Automatic phenotyping of emergency department patients with incidental hepatic steatosis: A machine learning clustering analysis

Background Incidental findings in the Emergency Department (ED) are often not acted upon due to acuity of care and lack of referral pathways. Artificial intelligence (AI) can be used to combine heterogeneous data collected during ED encounters to computationally identify patients with specific disease phenotypes allowing for directed care-paths. We identified ED patients with undiagnosed Metabolic Dysfunction-Associated Liver Disease (MASLD) to demonstrate the feasibility of automated algorithms for clinical phenotypes. Methods We identified adults in 5 EDs with abdominal imaging between 1/1/2018–12/31/2023. Using a large language model, we included patients with hepatic steatosis on imaging and liver enzyme measurements. We excluded patients with prior liver disease. The K-means algorithm, an unsupervised machine learning method, was used to cluster patients into clinical phenotypes. Results We identified 80,211 individuals with abdominal imaging, and 9103 (11.34%) met inclusion criteria. Clustering revealed three distinct phenotypes: Cluster 1 (Low Metabolic Burden Hepatic Steatosis), Cluster 2 (MASLD Dominant Hepatic Steatosis), and Cluster 3 (Non-MASLD Dominant Liver Disease). Cluster 2 ( n = 1762, 19.4%) showed increased incidence of hypertension (76.6%), type 2 diabetes mellitus (53.7%), and dyslipidemia (48.6%). Cluster 3 ( n = 520, 5.7%) had significantly elevated FIB-4 values (4.50 vs. 0.98 p < 0.001) but low incidence of MASLD risk factors. Finally, Cluster 1, the largest group ( n = 6821, 74.9%) showed low FIB-4 values and low incidence of MASLD risk factors. Conclusion An automatic AI-based algorithm identified a subset of patients with high risk factors for MASLD with low liver disease screening scores (FIB-4) allowing for future integration into health surveillance algorithms.