Comment on: Predicting mycophenolic acid exposure in pediatric lupus nephritis: a real-world study across three clinical scenarios

Dear Editor, Liu et al. should be congratulated for developing machine-learning models to predict mycophenolic acid (MPA) exposure in pediatric lupus nephritis (LN) across three clinically relevant scenarios [1]. Their study is clinically appealing because therapeutic drug monitoring of MPA is burdensome in children and because their own earlier Rheumatology study linked MPA exposure to induction outcomes in childhood LN [2]. We nevertheless believe that three methodological issues need clarification before these models are interpreted as ready for clinical use. First, the study appears to use the monitoring episode, rather than the patient, as the analytical unit. The authors state that each MPA concentration monitoring record was treated as an independent event, and that Scenarios 1 and 2 used a 7:3 training-testing split with 10-fold cross-validation [1]. In a dataset with repeated follow-up records from the same child, random allocation of visits can allow patient-specific information to appear in both training and test sets, inflating apparent model performance through within-patient correlation rather than true transportability. This issue is particularly relevant in pharmacokinetic prediction, where stable characteristics such as body size, renal function and treatment course are informative. A patient-level split, leave-one-patient-out analysis or, ideally, external validation would therefore be more convincing [3].