Description

Sepsis is a common and critical clinical syndrome encountered in intensive care units (ICUs), associated with high rates of mortality and morbidity. It is defined as a life-threatening organ dysfunction caused by a dysregulated host response to infection (Sepsis-3 definition). Early diagnosis and prompt initiation of appropriate treatment before the onset of organ failure are vital to reducing mortality and morbidity. Gram-negative bacteremia is a significant cause of sepsis cases. Early initiation of appropriate empirical or targeted antibiotic therapy in bacteremia cases plays a pivotal role in patient prognosis. However, the diagnosis of gram-negative bacteremia is generally based on blood culture results, which may take 24-72 hours. The delay during this period is considered a major contributor to increased mortality risk.

In recent years, machine learning-based prediction models have been increasingly used as decision support tools in healthcare. The GRADY prediction models, developed retrospectively in our hospital, aim to predict the risk of sepsis due to gram-negative bacteremia using vital signs and laboratory parameters obtained during routine clinical monitoring. However, prospective validation of these models is essential prior to their integration into clinical practice.

The rationale of this study is to facilitate early identification of ICU patients at risk for bacteremia or sepsis to enable prompt initiation of treatment. Reducing mortality and morbidity through early detection may help alleviate the burden on healthcare systems. Moreover, supporting current clinical practices with early prediction models may enhance decision-making efficiency. The use of early warning systems and machine learning-based algorithms may improve clinical predictive power and allow for timely interventions by clinicians. This study aims to evaluate the diagnostic accuracy and clinical applicability of GRADY models through prospective validation. In this regard, the findings may contribute to the development of new approaches for sepsis and bacteremia management in critical care settings.

In current clinical practice, scoring systems such as the Sequential Organ Failure Assessment (SOFA), Systemic Inflammatory Response Syndrome (SIRS), and National Early Warning Score 2 (NEWS2) are used to define sepsis and to identify high-risk patients early. However, these scoring systems are based on a limited set of clinical and laboratory parameters and have shown suboptimal sensitivity and specificity in early sepsis diagnosis according to various studies. This limitation may reduce the chance of early intervention and negatively affect patient outcomes. GRADY models aim to offer risk prediction based on routinely collected clinical and laboratory data using machine learning algorithms, providing an alternative to conventional scoring systems. This study will compare the diagnostic accuracy and clinical performance of GRADY models with widely used scoring systems such as SOFA, SIRS, and NEWS2.

Currently, there are only a limited number of validated and widely accepted scoring systems available for the early identification of bacteremia. The Pitt Bacteremia Score was developed to predict short-term mortality in patients diagnosed with bacteremia and has been validated in several studies. Unlike the Pitt score, GRADY models aim to predict the risk of bacteremia and sepsis in the early period before diagnosis using routine clinical and laboratory data. Although the two systems do not serve exactly the same purpose, Pitt Bacteremia Scores will be calculated for all patients in this study, and the potential relationship with high-risk classification by the GRADY model will be evaluated statistically.