0%, and 86.7%, respectively, and median time to decompensation of 37 (18–55), 33 (15–50), and 28 (17–48) months, respectively. Cumulative decompensation incidences differed significantly among risk groups (Gray’s test, P < 0.001). A user-friendly web-based LSM-Based Liver Decompensation Risk Score assessment tool was developed. Despite single-center retrospective design, hepatitis B focus, and lacking external validation, the LSM-based model effectively identified high-risk patients, providing valuable clinical decision support.

Introduction

Chronic liver disease represents a significant global public health challenge, with its associated morbidity and mortality continuing to rise [1]. Compensated advanced chronic liver disease (cACLD) constitutes a critical stage in chronic liver disease progression [2]. Although patients exhibit substantial liver fibrosis or early cirrhosis, they have not yet manifested clinical signs of liver decompensation [3]. This clinical silence often leads to underdiagnosis and delayed recognition, making population surveillance particularly challenging. The global burden of cACLD is substantial and continues to expand. In the United States, approximately 5 million individuals may suffer from cACLD, with the majority being male patients, those with diabetes, and obese individuals [4]. Conversely, in China, hepatitis B virus (HBV) infection represents the predominant cause of cACLD [5]. Despite the high prevalence, many cACLD patients remain unaware of their condition until disease progression occurs. Liver decompensation constitutes a pivotal milestone in this disease trajectory, representing the transition from a relatively stable compensated state to a clinically manifest, life-threatening condition [6]. The occurrence of decompensation events signifies substantial prognostic deterioration, accompanied by considerable increases in healthcare costs and severe impairment of quality of life [7–9]. Critically, following the initial decompensation event, patients face significantly elevated risks of recurrent decompensation, thereby establishing a detrimental cycle of accelerated disease progression. Therefore, accurate identification of high-risk patients and implementation of early intervention strategies are of paramount clinical importance for delaying disease progression, preventing decompensation events, and improving long-term patient outcomes.

Liver stiffness measurement (LSM) is an important non-invasive biomarker for evaluating hepatic fibrosis severity and guiding clinical decisions in cACLD patients [3]. Transient elastography can accurately reflect liver fibrosis progression and portal hypertension dynamics by quantifying hepatic tissue elasticity. The diagnostic utility of LSM is not limited to the reflection of fibrosis. When integrated with conventional biochemical markers, LSM demonstrates superior accuracy in identifying cirrhosis in patients with chronic hepatitis B [10]. Furthermore, LSM is not only a static diagnostic indicator, but its dynamic changes can also predict the risk of disease progression [11]. This dynamic monitoring capability has been valuable in clinical risk stratification. The combination of fibrosis-4 scores (FIB-4), LSM, and spleen stiffness measurement can effectively predict high-risk esophageal varices in patients with compensated cirrhosis [12]. Similarly, the AI-Safe-C score, incorporating LSM, successfully identifies non-cirrhotic individuals who are at high risk for liver-related events [13]. Patients with primary biliary cholangitis have a poor prognosis when LSM > 11 kPa [14]. Several prospective studies have shown that the higher the baseline LSM value, the greater the risk of liver-related events in patients, establishing LSM as a fundamental component of prognostic stratification [15–17]. However, Wong YJ, et al. have pointed out that current LSM can better predict prognostic risk than baseline LSM [18]. With the rising prevalence of metabolic dysfunction-associated steatotic liver disease (MASLD), there has been increasing focus on developing disease-specific prediction models. Combining LSM with other routine hematological tests enhances the ability to diagnose high-risk metabolic dysfunction-associated steatohepatitis and to predict liver-related complications [16]. Calzadilla Bertot L et al. proposed that the modified ABIDE model including LSM is more accurate than albumin–bilirubin score, FIB-4 and other models in predicting liver decompensation in patients with MASLD cirrhosis [19]. A two-step non-invasive approach that first assesses FIB-4 and then measures LSM can effectively stratify patients at different risks of liver-related events [20]. LSM is also associated with hepatocellular carcinoma risk in patients with MASLD, especially in non-cirrhotic MASLD patients with diabetes and LSM ≥ 10 kPa [21]. The heterogeneity of chronic liver disease etiologies necessitates the development of tailored prediction models that can account for disease-specific progression patterns and risk factors. Despite these advances, the occurrence of liver decompensation is a complex multifactorial process, and relying solely on LSM for risk assessment has certain limitations. Therefore, establishing a comprehensive prediction model that integrates multiple clinical parameters can more accurately assess the decompensation risk of individual patients and provide a more accurate basis for clinical decision-making.

