Sepsis remains a pressing challenge in intensive care medicine due to its high morbidity and mortality [1]. The early and appropriate use of antibiotics is thought to be fundamental for effective treatment of sepsis [2]. There is a strong rationale for personalizing antibiotic dosing, based on the association between antibiotic exposure and bacterial killing, and observed between-patient variability of antibiotic concentrations in plasma and at the site of infection [3]. This is further exemplified by risks of toxicity and adverse effects when patients are overdosed.

However, reaching and maintaining adequate antibiotic exposure is difficult. This is especially true for critically ill patients, as changing organ function and shifts in fluid balance can affect clearance and/or distribution of antibiotics [3]. This introduces within-patient variability during the course of disease, further complicating treatment. Consequently, studies show that conventional dosing can result in as much as a 500-fold variation in antibiotic concentrations in these patients [4]. This underlines the complexity of selecting appropriate dosages, which might be a daunting task for physicians during routine clinical practice, particularly without expert support [5].

Model-Informed Precision Dosing (MIPD, Fig. 1) has been introduced as an approach for adjusting treatment based on the characteristics of individual patients; optionally including information from antibiotic plasma concentration sampling [2, 6,7,8]. Pharmacokinetic (PK) models allow for the selection of individual dosing schedules which meet pre-specified target drug concentrations. These models describe how PK parameters such as drug clearance and volume of distribution vary between patients based on measured covariates. Unfortunately, both the development as well as the use of these types of models require considerable expertise and can be difficult for physicians to use without support from, for example, clinical pharmacists. To reduce this complexity, automated dosing software tools have been developed. These tools aim to bring MIPD closer to the bedside in a user-friendly manner. They directly interface with electronic health record (EHR) or patient data management systems (PDMS)—obviating the need for manual data entry—and offer visualizations of concentration–time curves to help physicians choose the optimal treatment strategy.

Fig. 1
figure 1

Model informed precision dosing requires accurate pharmacokinetic, machine learning or hybrid models, ideally deriving their input directly from the large amount of routinely collected patient data in the electronic health records and optionally from feedback from antibiotic plasma concentration sampling. To date, model informed precision dosing is mostly implemented by relying on pharmacist support for model interpretation and dosing advise. Making computerized decision support available to intensive care healthcare professionals directly has recently also been proven feasible and may facilitate adequate dosing right from the start of antibiotic therapy. In the future, these intensive care professionals might also be taken out of the loop as we may progress to fully automated and closed loop antibiotic dosing

Full size image

For the Right Dose Right Now randomized clinical trial, the AutoKinetics software (Amsterdam University Medical Center and OLVG Amsterdam) was developed and tested for the treatment of critically ill patients suffering from sepsis based on MIPD [6, 7].

The software collects real-time data of patient covariates from the EHR to predict the optimal dose to attain predefined PK target exposure right from the moment of admission in the intensive care unit (ICU) without requiring plasma concentration data. The authors found that the use of AutoKinetics was feasible, safe, and significantly improved PK target attainment for ciprofloxacin (but not for three other antibiotics) while obtaining excellent compliance by intensive care healthcare professionals [7]. However, improved target attainment for ciprofloxacin did not lead to significant differences in clinical outcomes, although this study was not powered for this finding. In the DOLPHIN study, MIPD relying on plasma concentration feedback (available 12–48 h after antibiotic initiation) was applied remotely by pharmacists for beta-lactam antibiotics and ciprofloxacin based on previously published PK models [8]. In this study, there was no significant difference in length of stay, or target attainment between standard dosing and PK-guided dosing.

The findings from these two studies point to some limitations of current approaches to implement MIPD. First, it is crucial to provide dosing guidance from the very start of antibiotic therapy. Second, evidence for the efficacy of MIPD is lacking and greatly relies on the accuracy of PK models, which is currently limited at best. Improving their accuracy might lead to better outcomes. PK models sometimes lack important covariates known to affect drug metabolism or include only simple representations of their effect. For example, serum creatinine is often used as a surrogate marker of kidney clearance. Several equations, such as CKD-EPI, exist to estimate globular filtration rate based on creatinine concentrations, and such estimates are often included in PK models. However, these estimates represent renal function at a single moment in time and will only affect predictions after this time point. To more accurately predict the changing PK, models might need to learn the continuous change in renal clearance, likely represented by complex combinations of multiple covariates. The first hours of treatment pose a specific challenge in this context as kidney function may improve after resuscitation, and biomarkers alone are not reliable to guide therapy [9].

In classical PK models, the equations representing the relationship between covariates and PK parameters need to be chosen manually during model development. Novel approaches involving Big Data represent a promising new avenue for refining PK models. Relationships between covariates and PK parameters can be learned directly from data, which are plentiful within the centralized EHR or PDMS systems. The HEROI2C research team developed a machine learning (ML) model for the prediction of piperacillin concentrations in critically ill patients [10]. The ML model was more accurate on an external data set compared to existing population PK models when no concentration measurements were available. However, these classical ML-based approaches are less reliable when extrapolated to unseen training examples or dosing schedules [11], and additional steps may be needed to improve performance. Several hybrid ML/PK models have been proposed that resolve this issue [11, 12]. Neural Ordinal Differential Equations (Neural-ODE)-based approaches are of particular interest as they can be used to learn “hidden” time-varying effects, such as the changing renal clearance in our prior example [13]. This may be the missing piece in the puzzle to improve the accuracy of antibiotic concentration predictions [14]. A natural extension of such software systems are target-controlled infusion systems where dose administration is fully controlled by a computer [15]. Such systems could be based on complex models leveraging the data from EHR systems to continuously adapt antibiotic dose to reach target concentrations. Current models used for MIPD do not yet support such an approach.

In conclusion, recent studies demonstrate the feasibility of using automated dosing software to assist ICU healthcare workers in antibiotic dosing decision making. The introduction of ML-based approaches potentially offers interesting opportunities for improving the prediction of antibiotic concentrations. More research is needed to evaluate its benefits and for assessing the reliability of MIPD of antibiotics and other antimicrobials. It is with great interest that we look forward to the developments in the field—watch this space!