Predicting the Progression of Non-Alcoholic Fatty Liver Disease Using Machine Learning and Clinical Laboratory Parameters: A Systematic Review

Abstract

Introduction: Non-alcoholic fatty liver disease (NAFLD) is a growing global health crisis, with a significant proportion of patients progressing to severe liver pathologies, including non-alcoholic steatohepatitis (NASH), fibrosis, and cirrhosis. The limitations and risks of invasive liver biopsy, the current gold standard for diagnosis, necessitate the development of accurate, non-invasive tools for risk stratification and disease monitoring. Machine learning (ML) models, which utilize routinely collected clinical laboratory data, have emerged as a promising and scalable solution for predicting disease progression. This review synthesizes the current evidence on the efficacy of these models.

Methods: A systematic literature search was conducted across PubMed, Google Scholar, Semanthic Scholar, Springer, Wiley Online Library databases in accordance with PRISMA guidelines. The search included studies that developed or validated ML models to predict NAFLD progression (to NASH, significant fibrosis, or advanced fibrosis using clinical laboratory parameters as primary predictors. Data on study design, population characteristics, ML algorithms, key predictors, and a full spectrum of performance metrics were extracted. The methodological quality of each study was rigorously assessed using the Prediction model Risk Of Bias ASsessment Tool (PROBAST).

Results: Sixteen studies met the inclusion criteria, encompassing a diverse range of populations and model architectures. The primary outcomes predicted were progression to NASH, significant fibrosis, and advanced fibrosis. Ensemble ML models, particularly eXtreme Gradient Boosting (XGBoost) and Random Forest (RF), consistently demonstrated superior predictive performance over traditional statistical models and other ML algorithms. For the critical endpoint of advanced fibrosis, these models frequently achieved Area Under the Receiver Operating Characteristic (AUROC) values exceeding 0.85 and, in some cases, approaching 0.92. A core set of laboratory parameters—including alanine aminotransferase (ALT), aspartate aminotransferase (AST), gamma-glutamyl transpeptidase (GGT), platelet count, triglycerides, and glycated hemoglobin (HbA1c)—were consistently identified as the most important predictors across multiple models, reflecting their central role in the pathophysiology of NAFLD.

Discussion: The evidence strongly indicates that ML models can effectively integrate complex, non-linear patterns from standard laboratory tests to generate a "digital signature" of NAFLD pathophysiology, enabling more accurate and individualized risk stratification than traditional scoring systems. These models hold significant potential for clinical application, from facilitating early identification of high-risk individuals in primary care settings to improving the efficiency of patient enrollment in clinical trials for emerging NASH therapies. However, the predominance of retrospective study designs, a lack of consistent external validation, and issues with model interpretability are key limitations of the current evidence base that must be addressed.

Conclusion: Machine learning models based on clinical laboratory parameters are powerful non-invasive tools for predicting NAFLD progression. Their high accuracy and reliance on readily available data position them as a transformative technology in hepatology. Future research must prioritize prospective validation in diverse, real-world clinical settings and focus on developing interpretable, longitudinally-informed models to facilitate their responsible and effective integration into routine clinical practice.