• Home  
  • Pancreatitis Clot Model Showed No Clear Gain Over One CT Measure
- Health

Pancreatitis Clot Model Showed No Clear Gain Over One CT Measure

A 360-patient pancreatitis study found a ten-predictor model moderately distinguished existing vein clots, but showed no clear advantage over pancreatic necrosis volume alone.

Radiologist assessing an abdominal CT scan in a modern hospital imaging room

A medical model should clear a high bar

For patients hospitalised with acute pancreatitis, inflammation can involve not only the pancreas but also the blood vessels around it.

One complication is splanchnic vein thrombosis, a clot affecting veins that drain abdominal organs.

Researchers in China investigated whether a statistical model combining ten clinical and imaging predictors could identify thrombosis already visible on a contrast-enhanced CT scan.

The comparison was important: the model was judged against much simpler ways of using the volume of pancreatic tissue affected by necrosis.

The result was more cautious than a dramatic diagnostic breakthrough.

The ten-predictor tool showed moderate ability to distinguish patients with and without thrombosis, but did not clearly outperform the single necrosis-volume measure on several central performance tests.

Why this particular clot matters

Splanchnic veins carry blood from the digestive organs toward the liver.

Thrombosis in those vessels may complicate the management of acute pancreatitis, a condition that can range from relatively mild inflammation to extensive tissue damage and systemic illness.

A finding on a scan therefore has clinical significance, but risk modelling in this setting comes with a demanding question: what additional information does a sophisticated model provide beyond what radiologists can already quantify?

An internally promising result is not enough on its own.

The model must also be accurate in new patients, appropriately calibrated, practical to use, and capable of improving decisions without unnecessarily exposing people to extra investigations or treatment.

Who was included in the study

The original research, led by Hong Xiang, retrospectively analysed 360 adults with acute pancreatitis at a single centre.

There were 276 men, and the median age was 41 years, with an interquartile range of 34 to 50.

Every patient had undergone non-contrast CT before the index contrast-enhanced CT assessment.

At that index scan, 47 people, or 13.1% of the cohort, had splanchnic vein thrombosis.

Because the clot was identified at the time of the index contrast-enhanced examination, it was counted as prevalent rather than an incident event that developed during follow-up.

This is a vital distinction: a retrospective study of existing findings cannot establish how accurately a model predicts future new clots.

How the researchers assembled the model

The investigators used a ten-predictor elastic-net logistic regression model, a form of statistical modelling that balances predictor selection and coefficient shrinkage.

Elastic-net methods can be helpful when variables overlap, because unregularised models may become unstable.

The paper compared this combined model with alternatives using pancreatic necrosis volume alone.

The simpler comparison was tested with both ordinal and nominal coding, rather than assuming that one way of representing the measure was unquestionably correct.

This design asked whether additional predictors meaningfully improved differentiation between patients with and without the clot.

More variables do not automatically produce more clinical value, particularly when the available sample includes relatively few positive cases.

Testing the model without a second hospital

No wholly independent external test cohort was available.

Instead, the researchers used repeated nested cross-validation and bootstrap resampling, approaches intended to estimate how the model might perform beyond the exact patients used to fit it while guarding against optimistic estimates from the training sample.

Nested cross-validation separates model development choices from evaluation within repeated data partitions.

Bootstrap correction offers another view of apparent performance.

These are thoughtful internal-validation methods, but they remain internal.

Data collected at the same hospital can share scanner practices, patient characteristics and clinical pathways that do not reproduce elsewhere.

A prospective multicentre assessment would be needed to establish whether the algorithm works reliably in different care systems.

What the performance numbers actually say

For the ten-predictor model, the reported area under the receiver operating characteristic curve, or AUC, was 0.765, with a 95% confidence interval from 0.679 to 0.838.

Its bootstrap-corrected AUC was 0.762.

An AUC of 0.765 is compatible with moderate discrimination: if one patient with thrombosis and one without were drawn at random from a comparable population, the model would tend to assign a higher score to the affected patient in about three-quarters of such pairs.

It is not a statement that the model correctly diagnoses 76.5% of patients.

Accuracy for a particular clinical decision depends on the chosen threshold and the balance of false positives and false negatives.

The trade-off behind a single threshold

At a threshold of 0.14 selected during training, the reported mean sensitivity was 0.66 and specificity was 0.77.

Sensitivity describes the share of actual thrombosis cases that the model flags, whereas specificity describes the share without thrombosis that it leaves unflagged.

The mean positive predictive value was 0.31: in the evaluated setting, many positive alerts would not correspond to an identified clot.

The mean negative predictive value was 0.94, but that figure depends on the prevalence of thrombosis in the population.

The authors also reported 15.9 missed clot events per 360-patient validation repeat.

No single threshold therefore makes both missed cases and unnecessary alerts disappear.

The main comparison was surprisingly close

When the ten-predictor model was compared with the ordinal pancreatic-necrosis-volume model on matched validation folds, their mean AUC values were 0.772 and 0.763 respectively.

The paired difference was just 0.009, with a 95% confidence interval spanning −0.027 to 0.046 and a reported P value of 0.608.

With nominal coding of pancreatic necrosis volume, the AUC difference was 0.013, again with uncertainty spanning both positive and negative differences and P=0.510.

These results did not establish a clear advantage in overall discrimination for the more complicated model.

Similarity does not prove that the two methods are identical, but it sharply limits claims that the extra variables deliver a meaningful improvement.

A secondary signal that warrants more testing

The ten-predictor tool did show higher mean average precision under both comparison codings, with reported differences of 0.074 and 0.089.

