Threshold‐Defined Liver Stiffness Trajectories and Liver‐Related Event Risk in Chronic Liver Disease: A Cohort Study

ABSTRACT

Background & Aims

Vibration‐controlled transient elastography (VCTE) predicts liver‐related events (LREs), but how serial liver stiffness (LS) changes should be translated into clinical risk—whether by their magnitude, direction or movement across established thresholds—remains uncertain. We tested whether a threshold‐defined trajectory framework anchored at the Baveno VII 10 kPa criterion would stratify LRE risk across chronic liver disease aetiologies.

Methods

From the retrospective V‐LINK registry (2006–2020), 6313 event‐free patients with baseline and 1‐year LS were stratified by the 10 kPa threshold into Stable Low, Improved, Worsened or Persistent High trajectories. A 3‐year landmark (LM3) subset (n = 1617) enabled trajectory reassessment over time. Cox models estimated adjusted hazard ratios (HRs).

Results

During a median follow‐up of 5.5 years, 472 LREs occurred. Compared with Stable Low, adjusted HRs were 1.5 (95% confidence interval [CI], 1.1–2.2) for Improved, 2.5 (1.5–4.3) for Worsened and 4.1 (3.1–5.5) for Persistent High, with 3‐year risks of 1.2%, 3.2%, 3.8%, and 15.0%. Persistent High (14.4% of the cohort) accounted for 61.9% of LREs; 65.7% showed a numerical decrease in LS yet remained ≥ 10 kPa. By aetiology, the HR for Persistent High was 6.4 (4.0–10.2) in non‐viral aetiologies versus 2.6 (1.8–3.8) in viral aetiologies (p for interaction < 0.001). At LM3, the trajectory‐based risk gradient was preserved (Persistent High HR, 3.5; 95% CI, 2.2–5.5).

Conclusions

Threshold‐defined LS trajectories identified a small subgroup concentrating most LRE risk, with the largest gradient in non‐viral aetiologies. Persistent threshold elevation was the most robust prognostic signal, supporting serial LS monitoring pending external validation.

Key Points

• A simple 10 kPa threshold classified serial liver stiffness into four clinically interpretable trajectories with clearly separated liver‐related event risks.

• Patients with Persistent High trajectories comprised 14.4% of the cohort but accounted for 61.9% of liver‐related events and remained at high risk even when liver stiffness decreased.

• The prognostic gradient across trajectories was substantially stronger in non‐viral than viral liver disease.

• Trajectory status changed over time, and serial trajectory information improved risk discrimination and clinical net benefit beyond conventional clinical variables and baseline liver stiffness.

Lay Summary

For people with chronic liver disease, repeated liver stiffness tests may provide useful information about the risk of future liver‐related complications. We found that patients whose liver stiffness remained above an important clinical threshold had the highest risk, even when their liver stiffness had decreased over time. This simple approach may help clinicians decide how closely patients should be followed and monitored.

Abbreviations

AUC

area under the receiver operating characteristic curve

BMI

body mass index

cACLD

compensated advanced chronic liver disease

CI

confidence interval

DM

diabetes mellitus

HBV

hepatitis B virus

HCC

hepatocellular carcinoma

HCV

hepatitis C virus

HR

hazard ratio

IQR

interquartile range

KM

Kaplan–Meier

LM1

1‐year landmark

LM3

3‐year landmark

LRE

liver‐related event

LS

liver stiffness

MASLD

metabolic dysfunction‐associated steatotic liver disease

NAFLD

nonalcoholic fatty liver disease

NR

not reported

SE

standard error

VCTE

vibration‐controlled transient elastography

V‐LINK

Vibration‐controlled Transient Elastography‐based Liver Fibrosis Investigation Network

1. Introduction

Liver stiffness (LS) measured by vibration‐controlled transient elastography (VCTE) is increasingly used for risk stratification in patients with chronic liver disease [1, 2, 3, 4, 5]. The Baveno VII consensus established LS < 10 kPa as a reliable criterion for ruling out compensated advanced chronic liver disease (cACLD) and introduced the ‘Rule of Five’ for graded risk stratification [6]. Beyond diagnosis, serial LS assessment is increasingly recommended for longitudinal monitoring because repeated measurements may capture treatment response, changes in disease activity and fibrosis dynamics over time [6, 7, 8, 9]. However, how serial LS changes should be translated into clinically actionable event risk remains uncertain. In particular, it is unclear whether prognosis is best captured by the most recent LS value, the absolute magnitude or direction of LS change, or movement across clinically established thresholds.

Several studies have explored the prognostic value of serial LS measurements. Joint modelling of repeated LS measurements in cACLD showed that individualized stiffness trajectories predicted hepatic decompensation and mortality [10], and disease‐specific studies have supported the prognostic relevance of LS dynamics in nonalcoholic fatty liver disease (NAFLD) [11], primary biliary cholangitis [12] and chronic hepatitis B [13]. In contrast, Wong et al. [14] reported that once the current LS value is known, previous measurements add little predictive value for LREs in cACLD. These apparently divergent findings may reflect differences in study population and analytic strategy: many studies have been restricted to cACLD populations [10, 14] or single aetiologies [11, 12, 13, 15], whereas others have relied on continuous delta‐LS metrics or complex joint models that are less easily implemented in routine practice. A simple threshold‐defined framework may therefore be useful, particularly if it can distinguish patients who remain above a clinically meaningful LS threshold despite numerical decreases from those who truly move into a lower risk range.

Whether such a threshold‐defined LS trajectory can provide clinically actionable risk stratification across disease stages and aetiologies remains unknown. In this study, we classified LS trajectories into 4 clinically intuitive categories—Stable Low, Improved, Worsened, and Persistent High—using the Baveno VII 10 kPa threshold at sequential landmark time points. Using a large registry spanning the full spectrum of chronic liver disease aetiologies, we evaluated whether these trajectory groups provided a clinically interpretable framework for 3‐year LRE risk stratification and examined aetiology‐specific differences in trajectory‐based prognostic performance.

