Employing artificial intelligence for optimising antibiotic dosages in sepsis on intensive care unit: a study protocol for a prospective observational study (KI.SEP) Authors: Britta Marko, Lars Palmowski, Hartmuth Nowak, Andrea Witowski, Björn Koos, Katharina Rump, Lars Bergmann, Julia Bandow, Martin Eisenacher, Patrick Günther, Michael Adamzik, Barbara Sitek, Tim Rahmel Published: 2024 ABSTRACT AbstractIntroductionIn sepsis treatment, achieving and maintaining effective antibiotic therapy is crucial. However, optimal antibiotic dosing faces challenges due to significant variability among patients with sepsis. Therapeutic drug monitoring (TDM), the current gold standard, lacks initial dosage adjustments and global availability. Even with daily TDM, antibiotic serum concentrations (ASCs) often deviate from the therapeutic range. This study addresses these challenges by developing machine learning (ML)-based ASC prediction models capable of handling variable data input and encompassing diverse clinical, laboratory, microbiological and proteomic parameters without the need for daily TDM.MethodsThis prospective observational study is conducted in a German university hospital intensive care unit. Eligible sepsis patients receive continuous antibiotic therapy with piperacillin/tazobactam (n=100) or meropenem (n=100) within 24 hours. Exclusion criteria include refusal, pregnancy, lactation and severe anaemia (haemoglobin <8 g/dL). Blood samples for TDM are collected from patients, along with clinical and laboratory parameters on days 1–8 and day 30 or on discharge. ML models predicting ASC between day 1 and day 8 serve as primary and key secondary endpoints. We will use the collected data to develop multifaceted ML-based algorithms aimed at optimising antibiotic dosing in sepsis. Our two-way approach involves creating two distinct algorithms: the first focuses on predictive accuracy and generalisability using routine clinical parameters, while the second leverages an extended dataset including a plethora of factors currently insufficiently explored and not available in standard clinical practice but may help to enhance precision. Ultimately, these models are envisioned for integration into clinical decision support systems within patient data management systems, facilitating automated, personalised treatment recommendations for sepsis.Ethics and disseminationThe study received approval from the Ethics Committee of the Medical Faculty of Ruhr-University Bochum (No. 23-7905). Findings will be disseminated through open-access publication in a peer-reviewed journal and social media channels.Trial registration numberDRKS00032970. STRENGTHS AND LIMITATIONS OF THIS STUDY This study employs a prospective cohort design with daily collection of clinical, laboratory and microbiological data, providing a robust dataset for machine learning (ML) model training and validation. This study employs a prospective cohort design with daily collection of clinical, laboratory and microbiological data, providing a robust dataset for machine learning (ML) model training and validation. Two ML algorithms will be developed: one using commonly available clinical parameters and another incorporating advanced proteomic and biomarker data, allowing for methodological comparison of predictive accuracy. Two ML algorithms will be developed: one using commonly available clinical parameters and another incorporating advanced proteomic and biomarker data, allowing for methodological comparison of predictive accuracy. The study design includes both internal validation using a 70/30 train-test split and external validation with an independent multicentre cohort (SepsisDataNet.NRW), ensuring methodological rigour in model performance assessment. The study design includes both internal validation using a 70/30 train-test split and external validation with an independent multicentre cohort (SepsisDataNet.NRW), ensuring methodological rigour in model performance assessment. A major limitation of the study is the potential for missing data due to variability in patient records, which will be addressed through the use of advanced ML techniques such as imputation methods. A major limitation of the study is the potential for missing data due to variability in patient records, which will be addressed through the use of advanced ML techniques such as imputation methods. It remains unclear to what extent the routinely available data may compromise the prognostic accuracy and the reliability of the results. It remains unclear to what extent the routinely available data may compromise the prognostic accuracy and the reliability of the results. Introduction Sepsis is characterised by a dysregulated immune response, caused by a severe infection.1 Approximately 48.9 million incident cases of sepsis are recorded worldwide each year, resulting in 11.0 million sepsis-related deaths, constituting 19.7% of all global mortality.2 Primarily, effective antibiotic treatment remains the cornerstone of