Abstract
BACKGROUND/AIM: Metabolic dysfunction-associated steatotic liver disease (MASLD) is often silent and progressive, affecting nearly one-fourth of the global population. However, liver biopsy remains the only accurate but invasive modality for diagnosis. We aimed to develop a machine-learning model to assess liver fibrosis in MASLD patients. METHOD: We retrospectively analyzed electronic medical records of biopsy-proven MASLD patients, incorporating demographic, clinical, and liver biopsy data (advanced fibrosis, F3/F4 as AF). Fourteen machine learning (ML) models were developed, with the best model compared against the fibrosis-4 (FIB-4) score and aspartate aminotransferase to platelet ratio index (APRI) for detecting AF. RESULTS: Among 452 patients (median age: 41 years; interquartile range (IQR) 33-49; 75% male), 222 (49.1%) had AF. The fibrosis assessment by extra trees classifier (FAET) demonstrated the best performance, achieving 84% accuracy, an area under the curve (AUC) of 0.82, an F1 score of 0.77, a negative predictive value (NPV) of 0.81, and a positive predictive value (PPV) of 0.91 using 13 features. Compared to FIB-4, FAET improved accuracy by 33%, AUC by 34%, F1 score by 45%, NPV by 15.7%, and PPV by 71%. Similarly, compared to APRI, FAET improved accuracy by 58.4%, AUC by 49%, F1 score by 45.2%, NPV by 20.8%, and PPV by 111%. Using the top five features, the model achieved an accuracy of 82%, an AUC of 0.80, an F1 score of 0.76, a PPV of 0.79, and an NPV of 0.83 on the external test set, enabling the development of a simplified web-based tool. CONCLUSION: FAET offers a significant advancement in the noninvasive assessment of advanced fibrosis in MASLD patients and has the potential for routine use pending further validation.