|
KMID : 1100820230130030189
|
|
Laboratory Medicine Online
2023 Volume.13 No. 3 p.189 ~ p.198
|
|
|
Liver Fibrosis Biomarker Validation Using Machine Learning Algorithms
|
|
Kim Min-Kyeong
Kwon Heon-Ju
Shin Kang-Su
Park Hyo-Soon
Woo Hee-Yeon
Park Chang-Hun
Kwon Min-Jung
|
|
|
Abstract
|
|
|
Background: Liver fibrosis which causes several liver diseases, requires early screening and management. The gold standard for fibrosis assessment, liver biopsy, has recently been replaced by noninvasive scores. In this study, we validated liver fibrosis-associated biomarkers using machine learning techniques applied in medical research and evaluated their prediction models.
Methods: Noninvasive scores were assayed in 144 patients who underwent transient elastography (TE). The patients were divided into three groups (<7 kPa, 7?10 kPa, ≥10 kPa) according to their TE results. Feature selection and modeling for predicting liver fibrosis were performed using random forest (RF) and support vector machine (SVM).
Results: Considering the mean decrease in impurity, permutation importance, and multicollinear analysis, the important features for differentiating between the three groups were Mac-2 binding protein glycosylation isomer (M2BPGi), platelet count, and aspartate aminotransferase (AST). Using these features, the RF and SVM models showed equivalent or better performance than noninvasive scores. The sensitivities of RF and SVM models for predicting ≥7 kPa TE results were higher than noninvasive scores (83.3% and 90.0% vs. <80%, respectively). The sensitivity and specificity of RF and SVM models for ≥10 kPa TE result was 100%.
Conclusions: We used machine learning techniques to verify the usefulness of established serological biomarkers (M2BPGi, PLT, and AST) that predict liver fibrosis. Conclusively, machine learning models showed better performance than noninvasive scores.
|
|
|
KEYWORD
|
|
|
Liver fibrosis, Stiffness measurements, M2BPGi, Machine learning, Random forest, Support vector machine
|
|
|
FullTexts / Linksout information
|
|
|
|
|
|
Listed journal information
|
|
|
|