causal sepsis therapy and exerts a significant impact on the outcome of patients with sepsis.3 4 The proper dosing of antibiotics and consequently the attainment of adequate antibiotic serum concentrations (ASCs) are crucial factors determining the effectiveness and the clinical success.5 However, ASCs are affected by the pharmacokinetics (PK) of the antibiotic substances in the complex and highly individual settings in sepsis patients.6 7 Therefore, especially sepsis patients are prone to outranged ASCs posing the risk of therapy failure and resistance development in cases of underdosing or severe organ damage in case of overdosing.5 This problem is particularly important for ß-lactam antibiotics such as piperacillin/tazobactam or meropenem.8 Here, the ASC should be dosed above the minimum inhibitory concentration (MIC) of the pathogen for a maximum duration, but at least 50% of the dosage interval (%fT>MIC), to achieve a sufficient antibiotic effect.9 Typically prolonged or in particular a continuous infusion of ß-lactam antibiotics allows to achieve the aforementioned concentration properties.10 However, there is a significant risk of continued misdosage, especially with continuous infusion regimens.11 Therefore, a dosing prediction implemented in a modern decision support tool for ß-lactam antibiotics like piperacillin/tazobactam or meropenem would be of utmost importance.12 To achieve optimised antibiotic treatment, the determination of antibiotic concentrations in the blood, known as therapeutic drug monitoring (TDM), is recommended for critically ill patients with sepsis.1315 However, TDM is not universally available and a costly procedure.5 16 Furthermore, a multicentre study also showed that despite the daily use of TDM, an adequate antibiotic dosage was achieved in only approximately 40% of sepsis patients.17 It, therefore, seems prudent, to reconcile the crude TDM concept in terms of integrating ASC measures in an intelligent prediction model, which in parallel considers comprehensive and multifaceted patient-related measures such as clinical data and individual laboratory values. Nevertheless, the output of the prediction model must ensure a rapid and intuitive interpretation of the results to enable a seamless transfer into clinical routine and easy decision support. Furthermore, decision support allowing an appropriate individualised dosage adjustment for the crucial first administration of antibiotics within the initial hour of sepsis, as recommended by current guidelines, is lacking.3 Consequently, despite the important progress associated with TDM, a suboptimal initial antibiotic dosing is still an unsolved problem, which substantially contributes to therapy failure, heightened antibiotic resistance and increased risk of toxic side effects, particularly in sepsis patients.17 In this prospective cohort study, we aim to analyse plasma concentrations of piperacillin/tazobactam or meropenem in patients with sepsis within the first 8 days. Additionally, we will investigate the influence of baseline characteristics, potential drug interactions, proteomic data and soluble biomarker concentrations (eg, cytokines and chemokines) on the pharmacodynamics of antibiotics. Subsequently, we intend to develop predictive algorithms using artificial intelligence (AI) on these comprehensive measurements, incorporating both clinical and molecular parameters. These advanced algorithms are designed to optimise initial dosing and ensure sustained therapeutic ASC levels supporting TDM-based dose adjustments. Furthermore, these algorithms are designed to act as ‘clinical decision support tools’ (CDSSs) aiding in the enhancement of dosage determination during antibiotic therapy, even in settings without TDM availability. Methods and analysis This study is a single-centre, prospective observational cohort study conducted in patients with sepsis undergoing antibiotic treatment with piperacillin/tazobactam or meropenem via continuous infusion, as defined by local treatment standards (piperacillin/tazobactam 18 g/24 hours and for meropenem 6 g/24 hours), with potential adjustments made by the attending physician independently of this study. Study population and conceptual overview In this study, a cohort of 200 patients with sepsis (100 with piperacillin/tazobactam and 100 with meropenem) will be prospectively enrolled to train the machine learning (ML) algorithms (derivation cohort with a 70/30 split for training and internal validation; see figure 1). Study flow of the trial. After screening and obtaining consent, blood samples are collected on days 1–8 and on day 30 (or at discharge if earlier). These samples undergo therapeutic drug monitoring and an extended assessment focusing on drug interactions and proteomics, along with an immunological characterisation. Two ML models will be developed. One model will be trained and validated using the prospectively recruited cohort in this study (n=200) and the existing Sepsis.DataNet cohort (n=500). The