Machine learning approach for dosage individualization of azithromycin in children with community-acquired pneumonia

The uncertainty about the efficacy and safety of currently used azithromycin dosing regimens in children warrants individualized therapy. The area under the plasma concentration-time curve over 24 h (AUC_0-24) of azithromycin correlates best with its effectiveness. The aim of this study was to evaluate the ability of machine learning (ML) to predict the AUC_0-24 of azithromycin in children with community-acquired pneumonia.

Various ML algorithms were used to build ML models based on simulated pharmacokinetic profiles from a published population pharmacokinetic model. A priori-ML model predicted AUC_0-24 using patients' characteristics and after the trough concentration (C₀) became available, a posteriori-ML model was built for improved prediction. Statistical methods and pharmacodynamic (PD) evaluation methods were used to evaluate the ML model's predictive accuracy in a real-world study. ML-optimized doses were evaluated by calculating the probability of PD target attainment in virtual trials compared with guideline-recommended doses.

The AUC_0-24 can be predicted by priori-ML model using the CatBoost algorithm with dosing regimen and two covariates as predictors (weight, alanine aminotransferase) before initial administration. A posteriori-ML model using CatBoost algorithm was built with adding C₀ as a predictor. In real-world validation, the mean absolute prediction error of the priori-ML and posteriori-ML models was less than 30%. The accuracy (determining whether the PD target is met) of the priori-ML model was 76.3%, whereas that of the posteriori-ML model increased to 90.4%.

ML models were established to predict the AUC_0-24 of azithromycin successfully and could be used for individual dose adjustment in children before treatment and after obtaining C₀.