HNHartmuth Nowak

Hartmuth Nowak

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Hartmuth Nowak has a total of 185 co-authors. Here are the top co-authors by number of shared publications:

  • Tim Rahmel

    Ruhr University Bochum, Knappschaft Kliniken University Hospital Bochum, Department of Anesthesiology, Intensive Care and Pain Therapy, 44892 Bochum, Germany

    21 shared publications
  • Michael Adamzik

    Ruhr University Bochum, Knappschaft Kliniken Universitätsklinikum Bochum, Klinik für Anästhesiologie, Intensivmedizin und Schmerztherapie, Universitätstraße 150, Bochum, Germany, 49 02343229242

    16 shared publications
  • Katharina Rump

    Ruhr University Bochum, Knappschaft Kliniken University Hospital Bochum, Department of Anesthesiology, Intensive Care Medicine and Pain Therapy, 44801 Bochum, Germany

    13 shared publications
  • Nina Timmesfeld

    Department of Medical Informatics, Biometry and Epidemiology, Ruhr University Bochum, Bochum, Germany

    13 shared publications
  • Lars Palmowski

    Ruhr University Bochum, Knappschaft Kliniken Universitätsklinikum Bochum, Klinik für Anästhesiologie, Intensivmedizin und Schmerztherapie, Universitätstraße 150, Bochum, Germany, 49 02343229242

    11 shared publications
  • Michael Adamzik

    Klinik für Anästhesiologie, Intensivmedizin und Schmerztherapie, Universitätsklinikum Knappschaftskrankenhaus Bochum, 44892 Bochum, Germany

    11 shared publications

Latest Publications

  • Open Access

    Development and evaluation of machine learning-based prediction-modelling for initial vancomycin serum concentrations in septic ICU patients using clinical health record data

    by Robin Grugel, Britta Westhus, Hartmuth Nowak, Michael Adamzik, Tim Rahmel, Martin Eisenacher

    Background: Sepsis remains a life-threatening condition with highly heterogeneous and dynamic pathophysiology, limiting the effectiveness of uniform therapeutic strategies. Beyond timely source control, antimicrobial therapy represents the only causal treatment option. Vancomycin is widely used for treatment of Gram-positive infections; however, optimal dosing in septic patients is challenging due to pronounced pharmacokinetic variability and substantial interindividual heterogeneity. Underdosing may promote antimicrobial resistance, whereas overdosing increases the risk of toxicity. This study aimed to develop and validate a machine learning-based prediction model to support individualized vancomycin dosing using routinely available clinical data. Methods: This single-center retrospective study included adult sepsis patients admitted to the intensive care unit, using routinely collected data from the hospital’s electronic medical records. Patients were eligible if they received a vancomycin loading dose followed by continuous infusion and had at least one measured serum concentration. Three machine learning models–elastic net regression, random forest, and XGBoost–were developed to predict the initial vancomycin serum concentration. To minimize bias and enhance generalizability, model training, hyperparameter tuning, and performance evaluation were conducted using a stratified nested cross-validation approach. Model performance was compared with seven commonly used population pharmacokinetic models. Results: The developed best performing elastic net model achieved a notable improvement with an average RMSE of 6.19, compared to 7.83 for the best pharmacokinetic model highlighting the potential of early and individualized dosing supported by a machine learning model. Final model analysis revealed that noradrenaline administration, together with classical pharmacokinetic parameters including body weight, serum creatinine, and the presence of chronic kidney disease, significantly influenced predictive performance. Conclusions: This machine learning-based approach for predicting initial vancomycin serum concentrations outperforms conventional PK models and enables more precise, individualized dosing prior to the availability of therapeutic drug monitoring results. By integrating key clinical variables, the model facilitates data-driven decision-making in sepsis care and underscores the potential of machine learning to advance personalized antimicrobial therapy.

  • Open Access

    AI-Supported, Integrative Prediction of Postoperative Delirium: Protocol for the CONFUSED Study

    by Katharina Rump, Hartmuth Nowak, Martin Eisenacher, Samuel Busch, Britta Westhus, Matthias Unterberg, Sara Sordon, Lars Palmowski, Andrea Witowski, Malte Bayer, et al.