other model will incorporate additional variables determined exclusively in our cohort. ASC, antibiotic serum concentration; ML, machine learning. Patients eligible for inclusion encompass those undergoing intensive care treatment, admitted within 72 hours following the diagnosis of sepsis as defined by sepsis-31 and included within 24 hours after commencing continuous piperacillin/tazobactam or meropenem treatment. Both male and female patients aged 18 years or older are considered for participation. Exclusions comprise patients declining study involvement (either personally or through their legal representative) and individuals in pregnancy or lactation. Patients with severe anaemia (haemoglobin levels below 8 g/dL) are excluded from this study due to ethical and safety concerns. This decision was made due to the frequent blood sampling required, which may exacerbate anaemia and potentially necessitate red blood cell transfusions. Such interventions would introduce additional risk to safety in patients with severe anaemia and were, therefore, defined as exclusion criterium. Key characteristics that will be collected include patient demographics (age, gender and body weight), organ function (renal and hepatic), comorbidities (such as diabetes and cardiovascular disease) and the baseline Sequential Organ Failure Assessment (SOFA) score to assess the severity of illness. Beyond these, PK parameters (such as volume of distribution and clearance), fluid balance, the presence of organ dysfunction and any concurrent medications will be monitored. To externally validate our ML models, we will access the SepsisDataNet.NRW dataset (DRKS00018871), which contains multidimensional data from 500 prospectively enrolled patients with sepsis. The SepsisDataNet.NRW project is a multicentre initiative that has already been granted ethical approval (19-6606-BR). The dataset is stored at the University Hospital Knappschaftskrankenhaus Bochum, and access to this dataset has already granted for use in this study. The data are fully anonymised and securely stored in accordance with local and national data protection regulations. Our research team has the necessary permissions to use this dataset for the intended external validation purpose. Patient and public involvement There was no involvement of patients or the public in the design, conduct, reporting or dissemination plans of this study. Study procedure and duration The recruitment is scheduled for January 2024 and is planned to last a total of 23 months until the end of November 2025 (see table 1). The individual start of the study is defined as the time of study inclusion. The formal end of study is defined as discharge from the intensive care unit (ICU), 30 days or death, whichever occurs first. Table 1Study schedule with planned timeline of the project regarding individual assessmentsStudy periodJanuary 2024–November 2025December 2025InclusionObservationClose outPostrecruitmentTime pointDay 1Days 2–8Days 9–29Day 30/dischargeEnrolmentScreening for eligibilityxInformed consentxAssessmentBiometrical/demographical dataxClinical parameterxxxxTDMxxMicrobiological analysisxxxDrug interactionsxxxImmune responsexxxProteomicsxxx30-day mortalityxAnalysisDevelopment of AI modelsxAIartificial intelligenceTDMtherapeutic drug monitoring Study schedule with planned timeline of the project regarding individual assessments AIartificial intelligenceTDMtherapeutic drug monitoring Due to the observational design of this study, no influence is exerted on the administered therapy. The selection of the initial antibiotic (piperacillin/tazobactam or meropenem) will adhere to local ICU treatment standards, current sepsis guidelines and antibiotic stewardship considerations. Importantly, the study will not influence the antibiotic choice made by the attending physician, as it is strictly observational. The only study-specific measure is the collection of blood samples at a total of eight defined time points, with a maximum of 200 mL of blood being taken in total. No further examinations or interventions are planned. Decisions regarding all diagnostic and therapeutic measures described in this study were made solely by the attending physician independent of the study. In keeping with the observational design of this study, no adjustments to antibiotic dosing will be made based on the ASC results. Attending physicians are blinded to the generated TDM data during the ICU stay. They have unlimited freedom to adapt or change the antibiotic therapy. The first ASC measurement will be conducted within 4–24 hours following the initial administration of piperacillin/tazobactam or meropenem. This early sampling will help determine if the initial dose achieves the desired therapeutic range and will provide valuable data for optimising subsequent dosing adjustments during the course of treatment. The magnitude of deviation of ASC from the target therapeutic range will be categorised into subtherapeutic, therapeutic and supratherapeutic levels