1714: A Novel Cloud Framework for Real-Time, Large Language Model-Based Data Synthesis in Critical Care

Introduction: Ongoing healthcare shortages necessitate robust remote surveillance models like the electronic Intensive Care Unit (eICU) to manage patients at risk of deterioration. While automated alerts can flag physiological decline, they often lack the clinical context needed for rapid and effective intervention, contributing to clinician cognitive burden and delays in decision-making. To address this gap, our framework utilizes a large language model (LLM), deployed on a scalable cloud platform, to synthesize patient data from a standardized, interoperable electronic health data source. Methods: The cloud based LLM using Gemini 2.5 was developed by integrating Google’s Vertex AI platform with our health system’s Fast Healthcare Interoperability Resources (FHIR)-native data store. We used the Successive Approximation Model (SAM) to guide the LLM development, and implemented a zero-shot, question-answering methodology, which circumvents the need for task-specific prompt fine-tuning. An interdisciplinary team of clinicians, artificial intelligence experts co-developed a user interface and optimized the prompt to generate patient summaries. Results: Triggered by deterioration alerts and using single-line, natural language queries, the LLM proved highly performant, generating summaries from (>45 documents) in less than one minute. The evaluation was conducted by a panel of clinicians and was structured using the Kirkpatrick model. For Level 1 (Reaction), clinicians reported a highly positive response to the summaries ease of use, while Level 3 (Behavior) feedback confirmed the tool’s potential to significantly enhance clinical workflow and decision-making speed. Completeness was rated as sufficient in 99% of cases, with all major diagnoses correctly identified. Crucially, zero instances of critical data hallucination were observed. High temporal accuracy (>99%) was observed with all major clinical events placed in the correct chronological order. Conclusions: This work establishes a robust, cloud-based framework capable of generating real-time clinical patient summaries. Unlike traditional LLM, this zero-shot, question-based method offers a significant leap in efficiency and scalability over fine-tuned models and provides a framework for multimodal LLM.