Progress in sepsis prediction models: from traditional scoring systems to multimodal intelligence and clinical translation

Sepsis is a leading cause of mortality and healthcare expenditures among patients in the intensive care unit (ICU). Its pathophysiology is complex and its clinical manifestations are highly heterogeneous; early identification and timely, targeted interventions are essential to improving outcomes. With the widespread adoption of electronic health records (EHRs) and the rapid expansion of critical care data, developing sepsis prediction models using machine learning (ML) and deep learning (DL) has become an active area of research. This review provides a systematic overview of advances in sepsis prediction, from clinical problem framing and outcome definitions to data sources, feature engineering, and methodological evolution. We summarize the progression from traditional scoring systems (e.g., SOFA, qSOFA) to modern ML algorithms (e.g., gradient boosting trees, random forests) and time series DL models (e.g., LSTM, Transformer models). We also outline reporting and evaluation standards (e.g., TRIPOD AI), and synthesize evidence on representative models for early warning, prognostic risk stratification, and prediction of organ dysfunction. Key translational challenges are discussed, including generalization, fairness, model drift, workflow integration, alarm fatigue, and real world utility. Finally, we highlight opportunities in multimodal data fusion, causal inference, federated learning, and digital twins for building next generation, clinically actionable sepsis intelligence, and we offer practical recommendations to help move from algorithmic accuracy to demonstrable clinical value, emphasizing that only models that are externally validated, well calibrated, prospectively evaluated, and tightly aligned with clinical workflows are likely to improve patient outcomes.