2. Methods

2.1. Study Design and Population

This retrospective longitudinal study used data from the V‐LINK 2025_v1.0_YUHS_SC, which is part of the Vibration‐controlled Transient Elastography‐based Liver Fibrosis Investigation Network (V‐LINK) established in 2025 as a multicenter, retrospective, longitudinal cohort for evaluating liver fibrosis using VCTE across a broad spectrum of chronic liver diseases (metabolic dysfunction‐associated steatotic liver disease (MASLD), chronic hepatitis B and C, autoimmune hepatitis and alcohol‐related liver disease). The primary aim of the V‐LINK cohort is to investigate the diagnostic and prognostic utility of noninvasive tests, particularly VCTE, in the assessment of liver fibrosis. The cohort is subject to regular updates, either through scheduled follow‐up for outcome ascertainment or on a proposal‐driven basis, in accordance with investigator‐initiated research needs.

This study included adults who underwent VCTE between April 2006 and December 2020. Patients were required to have both baseline and 1‐year LS data to be eligible for the 1‐year landmark (LM1) cohort. The exclusion criteria were missing baseline and/or 1‐year LS data, LREs before the 1‐year landmark, LM1 date beyond the study end, missing covariates and invalid follow‐up (Figure 1). A subset of LM1 patients who additionally had 3‐year LS data and remained event‐free at the 3‐year landmark constituted the LM3 cohort (Figure S1).

FIGURE 1.

Patient selection from the V‐LINK registry. Flowchart showing derivation of the 1‐year landmark (LM1) cohort (n = 6313) with stepwise exclusion criteria. LRE, liver‐related event; LS, liver stiffness. Among the 27 236 patients excluded because of missing liver stiffness data, 26 616 lacked only the 1‐year measurement, 286 lacked only the baseline measurement and 334 lacked both measurements.

This study was approved by the Institutional Review Board of Severance Hospital, Seoul, Republic of Korea (IRB No. 4‐2025‐0032). All research was conducted in accordance with the Declaration of Helsinki and the Declaration of Istanbul. The requirement for written informed consent was waived by the Institutional Review Board owing to the retrospective design.

2.2. VCTE

LS was assessed on a VCTE machine (FibroScan; EchoSens, Paris, France) by trained operators following a standardized protocol [16]. Patients were required to have at least 10 valid acquisitions with an interquartile range (IQR)/median ratio of less than 30% and a success rate of at least 60% for the measurements to be considered reliable. The LS reading acquired closest to each landmark was selected, with a preferred window of 6–24 months for the 1‐year and 24–48 months for the 3‐year landmark. The median interval from the index LS measurement to the 1‐year LS measurement was 12.3 months (IQR, 10.8–14.5; range, 5.1–18.9), and the median interval to the 3‐year LS measurement was 36.0 months (IQR, 33.1–38.8; range, 29.1–42.9).

2.3. Threshold‐Defined Liver Stiffness Trajectory Classification

LS trajectories were defined using the 10 kPa threshold, consistent with the Baveno VII cACLD rule‐out criterion [6]. LM1 trajectories were classified using the change from baseline to 1‐year LS, and LM3 transitions were classified using the change from 1‐year to 3‐year LS. Four categories were used: Stable Low (both time points < 10 kPa), Improved (≥ 10 to < 10 kPa), Worsened (< 10 to ≥ 10 kPa), and Persistent High (both time points ≥ 10 kPa).

2.4. Outcomes

The primary outcome was the occurrence of LREs, defined as a composite of hepatocellular carcinoma (HCC), hepatic decompensation (ascites, variceal haemorrhage, hepatic encephalopathy, or hepatorenal syndrome), liver transplantation and liver‐related death. The date of the earliest qualifying event was used as the LRE date. Patients were followed from the respective landmark date (LM1 or LM3) until the occurrence of an LRE or last follow‐up, whichever came first.

2.5. Prediction Model Development and Validation

To evaluate the incremental prognostic value of serial LS monitoring, dynamic prediction models for 3‐year LRE risk were developed in the LM3 cohort (n = 1617). A base model included clinical variables (age, sex, diabetes mellitus, serum albumin and platelet count). Sequential models were constructed by adding baseline LS alone, LM1‐based predictors (1‐year LS status as a binary variable [< 10 vs. ≥ 10 kPa] or LM1 trajectory as a 4‐category variable), and LM3‐updated predictors (3‐year LS status as a binary variable or LM3 transition as a 4‐category variable). Predicted 3‐year risk was estimated as 1 minus the Cox model–based survival probability at 3 years, with models fitted over the full follow‐up period.

Model discrimination was assessed using the Harrell C‐index, with 95% confidence intervals (CIs) estimated using the percentile bootstrap with 500 iterations. Time‐dependent discrimination was further evaluated using the time‐dependent area under the receiver operating characteristic curve (AUC), calculated at 6‐month intervals from LM3 using inverse probability of censoring weighting and applied to four nested prediction models (clinical only; clinical with baseline LS; clinical with baseline LS and LM1 trajectory; and clinical with baseline LS and LM3 transition). Calibration was assessed by comparing the predicted versus observed 3‐year risk across deciles of predicted risk, with the calibration slope and intercept derived from a logistic calibration regression. Observed risk in the calibration assessment was estimated using the Kaplan–Meier method at 3 years within each decile. A decision curve analysis [17] was performed in the LM1 cohort (n = 6313), comparing clinical‐only, clinical plus baseline LS, and clinical plus baseline LS plus LM1 trajectory models, using 3‐year predicted probabilities from each model to evaluate the clinical utility of the models by estimating the net benefit across a range of clinically relevant threshold probabilities (5%–20%), with 5%–15% highlighted as the prespecified clinically relevant range. Full regression coefficients of the final LM3‐updated prediction model are reported to enable external reproducibility (Table S1).