With the development of precision medicine concepts and digital healthcare system, network-based clinical decision support tools are becoming integral components of modern medical practice. This technological evolution addresses the growing need for accessible, evidence-based tools that can translate complex clinical data into actionable insights at the point of care. Interactive web risk assessment tools have offered simple operation and accurate calculation, and demonstrated remarkable efficacy across diverse medical specialties, such as breast cancer, Alzheimer’s disease, scleroderma [22–24]. These platforms exemplify the successful integration of evidence-based algorithms with user-centric design principles, enabling rapid patient stratification and facilitating the development of tailored monitoring protocols and therapeutic interventions. In the field of liver disease, although a variety of prognostic scoring systems have been developed and applied [25–27], they lack user-friendly interface design and comprehensive clinical decision support functions. Consequently, the development of a networked risk assessment tool that integrates multivariable prediction model with intuitive user interface design will provide important technical support for the precise management of liver disease.

This study aims to establish an LSM-based decompensation risk prediction model in cACLD patient, and to develop corresponding networked risk assessment tools, thereby providing a scientific foundation for individualized management and precision medicine approaches.

Methods

Study design

This study retrospectively enrolled cACLD patients who were treated in Beijing Ditan Hospital between January 2014 and May 2024. The study protocol received approval from the Ethics Committee of Beijing Ditan Hospital (DTEC-KY2024-069-01), and the research process strictly followed the ethical principles of the Declaration of Helsinki.

Inclusion criteria were: (1) age 18–65 years; (2) LSM ≥ 10 kPa [3]; (3) history of chronic liver disease ≥ 6 months, with clear etiology (including HBV, hepatitis C virus, nonalcoholic fatty liver disease, alcoholic liver disease, autoimmune liver disease, etc.); (4) absence of hepatic decompensation manifestations (ascites, varicose bleeding, hepatic encephalopathy); (5) complete baseline and follow-up data. Exclusion criteria included: (1) combined with other systemic malignancies; (2) severe cardiac, pulmonary, or renal dysfunction; (3) failed LSM acquisition or suboptimal measurement quality; (4) incomplete follow-up data; (5) previous history of liver transplantation or transjugular intrahepatic portosystemic shunt.

Data collection

Baseline clinical data of patients were collected, including: (1) demographic characteristics: age, gender; (2) medical history: cause of liver disease; (3) laboratory tests: hematological parameters [hemoglobin (HB), platelet (PLT), neutrophil-to-lymphocyte ratio (NLR)], liver function [alanine aminotransferase, aspartate aminotransferase, gamma-glutamyl transferase, total bilirubin (TB), albumin (ALB)], coagulation function [international normalized ratio (INR)]; (4) abdominal ultrasound: portal vein width and spleen diameter; (5) FibroScan test: LSM, controlled attenuation parameter (CAP).

All patients underwent liver ultrasound elastography (FibroScan®, Echosens, France) to measure LSM values. At least 10 valid measurements were obtained for each patient, and the median was taken as the final LSM value. The success criteria for the test were: success rate ≥ 60%, and interquartile range/median ≤ 30%. All laboratory tests were completed within 2 weeks of LSM and abdominal ultrasound.

Outcome events

Patients underwent regular outpatient surveillance at 3–6 monthly intervals, with systematic documentation of clinical symptoms, physical signs, laboratory and ultrasound test results. The primary outcome event was liver decompensation, defined as the occurrence of any of the following complications: (1) ascites: imaging confirmed abdominal effusion or positive clinical signs; (2) variceal bleeding: endoscopic confirmation of esophageal and gastric variceal bleeding; (3) hepatic encephalopathy: hepatic encephalopathy ≥ grade 2 diagnosed according to the West Haven criteria. Secondary outcome events included the occurrence of HCC and liver-related death. The follow-up deadline was May, 2024, or the first liver decompensation event, HCC, death, or the last follow-up (whichever occurred earliest).

Statistical analysis

Continuous variables were expressed as median (interquartile range), and categorical variables were expressed as frequency (percentage). Between-group comparisons utilized the student’s t-test, Mann–Whitney U test Mann–Whitney U test or the χ² test.

The Fine-Gray competing risk regression model was employed to analyze risk factors for liver decompensation, and univariate and multivariate regression analysis were performed to screen independent risk factors. Model performance was evaluated through the following methods: (1) Discrimination: calculation of Harrell’s concordance index (C index) and 95% confidence interval, where values > 0.7 indicate acceptable discrimination and > 0.8 indicate excellent discrimination; (2) Calibration: construction of 2-year, 4-year, and 6-year calibration plots with assessment of calibration slope and intercept, where perfect calibration is indicated by a slope of 1 and intercept of 0; (3) Overall accuracy: evaluation using Brier score, with lower values indicating better overall model performance; (4) Variable importance: likelihood ratio test (LRT) was used to assess individual variable contributions to overall model performance. Bootstrap validation analysis was performed using 1000 bootstrap samples to calculate C-index with 95% confidence intervals for both derivation and validation cohorts, providing robust estimates of model discrimination performance. Restricted cubic spline analysis was utilized to examine the dose–response relationship between LSM and liver decompensation risk.

