Interpretable prediction on treatment response of piperacillin–tazobactam for lower respiratory tract infections using machine learning

Empiric piperacillin/tazobactam (PIPT) therapy is commonly used in high-risk patients with severe community-acquired pneumonia. However, its overuse may lead to antibiotic resistance and unwanted side effects. To this end, we developed and validated a machine learning (ML) model to assess the effectiveness of PIPT in the treatment of septic lower respiratory tract infections and to explore relevant influencing factors. The study was based on data from hospitalized patients treated with PIPT, and a dataset of bacterial lower respiratory tract infections was constructed by retrospective analysis and divided into training and testing sets in a 7:3 ratio. After screening the key predictors using least absolute shrinkage and Least Absolute Shrink age and Selection Operator regression methods, 5 ML models, including logistic regression and random forest, were used to train these factors to predict efficacy. Model interpretation was performed using the SHapley Additive exPlanations technique. The results showed that the decision tree model had a performance score of 0.73 (95% CI 0.61–0.86) for prediction. The SHapley Additive exPlanations analysis identified several important factors for treatment failure, including low serum albumin levels, reduced drug dosage, and comorbidities such as chronic obstructive pulmonary disease and heart failure, in addition to an unfavorable neutrophil-to-lymphocyte ratio of ≥70%. This study demonstrates that the ML model is effective in predicting the outcome of PIPT therapy and helps to personalize medical regimens while adjusting strategies by identifying high-risk individuals, ultimately achieving the dual goals of optimizing patient care and reducing inappropriate antibiotic use.