2.6. Statistical Analysis

Baseline characteristics are summarized as the median (IQR) for continuous variables and number (percentage) for categorical variables. Differences between the LM1 and LM3 cohorts were compared using the Wilcoxon rank‐sum test and the chi‐squared test for continuous and categorical variables, respectively. A landmark analysis framework was used to avoid immortal time bias [18]. Kaplan–Meier survival curves for LRE‐free survival were constructed for each trajectory group, with differences assessed using the log‐rank test. Adjusted hazard ratios (HRs) with 95% CIs were calculated using Cox proportional hazards regressions over the entire follow‐up period, with the Stable Low group as the reference and adjusting for age, sex, diabetes mellitus, serum albumin, platelet count and baseline LS. Covariates were selected a priori based on their established prognostic relevance in chronic liver disease. Subgroup analyses assessed whether the prognostic effect of the LM1 trajectory varied by aetiology (viral vs. non‐viral), sex or diabetes mellitus status. Within each subgroup, separate Cox models were fitted with the same covariates, excluding the subgroup‐defining variable. Viral aetiology was defined as hepatitis B or hepatitis C; non‐viral aetiology included MASLD, alcohol‐related liver disease and other aetiologies. Aetiology was further stratified into HBV, HCV, alcohol‐related liver disease, MASLD and other aetiologies, classified hierarchically (HBV > HCV > alcohol‐related liver disease > MASLD > other) from index‐date aetiology codes. Interactions were assessed using the likelihood ratio test. In the subgroup analyses, standard Cox estimates were not reported for trajectory categories with fewer than five liver‐related events and were labelled NR. For aetiology‐specific cells with sparse events, Firth penalized‐likelihood Cox models were additionally fitted using the coxphf package.

To assess robustness of the threshold‐based framework to the choice of cutoff, we additionally re‐classified trajectories using alternative LS thresholds of 8 and 15 kPa and re‐estimated the trajectory hazard ratios. To address potential regression to the mean, we additionally performed (i) borderline‐exclusion sensitivity analyses removing patients with baseline LS within ±1, ±2, or ±3 kPa of the 10 kPa threshold and (ii) reclassification of trajectories using a Galton‐type correction in which the 1‐year LS was adjusted for its linear regression on baseline LS. Within the Improved group, we further examined a dose–response across the magnitude of LS change by stratifying patients into Small (|ΔLS| ≤ 2 kPa), Moderate (2 < |ΔLS| ≤ 5 kPa), and Large (|ΔLS| >5 kPa) categories and re‐estimating hazard ratios against the original Stable Low reference; a linear trend was tested using a continuous |ΔLS| term within the Improved subgroup. To gauge the influence of follow‐up completeness and outcome ascertainment, we additionally estimated the reverse Kaplan–Meier follow‐up, repeated the trajectory analysis in patients with adequate follow‐up (≥ 3 years of follow‐up or an observed event), and re‐fitted the model using incident hepatocellular carcinoma alone as the sole outcome, as HCC represents the most discretely defined and uniformly ascertained component of the composite endpoint. To complement the trajectory framework, we additionally computed the proportion of patients within each trajectory or transition group with a relative LS change of at least 30% in either direction and constructed a 4 × 4 reclassification matrix between the LM1 trajectory and the LM3 transition.

To assess the effects of the 1‐year LS measurement window and potential immortal‐time bias related to measurement timing, sensitivity analyses restricted the window to 6–18 months, re‐anchored follow‐up at the actual 1‐year LS measurement date, and excluded events occurring within 6 months after that measurement. Multivariable logistic regression models were used to identify baseline factors associated with Improved and Worsened LM1 trajectories. The models simultaneously included age, sex, diabetes mellitus, serum albumin, platelet count, baseline LS, viral aetiology, and BMI and were fitted among patients with complete covariate data. Among patients with baseline LS ≥ 10 kPa, Improved was modelled against Persistent High; among those with baseline LS < 10 kPa, Worsened was modelled against Stable Low.

All analyses were performed using R version 4.5.2 (R Foundation for Statistical Computing) and the survival, timeROC and coxphf packages. A 2‐sided p value of less than 0.05 was considered statistically significant.

3. Results

3.1. Study Population

After excluding 28 656 patients according to our exclusion criteria, 6313 met the eligibility criteria for the LM1 analysis (Figure 1). The median age was 53 years (IQR, 44–61), 59.9% were male, and the predominant aetiology was MASLD (46.4%), followed by hepatitis B (27.6%). The median baseline LS was 6.1 kPa (IQR, 4.6–9.2) (Table 1). During a median follow‐up of 5.5 years (IQR, 3.8–7.6) from LM1, 472 LREs occurred.

TABLE 1.

Baseline characteristics of the LM1 cohort ^a (n = 6313).

Variables	Values
Age, year, median (IQR)	53 (44–61)
Male sex, no. (%)	3781 (59.9)
BMI, kg/m2, median (IQR)	24.8 (22.7–27.5)
Diabetes mellitus, no. (%)	1498 (23.7)
Hypertension, no. (%)	2032 (32.2)
Aetiology, no. (%)
Hepatitis B	1744 (27.6)
Hepatitis C	247 (3.9)
Alcohol	409 (6.5)
MASLD	2929 (46.4)
Other	984 (15.6)
Autoimmune hepatitis	175 (2.8)
Primary biliary cholangitis	73 (1.2)
Wilson disease	53 (0.8)
Cryptogenic/unspecified	683 (10.8)
LS, kPa, median (IQR)	6.1 (4.6–9.2)
ALT, U/L, median (IQR)	31 (20–53)
AST, U/L, median (IQR)	30 (22–46)
Serum albumin, g/dL, median (IQR)	4.4 (4.2–4.6)
Platelet count, ×103/μL, median (IQR)	214 (168–262)

Note: Aetiologies were mutually exclusive and classified hierarchically as hepatitis B > hepatitis C > alcohol‐related liver disease > MASLD > other. Values are presented as median (IQR) for continuous variables and No. (%) for categorical variables.