A nomogram prediction model was constructed based on the regression coefficient. The X-tile software was used to determine the optimal cut-off value of the total points, and the patients were divided into three levels: low risk, medium risk, and high risk. The sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and area under the curve (AUC) were calculated for different cut-off thresholds. Gray’s test was used to compare the cumulative incidence of liver decompensation in patients with different risk stratification.

Based on the final prediction model, a web-based liver decompensation risk assessment tool was developed, named LSM-Based Liver Decompensation Risk Score (LBDRS). HTML, CSS and JavaScript technologies were used to achieve a user-friendly interface and automatic calculation function.

Statistical analysis was performed using R software (version 4.3.3), and P < 0.05 was considered statistically significant.

Results

Baseline characteristics

A total of 1064 patients with cACLD were included in this study, including 745 in the derivation cohort and 319 in the validation cohort (Table 1). The median age of the patients was 55 years (51–60 years), with a male predominance of 61.2%. HBV infection constituted the principal cause, accounting for 69.8%. The LSM was 18 kPa (13–30 kPa) in the overall cohort. During the follow-up time of 40 months (22–61 months), 328 patients (30.8%) developed liver decompensation with median decompensation time of 33 (16–50) months, 84 patients (7.9%) developed HCC, and 4 patients (0.4%) had liver-related deaths. There was no statistical difference between the derivation and validation cohort in all baseline characteristics (all P > 0.05).

Table 1.

Baseline characteristics of cACLD patients

	Overall cohort (n = 1064)	Derivation cohort (n = 745)	Validation cohort (n = 319)	P value
General information
Age, years	55 (51–60)	55 (51–60)	55 (50–59)	0.332
Sex (%male)	651 (61.2)	451 (60.5)	200 (62.7)	0.508
Etiology, n (%)				0.837
HBV	743 (69.8%)	513 (68.9%)	230 (72.1%)
Autoimmune liver disease	142 (13.3%)	102 (13.7%)	40 (12.5%)
HCV	68 (6.4%)	51 (6.8%)	17 (5.3%)
Alcoholic liver disease	79 (7.4%)	56 (7.5%)	23 (7.2%)
Others	32 (3.0%)	23 (3.1%)	9 (2.8%)
Outcome, n (%)				0.157
Decompensation	328 (30.8)	232 (31.1)	96 (30.1)
HCC	84 (7.9)	56 (7.5)	28 (8.8)
Liver-related death	4 (0.4)	3 (0.4)	1 (0.3)
Follow-up time, month	40 (22–61)	39 (22–61)	41 (22–63)	0.523
Clinical laboratory tests
AST, U/L	30 (24–43)	30 (24–44)	30 (24–41)	0.771
ALT, U/L	26 (19–40)	26 (19–40)	27 (18–42)	0.213
TB, μmol × L−1	18 (13–26)	18 (13–26)	18 (13–26)	0.689
ALB, g × L−1	43 (39–47)	43 (38.7–46.6)	43.1 (38–47)	0.828
GGT, U/L	33 (20–66)	33 (20–61)	33 (20–73)	0.438
Sodium, mmol/L	142 (140–144)	142 (140–143)	142 (140–144)	0.362
Creatinine, μmol × L−1	66 (56–75)	66 (56–74)	65 (57–76)	0.445
INR	1.16 (1.07–1.22)	1.16 (1.07–1.23)	1.14 (1.07–1.19)	0.213
NLR	1.89 (1.36–2.62)	1.91 (1.36–2.67)	1.86 (1.34–2.60)	0.445
HB, g × L−1	141 (125–156)	140 (124–155)	142 (128–156)	0.175
PLT, 109 × L−1	96 (64–142)	96 (64–139)	98 (64–148)	0.560
Imaging tests
LSM, kPa	18 (13–30)	19 (13–29)	18 (12–30)	0.572
CAP, dB/m	224 (196–256)	225 (196–258)	221 (199–254)	0.571
Portal vein Width, mm	12 (11–13)	12 (11–13)	12 (11–13)	0.543
Spleen diameter, mm	137 (119–156)	138 (120–156)	134 (118–156)	0.323

Medians (interquartile range, IQR) or counts (proportions). cACLD compensated advanced chronic liver disease, HCC hepatocellular carcinoma, HBV hepatitis B virus, HCV hepatitis C virus, AST aspartate aminotransferase, ALT alanine aminotransferase, TB total bilirubin, ALB albumin, GGT gamma glutamyltransferase, INR international normalized ratio, NLR neutrophil-to-lymphocyte ratio, HB hemoglobin, PLT platelets, LSM liver stiffness measurement, CAP controlled attenuation parameter

