Introduction
Advanced fibrosis is the primary prognostic determinant in patients with metabolic dysfunction-associated steatotic liver disease (MASLD). Current estimates suggest that more than 25% of adults worldwide are affected by MASLD, and it is expected to become the leading cause of liver transplants in the future [1]. Current non-invasive tools like Fibrosis-4 (FIB-4) lack optimal balance between sensitivity and specificity. We aimed to develop and compare two ML models: Extreme Gradient Boosting (XGBoost) and logistic regression to predict advanced fibrosis using routinely available clinical data, with liver stiffness measurement via transient elastography as the reference standard.
Aims & Methods
We used data from NHANES 2017-2018. Adults with valid liver stiffness measurement from transient elastography were included. Advanced fibrosis was defined as liver stiffness >=8.2 kPa. Input features included AST, ALT, platelet count, BMI, triglycerides, glucose, and diabetes status. A total of 2,625 participants were included. Data were split 80:20 into training and testing sets respectively using stratified sampling to preserve fibrosis prevalence (~9%). To address class imbalance, the XGBoost model was trained using a weighted loss function. Logistic regression was used as a baseline comparator. Classification thresholds were optimized using Youden's Index to determine the best trade-off between over- and under-classification of fibrosis. Model performance was evaluated using AUROC, accuracy, sensitivity, and specificity.
Results
Both models demonstrated good discriminatory performance. The XGB model achieved an AUROC of 0.73, with sensitivity of 78.9%, specificity of 61.2%, and accuracy of 77.3%. While logistic regression yielded a slightly higher AUROC of 0.77, greater specificity (81.6%), and lower sensitivity (60.7%) and accuracy (62.7%).
Conclusion
Both ML models using simple clinical and biochemical inputs demonstrated reasonable performance in detecting advanced fibrosis among patients with MASLD. The XGBoost model provided a more balanced profile with improved sensitivity, while logistic regression offered better specificity. To address class imbalance (~9% prevalence), XGBoost used a weighted loss function, and both models were calibrated using Youden's Index to optimize classification thresholds. Limitations include reliance on a non-invasive fibrosis surrogate (transient elastography), and lack of external validation. Future work should validate these models in prospective cohorts and explore integration into digital clinical tools.
References
[1] McTeer, M., Applegate, D., Mesenbrink, P., Ratziu, V., Schattenberg, J. M., Bugianesi, E., Geier, A., Romero Gomez, M., Dufour, J. F., Ekstedt, M., Francque, S., Yki-Jarvinen, H., Allison, M., Valenti, L., Miele, L., Pavlides, M., Cobbold, J., Papatheodoridis, G., Holleboom, A. G., Tiniakos, D., … LITMUS Consortium investigators (2024). Machine learning approaches to enhance diagnosis and staging of patients with MASLD using routinely available clinical information. PloS one, 19(2), e0299487. https://doi.org/10.1371/journal.pone.0299487
[2] Satapathy, S. K., Bernstein, D. E., & Roth, N. C. (2022). Liver transplantation in patients with non-alcoholic steatohepatitis and alcohol-related liver disease: the dust is yet to settle. Translational gastroenterology and hepatology, 7, 23. https://doi.org/10.21037/tgh-2020-15
