Can a six-protein AI score rule out serious fatty liver disease?
A multicohort study reports AUCs of 0.79–0.83 for a six-protein MASH score and a 0.99 negative predictive value in a low-prevalence UK cohort. Different reference standards, small biopsy cohorts and unavailable routine assays keep it from clinical readiness.
The short version
3 key points
- 1.Development used 142 biopsy-characterised participants from one Chinese hospital: 100 for training and 42 for internal validation. The score combines ACY1, CD59, CCL24, GPD1, LEP and IGFBP1.
- 2.Reported AUCs were 0.83 in training, 0.80 internally, 0.79 in 57 Israeli repository samples and 0.79 in 3,572 UK Biobank participants. The UK test used MRI-PDFF plus cT1 rather than biopsy as its reference.
- 3.A negative predictive value of 0.99 in the low-prevalence UK cohort may support further screening research, but it depends heavily on prevalence and does not show that using the score improves diagnosis, treatment or outcomes.
The answer: promising for ruling out disease, not ready to replace biopsy
A blood-based score built from six plasma proteins separated people with and without metabolic dysfunction-associated steatohepatitis, or MASH, with moderate discrimination across four cohorts. AUCs ranged from 0.79 to 0.83 rather than approaching perfect classification. In the largest low-prevalence UK test cohort, the authors report a negative predictive value of 0.99, suggesting the model rarely labelled an apparent low-risk person who met that cohort's MASH definition.
That is a useful signal for further screening research, not a clinical green light. Negative predictive value rises when disease is uncommon, and the UK test used imaging surrogates rather than liver biopsy. The six proteins do not yet have widely available routine assays, no prospective care pathway was tested, and a positive result still requires specialist evaluation. The study therefore supports the possibility of a first-stage rule-out tool—not replacement of pathology, clinical judgement or established assessment.[1]
How the score was built
Researchers recruited 142 participants at the First Affiliated Hospital of Nanjing Medical University between August 2023 and July 2024. Eighty-three had biopsy-confirmed MASH and 59 were controls. A random 7:3 division produced 100 training participants and 42 internal-validation participants. The team used LASSO and Boruta feature selection, compared ten machine-learning algorithms and selected partial least squares for the final score.
The six retained proteins were ACY1, CD59, CCL24, GPD1, LEP and IGFBP1. Proteomic measurements were processed with quality controls, including a protein-level false-discovery-rate threshold of 1% and a coefficient-of-variation threshold below 20% for quality-control proteins. Choosing features and algorithms within a small development cohort creates overfitting risk, which makes performance in untouched external data especially important.[1]
The external cohorts broadened geography but changed the target
The first external test comprised 57 samples from an Israeli public repository, including 37 MASH cases. It used histology and produced an AUC of 0.79, sensitivity of 0.62 and specificity of 0.75 at the chosen threshold. The authors themselves say its small size and missing clinical covariates mean it is preliminary rather than a fully qualified external validation set.
The second test used 3,572 UK Biobank participants, of whom 164—about 4.6%—met an imaging-based definition combining MRI proton-density fat fraction above 5.5% with corrected T1 above 800 milliseconds. The AUC was 0.79, sensitivity 0.80 and specificity 0.63. Because that reference is not biopsy, it may misclassify disease and is not directly equivalent to the Chinese or Israeli outcome. The large cohort improves precision but does not repair a changed ground truth.[1]
Why the 0.99 negative predictive value needs context
The UK cohort's negative predictive value of 0.99 sounds close to certainty, but predictive values depend on how common disease is in the tested population. With only about one participant in twenty meeting the imaging definition, most people will be negative before any model is applied. Sensitivity of 0.80 means the model still missed an estimated one-fifth of reference-positive cases, while specificity of 0.63 means a substantial share of reference-negative people could test positive and require follow-up.
For clinicians and commissioners, the relevant calculation is not whether 0.99 is impressive in isolation. It is how many people would avoid further testing, how many true cases would be missed, how many false positives would reach imaging or specialist clinics, and whether the protein assay is cheaper and more acceptable than alternatives. Those denominators must be estimated in the intended population, not transferred from UK Biobank to a diabetes clinic, liver service or general-practice register.[1]
Long-term associations are not a treatment-impact study
The researchers applied the score to a separate UK Biobank prognostic cohort of 47,575 people. Higher scores were associated with liver-related death, later MASH, cirrhosis, liver cancer, cardiovascular mortality and all-cause mortality. Reported hazard ratios included 8.18 for liver-related death, 12.07 for MASH, 3.37 for cirrhosis and 3.11 for liver cancer, with smaller associations for cardiovascular and all-cause mortality.
Those figures describe differences between score-defined groups in observational follow-up. They do not show that testing people, changing treatment because of the score or lowering the score would prevent events. Residual confounding is plausible because several proteins reflect inflammation, adiposity and metabolism beyond the liver. The authors also note that plasma origin is not necessarily liver-specific and that cohorts differed in collection, platforms and patient characteristics.[1]
What this could mean for clinicians and patients
For primary care and metabolic clinics, a reliable rule-out test could help identify which patients do not need invasive investigation and which should move to elastography or specialist review. For patients, a blood test may be more acceptable than biopsy. But the present model requires proteomic measurement of six proteins, several of which lack standard low-cost assays. A research-platform score cannot be assumed to fit ordinary laboratory turnaround, calibration and quality-assurance requirements.
