Routine clinical data formed a workable bleeding-risk signal in cirrhosis

Esophageal variceal bleeding remains one of the most feared complications of cirrhosis because it can turn compensated portal hypertension into an abrupt, life-threatening emergency. In a new proof-of-concept study, investigators built VA-BLED, a simple prediction model designed to estimate 1-year risk of esophageal variceal bleeding using information already available in routine care. The model was developed in a retrospective, outcome-balanced cohort of 108 patients with esophageal varices, split evenly between 54 patients with bleeding and 54 without bleeding.

The rationale is clinically practical. Current risk stratification tools focus heavily on portal hypertension stage, liver stiffness, platelet count, or endoscopic appearance, yet bedside decisions often hinge on a narrower question: who is likely to bleed in the coming year? This study addressed that gap by testing whether a small set of accessible variables could separate higher-risk from lower-risk patients without relying on invasive hemodynamic testing.

Child-Pugh status, ALT, alcohol use, and prior bleeding carried the signal

Two multivariable logistic models were developed. VA-BLED BASE included Child-Pugh score, normalized alanine aminotransferase, and current alcohol use. VA-BLED HISTORY added previous bleeding history. Both were internally validated with repeated stratified 5-fold cross-validation, and performance was assessed from out-of-fold predictions, a design intended to reduce optimism compared with a single training-test split.

The discriminative performance was moderate for the base model and improved when prior bleeding was added. VA-BLED BASE achieved an out-of-fold area under the receiver operating characteristic curve of 0.732, with a 95% confidence interval of 0.635 to 0.827. Adding previous bleeding increased the out-of-fold AUC to 0.798, with a 95% confidence interval of 0.705 to 0.883. Both models showed net benefit across clinically plausible threshold probabilities on decision curve analysis.

Prior bleeding carried particular weight because it captures a patient’s realized susceptibility, not just anatomic risk. A previous hemorrhage likely reflects a combination of more severe portal hypertension, more fragile variceal anatomy, and persistent exposure to modifiable drivers such as alcohol use or worsening liver function. In practical terms, the addition of this single historical variable shifted the model from general risk estimation toward a more selective enrichment tool for patients already on a higher-risk trajectory.

The score separated patients into usable bedside risk zones

At an assumed annual bleeding prevalence of 10%, VA-BLED BASE produced low-, intermediate-, and high-risk zones. VA-BLED HISTORY functioned more narrowly as a high-risk enrichment model, meaning it was better suited to identifying a subgroup in whom bleeding risk was sufficiently concentrated to justify heightened attention.

The study also recalibrated predictions to target prevalences of 5%, 10%, and 15%, because the derivation cohort was intentionally balanced at 50% cases and 50% controls. Without recalibration, the raw intercepts from such a design do not translate cleanly into real-world bleeding probabilities. By adjusting the intercepts to plausible prevalence settings, the investigators aimed to make the score more relevant to clinical populations where annual bleeding risk is far lower than 50%.

That approach matters because prognostic models in hepatology often fail at the point where they leave the derivation dataset. A model can rank patients correctly yet still misestimate absolute risk if the underlying event rate is not recalibrated to the target population. Here, the authors explicitly addressed that problem rather than leaving the score anchored to the artificial structure of the development sample.

Why these variables fit the biology of variceal hemorrhage

Each retained predictor aligns with the pathophysiology of portal-hypertension-related bleeding. Child-Pugh score summarizes synthetic dysfunction, ascites, and encephalopathy, all of which reflect worsening hepatic reserve and more advanced cirrhosis. Normalized ALT may capture ongoing hepatocellular injury or inflammatory activity, both of which can track with progressive disease. Current alcohol use continues to drive injury, worsen portal pressure, and complicate adherence to preventive therapy. Previous bleeding, meanwhile, marks a patient who has already crossed the threshold from structural risk to clinical instability.

This is the sort of model that may be useful precisely because it avoids dependence on data that are not always available in routine practice. HVPG remains the reference standard for portal hypertension, but it is invasive and not widely accessible. Endoscopic grading helps, yet it does not always translate into a short-term bleeding probability that can guide individualized follow-up. A score based on routinely documented variables can fit into outpatient hepatology workflows more easily, especially where specialized testing is unavailable.

Internal validation was reassuring, but the sample remains small and selected

The study was deliberately structured as a retrospective proof-of-concept project at a single center in Germany. Cases and controls were selected to balance the outcome, which supports model development but also means the cohort does not mirror routine prevalence. Patients without available clinical data 10 to 14 months before the patient-specific index date were excluded, and the bleeding outcome was retrieved from endoscopic reports with subjective interpretation. No blinded central adjudication was performed.

Those details matter because they define the limits of the evidence. Internal validation can show whether a model behaves consistently within the same data environment, but it does not establish transportability across other hospitals, different referral patterns, or broader cirrhosis populations. The investigators therefore framed the work as a development study, not a clinical implementation study.

The authors also note that external validation in cohort-design populations is required before the model can be applied in practice. That next step will determine whether the apparent signal holds when the score is tested against the lower, more realistic bleeding rates seen in everyday hepatology clinics. Until then, VA-BLED is best viewed as a promising candidate rule for risk enrichment rather than a replacement for portal-hypertension-based assessment or endoscopic judgment.

A screening adjunct, not a substitute for variceal surveillance

The most useful reading of this study is modest but concrete. VA-BLED may help clinicians identify which cirrhotic patients with varices are drifting toward a higher 1-year bleeding risk using data already present in the chart. It may also help distinguish patients who need closer follow-up from those whose immediate bleeding probability is lower, at least within the logic of the model.

Even so, the score does not erase the need for endoscopic assessment, variceal staging, or portal-hypertension management. It sits beside those tools, not above them. Before the model can move beyond proof of concept, it needs external validation, clearer performance in unselected cohorts, and testing in settings where decisions about prophylaxis and surveillance are made prospectively rather than reconstructed from records.

The final clinical question is not whether a simple score can be built. It already has been. The question is whether it will remain accurate when tested in broader cirrhosis populations where the event rate is lower, the data are messier, and the consequences of false reassurance are far more consequential.


Reporting by Dr. A. Nikhil Krishnan, MBBS for Medicine World Foundation, with additional data provided by PLOS ONE (Medicine).
PRIMARY CLINICAL EVIDENCE & REGISTRY DATA