Scholarlink

<p>Accurate survival prediction in chronic liver disease remains a vital yet challenging task, as many published models are narrowly optimized for specific datasets and lack transparency for clinical translation. This study presents a <i>reusable and auditable machine learning pipeline</i> for tabular clinical prognosis, designed to ensure reproducibility, calibration, and explainability from data ingestion to decision support. The workflow standardizes schema harmonization, missingness profiling and imputation, signal conditioning, leakage-safe resampling, and a consistent <i>model garden</i> comprising logistic regression (PCA, LASSO), decision tree, random forest, and XGBoost classifiers. A calibration and explainability layer combining reliability curves, Brier scores, SHAP, and LIME analyses enhances interpretability and confidence in model outputs. Using the Mayo primary biliary cirrhosis (PBC) dataset (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(n=418\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>n</mi> <mo>=</mo> <mn>418</mn> </mrow> </math></EquationSource> </InlineEquation>) as a case study, XGBoost achieved the highest discrimination (AUROC = 0.95) and robust calibration across probability thresholds, while ensemble models demonstrated superior stability compared to linear baselines. SHAP analysis identified bilirubin, prothrombin time, follow-up duration, and age as dominant predictors, consistent with hepatology insights. By integrating calibration, transparency, and reproducibility, this pipeline offers a generalized framework for clinical prognosis tasks beyond cirrhosis and supports reliable deployment of machine learning in clinical decision support systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

From Dataset to Deployment: A Reusable ML Pipeline for Clinical Prognosis with SHAP-Based Transparency

  • 期刊论文

  • 发表日期：

    2026年3月8日

查看全文
  • Estabrag Abaker,
  • Reema Alduhayan,
  • Abd-Elhamid M. Taha,
  • Nidal Nasser

摘要

Accurate survival prediction in chronic liver disease remains a vital yet challenging task, as many published models are narrowly optimized for specific datasets and lack transparency for clinical translation. This study presents a reusable and auditable machine learning pipeline for tabular clinical prognosis, designed to ensure reproducibility, calibration, and explainability from data ingestion to decision support. The workflow standardizes schema harmonization, missingness profiling and imputation, signal conditioning, leakage-safe resampling, and a consistent model garden comprising logistic regression (PCA, LASSO), decision tree, random forest, and XGBoost classifiers. A calibration and explainability layer combining reliability curves, Brier scores, SHAP, and LIME analyses enhances interpretability and confidence in model outputs. Using the Mayo primary biliary cirrhosis (PBC) dataset ( \(n=418\) n = 418 ) as a case study, XGBoost achieved the highest discrimination (AUROC = 0.95) and robust calibration across probability thresholds, while ensemble models demonstrated superior stability compared to linear baselines. SHAP analysis identified bilirubin, prothrombin time, follow-up duration, and age as dominant predictors, consistent with hepatology insights. By integrating calibration, transparency, and reproducibility, this pipeline offers a generalized framework for clinical prognosis tasks beyond cirrhosis and supports reliable deployment of machine learning in clinical decision support systems.