Risk Prediction in Patients With Metabolic Dysfunction-Associated Steatohepatitis Using Natural Language Processing

Metabolic dysfunction-associated steatohepatitis (MASH) is a highly heterogenous condition and a leading cause of end-stage liver disease. Understanding disease progression in real-world settings remains a major unmet need. We sought to define a real-world MASH cohort using natural language processing (NLP) and identify significant associations with all-cause mortality and progression to cirrhosis and liver transplantation.

We developed, validated, and applied a novel NLP algorithm, "NASHDetection," to identify patients at the University of California San Francisco who were diagnosed with MASH between 2012 and 2022. We used Cox regression with bidirectional stepwise variable selection to identify significant associations with outcomes.

NASHDetection was 86% accurate at identifying 2695 MASH patients. At the time of their diagnosis, the median age was 57 years; 55.4% had cirrhosis at baseline, with 34.0% having evidence of decompensation and 10.8% with hepatocellular carcinoma. The most common comorbidities were hypertension (61.9%), hyperlipidemia (47.4%), and type 2 diabetes mellitus (41.5%). Multiple comorbidities were associated with all-cause mortality, including type 2 diabetes mellitus (hazard ratio (HR): 1.36; confidence interval (CI): 1.07-1.73), heart failure (HR: 1.45; CI: 1.01-2.08), and peripheral artery disease (HR: 1.72; CI: 1.04-2.85). Significant laboratory-based predictors of mortality included high-low-density lipoprotein cholesterol (HR: 1.49; CI: 1.20-1.84) and high alkaline phosphatase (HR: 1.94; CI: 1.58-2.38).

We described a cohort of real-world MASH patients using a new NLP algorithm and found several potential predictors of progression to all-cause mortality, cirrhosis, and liver transplantation. The use of NLP to characterize these patients can help support the development of future interventional trials in MASH.