    Background: Postoperative delirium (POD) is a frequent and serious complication in older surgical patients, characterized by acute cognitive dysfunction and fluctuating levels of consciousness. POD is associated with prolonged hospitalization, long-term cognitive decline, reduced quality of life, and increased mortality. Despite its clinical relevance, the underlying pathophysiological mechanisms remain poorly understood, and reliable biomarkers for early prediction and prevention are lacking. Objective: The CONFUSED study aims to identify molecular and clinical predictors of POD by integrating clinical data with proteomic, transcriptomic, and epigenetic analyses. The primary objective is to develop predictive models for POD using multimodal data. Secondary objectives include the identification of delirium-associated genes, proteins, and epigenetic signatures, as well as the exploration of patient subgroups at increased risk for POD. Methods: CONFUSED is a prospective observational cohort study conducted at a German university hospital. Adult patients undergoing major surgery under general anesthesia will be enrolled until 100 cases of POD have been observed, which is expected to require a total sample size of approximately 200 to 300 patients. Blood samples are collected at 4 predefined time points: before premedication, immediately after surgery, and on postoperative days 2 and 5. Samples undergo comprehensive proteomic profiling, transcriptomic analysis using RNA microarrays, DNA methylation analysis, and genotyping of selected polymorphisms. Clinical data, including demographics, comorbidities, perioperative variables, medications, and delirium assessments using the Confusion Assessment Method (CAM) and CAM for the intensive care unit, are systematically recorded. Statistical analyses include univariate and multivariate methods, as well as machine learning approaches such as random forests and support vector machines, to identify relevant biomarkers and develop predictive models. The study protocol follows STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) and TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) guidelines and was approved by the responsible ethics committees. Results: The study was registered in the German Clinical Trials Register (DRKS00033854) on March 18, 2024. Recruitment started in January 2024 and is ongoing at the time of manuscript submission. As of now, 135 patients have been enrolled. Sample collection and laboratory analyses are ongoing. Data analysis began in January 2026, with first results anticipated in July 2026. Final data lock is anticipated after the completion of recruitment. Conclusions: By integrating multimodal molecular data with clinical parameters and applying advanced machine learning techniques, the CONFUSED study aims to improve the prediction and understanding of POD. The results are expected to support the development of personalized preventive strategies and contribute to improved perioperative care for patients at risk of POD.


  • Künstliche Intelligenz in der Intensivmedizin – Schwerpunkt Sepsis-Früherkennung

    by Hartmuth Nowak, Michael Adamzik

    Jede Stunde zählt: Eine Sepsis kostet täglich mehr als 150 Menschen in Deutschland das Leben – oft, weil wir nicht schnell genug reagieren können. Was wäre, wenn ein digitaler Kollege rund um die Uhr unterschwellige Warnzeichen erkennen würde, bevor es zu offensichtlichen klinischen Symptomen kommt? Künstliche Intelligenz (KI) verspricht genau das: eine drohende Sepsis Stunden vorher zu melden, um medizinische Maßnahmen frühzeitiger einsetzen zu können. Sepsis remains the third leading cause of death worldwide after cardiovascular and malignant diseases, claiming approximately 11 million lives annually. In Germany alone, around 280000 people develop sepsis each year, with more than 150 deaths daily. Despite advances in therapy, survival still depends primarily on one factor: time. Early initiation of antimicrobial therapy and source control remain critical. Artificial intelligence (AI) promises to identify subtle, pre-symptomatic physiological changes that precede overt clinical signs of sepsis. By continuously analyzing vast streams of data from monitors, laboratory results, and electronic health records, AI models can detect complex, subclinical patterns beyond human perception, potentially predicting sepsis hours in advance. Machine learning (ML) algorithms, particularly supervised learning models, are trained on historical patient data (vital signs, laboratory values, and clinical notes) paired with outcomes indicating whether and when sepsis occurred. Once validated on unseen cases, these models can provide real-time risk assessments in clinical practice. Compared with traditional scores such as SIRS, qSOFA, MEWS, or NEWS, AI systems can integrate multiple variables dynamically and adjust risk estimates continuously. Several studies have shown that ML-based models outperform classical scores in prediction accuracy, identifying sepsis significantly earlier. However, challenges remain: lack of generalizability across hospitals, risk of false alarms, limited interpretability, and insufficient clinical validation. While early implementations demonstrate improved patient outcomes, robust randomized controlled trials are still needed to confirm that AI-driven sepsis prediction truly saves lives and merits integration into clinical guidelines. Article published online: 23 March 2026 © 2026. Thieme. All rights reserved. Georg Thieme Verlag KG Oswald-Hesse-Straße 50, 70469 Stuttgart, Germany

  • Open Access

    Patient Benefits in the Context of Sepsis-Related AI-Based Clinical Decision Support Systems: Scoping Review

    by Pascal Raszke, Godwin Denk Giebel, Jürgen Wasem, Michael Adamzik, Hartmuth Nowak, Lars Palmowski, Philipp Heinz, Nina Timmesfeld, Marianne Tokic, Frank Martin Brunkhorst, et al.

