Machine learning-enabled early risk stratification of β-Iactam-induced electrolyte imbalances

초록

Aims: To estimate the incidence of beta-lactam/beta-lactamase inhibitor - associated electrolyte imbalances and develop an internally validated, interpretable prediction model for early risk identification. Patients and methods: We retrospectively analyzed 240 hospitalized adults treated with piperacillin - tazobactam or ampicillin - sulbactam. Electrolyte imbalance was defined as any post-treatment abnormality in sodium, potassium, chloride, calcium, or phosphate using CTCAE v5.0 and predefined reference ranges. Predictors were evaluated using logistic regression and machine learning - based feature selection across five algorithms, followed by repeated stratified cross-validation. Results: Electrolyte imbalances occurred in 71 patients (29.6%). Respiratory disease was associated with higher risk (AOR 2.616; 95% CI 1.072-6.386). Six consensus predictors were selected (age >= 65 years, ARBs, cardiovascular disease, endocrine disease, expectorants, respiratory disease). Logistic regression showed the best overall performance (AUROC 0.68; AUPRC 0.52) with high NPV (similar to 0.80). Conclusions: Electrolyte imbalances were common during beta-lactam/beta-lactamase inhibitor therapy. An interpretable model using routine clinical variables may support risk-informed monitoring; external validation is warranted.

키워드