Machine Learning Approach Enables Highly Accurate Identification of At-Risk Metabolic Dysfunction-Associated Steatohepatitis.

AIM: At-risk metabolic dysfunction-associated steatohepatitis (MASH), characterized by significant activity and fibrosis, increases the risk of liver complications. Liver stiffness measurement (LSM), commonly used to detect significant fibrosis, has limitations in terms of accessibility and performance in certain populations. We aimed to develop a simple-to-use machine learning (ML) model to identify at-risk MASH without LSM.

METHODS

We analyzed 884 patients with histologically confirmed metabolic dysfunction-associated steatotic liver disease from a nationwide multicenter cohort, divided into derivation (80%) and validation (20%) sets. Multiple ML algorithms (random forest [RF], logistic regression [LR], gradient boosting [GB], support vector machine [SVM], and deep learning [DL]) were trained using variables including age, sex, body mass index, hematological/biochemical parameters, and comorbidities.

RESULTS

In the validation cohort, the RF model showed superior discriminatory ability for predicting at-risk MASH (AUROC: 0.8405) compared to LR (0.7729), GB (0.8252), SVM (0.7816), and DL (0.7422). Using only seven routine clinical parameters, the RF model outperformed the fibrosis-4 index (AUROC: 0.7329; p < 0.001) and LSM (AUROC: 0.7428; p = 0.01) and showed comparable performance to the FibroScan-aspartate aminotransferase score (AUROC: 0.7914; p = 0.09) in the validation cohort. This RF model, termed the STEALTH-ARMS model, was implemented as an online application to enable patient-based individual risk assessment.

CONCLUSIONS

The RF-based ML model demonstrated high accuracy in identifying at-risk MASH using only seven routine clinical data, offering a highly accurate, noninvasive, and cost-effective alternative to LSM-based methods. This approach holds promise for broader clinical applications, particularly in resource-limited settings.

TRIAL REGISTRATION: UMIN-CTR 000049068.