Abbreviations: ALT, alanine aminotransferase; AST, aspartate amino transferase; BMI, body mass index; IQR, interquartile range; LM1, 1‐year landmark; LS, liver stiffness; MASLD, metabolic dysfunction‐associated steatotic liver disease.

^a The LM1 cohort included patients with both baseline and 1‐year LS data who were free of liver‐related events at the 1‐year landmark.

3.2. LRE According to LS Trajectory

At LM1, most patients were classified as Stable Low (n = 4698; 74.4%), followed by Persistent High (n = 907; 14.4%), Improved (n = 543; 8.6%) and Worsened (n = 165; 2.6%). In the multivariable Cox regression adjusting for age, sex, diabetes mellitus, serum albumin, platelet count and baseline LS, all trajectory groups had significantly higher HRs for LREs than the Stable Low group: Improved, 1.5 (95% CI, 1.1–2.2); Worsened, 2.5 (95% CI, 1.5–4.3); and Persistent High, 4.1 (95% CI, 3.1–5.5) (Table 2). The 3‐year LRE risks were 1.2%, 3.2%, 3.8%, and 15.0%, respectively, representing a more than 12‐fold gradient from the lowest‐ to highest risk group.

TABLE 2.

Trajectory‐based risk of liver‐related events in the LM1 cohort.

Trajectory group	No. (%)	Events	Crude incidence, /1000 PY	3‐year risk, %	5‐year risk, %	10‐year risk, %	Adjusted HR (95% CI)
Stable Low	4698 (74.4)	121	4.3	1.2	2.0	3.9	Reference
Improved	543 (8.6)	44	11.3	3.2	5.1	9.8	1.5 (1.1–2.2)
Worsened	165 (2.6)	15	13.2	3.8	6.1	11.7	2.5 (1.5–4.3)
Persistent High	907 (14.4)	292	52.3	15.0	22.7	38.3	4.1 (3.1–5.5)

Note: Crude incidence is the observed number of liver‐related events per 1000 person‐years. The 3‐, 5‐, and 10‐year risks and adjusted hazard ratios were derived from a Cox proportional hazards model adjusted for age, sex, diabetes mellitus, serum albumin, platelet count, and baseline LS, with Stable Low as the reference group; 3752 and 976 patients remained at risk at 5 and 10 years, respectively. LM3 transition results are presented in Table S15.

Abbreviations: CI, confidence interval; HR, hazard ratio; LM1, 1‐year landmark; LM3, 3‐year landmark; LS, liver stiffness; PY, person‐years.

Crude incidence rates and model‐derived long‐term risks showed the same ordering across the Stable Low through Persistent High groups, increasing from 4.3 to 52.3 per 1000 person‐years, from 2.0% to 22.7% at 5 years, and from 3.9% to 38.3% at 10 years (Table 2). Risk was strikingly concentrated: the Persistent High group, although comprising only 14.4% of the cohort, accounted for 292 of 472 LREs (61.9%). A Kaplan–Meier analysis confirmed clear separation of LRE‐free survival curves across the four LM1 trajectory groups (log‐rank p < 0.001; Figure 2). Notably, 596 of 907 patients (65.7%) in the Persistent High group showed a numerical decrease in LS from baseline (median ΔLS, −2.2 kPa) yet remained above 10 kPa, highlighting the prognostic importance of persistent threshold elevation rather than the direction of change alone. Among the threshold‐crossing groups, ≥ 30% relative LS change was observed in 75.1% of Improved patients and 81.2% of Worsened patients, compared with 21.6% in Stable Low and 38.8% in Persistent High, suggesting that most threshold‐crossing transitions were not explained solely by minor measurement fluctuation (Table S2). A sensitivity analysis using alternative thresholds confirmed that the trajectory effect was robust at 8 kPa (Persistent High: HR, 4.8; 95% CI, 3.6–6.5) and 15 kPa (HR, 2.5; 95% CI, 1.8–3.5), although the Improved group lost statistical significance at 15 kPa (HR, 1.2; p = 0.201), supporting the use of the Baveno VII 10 kPa threshold (Table S3). To address possible regression to the mean, trajectory effects were re‐estimated after excluding patients whose baseline LS fell within ±1 to ±3 kPa of the 10 kPa threshold and after Galton‐type correction of the 1‐year LS; the Persistent High effect persisted across all sensitivity analyses (HR, 3.3–6.9), whereas the Improved‐group HR was attenuated and lost statistical significance (HR, 1.2–1.5), suggesting that the Improved signal is partly attributable to regression to the mean (Table S4). Within the Improved group, a graded dose–response was observed across the magnitude of LS change: adjusted HRs were 3.9 (95% CI, 2.0–7.5) for the Small decline (|ΔLS| ≤ 2 kPa), 2.2 (95% CI, 1.4–3.5) for the Moderate decline (2 < |ΔLS| ≤ 5 kPa), and 0.7 (95% CI, 0.4–1.3) for the Large decline (|ΔLS| > 5 kPa) (p for trend = 0.014; Figure S2), indicating that residual risk in the Improved group is concentrated in patients with limited LS reduction. Reverse Kaplan–Meier median follow‐up was 5.74 years (Table S5); restricting to patients with adequate follow‐up (≥ 3 years or an observed event; n = 5262) reproduced the primary trajectory HRs (Persistent High HR, 4.0; 95% CI, 3.0–5.4), and an analysis using incident HCC alone as the sole outcome—the most discretely defined component of the composite endpoint—preserved the strong Persistent High signal (HR, 4.6; 95% CI, 3.2–6.7) (Table S6).

FIGURE 2.

