Ai-Augmented Clinical Handoff Systems: Integrating Large Language Models With Electronic Health Records For Safer Care Transitions

Abstract

Clinical handoffs are among the most vulnerable points in the chain of medical errors. For example, 80% of serious preventable adverse events in hospitals are attributed to communication failures. While the majority of organizations have implemented scripted tools (e.g., SBAR, I-PASS), the quality of handoffs remains inconsistent, and documentation remains a burden. We describe the design and deployment of AI-augmented clinical handoff systems that integrate LLMs and EHRs by transforming structured data into structured, evidence-based handoff notes with real-time risk stratification. Technical approaches included FHIR-based EHR integration, NLP-based clinical summary generation with LLMs, and ML-based deterioration detection, enabling automated descriptive documentation that is overseen by clinical personnel. Multi-hospital system implementations have observed that LLM-augmented documentation of patient handoff notes reduced documentation time by 50.2% to 55.1% per patient, saving between 474 and 981 hours across three hospitals monthly. Notes produced with LLMs demonstrated a BERTScore of 0.859 with a ROUGE-2 of 0.322. No critical patient safety risks were identified in 1600 emergency medicine notes. Machine learning identified deterioration 11 hours in advance with an area under the receiver operating curve (AUROC) of 0.895. Three quantitative measures are introduced: note efficiency, composite note quality, and AI risk stratification, with the purpose of defining benchmarks for documentation efficiency. In the augmentation model, the AI output is presented as one of many drafts, and the clinician remains the final decision-maker, maintaining safe care transitions for the patient.