Real-World Data Driven Early Warning System for Risk-Stratified Liver Injury in Hospitalized COVID-19 Patients: Machine Learning Models for Clinical Decision Support

Yuanguo Xiong¹Xu Cai¹Xin Lai²Wang Yuwen *³Hao Xin⁴Wei Song¹Feng Lv¹Xianxi Guo¹Ge Yang⁵Yue Wu^1*

• ¹Department of Pharmacy, Renmin Hospital of Wuhan University, Wuhan, China
• ²The First Affiliated Hospital of Nanchang University, Nanchang, Jiangxi Province, China
• ³Department of Pharmacy, Macheng People’s Hospital, Macheng, China
• ⁴Third People's Hospital of Qingdao, Qingdao, Shandong Province, China
• ⁵Department of Pharmacy,The First Hospital Affiliated to Army Medical University, Chongqing, Chongqing Municipality, China

To develop and validate a real-world evidence-driven early warning system for the risk-stratified prediction of COVID-19-associated hepatic dysfunction in hospitalized patients, leveraging interpretable machine learning models to provide clinically actionable decision support for timely intervation.Methods: A retrospective single-center cohort study was conducted utilizing high-resolution electronic health records (EHR) from 983 hospitalized COVID-19 patients. Clinical features (eg. laboratory results, medication exposures, and disease progression markers) were systematically analyzed. To mitigate class imbalance, we employed the Synthetic Minority Over-sampling Technique (SMOTE) prior to model development. Thirteen distinct machine learning (ML) algorithms were trained and benchmarked to construct an optimal risk stratification framework. Model performance was rigorously evaluated using metrics, including accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC). SHapley Additive exPlanations (SHAP) analysis was employed to enhance clinical interpretability and provided transparent decision-making insights.The SMOTEENN resampling strategy combined with Random Forest (RF) and Extra Trees (ET) models demonstrated superior predictive performance, achieving AUC values of 0.998±0.002 (RF) and 0.997±0.002 (ET), respectively. The SHAP-based interpretability analysis identified glutathione administration and hepatic enzymes (e.g., GGT, ALT) as the most influential predictors. The online prediction platforms were developed for liver injury early warning risk stratification (low-, and high-risk) based on predicted probabilities classification. Conclusions: This research successfully established machine learning-powered early warning system capable of real-time risk stratification for COVID-19-associated liver injury through dynamic clinical data integration. The ensemble RF/ET-based models demonstrated significant clinical utility as decision support tools, particularly through ability to identify high-risk patients requiring intensified monitoring and optimize hepatoprotective. By emphasizing drug-induced injury markers and disease progress process, ML models establish personalized monitoring framework that could potentially transform clinical management for target patients.