Analysis of risk factors for liver decompensation

Competing risk regression analysis was performed in the derivation cohort (Table 2). Univariate analysis identified some variables significantly associated with liver decompensation risk, including age, male, LSM, CAP, portal vein width, spleen diameter, TB, ALB, INR, NLR, HB and PLT. Subsequent multivariate analysis showed that age (aSHR = 1.04, 95% CI: 1.01–1.06, P = 0.002), LSM (aSHR = 1.03, 95% CI: 1.02–1.04, P < 0.001), spleen diameter (aSHR = 1.01, 95% CI: 1.00–1.01, P = 0.025), INR (aSHR = 2.50, 95% CI: 1.02–6.11, P = 0.045), HB (aSHR = 0.99, 95% CI: 0.98–1.00, P = 0.047) and PLT (aSHR = 1.00, 95% CI: 0.99–1.00, P = 0.007) were independent risk factors for liver decompensation. Based upon these six prognostic variables, an LSM-based risk stratification score was subsequently developed and validated.

Table 2.

Risk factors for first decompensation events in derivation cohort

	Univariate analysis		Multivariate analysis
	SHR (95% CI)	P value	aSHR (95% CI)	P value
Age	1.05 (1.02–1.07)	< 0.001	1.04 (1.01–1.06)	0.002
Sex, male	0.78 (0.61–1.00)	0.048	0.85 (0.63–1.15)	0.290
LSM	1.03 (1.03–1.04)	< 0.001	1.03 (1.02–1.04)	< 0.001
CAP	1.00 (0.99–1.00)	0.033	1.00 (0.99–1.00)	0.140
Portal vein Width	1.14 (1.08–1.21)	< 0.001	1.03 (0.96–1.10)	0.450
Spleen diameter	1.01 (1.01–1.02)	< 0.001	1.01 (1.00–1.01)	0.025
AST	1.00 (1.00–1.00)	0.100
ALT	1.00 (1.00–1.00)	0.230
TB	1.01 (1.00–1.01)	0.012	1.00 (0.99–1.01)	0.055
ALB	0.95 (0.92–0.97)	< 0.001	1.01 (0.98–1.05)	0.360
GGT	1.00 (1.00–1.00)	0.330
Na	0.98 (0.94–1.03)	0.480
Creatinine	1.00 (0.98–1.01)	0.360
INR	6.67 (3.24–13.75)	< 0.001	2.50 (1.02–6.11)	0.045
NLR	1.15 (1.10–1.20)	< 0.001	1.05 (1.00–1.10)	0.062
HB	0.98 (0.98–0.99)	< 0.001	0.99 (0.98–1.00)	0.047
PLT	0.99 (0.99–0.99)	< 0.001	1.00 (0.99–1.00)	0.007

SHR subdistribution hazard ratios, aSHR adjusted subdistribution hazard ratios, CI confidence intervals, LSM liver stiffness measurement, CAP controlled attenuation parameter, AST aspartate aminotransferase, ALT alanine aminotransferase, TB total bilirubin, ALB albumin, GGT gamma glutamyltransferase, INR international normalized ratio, NLR Neutrophil-to-lymphocyte ratio, HB hemoglobin, PLT platelets. Bold values indicate statistical significance (P < 0.05).

Performance evaluation of LSM-based risk score

The performance of LSM-based risk score was evaluated by the calibration curves of liver decompensation risk at 2, 4, and 6 years (Fig. 1). In the derivation cohort (Fig. 1a) and the validation cohort (Fig. 1b), the calibration plots showed that the model-predicted probability was well consistent with the actual observed probability, indicating that the model had good calibration in the two cohorts. Our results demonstrate excellent overall accuracy with Brier scores of 0.139 and 0.153 in the derivation and validation cohort, respectively. Bootstrap validation analysis demonstrated robust model performance in the derivation cohort with C-index of 0.861 (95% CI: 0.804–0.913). In the validation cohort, the bootstrap C-index was 0.828 (95% CI: 0.727–0.918), indicating consistent discrimination performance across both cohorts.

Fig. 1.