using the non-species-related EUCAST breakpoints for meropenem and piperacillin/tazobactam. This approach will help explore whether the magnitude of deviation has a clinically significant impact. If antibiotic therapy is modified before the completion of the 8-day observation period, data collection will continue for the new antibiotic regimen. All changes in antibiotic therapy will be documented to ensure continuous monitoring of ASCs and clinical outcomes, allowing for a comprehensive analysis of treatment patterns throughout the study period. Patient recruitment The recruitment of study participants will be conducted in the ICU of a University Hospital in Germany considering all patients who fulfil the inclusion and exclusion criteria. If the patient is capable of providing informed consent, they will be informed about the study prior to enrolment. In the case of patients who lack the capacity to provide consent, the legal guardian will be informed. The information will be conveyed both orally and in writing using a standardised informed consent form. However, it is anticipated that a significant number of patients may not be communicative or capable of understanding due to the severity of their illness and analgosedation that is often required. To ensure timely inclusion of patients in the studies, preliminary consent will be obtained from an independent physician not associated with the trial (consultant physician procedure). Once the patient regains the capacity to provide informed consent, they will be promptly reinformed about study participation. The inclusion of the patient in the study will only occur after a positive evaluation, informed consent and approval by the patient, their legal representative, or the independent physician not associated with the study (see onlinesupplemental files 13). For each enrolled study participant, a random four-character combination of numbers and letters will be assigned as a pseudonym. These pseudonyms will be maintained in a decoding list accessible only to the participating investigators. Study endpoints Primary endpoint Prediction of ASC for piperacillin/tazobactam or meropenem during therapy on day 1. Prediction of ASC for piperacillin/tazobactam or meropenem during therapy on day 1. Secondary endpoints Prediction of ASC for piperacillin/tazobactam or meropenem over the course of therapy on days 2, 3, 4, 5, 6, 7 and 8. Prediction of ASC for piperacillin/tazobactam or meropenem over the course of therapy on days 2, 3, 4, 5, 6, 7 and 8. Assessment of the impact of ASC on 30-day mortality. Assessment of the impact of ASC on 30-day mortality. Evaluation of the influence of ASC on recovery from the SOFA score (Delta-SOFA). Evaluation of the influence of ASC on recovery from the SOFA score (Delta-SOFA). Monitoring the occurrence of antibiotic resistance during therapy. Monitoring the occurrence of antibiotic resistance during therapy. Investigation of the effect of ASC on the incidence of acute kidney injury. Investigation of the effect of ASC on the incidence of acute kidney injury. Analysis of the impact of ASC on the incidence of acute liver damage. Analysis of the impact of ASC on the incidence of acute liver damage. Sample size calculation Based on the specified inclusion and exclusion criteria, a total of 200 patients with sepsis are intended to be prospectively enrolled in this study during the stated recruitment period (January 2024 to November 2025), with 100 receiving treatment with piperacillin/tazobactam and 100 with meropenem. A conventional sample size planning for the primary endpoint is not applicable due to the lack of data on ASC and the complexity of the planned AI models. Therefore, the sample size planning considers a patient number that, based on our own assessment and experience regarding the research question, is deemed sufficient and comparable to the sample size in existing studies on TDM in patients with sepsis.17 Furthermore, the sample size planning also accounts for formal requirements of a certain sample size deemed as adequate for training the ML algorithms. Therefore, sample size planning for a multiple regression is conducted. It is assumed that variables will be removed if they exhibit high correlation, have many missing values or show very low variance. Additionally, based on our experience with similar datasets, it is expected that no more than 15 variables will remain in the final model during stepwise variable selection (see figure 2). For the sample size planning, a power of 80 %, a significance level of 5 % and an effect size of Cohen’s f2=0.35 are assumed, representing a high effect size,18 which is crucial for the clinical utility of the model. The sample size was calculated using R V.4.3.1 (R Core Team, 2023) and the pwr package V.1.3-0 (Champely, 2020). Depending on the number of variables, the sample size ranges from 25 to 68. ML methods usually require a higher sample size than linear regression. In this project, we plan to implement transfer