Kaplan–Meier curves for liver‐related events by LM1 trajectory. Kaplan–Meier estimates of liver‐related event‐free survival according to the LM1 trajectory groups (n = 6313; 472 events). Each trajectory was defined using a 10 kPa threshold: Stable low (both time points < 10 kPa), Improved (≥ 10 to < 10 kPa), Worsened (< 10 to ≥ 10 kPa), and Persistent high (both time points ≥ 10 kPa). Differences between groups were compared using the log‐rank test. LRE, liver‐related event; LM1, 1‐year landmark.

The associations remained similar after the 1‐year LS measurement window was restricted to 6–18 months, follow‐up was re‐anchored at the actual 1‐year LS measurement date, or a 6‐month lag was applied after that measurement (Tables S7 and S8).

3.3. Subgroup Analyses

Interaction analyses stratified by aetiology were performed (Table 3); subgroup analyses by sex and diabetes status are presented in Table S9. A significant interaction was observed between trajectory and aetiology (p for interaction < 0.001). In patients with non‐viral aetiologies (n = 4322), the trajectory effect was substantially amplified: adjusted HRs were 4.4 (95% CI, 2.0–9.9) for Worsened and 6.4 (95% CI, 4.0–10.2) for Persistent High. By contrast, in patients with viral hepatitis (n = 1991), the corresponding HRs were attenuated: 1.4 (95% CI, 0.7–2.9) for Worsened and 2.6 (95% CI, 1.8–3.8) for Persistent High. No significant interaction was found for sex (p = 0.093) or diabetes (p = 0.113) (Table S9). With further stratification by aetiology, the HR for Persistent High ranged from 2.6 in HBV to 6.2 in other aetiologies; corresponding estimates were 3.4 in HCV, 3.7 in alcohol‐related liver disease, and 5.6 in MASLD (Table 3). Baseline characteristics stratified by aetiology are summarized in Table S10. Event counts by trajectory category and aetiology are presented in Table S11, and Firth penalized‐likelihood estimates for sparse cells are provided in Table S12. In complete‐case multivariable logistic regression models, lower baseline LS, lower BMI, lower serum albumin, and higher platelet count were associated with an Improved trajectory among patients with baseline LS ≥ 10 kPa. Among patients with baseline LS < 10 kPa, higher baseline LS, higher BMI, younger age, female sex and lower serum albumin were associated with a Worsened trajectory (Table S13).

TABLE 3.

Subgroup analyses of the association between LM1 trajectory and liver‐related events by aetiology.

Subgroup	N	Events	Stable Low (n)	Improved HR (95% CI)	Worsened HR (95% CI)	Persistent High HR (95% CI)	p for interaction
Overall	6313	472	4698	1.5 (1.1–2.2)	2.5 (1.5–4.3)	4.1 (3.1–5.5)	—
Aetiology
Viral	1991	303	1126	1.0 (0.7–1.6)	1.4 (0.7–2.9)	2.6 (1.8–3.8)	< 0.001
Hepatitis B	1744	272	973	0.9 (0.6–1.5)	1.1 (0.5–2.5)	2.6 (1.7–3.7)
Hepatitis C	247	31	153	NR	NR	3.4 (1.0–11.4)
Non‐viral	4322	169	3572	1.8 (0.9–3.5)	4.4 (2.0–9.9)	6.4 (4.0–10.2)
Alcohol	409	42	266	NR	NR	3.7 (1.2–11.3)
MASLD	2929	57	2536	NR	NR	5.6 (2.4–13.2)
Other	984	70	770	1.6 (0.6–4.7)	NR	6.2 (2.9–13.5)

Note: Adjusted HRs were calculated using Cox proportional hazards regression adjusted for age, sex, diabetes mellitus, serum albumin, platelet count, and baseline LS. Stable Low served as the reference group. p for interaction was calculated using a likelihood ratio test comparing models with and without the interaction term between LM1 trajectory and aetiology. Subgroup analyses by sex and diabetes mellitus are presented in Table S9. Aetiology was classified hierarchically from index‐date aetiology codes in the following order: HBV > HCV > alcohol‐related liver disease > MASLD > other. Viral and non‐viral subtotals are presented above each aetiology cluster. NR denotes trajectory categories with fewer than five events, for which standard Cox estimates were not reported owing to imprecision; Firth penalized‐likelihood estimates for these cells are provided in Table S12.

Abbreviations: CI, confidence interval; HBV, hepatitis B virus; HCV, hepatitis C virus; HR, hazard ratio; LM1, 1‐year landmark; LS, liver stiffness; MASLD, metabolic dysfunction‐associated steatotic liver disease; NR, not reported.

3.4. LS Trajectory Transitions at the 3‐Year Landmark

Among the 6313 LM1 patients, 1617 who had available 3‐year LS data and remained event‐free at LM3 constituted the dynamic prediction cohort (Figure S1). Compared with the LM1 cohort, this subset had a higher proportion of viral aetiologies (51.3% vs. 31.5%) and a higher median baseline LS (7.1 vs. 6.1 kPa), reflecting the greater likelihood of continued surveillance among patients with established liver disease (Table S14). Reclassification based on the 1‐year to 3‐year LS change yielded a similar overall distribution (Stable Low, 72.0%; Persistent High, 15.8%; Improved, 9.0%; Worsened, 3.2%). Notably, trajectory status was not static: a substantial proportion of patients transitioned between categories from LM1 to LM3, underscoring the dynamism of LS changes over time (Figure 3). Among the 1617 patients who reached LM3 event‐free, 170 LREs occurred during subsequent follow‐up. The LM3 transition groups showed a prognostic gradient consistent with that observed at LM1. Compared with the Stable Low group, the adjusted HRs were 2.2 (95% CI, 1.3–3.6) for Improved, 2.6 (95% CI, 1.3–5.2) for Worsened, and 3.5 (95% CI, 2.2–5.5) for Persistent High; the corresponding 3‐year LRE risks were 2.0%, 6.3%, 7.4%, and 13.1%, respectively (Table S15). A 4 × 4 reclassification matrix comparing the LM1 trajectory with the LM3 transition further quantified the dynamic changes illustrated in Figure 3 (Table S16). Among LM3‐eligible patients initially classified as Persistent High at LM1, 31.2% moved to the Improved category at LM3; conversely, 30.2% of those initially classified as Worsened at LM1 became Persistent High at LM3, confirming that trajectory status was not fixed over time.

