JMIR Publications

ABSTRACT

Background:

Metabolic dysfunction-associated fatty liver disease (MAFLD) is a leading cause of chronic disease and may develop into liver fibrosis or hepatocellular carcinoma. It has distinct subtypes—obesity, diabetic, and lean—each associated with divergent fibrosis burden and complications. However, existing analytics methods often overlook the multisystem nature, intra-phenotype heterogeneity and disease dynamics. These limitations hinder accurate risk stratification and personalized intervention planning for MAFLD patients.

Objective:

This study develops a novel, two-stage, contrastive learning–based method to estimate the phenotype of metabolic dysfunction-associated fatty liver disease (MAFLD) among adults. The proposed method leverages multi-view contrastive learning; it models individual heterogeneity and important relationships in clinical and survey-based data to make effective phenotype predictions in support of clinical decision-making and personalized care.

Methods:

Demographic, clinical, lifestyle-related, and family genetics history–oriented data of 4,408 adults reveal how capturing important relationships in the data from different sources can transform individual-level representations into multiple, complementary views. The comparative evaluation of the predictive efficacy of the proposed method, in comparison with eight prevalent methods, relies on recall, precision, F-measure, and area under the curve.

Results:

The proposed method consistently and significantly outperforms all benchmark methods. It attains the highest F-measure value, with a 34.0% improvement for the non-diabetics phenotype and 22.2% improvement for the diabetes phenotype than the respective best-performing benchmarks. These results demonstrate the clinical value and utilities of integrating clinical and survey-based data to better estimate MAFLD phenotypes among adults.

Conclusions:

The proposed method provides a viable means for MAFLD phenotype estimates. It is more efficacious for identifying at-risk adults than many prevalent data-driven methods and thereby can enhance clinical decision-making and personalized care.