Can multimodal AI guide liver-cancer care yet? A 28-model review says not yet
Models combining scans, clinical records and molecular data often reported promising predictions, but a new peer-reviewed review found no prospective clinical-impact study or randomised trial—and basic reporting gaps remain common.
The short version
- What happened
- The peer-reviewed narrative review appraised 28 multimodal AI prediction models spanning diagnosis, grading or microvascular invasion, treatment response, and prognosis or recurrence.
- Who it affects
- People with liver cancer need evidence that a prediction changes care for the better, not only that it fits retrospective records. Premature use could misclassify risk or steer treatment using models that have not been tested in the population receiving care.
- What remains unproven
- This is a peer-reviewed narrative review rather than a preregistered systematic review or meta-analysis; it searched four databases and literature published through 2025.
Keep in mind
This is a peer-reviewed narrative review rather than a preregistered systematic review or meta-analysis; it searched four databases and literature published through 2025.
The 28-model audit mixes different tasks, data types and endpoints, so reported performance cannot be pooled or compared as if all models answered the same question.
The review team was based in one Chinese university system, received Hunan provincial public funding and declared no competing interests. The authors disclosed ChatGPT-4o use for writing and language polishing, not for the literature collation or analysis.
No included study demonstrated prospective clinical benefit, and evidence was heavily weighted toward East Asian viral-hepatitis populations.
Source analysis · 1 linked source. Primary analysed source date: 22 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
The review asks a clinical question performance scores cannot answer
Multimodal AI tries to combine two or more sources—such as CT or MRI scans, digital pathology, genomic measurements and ordinary clinical records—to characterise a liver tumour more completely than one source alone. A peer-reviewed narrative review published on 22 October examined how far that promise has travelled toward care for hepatocellular carcinoma, the most common primary liver cancer.
The authors searched PubMed, Web of Science, the Cochrane Library and China National Knowledge Infrastructure for work published through 2025. They then appraised 28 prediction models across four tasks: diagnosis, grading or microvascular invasion, treatment-response assessment, and prognosis or recurrence. This was not a systematic review or meta-analysis, so the set should be read as a structured evidence map rather than a guaranteed census or pooled estimate of performance.[1]
The validation ledger is stronger than a prototype—but short of clinical evidence
Twenty-seven models—96.4%—came from retrospective studies; the remaining one was an animal experiment. Nineteen used multicentre data and 19 reported external validation, while 21 used an independent test set. Those are useful safeguards against a model merely memorising one hospital's data. Yet none of the 28 reached the review's highest evidence level: a prospective study or clinical-impact test. No randomised trial showed that using a model improved a decision, survival, quality of life or another patient outcome.
Important reporting checks were uneven. Twenty-seven studies gave confidence intervals, but only 10 reported calibration—whether predicted risks match observed risks. Ten reported decision-curve analysis and predefined endpoints. Just three addressed class imbalance, while seven described how missing data were handled. None explicitly followed any of five AI-specific reporting or bias guidelines checked by the reviewers. Strong headline accuracy or area-under-the-curve results cannot repair those gaps.[1]
The distinction that matters
What the 28-model evidence map actually shows
The review separates promising development signals from the evidence needed to change care.
Inputs
Combine imaging, pathology, molecular and clinical data
Build a richer tumour representation
Validation
Use multicentre data, an external cohort and an independent test set
Reduce—but do not remove—overfitting risk
Clinical test
Prospectively compare AI-assisted care with current practice
No included model reached this level
For patients, the useful message is restraint rather than rejection
Combining routinely collected information could eventually help clinicians detect atypical tumours, estimate microvascular invasion before surgery, anticipate treatment response or tailor follow-up. But the review does not justify choosing therapy from a multimodal score today. Models trained mainly on East Asian viral-hepatitis cohorts may also travel poorly to populations where metabolic liver disease is more common; the authors found no multimodal study specifically focused on Western metabolic-disease-related liver cancer.
The assessment would change with preregistered prospective, multinational studies that compare AI-assisted decisions with standard care, report calibration and missing-data handling, and measure patient outcomes as well as model performance. Until then, these systems are research tools—not replacements for imaging, pathology, liver-function assessment or multidisciplinary clinical judgement.[1]