Both unadjusted secondary comparisons had P values of 0.006.

Average precision focuses on performance across precision-recall trade-offs and can be informative when the outcome is relatively uncommon.

Yet the researchers emphasised that this was a secondary finding, not a confirmed clinical win.

The study did not demonstrate an improvement in the Brier score, which captures aspects of prediction error, or in decision-curve net benefit.

A secondary statistical advantage can motivate a new independent test without overturning the larger conclusion about uncertain added value.

What pancreatic necrosis volume contributes

Pancreatic necrosis volume describes the amount of pancreatic tissue showing necrotic changes on imaging.

In this cohort it was associated with thrombosis detected on the index contrast-enhanced CT.

Its descriptive ability to distinguish cases was higher than that of the recorded composite imaging scores, according to the paper.

That does not establish the superiority of necrosis volume as a universal diagnostic test.

It may be correlated with underlying disease severity, inflammatory effects or other characteristics that accompany vascular complications.

Nor does an association show that more necrosis causes a clot through a single known biological route.

The result is best understood as a useful imaging signal deserving more careful replication.

Why the timing of CT matters

Patients in this study had a non-contrast scan before the index contrast-enhanced scan.

That sequencing matters because the model is being evaluated against an outcome identified at the index examination, not against a clot that developed months later.

A model associated with an existing finding may have value in supporting clinical attention or deciding what further assessment to consider, but the present work does not demonstrate such an effect.

It would be misleading to present it as an established early-warning system that prevents thrombosis.

Any proposed use must specify which information is available before the clinical decision and what clinicians would do differently once the estimate is produced.

Calibration, usefulness and false reassurance

Model discrimination is only one aspect of quality.

A score can rank people reasonably well while still giving probabilities that are too high or too low for a new hospital population.

The paper’s Brier score of 0.099 provides an error summary within internal evaluation, but it does not prove calibration in patients with a different mix of severity or disease prevalence.

The 0.94 negative predictive value at the training-derived threshold may sound reassuring, yet it does not justify ruling out thrombosis in an individual without proper clinical assessment.

When a false negative could affect care, an apparently strong negative predictive value must be interpreted alongside the outcome frequency, the testing environment and the limitations of retrospective data.

How this fits into the broader evidence

Research Today has covered related problems in evaluating medical prediction tools.

A report on diagnostic AI and false alarms illustrates why gains on one statistical measure need to be weighed against unnecessary positive flags.

Another article on data leakage in medical-image AI evaluation shows why the separation between development and evaluation data matters.

These articles concern different models and populations, so they cannot validate this pancreatitis result.

Their relevance is methodological: model performance should be reported in terms of the decisions it might affect, and external testing is essential before considering deployment.

Limits of a retrospective single-centre cohort

The dataset contained 47 thrombosis cases, limiting the precision with which a model containing ten predictors can be evaluated.

Patients came from one institution, and decisions about CT scanning, treatment and admission may differ in other settings.

Retrospective datasets also depend on what clinicians recorded, how reliably imaging measures were available and which eligible patients received each investigation.

The study did not randomly assign imaging methods or clinical interventions.

Its cross-validated results help reduce the risk of overfitting but cannot eliminate selection bias, unrecorded confounding or the possibility that model performance degrades in new populations.

What stronger confirmation would require

A convincing next step would be an independent cohort from hospitals with different patient demographics, disease-severity profiles and imaging protocols.

Investigators could lock the model before testing, report calibration as well as discrimination, and examine whether any improvement over necrosis volume alone persists under clinically relevant thresholds.

Prospective work would also need to determine when the score is available and how it changes clinical actions.

If a model adds cost and workflow complexity without reducing missed clots or unnecessary procedures, even statistically interesting performance may have little practical value.

Conversely, a reproducible incremental gain could matter if it helps solve a clearly defined decision problem.

The measured conclusion

The study provides evidence that a ten-predictor clinicoradiologic model can moderately distinguish patients whose index CT shows splanchnic vein thrombosis, and that pancreatic necrosis volume contains related information.

The main result is not that a sophisticated system has solved diagnosis: its AUC, Brier score and estimated net benefit did not show a clear advantage over the simpler necrosis-volume approaches.

Higher average precision is an interesting secondary observation that remains hypothesis-generating.

Patients should not rely on this exploratory model to make decisions about symptoms or treatment.

The appropriate scientific conclusion is to validate carefully before considering clinical use.

Source Information

Study title: Exploratory derivation of a clinicoradiologic model for splanchnic vein thrombosis in acute pancreatitis

Authors: Hong Xiang, Jiayin Che, Caiqiong Wang, Jia Lv, Hongmei Jiang, Yaqiong Tao, Zhihui Fang, Zhe Chen, Xiahong Yang, Xiaolong Liu

Journal: Scientific Reports

Publication date: 10 November 2026

Study design: Single-centre retrospective exploratory derivation and internal validation using repeated nested cross-validation and bootstrap resampling

Sample or scope: 360 adults with acute pancreatitis; 47 (13.1%) had splanchnic vein thrombosis on index contrast-enhanced CT

Key finding: Ten-predictor model AUC 0.765 (95% CI 0.679–0.838), with no clear AUC, Brier-score or net-benefit advantage over pancreatic-necrosis-volume-only models

DOI: 10.1038/s41598-026-80340-4

Contact Us

Research Today is a South African digital publication that makes credible research easier to understand.

TERMS OF USE & PRIVACY POLICY

follow us