Calibration performance of the LSM-based risk prediction model a Calibration curves for the liver decompensation risk prediction at 2 years (purple), 4 years (orange), and 6 years (blue) in the derivation cohort. b Calibration curves for the liver decompensation risk prediction at 2 years (purple), 4 years (orange), and 6 years (light blue) in the validation cohort. The gray diagonal line represents the ideal calibration line. LSM, liver stiffness measurement (Color figure online)

To evaluate the contribution of each variable to the performance of the prediction model, the LRT was used to analyze the importance of the model variables (Table 3). The C index of the complete six-variable model was 0.779 (95% CI: 0.714–0.845). Sequential removal analysis revealed that the C index of the model after removing LSM was significantly reduced to 0.739 (95% CI: 0.666–0.811, LRT = 47.99, P < 0.001), indicating that the removal of LSM had the most significant impact on the prediction performance of the model. In contrast, the effect of removing the remaining variables on model performance was relatively small: removing spleen diameter, the C index was 0.776 (LRT = 5.37, P = 0.021); removing PLT, the C index was 0.768 (LRT = 17.05, P < 0.001); removing HB, the C index was 0.767 (LRT = 7.76, P = 0.005); removing INR, the C index was 0.777 (LRT = 8.13, P = 0.004); and removing age, the C index was 0.779 (LRT = 14.04, P < 0.001).

Table 3.

Assessment of the impact of variables on the model’s effectiveness

	C-index (95% CI)	LRT	P value
Age, LSM, Spleen diameter, PLT, HB, INR	0.779 (0.714–0.845)	Reference	Reference
Age, Spleen diameter, PLT, HB, INR	0.739 (0.666–0.811)	47.99	< 0.001
Age, LSM, PLT, HB, INR	0.776 (0.710–0.841)	5.37	0.021
Age, LSM, Spleen diameter, HB, INR	0.768 (0.698–0.838)	17.05	< 0.001
Age, LSM, Spleen diameter, PLT, INR	0.767 (0.700–0.835)	7.76	0.005
Age, LSM, Spleen diameter, PLT, HB	0.777 (0.744–0.809)	8.13	0.004
LSM, Spleen diameter, PLT, HB, INR	0.779 (0.717–0.842)	14.04	< 0.001

LRT Likelihood ratio test, INR international normalized ratio, PLT platelet, HB hemoglobin, LSM liver stiffness measurement

Given the pivotal contribution of LSM to the prediction model, the dose–response relationship between LSM and liver decompensation risk was further analyzed. In the derivation cohort (Fig. 2a), LSM was nonlinearly related to the risk of liver decompensation (P for overall = 3.13 × 10⁻¹⁰, P for nonlinear = 0.035). When the LSM value was lower than 20 kPa, the SHR increased relatively slowly. However, once LSM exceeded 20 kPa, the SHR showed a steeper upward trend, indicating that liver decompensation risk in patients with high LSM values was significantly increased. In the validation cohort (Fig. 2b), the significant nonlinear relationship between LSM and liver decompensation risk was also confirmed (P for overall = 1.45 × 10⁻¹⁰, P for nonlinear = 0.020).

Fig. 2.

Relationship between LSM and liver decompensation risk a Restricted cubic spline plot illustrating the nonlinear association between LSM and risk of decompensation in the derivation cohort. b Restricted cubic spline plot between LSM and risk of decompensation in the validation cohort. LSM, liver stiffness measurement; SHR, subdistribution hazard ratio; CI, confidence interval

Construction of LSM-based risk score for predicting liver decompensation risk

Utilizing the six independent predictors identified through multivariable competing risk regression modeling, a nomogram for predicting liver decompensation risk was constructed (Fig. 3a). The nomogram integrates age, LSM, spleen diameter, HB, PLT, and INR variables generate individualized risk estimates at 2-, 4-, and 6-year. In the derivation cohort (Fig. 3b) and validation cohort (Fig. 3c), the total points of patients in the decompensated group were significantly higher than those in the non-decompensated group (all P < 0.001).

Fig. 3.

LSM-based nomogram for liver decompensation risk prediction a Predictive nomogram incorporating six independent risk factors to estimate decompensation probability at 2-, 4-, and 6-year. b Total points distribution comparison between patients with and without decompensation in the derivation cohort. c Total points distribution comparison between patients with and without decompensation in the validation cohort. LSM, liver stiffness measurement; HB, hemoglobin; PLT, platelet count; INR, international normalized ratio

To establish a risk stratification system based upon total points of the nomogram, X-tile software was used to identify optimal cut-off values. Based on the X-tile analysis (Fig. 4a, b), patients were divided into three risk levels: low-risk group (≤ 147.7 points), medium-risk group (147.7–206.6 points), and high-risk group (≥ 206.6 points), with decompensation rates of 13.4%, 58.0%, and 86.7%, respectively, and median time to decompensation of 37 (18–55), 33 (15–50), and 28 (17–48) months, respectively. Performance characteristics of these threshold values were subsequently evaluated across 2-, 4-, and 6-year liver decompensation prediction in the derivation and validation cohort (Table 4). For the derivation cohort, the 147.7-point cut-off demonstrated sensitivities of 0.81, 0.81, and 0.78 at 2-, 4-, and 6-year, with corresponding specificities of 0.63, 0.63, and 0.67. The respective AUC values were 0.72 (95% CI: 0.67–0.77), 0.72 (95% CI: 0.68–0.76), and 0.73 (95% CI: 0.67–0.79), respectively. When applying the 206.6-point cut-off, the sensitivity and AUC were relatively reduced, but the specificity was improved. In the validation cohort, the predictive performance of the 147.7-point cut-off remained comparable to the derivation cohort, while the performance of the cut-off of 206.6 points was poor.