FIGURE 3.

Liver stiffness trajectory transitions and the associated 3‐year risk of liver‐related events. Sankey diagram illustrating the distribution of the liver stiffness trajectory groups at LM1 (n = 6313) and transitions at LM3 (n = 1617), defined using a 10 kPa threshold: Stable Low (both time points < 10 kPa), Improved (≥ 10 to < 10 kPa), Worsened (< 10 to ≥ 10 kPa), and Persistent High (both time points ≥ 10 kPa). Flow widths are proportional to patient numbers. The three‐year liver‐related event risk for each LM3 group is shown on the right. LM1, 1‐year landmark; LM3, 3‐year landmark.

3.5. Dynamic Risk Prediction With Serial LS Monitoring

We next assessed whether incorporating trajectory information improved risk prediction beyond conventional clinical variables. A base model including age, sex, diabetes mellitus, serum albumin, and platelet count yielded a C‐index of 0.778 (95% CI, 0.749–0.810). Adding the LM1 trajectory further improved the C‐index to 0.795 (95% CI, 0.767–0.830), and updating the model with the LM3 transition increased it to 0.803 (95% CI, 0.771–0.839). Calibration slopes were near unity, and intercepts were close to zero across all models (Table S17; Figure S3; Table S18). Time‐dependent AUC analyses confirmed that this discrimination ranking was preserved across the 3‐year follow‐up window from LM3 (Figure S4).

3.6. Clinical Utility of Trajectory‐Based Prediction

In the LM1 cohort, decision curve analysis showed that adding LM1 trajectory to a model containing clinical variables and baseline LS provided incremental net benefit across threshold probabilities of 5%–15%, a range relevant to decisions about intensified HCC surveillance or portal hypertension screening (Figure S5). The trajectory‐enhanced model showed higher net benefit than both the clinical‐only model and the clinical + baseline LS model across this range, supporting the clinical utility of serial LS trajectory information beyond baseline LS and routine clinical variables.

4. Discussion

In this registry‐based study of 6313 patients spanning the full spectrum of chronic liver disease aetiologies, we demonstrate that a simple threshold‐defined classification of LS trajectory based on the Baveno VII 10 kPa criterion provides a clinically interpretable framework for longitudinal LRE risk stratification. The four trajectory groups yielded 3‐year LRE risks ranging from 1.2% to 15.0%, and the gradient persisted at 5 and 10 years, with progressively greater absolute separation between groups. Updating trajectory status at the 3‐year landmark confirmed that risk profiles are dynamic and can be meaningfully reclassified. A significant aetiology–trajectory interaction revealed that trajectory‐based stratification was particularly prognostic in non‐viral liver diseases, offering new insight into the discrepant findings reported across aetiology‐restricted studies [10, 14, 19]. Together, these findings support threshold‐defined LS trajectories as a practical framework for longitudinal risk assessment across chronic liver disease aetiologies, with the strongest prognostic gradient in non‐viral disease.

Our study has several clinical implications. First, the prognostic value of serial LS has been demonstrated in cACLD [10], primary biliary cholangitis [12], MASLD [19, 20] and hepatitis C [15]. However, those studies share two features that limit their generalizability: reliance on continuous delta‐LS metrics or computationally intensive joint models and restriction to single‐aetiology or cACLD‐only populations. Our study addresses both gaps. The threshold‐defined trajectory approach requires only two LS values and one binary threshold, offering an accessible alternative that retains meaningful risk discrimination. Equally important, the inclusion of the full fibrosis spectrum across multiple aetiologies enabled us to detect aetiology‐specific prognostic patterns that the single‐aetiology designs could not capture.

Second, our findings could appear to contrast with those of Wong et al., who reported that prior LS values add little predictive value over the current measurement in cACLD [14]. These apparently divergent findings may reflect important design and analytic differences. First, their cohort was restricted to established cACLD (LS ≥ 10 kPa; n = 480), whereas 77% of our patients were below the threshold, enabling the detection of prognostically meaningful transitions across the 10 kPa cutoff. In addition, their use of continuous LS slope might have diluted threshold‐defined risk states that carry distinct clinical significance [6]. Indeed, in our Persistent High group, 65.7% of patients showed a numerical decrease in LS yet remained above 10 kPa (Table S2); a continuous delta‐LS approach would treat these patients as having improved, obscuring the prognostic significance of their persistent threshold elevation. Among the threshold‐crossing groups, most transitions represented substantial relative change rather than minor fluctuation, with ≥ 30% LS decrease in 75.1% of Improved patients and ≥ 30% LS increase in 81.2% of Worsened patients (Table S2). Lastly, the limited event count (n = 28 LREs in Wong et al. vs. 472 in this study) might have constrained their statistical power. These observations suggest that trajectory information is most informative when the study population encompasses the full fibrosis spectrum and the analytic framework incorporates clinically established thresholds.

Nevertheless, an absolute cutoff does not fully capture the magnitude of LS change, because small changes crossing the 10 kPa threshold and larger decreases that remain above 10 kPa are classified differently. In our cohort, 24.9% of patients in the Improved group had a relative LS decrease of < 30%, whereas 26.9% of those in the Persistent High group had a relative LS decrease of ≥ 30% but remained at or above 10 kPa (Table S2). Therefore, threshold‐defined trajectories should be interpreted together with the magnitude of change, particularly for the Improved category. By contrast, the prognostic signal associated with Persistent High remained robust in analyses addressing regression to the mean (Table S4).

