Introduction and background

Sepsis is a time-sensitive syndrome characterized by a dysregulated host response to infection that can lead to life-threatening organ dysfunction, prolonged hospitalization, and substantial mortality. It remains a major challenge for health systems worldwide because early clinical manifestations are often nonspecific and may overlap with other forms of physiological deterioration [1]. Delayed recognition is associated with worse outcomes, whereas timely administration of antimicrobials, source control, hemodynamic support, and structured care bundles has been linked to improved survival in appropriately selected patients. Despite advances in critical care and the implementation of standardized management pathways, early identification of sepsis continues to be difficult in routine clinical practice, particularly in settings with high patient acuity, complex comorbidity profiles, and large volumes of continuously generated clinical data [2,3].

Conventional approaches to sepsis detection have relied on bedside assessment, laboratory trends, and rule-based screening tools such as the Systemic Inflammatory Response Syndrome (SIRS) criteria, the quick Sequential Organ Failure Assessment (qSOFA), the National Early Warning Score (NEWS), and institution-specific early warning systems [4,5]. Although these methods can support situational awareness, they are limited by modest specificity, variable sensitivity, delayed triggering, and reliance on static thresholds that do not adequately capture nonlinear and time-dependent patterns of physiological deterioration. In recent years, artificial intelligence (AI) and machine learning (ML) methods have been increasingly applied to electronic health records (EHRs), bedside monitoring streams, and longitudinal clinical data to enhance early recognition. These models are capable of integrating multiple variables simultaneously, updating risk estimates dynamically, and identifying subtle trajectories that may precede overt clinical decline [6]. However, improved statistical performance does not necessarily translate into clinical utility, and important concerns persist regarding false-alert burden, calibration, generalizability across institutions, fairness, and the interpretability of complex models.

The concept of explainable or interpretable AI has therefore gained increasing relevance in sepsis prediction research. Clinicians are more likely to engage with decision-support systems when the underlying basis of a prediction can be understood, contextualized, and translated into actionable clinical decisions. At the same time, many published sepsis prediction studies remain retrospective in design, employ heterogeneous outcome definitions, evaluate variable prediction windows, and report performance metrics inconsistently [7]. Consequently, although the evidence base is expanding rapidly, it remains difficult to synthesize in a manner that directly informs bedside implementation. A focused appraisal of real-time systems that utilize continuously updated clinical data may help identify the most promising methodological approaches and highlight critical gaps in the current literature [8].

The objective of this systematic review was to evaluate the performance, interpretability, and clinical relevance of real-time AI models designed for early sepsis prediction using continuously updated hospital data. Specifically, this review aimed to synthesize evidence regarding predictive accuracy, lead time before sepsis recognition, approaches to explainability, external validation, and factors influencing successful translation into routine clinical practice.