B54-25 Risk Stratification and Heterogeneity for Organ Damage in Acute Diquat Poisoning: A Multi-Center Cohort Study

Objective This study aimed to characterize the diquat toxicity and, in parallel, to develop interpretable machine-learning models to enable early prediction of organ damage in acute diquat poisoning. Methods This multi-center retrospective cohort study included 193 patients with acute diquat poisoning admitted between February 2022 and July 2023. We evaluated 8 types of organ damage: toxic encephalopathy, circulation collapse, myocardial injury, acute kidney injury, skeletal muscle injury, liver injury, lung injury, and gastrointestinal damage. To characterize the relationship between initial plasma diquat concentration and the risk of organ damage, logistic regression models with restricted cubic splines were fitted. The inflection point, representing a clinically meaningful threshold, was defined as the plasma concentration at which the spline-estimated odds ratio first exceeded 1.0, indicating the onset of increased risk relative to baseline. For each organ damage outcome, prediction models were developed using data from the primary study center (development set) and externally validated using data from 28 other hospitals (validation set). Model development followed an iterative feature selection and algorithm comparison framework. Feature importance was quantified using SHapley Additive exPlanations (SHAP) values derived from an XGBoost model fitted to the development dataset. Three machine learning algorithms (Random Forest, CatBoost, and XGBoost) were evaluated. Features were sequentially incorporated according to their SHAP-based importance rankings. Starting from the most important feature, models were iteratively retrained with additional features, and performance was evaluated by the 5-fold cross-validation on the development set. The feature set yielding the highest mean area under the receiver operating characteristic curve (AUROC) was selected for each algorithm. The algorithm achieving the highest cross-validated AUROC was designated as the optimal model for that specific organ damage outcome. Results Restricted cubic spline analysis revealed distinct plasma diquat concentration thresholds associated with increased organ damage risk. The inflection points varied across organ systems: acute kidney injury demonstrated the lowest threshold at 124 ng/mL, followed by liver injury at 168 ng/mL, skeletal muscle injury at 313 ng/mL, myocardial injury at 388 ng/mL, toxic encephalopathy at 551 ng/mL, and circulation collapse at 662 ng/mL. Lung and gastrointestinal injuries exhibited a unique pattern, with no significant inflection point identified for lung injury and risk increasing from the lowest detectable concentrations for gastrointestinal injury. External validation confirmed the generalizability of all prediction models, with AUROCs ranging from 0.752 to 0.904. Conclusions This study provided a practical framework for risk stratification and targeted organ-specific monitoring in acute diquat poisoning. This abstract is funded by: National Natural Science Foundation of China, grant number [No. 82372214]