Third, beyond its cross‐sectional prognostic value, trajectory status was not fixed over time; the LM1‐to‐LM3 reclassification matrix showed that 31.2% of patients initially classified as Persistent High moved to Improved, whereas 30.2% of those initially classified as Worsened became Persistent High. This dynamism suggests that LS trajectories may reflect modifiable disease processes, including treatment response, metabolic control and lifestyle modification, rather than a fixed trait [3, 19]. From a clinical standpoint, these findings reinforce the rationale for repeated LS assessment at defined intervals because updated trajectory status enables meaningful risk reclassification that can inform decisions to escalate or de‐escalate surveillance intensity. Equally noteworthy is the residual risk among patients whose LS decreased below the 10 kPa threshold. In the LM1 cohort, the Improved group retained an adjusted hazard ratio of 1.5 (95% CI, 1.1–2.2) versus Stable Low, and this hazard ratio rose to 2.2 (95% CI, 1.3–3.6) when normalization occurred between the 1‐ and 3‐year landmarks (LM3 Improved). The higher hazard associated with late versus early improvement suggests that delayed normalization may reflect more advanced or prolonged prior injury, with cumulative tissue damage that persists despite achieving a sub‐threshold LS reading. Clinically, this suggests caution against de‐escalating surveillance solely on the basis of LS dropping below 10 kPa, particularly when improvement occurs after prolonged elevation.

Fourth, a notable finding is the significant aetiology–trajectory interaction. The adjusted hazard ratio for the Persistent High group reached 6.4 in non‐viral liver diseases, compared with 2.6 in viral hepatitis, likely reflecting the divergent biological meaning of LS changes across aetiologies. In viral hepatitis, treatment‐related reductions in necroinflammation may attenuate the prognostic meaning of LS change, and residual risk may be driven partly by HCC rather than decompensation alone [21, 22]. In non‐viral diseases, LS changes may more closely reflect ongoing fibrogenic or metabolic injury and its resolution [11, 20, 23, 24], consistent with the four‐fold LRE risk increase with LS progression to ≥ 10 kPa reported by Gawrieh et al. [19] in NAFLD. Our data extend this evidence to a multi‐aetiology setting and provide a unifying explanation for conflicting findings in prior aetiology‐restricted studies [10, 14, 19], in which attenuated trajectory effects in viral populations contrast with the strong prognostic impact observed in metabolic liver disease.

Fifth, from a model performance perspective, our decision curve analysis demonstrated that adding LS trajectory to clinical variables and baseline LS yielded incremental net benefit across threshold probabilities of 5%–15%, the range most relevant to clinical decisions about intensified HCC surveillance or portal hypertension screening [6, 7], suggesting that trajectory information could improve clinical decision‐making compared with models based on clinical variables or baseline LS alone. Calibration was acceptable in the LM3 models, supporting the internal consistency of this framework while underscoring the need for external validation. The simplicity of the threshold‐defined framework, which requires only two LS values and the established 10 kPa threshold, facilitates integration into routine VCTE‐based monitoring, which has demonstrated prognostic value across diverse clinical settings [6, 7, 8, 9, 25], and the benefit of updating the trajectory at the 3‐year landmark reinforces that risk assessment should be an iterative process [3].

Sixth, the threshold‐defined trajectory framework remained robust across multiple sensitivity analyses that addressed potential concerns about regression to the mean and outcome ascertainment. The Persistent High hazard persisted—and increased rather than attenuated—when patients near the 10 kPa threshold were sequentially excluded (Table S4), the opposite pattern from what regression to the mean would produce. Within the Improved group, a graded dose–response between magnitude of LS reduction and residual risk (Figure S2) further indicates that the threshold‐based framework captures genuine biological signal rather than a statistical artefact. We acknowledge that the Improved group's overall hazard was attenuated in borderline‐exclusion analyses, suggesting that a portion of this smaller signal may reflect regression to the mean—an important caveat that does not extend to the Persistent High group, which accounts for the majority of clinical risk. Restriction to patients with adequate follow‐up duration preserved the trajectory gradient, and restriction to incident HCC alone preserved the strong Persistent High signal (Table S6), supporting that differential outcome ascertainment is an unlikely explanation for the observed findings. The trajectory gradient remained similar after the 1‐year LS measurement window was restricted to 6–18 months, follow‐up was re‐anchored at the actual 1‐year LS measurement date, or a 6‐month lag was applied after that measurement. These findings reduce concern that the results were driven by the width of the measurement window or the timing of the landmark (Tables S7 and S8).

Despite these clinical implications, this study has several limitations. First, the retrospective design might have introduced selection bias because patients with serial LS measurements are more closely monitored than the general population. Second, the fixed landmark intervals of 1 and 3 years might not capture the optimal measurement timing for all patients. Third, this retrospective analysis was conducted in a Korean population, and external validation in ethnically diverse cohorts is needed to confirm generalizability. Fourth, the LM3 cohort was enriched for higher baseline LS and viral aetiologies, which could limit the generalizability of the dynamic prediction model to unselected populations. Fifth, several aetiology‐specific trajectory categories contained few events. Therefore, subgroup estimates for these sparse cells, including the Worsened categories within individual non‐viral aetiologies, should be considered exploratory, although Firth penalized‐likelihood estimates are provided in Table S12.

In conclusion, threshold‐defined LS trajectories using the Baveno VII 10 kPa criterion identified a small high‐risk subgroup, with Persistent High patients comprising 14.4% of the cohort but accounting for 61.9% of LREs. Because the strongest risk signal was persistent threshold elevation rather than numerical decrease alone, these findings support serial LS monitoring as an iterative risk‐stratification tool in chronic liver disease, particularly in non‐viral aetiologies.