A positive result should not be presented as a diagnosis. The paper proposes follow-up with liver-stiffness measurement and clinical assessment, which is a safer interpretation than direct treatment. Services would also need thresholds adapted to prevalence, clear handling of indeterminate results and monitoring for performance differences by age, sex, ethnicity, obesity, diabetes and comorbidity. None of those implementation questions was tested prospectively here.[1]
Funding and evidence limits
The study was funded by the National Natural Science Foundation of China, the Endocrinology and Metabolism Talent Program of the China International Medical Exchange Foundation and two Huai'an research grants. The authors reported no commercial conflict in the paper. Data came from the Chinese hospital, a public Israeli repository and UK Biobank, but the authors could not independently verify all selection and measurement errors in the external public datasets.
Other limitations include the small biopsy-confirmed development and Israeli cohorts, different reference standards across cohorts, inability to pool data because of incompatible platforms and populations, and no evidence from prospective clinical use. Continuous monitoring and response to treatment were not tested. External validation across multiple countries is a strength, but geographic diversity should not be confused with identical measurement or a completed clinical validation pathway.[1]
What would change the assessment
Confidence would rise with a preregistered multicentre study using the same six-protein assay, locked model and biopsy or a clearly justified reference standard in every site. It should recruit enough participants to estimate missed cases and false referrals precisely in primary care and specialist populations, and report calibration as well as AUC. Independent laboratories should reproduce measurements with a practical assay.
The decisive evidence would be a prospective impact study comparing current pathways with score-assisted triage. Outcomes should include biopsies or imaging avoided, missed MASH, referral burden, time to diagnosis, patient experience, costs and treatment decisions. Until that evidence exists, MASHPredict is a credible research candidate with useful external signals, but not a clinically established rule-out test.[1]
What this means for people
- Patients might avoid some invasive investigation if a future blood test reliably rules out MASH, but this score has not yet been tested in a clinical pathway.
- GPs and liver teams need absolute false-negative and false-positive counts at their local disease prevalence, not only AUC and negative predictive value.
- Laboratories would need affordable, standardised assays for all six proteins before routine deployment is realistic.
Global context
The study spans a Chinese biopsy cohort, an Israeli repository and UK Biobank, but reference standards and measurement platforms differed. That geographic breadth is valuable for stress-testing the signal; it does not make the cohorts interchangeable or establish performance in lower-resource health systems and underrepresented populations.
What the evidence does not yet show
- Development and internal validation used only 142 participants from one Chinese hospital.
- The 57-person Israeli test cohort was small and lacked many clinical covariates.
- The 3,572-person UK test used MRI-PDFF plus cT1 rather than biopsy as the MASH reference.
- A 0.99 negative predictive value partly reflects the low prevalence of reference-positive disease in the UK cohort.
- The six proteins lack widely available routine assays, and no prospective clinical-impact study was conducted.
- Prognostic associations do not prove that score-guided care changes outcomes.
What to watch next
- Locked multicentre validation using a consistent reference standard
- Low-cost reproducible assays for the six-protein panel
- Calibration and error counts in primary care, diabetes and liver clinics
- Prospective evidence on avoided investigations, missed cases and referral burden
- Independent comparison with elastography and established non-invasive scores
Living evidence record
Impact record IAI-18B1CHJ
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
15 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Frontiers in Immunology published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 15 October 2026
Research question, topics and regions
Whether a six-protein machine-learning score can identify metabolic dysfunction-associated steatohepatitis across Chinese, Israeli and UK cohorts
Topics: MASH · Fatty liver disease · Proteomics · Machine learning · External validation · Screening
Regions: China · Israel · United Kingdom · Asia · Middle East · Europe
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Watch the background, then check the evidence
The Royal Society
AI-generated illustration · not a video frame
YouTube loads only when you choose. Press play in the player to start. If playback is unavailable, use “Watch on YouTube” below.
Healthcare, culture & human impact · Background explainer
Published 2026-07-01 · 44 min 28 sec
Is this the AI future we want? | The Royal Society
A Royal Society discussion with Madhumita Murgia and biomedical engineer Alison Noble on AI’s potential in healthcare, science and culture, recorded at the 30 June 2026 event.
Evidence check: A discussion of opportunities and risks, including future possibilities. It is not evidence that every application described is available or clinically proven.
Source checked 14 October 2026. Check available captions on the publisher’s video page.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Continue the story
There’s more to this story
Health & Life Sciences
Did a breast-cancer AI score survive external validation?
A 50-gene machine-learning signature looked moderately accurate in 101 development cases, but its transferable component fell to an AUC of 0.55 in 320 external cases and added no value beyond routine clinical factors.
7 min · 1 source
Health & Life Sciences
Can AI prioritise cognitive assessment in sleep apnoea?
A random-forest model separated concurrent mild cognitive impairment in an external Chinese cohort, but overestimated probabilities and was evaluated almost entirely in men. It is not a diagnostic or prognostic tool.
6 min · 1 source
Health & Life Sciences
Can genomic AI reduce missed antibiotic resistance?
A peer-reviewed model trained on 5,942 Klebsiella pneumoniae genomes made fewer false-susceptible predictions than two established genomic tools in retrospective global, African and Thai benchmarks. It has not yet shown that it improves prescribing or patient outcomes in live care.
9 min · 1 source
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.