Fig. 4.

Establishment of optimal threshold values and cumulative decompensation incidence of risk stratification a X-tile software analytical output employed to establish the optimal threshold value for the nomogram total points. b Distribution histogram displaying patient total points based on the established optimal threshold value. c Cumulative decompensation incidence curves of patients with different risk stratifications in the derivation cohort. d Cumulative decompensation incidence curves of patients with different risk stratifications in the validation cohort

Table 4.

Accuracy for prediction of liver decompensation using the LSM-based score cut-off values of 147.7 and 206.6

		Derivation cohort		Validation cohort
	Cut-off value	147.7	206.6	147.7	206.6
2-year prediction	Sensitivity	0.81	0.43	0.68	0.22
	Specificity	0.63	0.91	0.67	0.93
	PPV	0.22	0.39	0.19	0.26
	NPV	0.96	0.92	0.95	0.91
	AUC	0.72	0.67	0.67	0.57
4-year prediction	Sensitivity	0.81	0.31	0.67	0.21
	Specificity	0.63	0.91	0.68	0.91
	PPV	0.41	0.53	0.39	0.43
	NPV	0.91	0.80	0.87	0.79
	AUC	0.72	0.61	0.67	0.56
6-year prediction	Sensitivity	0.78	0.28	0.66	0.18
	Specificity	0.67	0.97	0.67	0.93
	PPV	0.70	0.89	0.63	0.69
	NPV	0.76	0.58	0.70	0.58
	AUC	0.73	0.62	0.67	0.56

LSM liver stiffness measurement, AUC area under curve, PPV positive predictive value, NPV negative predictive value

According to risk point stratification, patients were divided into three groups, with subsequent analysis of cumulative decompensation incidence across different time points in each group. In the derivation cohort (Fig. 4c) and validation cohort (Fig. 4d), the cumulative incidence of decompensation in the low-risk, medium-risk, and high-risk group was statistically significant (Gray’s test P < 0.001). In the derivation cohort (Table 5), the cumulative decompensation incidence in patients in the low-risk group at 2, 4, and 6 years was 3.8% (95% CI: 1.8–5.8%), 9.0% (95% CI: 5.6–12.3%), and 21.3% (95% CI: 14.7–28.0%), respectively. The cumulative incidence of patients in the intermediate-risk group was significantly increased, which were 17.4% (95% CI: 11.8–23.0%), 40.1% (95% CI: 32.4–47.8%), and 70.3% (95% CI: 62.0–78.7%), respectively. Patients in the high-risk group showed the highest risk of decompensation, with cumulative incidence of 43.2% (95% CI: 32.0–54.5%), 61.1% (95% CI: 49.7–72.5%), and 90.9% (95% CI: 83.5–98.3%) at 2, 4, and 6 years, respectively, with significant differences among the three groups (P < 0.001). In the validation cohort (Table 5), the cumulative incidence patterns largely concordant with the derivation cohort across all risk strata (P < 0.001).

Table 5.

Cumulative incidence of first decompensation events

	2-year cumulative incidence (%)	4-year cumulative incidence (%)	6-year cumulative incidence (%)	P value
Derivation cohort				< 0.001
Low risk	3.8 (1.8–5.8)	9.0 (5.6–12.3)	21.3 (14.7–28.0)
Median risk	17.4 (11.8–23.0)	40.1 (32.4–47.8)	70.3 (62.0–78.7)
High risk	43.2 (32.0–54.5)	61.1 (49.7–72.5)	90.9 (83.5–98.3)
Validation cohort				< 0.001
Low risk	5.3 (1.9–9.6)	12.7 (6.8–18.6)	26.5 (16.1–36.9)
Median risk	19.2 (10.1–28.2)	43.6 (30.3–56.9)	69.8 (56.4–83.3)
High risk	31.5 (12.1–51.0)	60.9 (39.9–81.8)	78.6 (59.4–97.8)

Web-based tool for liver decompensation risk assessment: LBDRS

To improve the clinical practicality and accessibility of the prediction model, a web-based liver decompensation risk assessment tool (LBDRS, LSM-Based Liver Decompensation Risk Score) was developed (Fig. 5). LBDRS (https://lbdrs.netlify.app/) has a user-friendly interface with straightforward functionality. It incorporates two primary functional modules: risk evaluation and outcome presentation. On the risk evaluation page (Fig. 5a), users need to input six essential clinical parameters: age (years), LSM (kPa), spleen diameter (mm), PLT (× 10⁹/L), HB (g/L), and INR. The platform automatically calculates the total points and performs risk stratification based on the input parameters. The results page (Fig. 5b) clearly displays the patient’s total points, risk classification, and corresponding 2-, 4-, and 6-year decompensation probability estimates. In addition, LBDRS also provides targeted clinical recommendations for each risk stratum, providing a reference for clinicians to develop individualized management strategies.

