Predictive performance of multi-model approaches for model-informed precision dosing of piperacillin in critically ill patients

•Multi-model approaches potentially increase the applicability of MIPD.•No need for individual model selection if at least one TDM-sample is available.•Comparable or better predictive performance compared to best single models.•Model averaging algorithm performs better compared to model selection algorithm. Piperacillin (PIP)/tazobactam is a frequently prescribed antibiotic; however, over- or underdosing may contribute to toxicity, therapeutic failure, and development of antimicrobial resistance. An external evaluation of 24 published PIP-models demonstrated that model-informed precision dosing (MIPD) can enhance target attainment. Employing various candidate models, this study aimed to assess the predictive performance of different MIPD-approaches comparing (i) a single-model approach, (ii) a model selection algorithm (MSA) and (iii) a model averaging algorithm (MAA). Precision, accuracy and expected target attainment, considering either initial (B1) or initial and secondary (B2) therapeutic drug monitoring (TDM)-samples per patient, were assessed in a multicentre dataset (561 patients, 11 German centres, 3654 TDM-samples). The results demonstrated a slight superiority in predictive performance using MAA in B1, regardless of the candidate models, compared to MSA and the best single models (MAA, MSA, best single models: inaccuracy ±3%, ±10%, ±8%; imprecision: 73%). The inclusion of a second TDM-sample notably improved precision and target attainment for all MIPD-approaches, particularly within the context of MSA and most of the single models. The expected target attainment is maximized (up to >90%) when the TDM-sample is integrated within 24 h. In conclusion, MAA streamlines MIPD by reducing the risk of selecting an inappropriate model for specific patients. Therefore, MIPD of PIP using MAA implicates further optimisation of antibiotic exposure in critically ill patients, by improving predictive performance with only one sample available for Bayesian forecasting, safety, and usability in clinical practice. [Display omitted]