Problem
The Cardiothoracic Intensive Care Unit (CTICU) at Duke University Hospital is a high-volume, high-acuity environment where patient care demands are both complex and dynamic. Each day, attending physicians must perform critical shift turnovers, communicating essential patient information to the incoming team. Currently, this handover process relies on manual summarization of patient status, typically via email or verbal communication. These summaries are highly variable in content, length, and quality, and lack consistent semantic stratification. In other words, the urgency, severity, and prioritization of clinical concerns are not reliably conveyed. The manual nature of this process is further complicated by physician fatigue, especially at the end of long shifts, and by the sheer volume of unstructured data embedded in clinical notes. As a result, vital information may be omitted, miscommunicated, or lost in translation, potentially impacting patient safety and care continuity.
The affected demographic included 15 attending CTICU physicians who rotate through shifts and are responsible for the care of critically ill patients, including those undergoing heart and lung transplants, mechanical circulatory support, and complex cardiac or thoracic surgeries. The lack of a standardized, efficient, and comprehensive handover process not only increases cognitive burden on clinicians but also introduces risk to patient outcomes due to inconsistent communication.
Solution
Constructed an AI-powered assistant summarizing selected clinician CTICU notes within the EHR to facilitate concise but accurate/comprehensive turnover communication at shift turnover.
To address these challenges, the CTICU Handover project proposed the development and implementation of an AI-powered assistant designed to automate and standardize the CTICU handover process. Leveraging advanced large language modeling (LLM) and generative AI, the tool extracts and summarizes CTICU daily event logs, assessment, and plan sections within the electronic health record (EHR) handoff documentation spaces. This material often contains nuanced clinical reasoning, intent, and concern that are not captured in structured data fields.
The AI solution was trained on data from the CTICU’s attending physicians, ensuring that its output aligns with clinical expectations and workflow needs. More than 97 cases with 23 measures over four runs, meaning 8,924 outputs, were validated by two critical care physicians and checked by two medical students. The DIHI Handover system generated relevant, accurate, and comprehensive shift summaries for each patient, incorporating predefined key components to ensure consistency. 99% of the outputs contained the true/false checks required, 99.6% contained the desired text order and volume, and 98.4% met the manual fact accuracy checks. For context, in the world of biostatistics and generative models in the year 2025, this is exceptional.
In application, at the end of each shift, the CTICU Handover AI automatically retrieves relevant notes and generates a standardized summary for each patient. The summary is then securely delivered to the incoming attending physician, either via email or a Duke Microsoft Team channel.
The phased implementation included staff training and full-scale deployment, with ongoing monitoring and iterative improvement based on user input. Full-scale deployment has included interest from several other critical care units and derivates for other hospital services. The Microsoft Team channel has not been an agreeable interface for physicians, so we are pursuing alternative interfaces and Epic integrations to enable full scale use.
Impact (Process)
The introduction of the DIHI AI Handover Solution is poised to transform the Duke University Health Systems’ hospital unit shift-handover process in several key ways:
- Standardization and Consistency through strictly prompted, physician-designed structure. This reduces variability in content and quality, mitigating the risk of omissions or miscommunication.
- Efficiency and Clinician Well-being: Automation alleviates the cognitive and administrative burden on care practitioners, particularly at the end of demanding shifts. This not only saves time but also reduces the impact of fatigue, enabling clinicians to focus on direct patient care and decision-making.
- Improved Documentation Quality: Knowing that daily notes will be used for automated handover summaries incentivizes physicians to maintain high-quality, comprehensive documentation, further enhancing the reliability of the process.
- Feedback and Continuous Improvement: The system incorporated many early feedback loops and was created through a deep partnership entirely between Duke engineers, data scientists, physicians, and nurses. This ensured relevant and effectiveness. Pre- and post-deployment surveys will measure satisfaction and completeness of handovers, guiding ongoing optimization.
Potential Business Impact
Beyond immediate process improvements, the DIHI Handover solution has the potential to deliver significant business value to Duke Health and the broader healthcare system:
- Patient Safety and Outcomes: By reducing communication errors and ensuring that critical information is consistently conveyed during handover.
- Operational Efficiency: Measurable time savings for physicians can free up resources for other clinical duties and potentially reduce overtime costs. It may also enhance staff satisfaction and retention.
- Scalability and System-wide Adoption: While initially focused on the CTICU, the DIHI Handover platform serves as a template for AI-driven clinical note summarization across other units and specialties. This amplifies impact on clinician efficiency and care coordination.
- Competitive Advantage and Innovation Leadership: By pioneering the use of generative AI in clinical handovers, Duke Health positions itself at the forefront of healthcare innovation. This not only supports recruitment and retention of top clinical talent but also strengthens the institution’s reputation as a leader in digital health transformation.
- Cost Savings and ROI: Over time, reductions in adverse events, improved workflow efficiency, and enhanced documentation quality may translate into tangible cost savings. The initial investment in development and implementation is modest relative to the potential long-term benefits in quality, safety, and operational performance.
Conclusion
The DIHI Handover represents a strategic investment in AI-driven clinical workflow optimization, addressing a critical need in the CTICU for standardized, efficient, and reliable shift handovers. By leveraging advanced LLMs and generative AI, the tool not only improves the quality and consistency of communication among attending physicians but also laid groundwork for broader adoption of AI solutions in Duke hospitals. For example, we worked immediately with the Surgical Intensive Care Unit (SICU) to replicate our methods and we received approval to pilot the handover there. Furthermore, the 2026 request for applications resulted in two funded Duke Health DIHI projects that scale this service across our health system: a Nurse Handoff Tool and a Pediatric Psychiatry Consultation-Liaison Handoff.