Fig. 5.

Web-based LSM-derived liver decompensation risk score (LBDRS) assessment platform interface An example outlining the process and results of developing a web-based risk assessment tool based on input and output. a Risk assessment test page, showing the user’s data input interface. b Results display page, showing the patient’s total score, risk stratification, corresponding 2-year, 4-year, and 6-year decompensation risk prediction values and clinical management recommendations

Discussion

Based on a cohort of 1064 patients with cACLD, this study constructed a multifactorial liver decompensation risk prediction model incorporating LSM as a core component, and developed a corresponding networked risk assessment tool (LBDRS). The LSM-based risk score performed well in identifying high-risk patients.

Conventional liver prognostic assessment models have a certain role in risk stratification of compensated patients. Model for End-Stage Liver Disease (MELD) 3.0 performed best in predicting mortality and liver-related complications in patients with cirrhosis, with an AUC of 0.851. This performance metric significantly surpasses that of established scoring systems, including Child–Pugh classification, albumin–bilirubin grade, original MELD, and MELD-Na [28]. MELD 3.0 and MELD-Na are reliable non-invasive tools for predicting long-term mortality and rebleeding risk after acute variceal bleeding in patients with cirrhosis. Notably, MELD 3.0 maintained a mean AUC of 0.789 that remained above 0.8 for up to 18 months [29]. In addition, there are some studies evaluating other non-invasive indicators. The FIB-4 index can better evaluate the degree of liver fibrosis by integrating age, aspartate aminotransferase, alanine aminotransferase and PLT [30, 31]. Similarly, the aspartate aminotransferase to platelet ratio index (APRI) is another widely used fibrosis score in chronic liver diseases [32, 33]. However, these non-invasive fibrosis scores focus mainly on quantifying fibrosis severity rather than prognosticating decompensation risk, highlighting the clinical imperative for more sophisticated prognostic approaches. Dynamic LSM changes can effectively predict clinical outcomes in patients with MASLD and primary biliary cholangitis [34–36]. A multicenter study suggested that the ABID-LSM model has higher accuracy in predicting liver decompensation in patients with MASLD-induced cirrhosis [19]. LSM is also associated with HCC risk in MASLD and chronic hepatitis patients [21, 37]. LSM-based prediction models for decompensation events in patients with cACLD of different etiologies require further investigation.

LSM represents a pivotal biomarker in predicting liver decompensation. Our likelihood ratio analysis demonstrated that removal of LSM had the most significant impact on the model’s prediction performance, with the C index declining significantly from 0.779 to 0.739. This significant reduction fully confirmed the core position of LSM in risk prediction models. These findings align closely with emerging evidence from research investigating non-invasive assessment strategies across diverse liver pathologies. In patients with MASLD, a two-step non-invasive approach using FIB-4 and LSM can effectively stratify the risk of different liver disease-related events [20]. Similarly, among patients with autoimmune hepatitis, LSM can predict the high-risk subgroups predisposed to developing cirrhosis and decompensation [38]. Furthermore, in compensated alcoholic liver disease, LSM can be used to monitor the risk of decompensation and mortality [15]. LSM is also an independent risk factor for liver decompensation, liver transplantation, or liver-related death in patients with primary biliary cholangitis [39]. More importantly, our study revealed a nonlinear dose–response relationship between LSM and liver decompensation risk. When the LSM value was lower than 20 kPa, the risk increased relatively slowly; however, when the LSM value exceeded 20 kPa, the risk showed a steeper upward trend, providing an important threshold reference for clinical risk stratification.