learning,19 which involves using pretrained models, such as those from the SepsisDataNet.NRW project, and specifically training them on the newly collected data and the novel research question. This approach requires significantly fewer cases than training a comparable model from scratch. Therefore, it is anticipated that approximately a 1.5–4 times higher sample size compared with the linear model should be sufficient. Calculated sample size for a linear regression model with a significance level of 5%, power of 80% and effect size f2=0.35 for varying numbers of variables in the model. Study-specific measures In the prospective cohort of this project, samples are collected daily in the initial days (days 1, 2, 3, 4, 5, 6, 7 and 8—up to 8 time points—to optimally capture the evolution of ASC. Divided into seven work packages (WPs) (see figure 3), WP1 focuses on standardised antibiotic administration, blood sampling and the systematic recording of comprehensive clinical parameters. Subsequently, these samples are processed in WPs 2–4, employing various analytical approaches to characterise the samples and provide relevant information that may impact antibiotic dosing. Data from WPs 1–4 will be harmonised in WP 5. In WP 6, there will be the utilisation of the Sepsis.DataNet Cohort, integrative data analysis and AI model development. The focus lies on creating multiple AI models from both the main dataset and the extended dataset. During this process, variables will be selected through feature selection. These models are intended to be utilised in WP seven as prototypes for an Antibiotic Dosage Recommendation Tool. AI, artificial intelligence; TDM, therapeutic drug monitoring; WP, work package. Microbiological investigations (WP2) address bacterial load, β-Lactamase activity in the blood, susceptibility of pathogens isolated from patient samples to administered antibiotics and the influence of administered sedatives and pain medications on antibiotic efficacy. The focus here is on interactions between antibiotics and medications that could significantly improve or impair the success of treatment. Quantification of antibiotic, analgesic and β-Lactamase concentrations, as well as their breakdown products in blood samples, is performed using liquid chromatography coupled with mass spectrometry in WP3. Further characterisation in WP4 will focus on assessing the immune response by measuring soluble biomarkers and analysing gene polymorphisms related to cytokines, chemokines and bacterial defence. Cytokine and chemokine levels, such as IL-6, IL-8 and TNF-α, will be quantified using ELISA and FACS-Immunoassay, serving as critical indicators of inflammation and immune activation. Additionally, proteomic profiling will be conducted to identify and quantify immune-related proteins involved in the acute phase response and complement system using mass spectrometry, providing a comprehensive view of immune regulation and its impact on antibiotic pharmacodynamics. Results from these three subprojects will be transmitted to WP5 for further processing. Clinical and TDM data were collected in WP1 along with the pre-existing Sepsis.DataNet cohort (n=500, independent, multicentre validation cohort), form the basis for developing and validating AI models in WP6. Here, the focus will be on the development and validation of ML models for predicting ASCs. In principle, two different approaches are pursued: First, we will develop a model solely using standard clinical and laboratory parameters that are commonly available in routine clinical setting to primarily focusing on generalisability. The second model will incorporate an extended dataset, including proteomic, microbiological and biomarker data, to enhance predictive accuracy. Feature selection methods, such as LASSO or random forest, will be employed to identify the most relevant variables. The dataset from the prospective cohort will be split into a 70/30 ratio for training and internal validation. Additionally, external validation will be conducted using the SepsisDataNet.NRW cohort to assess model performance in an independent patient population. The models will be evaluated based on metrics such as accuracy, sensitivity, specificity and Area Under the Receiver Operating Characteristics (AUROC). The ML model training procedure will involve several key steps, starting with data preprocessing. Data will be cleaned and standardised to ensure consistency across all variables. Missing data will be addressed using imputation techniques, such as mean or median imputation for continuous variables, and mode imputation for categorical data, as well as more advanced methods like multiple imputation depending on the extent and nature of the missing values. Feature scaling will be applied where necessary to ensure compatibility across variables. To improve model accuracy, we will explore a variety of ML algorithms, including random forests, gradient boosting and support vector machines. We will also explore the