| Autor: |
Jin Ge, Jean C. Digitale, Cynthia Fenton, Charles E. McCulloch, Jennifer C. Lai, Mark J. Pletcher, Efstathios D. Gennatas |
| Rok vydání: |
2023 |
| DOI: |
10.1101/2023.03.03.23286729 |
| Popis: |
BackgroundLiver transplantation (LT) is a treatment for acute-on-chronic liver failure (ACLF) but up to 40% mortality post-LT has been reported. Existing post-LT models in ACLF have been limited by small samples. In this study, we developed a novel Expert-Augmented Machine Learning (EAML) model to predict post-LT outcomes.MethodsWe identified ACLF patients in the University of California Health Data Warehouse (UCHDW). We used EAML, which uses the RuleFit machine learning (ML) algorithm to extract rules from decision-trees that are then evaluated by human experts, to predict post-LT outcomes. We compared EAML/RuleFit’s performances versus other popular models.ResultsWe identified 1,384 ACLF patients. For death at one-year: areas-under-the-receiver-operating characteristic curve (AUROCs) were 0.707 (Confidence Interval [CI] 0.625-0.793) for EAML and 0.719 (CI 0.640-0.800) for RuleFit. For death at 90-days: AUROCs were 0.678 (CI 0.581-0.776) for EAML and 0.707 (CI 0.615-0.800) for RuleFit. In pairwise comparisons, EAML/RuleFit models outperformed cross-sectional models. Divergences between experts and ML in rankings revealed biases and artifacts in the underlying data.ConclusionsEAML/RuleFit outperformed cross-sectional models. Significant discrepancies between experts and ML occurred in rankings of biomarkers used in clinical practice. EAML may serve as a method for ML-guided hypothesis generation in further ACLF research. |
| Databáze: |
OpenAIRE |
| Externí odkaz: |
|