Our study has constructed a multifactorial prediction model, integrating age, LSM, spleen diameter, HB, PLT, and INR. This integrated approach had better prediction accuracy for liver decompensation. Compared with the currently widely used non-invasive fibrosis scores, our model showed obvious advantages. By integrating LSM with other clinical parameters, LSM-based risk score can not only assess the degree of fibrosis, but more importantly, accurately predict decompensation risk. The clinical utility of our prognostic algorithm is exemplified through its implementation as a tripartite risk stratification framework. Patients are categorized into three risk levels: low risk, intermediate risk, and high risk. This three-classification system not only ensures the significance of risk differences between groups, but also maintains the operability of clinical decision-making. This risk stratification system performed well in predictive performance with high sensitivities and AUC, showing good discrimination ability. Although the cut-off of 206.6 reduced sensitivity, it significantly improved specificity, providing a more accurate tool for identifying high-risk patients in the clinic. The cumulative decompensation rates of patients in the low-risk group were 3.8%, 9.0%, and 21.3% at 2, 4, and 6 years, respectively, showing a relatively good short- to medium-term prognosis. Conversely, patients in the high-risk group showed markedly elevated decompensation propensity, with cumulative decompensation rates of 43.2%, 61.1%, and 90.9% at 2, 4, and 6 years, respectively. This risk stratification system provides a valuable clinical decision support tool for the individualized management of cACLD patients. The robust validation results provide strong evidence for the clinical reliability and generalizability of our prediction model. Both derivation and validation cohorts demonstrated C-index values exceeding the 0.7 threshold for acceptable clinical prediction models, while Brier scores below 0.25 across all time points indicated excellent overall accuracy. The narrow bootstrap confidence intervals further demonstrated stable model performance with minimal uncertainty, strongly supporting the clinical utility and practical implementation of our LBDRS tool in diverse healthcare settings.

The LBDRS networked risk assessment tool developed in our study represents an important advancement in the transformation of liver disease management to precision medicine. Compared with existing scoring systems, LBDRS has the advantages of simple operation, accurate calculation, and real-time update. These features collectively empower clinicians to expedite risk stratification processes and develop individualized surveillance protocols and treatment strategies. The development of LBDRS aligns with contemporary trends emphasizing the expanding role of non-invasive diagnostic modalities in hepatic fibrosis management. By converting complex statistical models into user-friendly network tools, LBDRS are expected to be widely used in clinical practice, providing standardized risk assessment methods for diverse healthcare settings. Furthermore, the widespread deployment of LBDRS is projected to more informed clinical decision-making, optimized resource allocation, and improved patient counseling regarding disease trajectory and therapeutic options.

However, this study also has some limitations. First, this is a single-center retrospective study, which may have selection bias. Although the results of the validation cohort are basically consistent with the derivation cohort, multicenter prospective studies are still needed to further verify the external validity of the model. Secondly, the median follow-up time of this study was 40 months, and the ability to predict long-term prognosis still needs further observation. In addition, this study mainly included patients with hepatitis B virus infection, accounting for 69.8%. Given the distinct epidemiological patterns between Asian and Western populations, the applicability for cACLD patients with other etiologies, particularly MASLD patients who may have different metabolic profiles and progression patterns, needs further external validation and potential model recalibration incorporating MASLD-specific metabolic parameters. Based on these limitations, future studies should conduct large-scale multicenter prospective cohort studies to further verify and optimize the prediction model; explore the relationship between dynamic LSM changes and disease progression; and conduct intervention studies to evaluate the improvement effect of individualized management strategies based on risk stratification on patient prognosis.

Conclusion

This study constructed an LSM-based risk score for predicting liver decompensation in cACLD. The model demonstrated excellent discrimination and calibration properties, enabling effective identification of high-risk patients. The novel LBDRS network risk assessment tool provides crucial clinical decision support for the individualized management of cACLD patients. These findings may have broad clinical applicability, offering valuable guidance for delaying disease progression and improving patient prognosis. Future improvements should focus on multicenter prospective validation across diverse populations and etiologies, integration with emerging biomarkers and artificial intelligence technologies, and prospective intervention studies to validate clinical utility in improving patient outcomes.

Author contributions

Xianbo Wang, Ying Feng designed the manuscript. Yanqiu Li drafted the manuscript. Zihang Qiao and Jinze Li carefully reviewed the manuscript. Yongqi Li drew the figures. All authors approved the final version of the manuscript.

Funding

This research was funded by the High-level Chinese Medicine Key Discipline Construction Project (No. zyyzdxk-2023005, to XBW); Capital’s Funds for Health improvement and Research (No. 2024–1-2173, to XBW); National Natural Science Foundation of China (No. 82474419, to XBW and 82474426, to YF); Beijing Municipal Natural Science Foundation (No. 7232272, to YF); and Beijing Traditional Chinese Medicine Technology Development Fund Project (No. BJZYZD-2023–12, to XBW).

Data availability

The datasets used and analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Conflicts of interest

The authors declare no competing interests.

Ethics approval

All procedures followed were in accordance with the 1975 Helsinki Declaration, as revised in 2008. The study protocol was approved by the Ethics Committee of Beijing Ditan Hospital (approval number: DTEC-KY2024-069–01).

Consent to participate

Informed consent was obtained from all individual participants included in the study.