034: Application of Hybrid Pharmacometric/Machine-Learning Approach to Inform Moxifloxacin Therapeutic Drug Monitoring in Children

Statement of Purpose: Moxifloxacin is a fluoroquinolone antibiotic active against drug-resistant bacteria. Pediatric use is limited by lacking data. We sought to establish a limited sampling strategy of plasma concentrations to support safe and effective moxifloxacin dosing in children.

Description of Methods & Materials: Moxifloxacin plasma concentrations were simulated in an in-silico pediatric population (n=500 per dose) from a population pharmacokinetic model derived from children 0-18 years with tuberculosis (PMID:34286843). Briefly, the model followed 2-compartment distribution with allometrically-scaled clearance and volume and clearance~height-for-age z-score (HAZ) effect. Dosing simulations included 100 mg, 200 mg, 300 mg and 400 mg. C_max was the maximum concentration in each curve. AUC₂₄ was integration of the curve from 0 to 24 hours. Random forest (RF) models were built to predict pharmacokinetic outcomes (C_max, AUC₂₄) using 75/25 train/test split. Features were plasma concentrations at times (C-t) 1-24 hours after dose, dose, weight and HAZ. C-t for limited sampling models were selected from RF feature importance. Models were evaluated by root mean squared error (RMSE) of test predictions and externally validated with clinical trial data. All simulations and modeling were performed in R.

Data & Results: The in-silico population had median (range) weight of 27 (5-49) kg and corresponding dose range of 2-80 mg/kg. Important C-t features (rank high to low) were 6, 7, 5, 4, and 8 h for AUC₂₄ and 3, 4 and 2 h for C_max. Models including one C-t feature predicted AUC₂₄ with 13.8% (C-t4), 11.7% (C-t5), 11.6% (C-t6), 11.8% (C-t7) and 11.9% (C-t8) RMSE and C_max with 8.1% (C-t2), 9.3% (C-t3), and 13.7% (C-t4) RMSE. Better predictions were obtained using two C-t features: RMSE was 9.1% (AUC₂₄) and 6.8% (C_max) for C-t2+C-t6, 7.9% (AUC₂₄) and 7.3% (C_max) for C-t2+C-t8 and 8.9% (AUC₂₄) and 13.0% (C_max) for C-t4+C-t8. Clinical data for validation were available at 2, 4, and 8 h after dose in n=32 children dosed at ~10 mg/kg. AUC₂₄ predicted with 14.4% (Ct-8), 20.9% (Ct-4), 13.8% (Ct-2 + C-t8) and 12.0% (Ct-4+C-t8) RMSE.

Interpretation, Conclusion or Significance: Moxifloxacin AUC₂₄ and C_max were accurately predicted with one or two sampling points. The best one-concentration sample was 2 h for C_max and 6 h for AUC₂₄. Sampling at 2 h and 8 h allows more accurate prediction of C_max and AUC₂₄. This work demonstrates how machine-learning can be applied with pharmacometrics models to inform clinical drug monitoring and accurately predict pharmacokinetic indices. This hybrid method was computationally efficient, simple, and could be easily implemented into a clinical tool.

Disclosures: No disclosures.