    Background: Global digitalization continues to advance, extending its influence into medicine and health care systems worldwide. In recent years, substantial advancements have been made in the research and development of artificial intelligence (AI), raising questions about its potential in medicine. The integration and application of AI in intensive care medicine, particularly in sepsis treatment, presents significant potential for advancing patient outcomes and enhancing patient-relevant benefits. However, a comprehensive and systematic overview of the full spectrum of patient-relevant benefits associated with AI-based clinical decision support systems (CDSS) remains lacking. Objective: This scoping review aimed to identify and categorize evidence on patient-relevant benefits of AI-based CDSS in sepsis care. Methods: Systematic research was conducted in 4 electronic databases: MEDLINE via PubMed, Embase, the ACM Digital Library, and IEEE Xplore. In addition, a comprehensive search on the websites of relevant international organizations, along with a citation search of the included articles, was conducted. Articles were included if they (1) focused on sepsis and (2) described patient-relevant benefits of AI-based CDSS. Articles published between January 1, 2008, and March 2, 2023, were considered for inclusion. Study selection was performed independently by 2 reviewers. The manuscript was drafted in accordance with the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) checklist. The analysis of the included articles was conducted using the program MAXQDA (VERBI Software GmbH), with systemization finalized in a consensus workshop. Results: A total of 3368 records were identified across the 4 databases, of which 24 met the inclusion criteria and were included in the scoping review. The additional search on international websites and in reference lists identified 6 more relevant articles, resulting in 30 included studies. Of these, 20 were quantitative, comprising 7 prospective and 13 retrospective designs. In addition, 1 qualitative study, 1 mixed methods study, 6 review articles, and 2 articles from institutional websites were included. Patient-relevant benefits were systematized in six main categories: (1) prediction, (2) earlier treatment and prioritization of high-risk patients, (3) individualized therapy, (4) improved patient outcomes (including improved Sequential Organ Failure Assessment score, reduced length of stay, and reduced mortality), (5) general improvements in care, and (6) reduced readmission rate. Conclusions: This scoping review underscores the potential of AI-based CDSS to positively impact patient-relevant benefits, particularly in sepsis care, where they demonstrate considerable promise for improving intensive care. However, the majority of the identified studies rely on retrospective database analyses. Future research should focus on validating these findings through prospective studies.


  • Open Access

    Human Cytomegalovirus Serostatus Defines Cytokine-Based Predictive Signatures in Sepsis

    by Frederik Krause, Birte Dyck, Kerstin Kappler, Matthias Unterberg, Hartmuth Nowak, Tim Rahmel, Lars Bergmann, Lars Palmowski, Britta Westhus, Alexander Wolf, et al.

    (1) Background: Sepsis is characterized by profound heterogeneity of immune responses, complicating biomarker-based prediction of clinical outcomes. Latent human cytomegalovirus (HCMV) infection is one of the strongest modulators of the human immune system and may influence cytokine-mediated signaling during sepsis. (2) Methods: In this post hoc analysis of 331 patients from the prospective multicenter SepsisDataNet.NRW cohort (German Clinical Trial Registry No. DRKS00018871), we quantified 13 serum cytokines on day 1 after sepsis diagnosis and determined HCMV IgG serostatus via ELISA. Using nested cross-validated logistic regression with exhaustive feature selection, we identified cytokine panels predictive of 30-day survival in the total cohort and in subgroups stratified by HCMV serostatus. (3) Results: In the total cohort, a four-cytokine panel (IL-6, IL-10, TNF-α, IL-12p70) predicted 30-day survival with a cross-validated area under the curve (AUC) of 0.66 [95% CI: 0.59–0.72]. Stratification by HCMV serostatus revealed distinct predictive profiles: in HCMV-seropositive patients, a two-cytokine model (IL-10, IL-23) achieved an AUC of 0.69 [95% CI: 0.61–0.77], whereas in seronegative patients, a model based on IL-8 and IL-17A failed to generalize (AUC = 0.47 [95% CI: 0.33–0.61]). Kaplan–Meier analysis confirmed a significant separation of survival curves for the HCMV-seropositive group (p < 0.001) but not for seronegative patients (p = 0.282). (4) Conclusions: HCMV serostatus defines an immunological context in which cytokine-based prediction of sepsis outcome becomes feasible. These data suggest that viral serostatus should be systematically incorporated into biomarker discovery and immunophenotyping approaches to improve the reproducibility and biological interpretability of sepsis endotyping.

  • Open Access

    Association Between Differential Heterogeneity of Antibiotics Consumption and Share of Resistant Pathogens and Its Implication for Antibiotic Stewardship in a German Hospital Intensive Care Unit

    by Hans H. Diebner, Pierre Schumacher, Tim Rahmel, Michael Adamzik, Nina Timmesfeld, Hartmuth Nowak

    Background: The rapid rise in antimicrobial resistance has become one of the 10 most pressing health problems worldwide in recent years. Antibiotic stewardship offers hope in the fight against antibiotic resistance, but it is currently still falling short of expectations. A better understanding of the dynamics of the interaction between antibiotic consumption and the emergence and spread of resistance is urgently needed. Methods: We discuss a simple dynamic model based on a differential equation to describe the increase in the proportion of a pathogen’s antimicrobial resistance to an antibiotic as a function of the time-dependent consumption of that antibiotic. Furthermore, we investigate the association of heterogeneity in the consumption of antibiotics with the rate of resistant pathogens. Data basis is the hospital information system and the patient data-management system of a German hospital, restricted to the intensive care unit. To quantify heterogeneity, we discuss and compare different entropy measures. Results: For some pathogen–antibiotic pairs, the consumption-dependent dynamic model for the growth in the proportion of antimicrobial resistance provides acceptable predictions, while for others, the model is less suitable. Cross-resistance and complex interactions with other pathogens and antibiotics may be responsible for this, suggesting that the observed dynamic behavior should be complementary, described using heterogeneity models. Time courses of Shannon entropy, the Antibiotic Heterogeneity Index, and the negative Gini Index correlate positively with the time series of the resistance rate. Thus, an increase in heterogeneity correlates with a decreasing resistance rate. However, a time-delayed cross-correlation of a differential entropy measure with resistance share suggests a functional dependence that can be utilized for antibiotic stewardship. Conclusions: Evidence is provided that the amount of consumption of certain antibiotics drives the corresponding proportions of pathogens’ resistance to these antibiotics; however, the model predictions of these univariable models are generally not sufficiently good, pointing to a more complex interaction dynamics. Therefore, we switch to the level of structural features and show that the degree of constantly mixing of the shares of antibiotic consumption has a control function regarding the incidence of resistance. Controlling differential consumption heterogeneity, therefore, appears to be a feasible operational basis for antibiotic stewardship. Experimental studies are demanded to identify functional dependencies; however, the integration of clinical expertise with model-based prediction appears to be a feasible antibiotic stewardship strategy.

  • Open Access

    AI-supported, integrative prediction of postoperative delirium- study protocol of the CONFUSED study (Preprint)

    by Katharina Rump, Hartmuth Nowak, Martin Eisenacher, Samuel Busch, Britta Westhus, Matthias Unterberg, Sara Sordon, Lars Palmowski, Andrea Witowski, Malte Bayer, et al.

