Sepsis Watch: The AI System Saving Lives and Cutting ICU Costs

Overview: How Sepsis Watch’s AI-Driven Real-Time Clinical Decision Support Platform Transforms Hospital Sepsis Care

Sepsis Watch (Duke) is an AI-driven early warning system that provides real-time clinical decision support for sepsis by monitoring emergency department patients and flagging those at high risk so care teams can act earlier. It addresses the challenge that sepsis often progresses rapidly and is difficult to detect early using rule-based screening alone, which can lead to delayed treatment and poor outcomes despite existing protocols. The system uses a deep learning model trained on millions of EHR data points, including vital signs, labs, medications, and medical history, to generate continuous risk scores and present them on a dashboard for a rapid response team that coordinates timely assessment and initiation of sepsis bundles.

By integrating into Duke’s ED workflows and updating risk predictions every few minutes, Sepsis Watch helps surface high‑risk patients several hours before traditional clinical recognition, providing a more actionable timeframe for intervention. This has been associated with substantial improvements in SEP‑1 bundle compliance and reductions in observed versus expected mortality, demonstrating measurable impact on both quality metrics and clinical outcomes when the model is coupled with a structured response workflow. For operations and quality teams, the surrounding infrastructure—which includes real-time performance dashboards and quality improvement tools—also supports continuous monitoring of model performance, treatment timeliness, and utilisation patterns, enabling iterative refinement of both the algorithm and the associated care processes.

Last checked on May 19, 2026: Sepsis Watch remains an active Duke Health AI program, with a 2025 multisite external validation study confirming strong generalisability of the sepsis risk model to community emergency departments.

What is Sepsis Watch?

Sepsis Watch (Duke) is an AI-driven clinical decision support system that uses a deep learning model over streaming EHR data to provide real-time early warning for patients at risk of sepsis in emergency departments. It is used by hospital clinical teams and rapid response nurses to prioritise high‑risk patients, coordinate timely assessment, and support completion of sepsis treatment bundles. Sepsis Watch is differentiated by its recurrent neural network–based risk model, tight integration into ED workflows via dashboards and alerts, and published evaluations showing earlier sepsis prediction and improved bundle compliance compared with traditional rules-based screening.

Why Do Leading Healthcare Teams Trust Sepsis Watch?

• Sepsis Watch was developed and is operated by the Duke Institute for Health Innovation at Duke Health, giving it backing from a major academic health system with an established AI and digital innovation program.

• The underlying deep learning model was licensed from Duke to a commercial partner (Cohere-Med Inc.), and a recent multi‑site external validation study at four Summa Health emergency departments found that the model’s performance generalized well outside Duke (AUROC 0.906–0.960).

• Published implementation studies and case reports describe real‑world integration of Sepsis Watch into ED workflows, including a federally registered clinical trial and analyses of its impact on early detection and treatment of sepsis.

• Duke reports that during the pilot period, Sepsis Watch was associated with approximately a two‑fold improvement in 3‑hour sepsis bundle compliance at Duke University Hospital compared with the prior two‑year average, indicating measurable quality gains when combined with process redesign.

• External commentary from MIT Technology Review and other outlets has profiled Sepsis Watch as an example of an AI system that demonstrated reduced sepsis-related mortality in routine practice, adding third‑party narrative evidence of impact.

• Academic and policy analyses treat Sepsis Watch as a socio‑technical system rather than a standalone algorithm, documenting governance structures, human oversight, and workflow adaptations needed for safe deployment, which may help reassure buyers about implementation maturity.

• A recent Nature Digital Medicine paper evaluated the generalizability of the Sepsis Watch model across multiple community emergency departments, concluding that performance and calibration remained strong across different patient populations and settings.

• The solution is framed as an adjunctive decision-support tool that augments, rather than replaces, clinician judgment, consistent with regulatory expectations for AI-driven early warning systems for sepsis.