Key result
Machine learning predicts piperacillin/tazobactam-associated hypokalemia with ~0.76 AUC driven by baseline potassium and diuretics.
Why the study?
Hypokalemia is a common adverse event during piperacillin/tazobactam therapy in critically ill patients, and predictive models using baseline variables are needed.
Can machine learning models accurately predict piperacillin/tazobactam-associated hypokalemia in critically ill patients?
Cohort (n=1,766)
Yes
Can machine learning models accurately predict piperacillin/tazobactam-associated hypokalemia in critically ill patients?
Machine learning models utilizing baseline variables, particularly baseline potassium and potassium-wasting diuretic use, can moderately predict the risk of piperacillin/tazobactam-associated hypokalemia in critically ill patients.
Answers are grounded in this paper and cite their sources.
Moderate ML prediction supports targeted monitoring in critical care; leaves open clinical utility pending prospective validation.
Background: Hypokalemia is a common adverse event during piperacillin/tazobactam (TZP) therapy in critically ill patients. This study developed and externally validated machine learning models to predict hypokalemia using baseline variables. Methods: Data were obtained from the Medical Information Mart for Intensive Care (MIMIC)-IV database and an external cohort from the Third Affiliated Hospital of Zunyi Medical University. Hypokalemia was defined as the first serum potassium <3.5 mmol/L within 14 days after TZP initiation. Variables with >20% missing data were excluded; remaining missing values were imputed. Least Absolute Shrinkage and Selection Operator (LASSO) regression with 10-fold cross-validation based on binomial deviance was used for feature selection. Logistic Regression, Random Forest, Extreme Gradient Boosting Model (XGBoost), Adaptive Boosting (AdaBoost), and Gradient Boosting Machine models were developed. Performance was assessed with receiver operating characteristic curves, the area under the receiver operating characteristic curve (AUC), the area under the precision-recall curve (AUPRC), calibration, precision-recall analysis, and decision curve analysis. Results: The internal cohort included 1,345 patients (594 hypokalemia), and the external cohort included 421 patients (159 hypokalemia). Internal AUCs were 0.759 (Random Forest), 0.738 (GBM), 0.735 (XGBoost), 0.716 (AdaBoost), and 0.694 (Logistic Regression). External AUCs were 0.668 (GBM), 0.654 (XGBoost), 0.648 (Random Forest), 0.645 (AdaBoost), and 0.620 (Logistic Regression). The Random Forest model achieved an AUPRC of 0.701 in the internal cohort, while GBM achieved the highest AUPRC in the external cohort (0.554). Calibration and decision curve analyses demonstrated acceptable agreement and potential clinical utility. SHapley Additive exPlanations (SHAP) analysis identified baseline potassium, potassium-wasting diuretic use, serum creatinine, age, renal disease, serum sodium, and albumin as the most influential predictors of TZP-associated hypokalemia. Conclusion: Machine learning models achieved moderate predictive performance. Baseline potassium and potassium-wasting diuretic use were the strongest predictors.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2026) conducted a cohort in Piperacillin/tazobactam-associated hypokalemia (n=1,766). Piperacillin/tazobactam was evaluated on First serum potassium <3.5 mmol/L within 14 days after TZP initiation (AUC 0.759, 95% CI 0.733-0.784). Machine learning models achieved moderate predictive performance for piperacillin/tazobactam-associated hypokalemia, with baseline potassium and potassium-wasting diuretic use as the strongest predictors.
Synapse has enriched 10 related papers. Consider them for comparative context: