About this project

Sepsis is a life-threatening condition where the body’s response to infection harms its own organs. It remains difficult to detect early, and outcomes vary widely between patients. Current diagnostic tools often identify sepsis too late, reducing the effectiveness of treatment - especially as multidrug-resistant bacteria become more common.

The Personalized Swiss Sepsis Study (PSSS) and Personalized, data-driven prediction and assessment of Infection related outcomes in Swiss ICUs (IICU) projects bring together hospitals, research groups, and data science experts across Switzerland to change this. Our goal is to recognize sepsis earlier and to predict patient outcomes more accurately by combining digital health data, molecular profiling, and advanced machine-learning methods.

By building a large, nationwide research network and collecting high‑quality clinical and biological data, we aim to identify new digital, molecular, and hybrid biomarkers that support personalized diagnostics and treatment strategies. Ultimately, our project seeks to improve survival, reduce long‑term complications, and advance precision medicine in sepsis care.

Data acquisition, processing and quality assurance:

In both projects funded by the Swiss Personalized Health Network (SPHN) and Personalized Health Related Technologies (PHRT), each University Hospital (Basel, Bern, Geneva, Lausanne and Zurich) and University of Bern and Zurich provide data from critically ill patients hospitalized in intensive care units (ICUs).

The locally generated data is collected from various hospital information systems (IS) such as clinical and laboratory IS, devices such as ventilators and dialysis machines, ICU monitors, and treatment-related data, e.g. drug doses and transferred and stored in local clinical data warehouses (CDWH). To follow Sepsis 3.0 criteria, we regularly tag patient with suspected infection and collect information in the clinical IS. This will allow to compute Sepsis-3 criteria and compare data patterns to this international accepted definition of sepsis. The data is structured and interoperable between centers following SNOMED CT and LOINC ontologies. Data is then transferred in a resource description framework (RDF) format after encoding and encryption via the BioMedIT nodes to the ETH domain for further analysis. The exchanged follows secured routes of BioMedIT to the ETH Domain for quality control and analysis. A shared data model will be used for machine-learning based analysis.

PSSS Studies:

There are two observational trials conducted within this project.
Study 1:

Digital biomarker study: includes every patient admitted to an ICU. Following specific inclusion/exclusion criteria, we include sepsis (cases) and non-sepsis (controls) upon entry of the ICU. Of all patients with an ICU admission also available pre/post-ICU data will be integrated into the data analysis. We compare data patterns between cases and controls in regards of the previously defined outcomes. For this study we have obtained ethical approval to collect data from 15,000 ICU patients from participating centers. In addition, a data transfer and use agreement as well as a consortium agreement have been generated. The legal process to allow data sharing was complex and involved a multiple round assessment and discussion of the legal framework with various stakeholders.

A list of clinical, physiological and laboratory variables, covering more than 500 attributes of interest was generated and linked to ontologies such as time points, application route and 4. dosage of specific drugs, etc.

At February 2021 we have collected data from almost 11’140 patients from the participating ICUs. Data transfer and exchange protocols have been set up with all CDWH, BioMedIT, and ETH domain teams and continuously improved.

clinicaltrial.gov

Study 2:

‘-omics biomarker study: includes a subgroup of the patients from the ‘digital biomarker project’. In all five hospitals a total of 300 patients with severe community acquired pneumonia (cases) and 100 patients with a severe systemic inflammation (controls) will be included. We will collect additional patient material such as serum, whole blood for DNA and RNA, respiratory and stool material, immune cells and bacterial isolates. These samples will be in-depth characterized using metabolomics, metagenomics, whole genome sequencing and immune phenotyping technologies. Digital and -omics data will then be merged and analyzed to form hybrid biomarkers. Again, case and control patients will be compared in regards of the previously defined outcomes.

This study has been stopped. The ethical proposal has been submitted and approved. However, the study could not be conducted due to resource issues in different centers.

IICU Studies:

There are five observational trials conducted within this project.
Study 1: Digital Biomarkers

The project aims to develop (semi-)automated methods for annotating clinical treatments using machine learning and natural language processing. By analyzing semi-structured domain knowledge collected during medical decision-making, the project will use pretrained language models and clinical embeddings to eventually suggest treatment annotations to physicians. The project will also identify which types of information hold clinical relevance.

The project aims to develop algorithms for biomarker discovery and clinical outcome prediction using multimodal patient data. It will identify associations between pathogen or patient characteristics and disease progression using time-series analyses. The project will apply causal inference techniques to uncover treatment effects and integrate diverse data sources to build early outcome prediction models using classical or deep learning methods.

This study has not yet started. The ethical proposal has been approved. Study data collection are progressing at all centers.

Study 2: Endotypes

Endotypes in sepsis reflect the complex interactions between pathogens, host responses, and hidden population differences, which makes predicting outcomes and identifying patients who may benefit from targeted therapies difficult. Although several protein and gene expression biomarkers associated with poor prognosis have been identified, clinical decision‑making still relies mainly on observable signs because current biomarkers do not fully explain sepsis heterogeneity.

The project aims to define new clinical phenotypes and biological endotypes of septic shock. We aim to model phenotypes through careful data preparation and both supervised and unsupervised machine‑learning methods. This will help uncover clinically relevant patient clusters and associated biological markers.  Tose will be introduced as standard measurements and outcomes from patients treated before and after their introduction will be compared to evaluate their impact on management and prognosis, while accounting for factors such as age, organ involvement, and chronic dysfunction.

It is expected either to identify clear phenotypes with distinct biological signatures or to highlight the need for deeper molecular endotyping and expanded biomarker discovery.

This study has not yet started. The ethical proposal has been approved. Study data collection are progressing at all centers.

Study 3: Genotypic and phenotypic virulence and resistance of bacteria

Antimicrobial resistance significantly affects patient treatment and outcomes. Understanding the molecular mechanisms behind resistance and how they evolve in vivo is essential for optimizing antibiotic choices. Bacterial virulence factors—such as adhesion, invasion, immune evasion, apoptosis regulation, serum resistance, quorum sensing, and biofilm formation—also play a crucial role in infection severity. While genome sequencing can identify virulence genes, it provides limited insight into actual bacterial behavior, and phenotypic tests for virulence remain scarce and time‑consuming.

Phenotypic resistance testing is the current gold standard, but traditional culture methods are slow. New rapid antimicrobial susceptibility testing (RAST) methods and EUCAST guidelines for blood cultures have improved speed, though their impact on clinical outcomes is still uncertain.

The planned contributions include linking bacterial genotypes to phenotypes, studying in vivo evolution of resistance, performing bacterial genome‑wide association studies, and correlating MALDI‑TOF MS spectra with resistance patterns. Integrating genomic and proteomic data with clinical information within the NDS framework aims to deepen understanding of how virulence and resistance genes influence patient outcomes.

This study has started. The ethical proposal has been approved. Study data collection are progressing at all centers. Bacterial genome sequencing was initiated in all centers and the active data collection part finished in December 2025. The data is currently analyzed.

Study 4: Infection with viruses and fungi

Invasive fungal infections (IFI) are increasingly common and represent a major cause of life‑threatening opportunistic infections in critically ill and ICU patients, comprising 5–10% of healthcare‑associated infections. Fungi are second only to bacteria as a cause of sepsis and septic shock. Candida species account for 70–80% of IFI and Aspergillus species for 5–10%. Invasive candidiasis (IC) shows high mortality (35–40%), and one third of cases occur in the ICU. However, risk factors are nonspecific, predictive scores perform poorly, and culture‑based diagnostics are slow. Novel non‑culture diagnostic methods—PCR and serological markers such as β‑D‑glucan or mannan—remain investigational.

Invasive aspergillosis (IA) has increasingly been observed as a severe complication of viral respiratory infections such as influenza and COVID‑19. Key mechanisms include virus‑induced damage to the respiratory epithelium, impaired innate immune responses (e.g., neutrophil dysfunction via STAT1 activation), and immune dysregulation with altered cytokine expression.

The project aims to profile immunity and inflammation in ICU patients with severe influenza or COVID‑19, correlating cytokine signatures with the occurrence of IA. Analyses will incorporate clinical risk factors such as demographic characteristics, severity scores, immunosuppression, treatments, and mechanical ventilation duration. Within the NDS framework, the project will add semantic terminologies tailored to viral and fungal infections and aims to identify endotypes predisposing to IC and IA. Such insights may support individualized preventive measures, including targeted antifungal prophylaxis.

This study has not yet started. The ethical proposal has been approved. Study data collection are progressing at all centers.

Study 5: Antibiotic Stewardship

Multiple factors influence the success of antibiotic therapy. Timely administration of an effective antibiotic is crucial for patient survival, and early bacterial identification enables more precise, targeted treatment. Personalized antibiotic regimens help reduce side effects and limit the emergence of resistance. Treatment outcomes are also affected by dosage, choice of antibiotic, route and mode of administration (e.g., bolus vs. continuous infusion), and how resistance findings are communicated to clinicians.

Ventilated patients are at risk of developing ventilator‑associated pneumonia (VAP), often caused by multidrug‑resistant organisms, which worsens outcomes. Although some clinical risk factors for VAP are known, recent machine‑learning‑based approaches have improved predictive capabilities.

The project aims to assess how dosage, antibiotic type, route and mode of administration, and timing of treatment recommendations influence outcomes, as well as factors driving rapid or delayed therapy adjustments. Machine learning, including reinforcement learning, will be used to identify optimal personalized antibiotic strategies. Integrating this work with other nested projects will enhance understanding of how factors such as viral infections and antibiotic exposure impact resistance development.

Within the NDS framework, the project contributes by improving insights into antibiotic consumption and stewardship. A better understanding of risk factors for prolonged ventilation and VAP may support preventive strategies and optimize antibiotic use.

This study has not yet started. The ethical proposal has been approved. Study data collection are progressing at all centers.