Machine learning can predict advanced but not early liver disease on brain MRI

Background: Steatotic liver disease (SLD) is linked to cognitive decline and an increased risk of dementia. Metabolic and alcohol-related liver disease (MetALD) further impacts brain health due to alcohol's neurotoxicity. Advanced liver diseases like metabolic dysfunction–associated steatohepatitis (MASH) or cirrhosis can lead to increased inflammation. This study examines the effects of SLD on brain structure using UK Biobank MRI data.

Methods: Individuals with metabolic-associated steatotic liver disease (MASLD; n=4584), MetALD (n=411), at-risk MASH (n=104), and hepatic failure/liver cirrhosis (ICD-10, K72+K74; n=26) were propensity-score matched to same-sized healthy control groups. Diagnoses were based on liver MRI and serological markers. Patients with diseases that affect the brain were excluded. We used random forest classifiers (RFCs) for class prediction, utilising imaging-derived phenotypes (IDPs) from brain MRIs. We also analysed residuals obtained from a normative model previously trained on separate healthy patients (n=14810).

Results: The trained RFCs achieved average AUCs of 0.57 for MASLD, 0.61 for MetALD, 0.59 for at-risk MASH, but 0.79 for hepatic failure/liver cirrhosis on a 4-fold cross-validation . For hepatic failure/liver cirrhosis, the RFC consistently identified the left inferior parietal lobule as the most important feature in every fold on the residuals.

Conclusion: Our study reveals distinct brain features in advanced liver disease. Early SLD does not show significant single IDP differences from controls, but RFCs can still predict the disease, indicating that multiple IDPs may contribute, despite milder brain effects compared to severe SLD.

Publication History

Article published online:
20 January 2025