value of stacked ensemble modelling, which combines predictions from multiple models to improve accuracy by leveraging the strengths of different algorithms.20 The Common Data Set is available as online supplemental file to ensure transparency and provide a detailed outline of all variables included for training and validation of the ML models (see online supplemental file 4). WP7 will focus on developing a prototype for a CDSS based on the validated models from WP6. This CDSS will be integrated into the Patient Data Management System of the ICU. The system will provide real-time, non-binding recommendations for individualised antibiotic dosing, using patient-specific clinical and laboratory data. It will also suggest optimal time points for TDM to refine predictions. Rigorous testing of the prototype in a simulated clinical environment will be conducted before considering broader implementation. Ethics and dissemination Ethical considerations All study subjects, their guardians or their legal representatives, as appropriate, provided written informed consent prior to enrolment. The study has received the approval of the Ethics Committee of the Medical Faculty of Ruhr-University Bochum (No. 23-7905). In addition, we have ethics approval to use the data from participants in the external validation cohort, SepsisDataNet.NRW (DRKS00018871). This external cohort was approved by the Ethics Committee of the Medical Faculty of Ruhr-University Bochum (Approval No. 19-6606-BR). All participants in the SepsisDataNet.NRW study provided written informed consent for the use of their anonymised data for future research purposes, including external validation studies. Participant confidentiality will be ensured throughout the study, and all external validation data will remain anonymised and securely stored. Declaration of Helsinki The study is conducted in accordance with the Declaration of Helsinki in its current version. Publication of study results On completion of the study, all findings will be published in one or more open-access, peer-reviewed international scientific journal. After the acceptance of the first publication, a complete anonymised dataset will be made available to all interested individuals. This dataset will be deposited in public repositories. The results will also be shared through presentations at national and international conferences, as well as via relevant social media channels and the official websites of our research institutions, to reach a broader audience, including patients, relatives and a broad general audience. Risk–benefit assessment In this prospective, observational cohort study, obtaining an additional volume of approximately 15–35 mL of blood per sampling time point is necessary during routine blood collections. This corresponds to a total study blood volume of <200 mL, resulting in a neglectable impact on participants’ haemoglobin levels and consequently, their oxygen supply. As an additional safety measure, patients with critically low haemoglobin levels (<8 g/dL) are not enrolled in the study. Blood sampling for the study is always integrated with routine blood collections; no separate punctures are conducted. Therefore, there is no additional risk to patients compared with standard blood collections, with no added risk of infection, nerve trauma or swelling at the puncture site. These factors should be considered in relation to the associated benefits. The subsequent analysis is expected to yield new and crucial insights into sepsis treatment and appropriate antibiotic dosages, potentially benefiting future sepsis management. The potential improvements in survival and prognosis resulting from this work outweigh the potential risks associated with the measures implemented. Discussion and outlook The current state of antibiotic dosing in sepsis highlights a critical and urgent need for improvement. Even with TDM, recent findings show that only 30%–40% of patients receive correct dosing, compared with just 15% without TDM.7 8 This is highly concerning, given that adequate antibiotic dosing is the cornerstone of causal sepsis therapy.9 Addressing this gap is essential, as antibiotics are the primary intervention available in sepsis management, yet current tools and strategies fail to ensure adequate and timely dosing for the majority of patients. Our study introduces a novel approach using AI to develop ML-based algorithms for optimising antibiotic dosing in sepsis. Unlike traditional approaches, which primarily rely on fixed dosing or isolated PK measurements, our model combines routinely available clinical data with more complex variables, such as immune response markers and proteomic profiles. The goal is to create an integrated tool that can provide precise dosing recommendations, which can be seamlessly incorporated into CDSS in ICUs, thereby bridging the gap between conventional TDM and a comprehensive, data-driven strategy. The aim of this study is not to replace TDM but to significantly enhance its