    BACKGROUND Postoperative delirium (POD) is a prevalent and serious complication affecting older surgical patients, characterized by cognitive disturbances and fluctuating consciousness. Despite its impact on quality of life and increased mortality, POD’s causes remain inadequately understood and effective diagnostic and preventive measures are lacking. Objektive: The CONFUSED study aims to develop predictive models for POD using an integrative approach combining clinical, proteomic, transcriptomic and epigenetic data. OBJECTIVE The CONFUSED study aims to develop predictive models for POD using an integrative approach combining clinical, proteomic, transcriptomic and epigenetic data. METHODS The CONFUSED study is a prospective observational research conducted at a German university hospital, involving 200 patients undergoing major surgeries under general anesthesia. Key objectives include: 1) Identifying significantly altered genes (transcriptomics) and proteins (proteomics) through comprehensive analyses of blood samples; 2) Applying univariate and multivariate analyses, including AI techniques, to develop predictive models for POD on base of clinical and laboratory data. RESULTS Ethics and dissemination: The study was approved by the Ethics Committee of the Medical Faculty of Ruhr-University Bochum (23-7794) and the Ethics Committee Westfalen-Lippe (2024-082-f-S). Participants are recruited from the anesthesia outpatient clinic and provide informed consent to participate in the study. Blood samples are collected at four critical time points: premedication, immediately postoperatively, and on postoperative days two and five. These samples undergo extensive analysis, including proteomic profiling, RNA microarrays, DNA methylation analysis, and genotyping of potential polymorphisms. Clinical data, such as vital signs, medications, and delirium assessments using CAM and CAM-ICU, are systematically documented. The protocol followed the SPIRIT guidelines. CONCLUSIONS The study employs advanced data integration and AI techniques to analyze the collected data. Initial analyses include univariate statistical methods to identify significant proteins and genes and multivariate methods to discover patient subgroups and develop predictive models. AI models, such as random forests and support vector machines, are used to predict the likelihood of POD based on the significant genes and proteins and on clinical data. By integrating these diverse data types and utilizing AI for predictive modeling, the study aims to enhance early diagnosis and personalized management of POD. The results are anticipated to lead to improved preventive strategies and therapeutic approaches, ultimately reducing the incidence and impact of postoperative delirium in surgical patients. CLINICALTRIAL The study is registered in the German Clinical trials register (DRKS) under the registration number DRKS00033854 on march 18th 2024. This is Version 1.0

  • Open Access

    Association Between Differential Heterogeneity of Antibiotics Consumption and Share of Resistant Pathogens and Its Implication for Antibiotic Stewardship in a German Hospital Intensive Care Unit

    by Hans H. Diebner, Pierre Schumacher, Tim Rahmel, Michael Adamzik, Nina Timmesfeld, Hartmuth Nowak

    Background: The rapid rise in antimicrobial resistance has become one of the 10 most pressing health problems worldwide in recent years. Antimicrobial stewardship is a beacon of hope with regard to measures to combat resistance, but is currently still falling short of expectations. A better understanding of the dynamics of the interaction between antibiotic consumption and the emergence and spread of resistance is urgently needed. Methods: First, we discuss a simple dynamic model based on a differential equation to describe the increase in the proportion of a pathogen’s antimicrobial resistance to an antibiotic as a function of the time-dependent consumption of that antibiotic. Second, we investigate the association of heterogeneity in the consumption of antibiotics with the rate of resistant pathogens. The data basis for this is the hospital information system and the patient data management system of a German hospital, restricted to the intensive care unit. To quantify heterogeneity, we discuss and compare different entropy measures, including differential entropy. Results: For some pathogen-antibiotic pairs, the consumption-dependent dynamic model for the growth in the proportion of antimicrobial resistance provides surprisingly good predictions, while for others, the model is less suitable. Cross-resistance and other complex interactions with other pathogens and antibiotics may be responsible for this, suggesting that the observed dynamic behavior should be complementary described using heterogeneity models. It turns out that the three time courses of Shannon entropy, the so-called Antibiotic Heterogeneity Index, and the negative Gini Index correlate positively with the time series of the resistance rate. Thus, an increase in heterogeneity correlates with a decreasing resistance rate. However, a time delayed cross correlation of a differential entropy measure with resistance share suggests a functional dependence that can be utilized for antibiotic stewardship. Conclusions: We provide evidence that the amount of consumption of certain antibiotics drives the corresponding proportions of pathogens’ resistances to these antibiotics, however, the model predictions of the mono-causal models used are generally not sufficiently good, pointing to a more complex interaction dynamics. Therefore, we switch to the level of structural features and are thus able to show that the degree of constantly mixing the shares of antibiotic consumption has a control function with regard to the incidence of resistance. Controlling differential consumption heterogeneity therefore appears to be a feasible operational basis for antibiotic stewardship. Our results speak in favor of an experimental study or at least inter-clinical comparative studies, as this is the only way to identify functional dependencies. In summary, the integration of clinical expertise with model-based prediction appears to be a feasible antibiotic stewardship strategy.

  • Open Access

    Computer-Interpretable Quality Indicators for Intensive Care Medicine: Development and Validation Study

    by Falk von Dincklage, Viktor Karl Bublitz, Oliver Kumpf, Carlo Jurth, Reimer Riessen, Maria Deja, Christiane Maria Schewe, Dirk Schädler, Christian Fuchs, Sebastian Gibb, et al.

    Background: Quality indicators (QIs) can help assess intensive care quality, identify potential for improvement, and ultimately enhance patient outcomes. Therefore, the German Interdisciplinary Association of Critical Care and Emergency Medicine (DIVI) has developed QIs for intensive care medicine. However, variability in how these are technically implemented across health care facilities currently limits their comparability. Objective: The aim of the study is to develop unambiguous computer-interpretable representations of the DIVI QIs for intensive care medicine using Fast Healthcare Interoperability Resources (FHIR) and to establish a replicable process for translating narrative QIs into standardized digital formats. Methods: We first decomposed the narrative DIVI intensive care medicine QIs into two sets of semantic concepts that characterize (1) the targeted patient population and (2) the care aspect specified by each indicator. We mapped the concepts to international vocabularies, defining a supplementary code system for concepts not appropriately represented in existing vocabularies. The decomposed and semantically mapped QIs were then implemented in FHIR using an implementation guide we previously developed to represent clinical practice guideline recommendations. As the translation process holds risks of inducing logical and semantic deviations, the final FHIR representations were back-translated into a narrative form and reviewed with clinical experts, including the authors of the original QIs. The decomposition and semantic mapping were iteratively adjusted based on the experts’ feedback until the results accurately reflected the original intent of the QIs. Results: The 10 DIVI QIs were decomposed into 31 separately measurable indicators, including 9 structural indicators, 17 process indicators, and 5 outcome indicators. All process and outcome indicators were successfully specified as computer-interpretable representations in FHIR. In total, 58 unique medical concepts were used, of which 52 (90%) could be mapped to concepts from international vocabularies. The remaining 6 concepts—mostly intensive care unit–specific scores or roles—were defined in a supplementary code system. Nested Boolean logic and temporal conditions were fully supported using standard FHIR mechanisms. After iterative adjustments, the final representations were approved as accurate representations of the DIVI QIs by the clinical expert panel. Conclusions: Our work demonstrates that the structured process developed here enables the unambiguous, computer-interpretable representation of QIs for intensive care. These representations can be used in automated quality management systems to standardize quality assessments across health care facilities. Our newly defined structured process can serve as a blueprint for similar efforts in other specialties. The here-developed computer-interpretable QIs are openly available for reuse and ongoing maintenance. Future work will focus on piloting these indicators in real-world clinical systems and extending the framework to include structural indicators.


  • Open Access

    Machine learning identifies clinical sepsis phenotypes that translate to the plasma proteome

    by Thilo Bracht, Maike Weber, Kerstin Kappler, Lars Palmowski, Malte Bayer, Karin Schork, Tim Rahmel, Matthias Unterberg, Helge Haberl, Alexander Wolf, et al.

    Background: Sepsis therapy is still limited to treatment of the underlying infection and supportive measures. To date, various sepsis subtypes were proposed, but therapeutic options addressing the molecular changes of sepsis were not identified. With the aim of a future individualized therapy, we used machine learning (ML) to identify clinical phenotypes and their temporal development in a prospective, multicenter sepsis cohort and characterized them using plasma proteomics. Methods: Routine clinical data and blood samples were collected from 384 patients. Sepsis phenotypes were identified based on clinical measurements and plasma samples from 301 patients were analyzed using mass spectrometry. The obtained data were evaluated in relation to the phenotypes, and supervised ML models were developed enabling prospective phenotype classification. Results: Three sepsis phenotypes were identified. Cluster C was characterized by the highest disease severity and multi-organ failure with leading liver failure. Cluster B showed relevant organ failure, with renal damage being particularly prominent in comparison to cluster A. Time course analysis showed a strong association of cluster C with mortality, while patients in cluster B were likely to change the cluster until day 4. The plasma proteome reflected the clinical features of the phenotypes and showed gradual consumption of complement and coagulation factors with increasing sepsis severity. Supervised ML models allowed the assignment of patients based on only seven widely available features (alanine transaminase (ALT), aspartate transaminase (AST), base excess (BE), international normalized ratio of thrombin time (INR), diastolic arterial blood pressure, systolic arterial blood pressure (BPdia, BPsys) and activated partial thromboplastin time (aPTT)). Conclusions: The identified clinical phenotypes reflected varying degrees of sepsis severity and were mirrored in the plasma proteome. Proteomic profiling offered novel insights into the molecular mechanisms underlying sepsis and enabled a deeper characterization of the identified phenotypes.