Author Contributions

Dong Yun Kim: conceptualization, methodology, formal analysis, data curation, investigation, funding acquisition, writing – original draft, writing – review and editing. Seung Up Kim: conceptualization, supervision, project administration, writing – review and editing. Jae Seung Lee: investigation, writing – review and editing. Hye Won Lee: investigation, writing – review and editing. Mi Na Kim: investigation, writing – review and editing. Beom Kyung Kim: investigation, writing – review and editing. Jun Yong Park: investigation, writing – review and editing. Do Young Kim: investigation, writing – review and editing. Sang Hoon Ahn: investigation, writing – review and editing. All authors approved the final version of the manuscript and agree to be accountable for all aspects of the work.

Funding

This research was supported by a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (grant number: RS‐2025‐25459146).

Ethics Statement

This study was approved by the Institutional Review Board of Severance Hospital, Seoul, Republic of Korea (IRB No. 4‐2025‐0032).

Consent

The requirement for written informed consent was waived by the institutional review board owing to the retrospective design. Permission to reproduce material from other sources: No material from other copyrighted sources was reproduced in this manuscript.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Derivation of the 3‐year landmark (LM3) cohort. Among 6313 patients in the LM1 cohort, 1617 who had available 3‐year liver stiffness and remained event‐free constituted the LM3 analysis set. LM1, 1‐year landmark; LM3, 3‐year landmark; LRE, liver‐related event; LS, liver stiffness.

Figure S2: Liver‐related event risk by magnitude of liver stiffness change within the Improved trajectory group. Adjusted hazard ratios for liver‐related events within the Improved group in the LM1 cohort (n = 543; 44 events), stratified by the magnitude of LS change from baseline to the 1‐year landmark: small (|ΔLS| ≤ 2 kPa; n = 43, 10 events), moderate (2 < |ΔLS| ≤ 5 kPa; n = 218, 22 events) and large (|ΔLS| > 5 kPa; n = 282, 12 events). The Stable Low group served as the reference. Cox models were adjusted for age, sex, diabetes mellitus, serum albumin, platelet count and baseline LS. Points indicate hazard ratios, and vertical bars indicate 95% confidence intervals on a log scale. The horizontal dashed line denotes HR = 1; the dashed connecting line connects point estimates (p for trend = 0.014). LM1, 1‐year landmark; LS, liver stiffness.

Figure S3: Calibration of 3‐year liver‐related event risk prediction models. Calibration plots comparing predicted and observed 3‐year liver‐related event risk in the LM3 cohort (n = 1617; 170 events). Patients were grouped into deciles of predicted risk. The dashed diagonal line represents ideal calibration. Models are shown as follows: (A) base model plus 1‐year LS status; (B) base model plus LM1 trajectory; (C) base model plus 3‐year LS status; and (D) base model plus LM3 transition. Calibration slope and intercept are shown in each panel. Observed event rates were estimated using the Kaplan–Meier (KM) method. LM1, 1‐year landmark; LM3, 3‐year landmark; LS, liver stiffness.

Figure S4: Time‐dependent discrimination across sequential prediction models. Time‐dependent area under the receiver operating characteristic curve (AUC) was calculated at 6‐month intervals during the 3 years after LM3 in the LM3 cohort (n = 1617; 170 events). Four nested Cox prediction models were compared: clinical only; clinical + baseline LS; clinical + baseline LS + LM1 trajectory; and clinical + baseline LS + LM3 transition. The clinical model included age, sex, diabetes mellitus, serum albumin, and platelet count. Shaded ribbons represent 95% confidence intervals; AUC values are annotated at 3 years. LM1, 1‐year landmark; LM3, 3‐year landmark; LS, liver stiffness.

Figure S5: Decision curve analysis of LM1 trajectory–based risk prediction. Decision curve analysis comparing three models for 3‐year liver‐related events in the LM1 cohort (n = 6313; 472 events): clinical only; clinical + baseline LS; and clinical + baseline LS + LM1 trajectory. The clinical model included age, sex, diabetes mellitus, serum albumin and platelet count. The shaded area indicates the prespecified clinically relevant threshold probability range of 5%–15%. Treat‐all and treat‐none reference strategies are shown for comparison. LM1, 1‐year landmark; LS, liver stiffness.

Table S1: Regression coefficients of the LM3‐updated prediction model.

Table S2: Liver stiffness values by trajectory group.

Table S3: Sensitivity analysis using alternative LS trajectory thresholds.

Table S4: Sensitivity analyses for regression to the mean in the LM1 cohort.

Table S5: Follow‐up completeness in the LM1 cohort.

Table S6: Outcome‐specific sensitivity analyses in the LM1 cohort.

Table S7: Sensitivity analysis restricting the 1‐year LS measurement window to 6–18 months.

Table S8: Sensitivity analyses for immortal‐time bias and 1‐year LS measurement timing.

Table S9: Subgroup analyses of LM1 trajectory effects by sex and diabetes mellitus.

Table S10: Baseline characteristics of the LM1 cohort by aetiology.

Table S11: Number of liver‐related events by LM1 trajectory category and aetiology subgroup.

Table S12: Firth penalized‐likelihood Cox regression for liver‐related events by aetiology subgroup and LM1 trajectory.

Table S13: Baseline factors associated with Improved or Worsened LM1 trajectories.

Table S14: Comparison of baseline characteristics between LM1 and LM3 cohorts.

Table S15: Trajectory‐based risk of liver‐related events in the LM3 cohort (1‐year to 3‐year transition).

Table S16: Reclassification matrix from LM1 trajectory to LM3 transition in the LM3 cohort (n = 1617).

Table S17: Dynamic prediction model performance in the LM3 cohort.

Table S18: Observed vs. predicted 3‐year risk by decile for the LM3‐updated model.

Kim D. Y., Lee J. S., Lee H. W., et al., “Threshold‐Defined Liver Stiffness Trajectories and Liver‐Related Event Risk in Chronic Liver Disease: A Cohort Study,” Liver International 46, no. 8 (2026): e70819, 10.1111/liv.70819.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request, subject to institutional regulations and approval.