utility and interpretation. By offering individualised dosing recommendations both at the initiation of therapy and throughout its course, the developed tool can provide clinicians with a reliable, real-time support system, reducing the dependency on frequent sampling and overcoming the practical limitations of TDM. This proactive approach addresses a major shortcoming in current sepsis treatment by significantly increasing the likelihood of achieving therapeutic concentrations in a timely manner. The study employs a two-way algorithmic approach. The first algorithm aims to develop a model based on a concise set of widely available clinical and laboratory parameters, focusing on maximum predictive accuracy and generalisability. This ensures that the tool can be easily applied in a typical ICU setting, making it suitable for broad adoption. In contrast, the second algorithm, built on an extended dataset, incorporates more complex factors such as potential drug interactions, immune status assessments and proteomic analyses, with the goal of enhancing model precision. While proteomic data are generally more difficult to obtain in routine clinical practice compared with TDM, in our study, these data will primarily be used to identify novel biomarkers or surrogates and refine experimental models to improve predictive accuracy. Thus, the main goal of the extended approach is to identify key novel biomarkers or surrogates that provide crucial information about antibiotic dosing, significantly enhancing the model’s predictive accuracy. Regarding the efficacy of ß-lactam antibiotics, such as piperacillin/tazobactam and meropenem, it is closely tied to maintaining drug levels above the MIC (%fT>MIC). Extended or continuous infusions are two equivalent effective strategies to optimise %fT>MIC, especially in critically ill patients with sepsis, considering the current literature.21 From a purely theoretical point of view, however, continuous administration could be superior if the dosage is constant and sufficient, as it formally allows the time above the MIC to be maximised. In addition, timing of sampling is easier under continuous infusion. Taking these two aspects into account, we deliberately decided to use continuous infusion as a standard for our study. Addressing missing values will pose a notable challenge. The authors need to distinguish between two categories of missing values. First, it will be essential for models to handle potential gaps in documentation, reflective of clinical reality. Second, the absence of certain variables (eg, ventilation parameters in the case of a non-ventilated patient) can still convey valuable information (eg, indicating a non-ventilated status) and, therefore, requires specific consideration. A plausible algorithm capable of addressing these challenges is eXtreme Gradient Boosting, commonly known as XGBoost.22 Furthermore, the authors admit the practical constraints associated with TDM, which is often limited to large hospitals in high-resource countries. The success of the algorithm could potentially democratise access to tailored antibiotic dosing, offering a cost-effective and non-invasive solution for low-resource countries. While this study primarily focuses on developing an ML model for accurate prediction of ASCs, we recognise that an important next step will be to evaluate the model’s effectiveness for continuous monitoring. Future research will have to aim integrating the model into real-time clinical settings, thus also effectiveness for monitoring of the ML approach is crucial. Our long-term goal is to enable dynamic therapeutic monitoring, allowing clinicians to make timely adjustments to antibiotic dosing based on patient responses and changing clinical conditions. Such integration would not only enhance predictive accuracy but also facilitate real-time decision-making, extending the model’s utility beyond prediction to actively improving antibiotic management and patient outcomes. Looking forward, the authors envision the integration of their findings into CDSS within the PDMS of ICUs. This integration would facilitate automated, non-binding recommendations to physicians on the admission of sepsis patients, using baseline characteristics and clinical parameters. Additionally, it could provide guidance on specific time points at which TDM measurements should be taken to maximise the accuracy of ASC. Moreover, these models could be continuously enhanced through the use of TDM. This allows for not only quality management but also anticipates a continual improvement in our predictions under real clinical conditions. This pragmatic approach anticipates a potential paradigm shift in sepsis management, emphasising personalised treatment strategies with potential benefits for patient outcomes and enhancing survival in the long term. Trial status The first patient was recruited on 15 January 2024. Enrolment of participants is ongoing and is expected to continue until November 2025. supplementary material