Paper Type
Complete
Paper Number
PACIS2025-1785
Description
Clinicians increasingly struggle to locate critical insights hidden within sprawling electronic health records (EHRs). We design, implement, and evaluate an agentic artificial-intelligence prototype that enables voice-driven, conversational exploration of EHRs. Built on a domain-adapted BART transformer and micro-services architecture, the system parses natural-language queries, retrieves structured laboratory values and unstructured note excerpts, and proactively suggests complementary information. Using a de-identified multi-institutional corpus we fine-tune medical vocabulary embeddings and implement a query decomposition engine that links narrative text with temporal data trajectories. Simulated user sessions involving practising physicians achieved 97 % retrieval accuracy, median 2.5-second response time, and a 40 % reduction in information-seeking time compared to conventional keyword search. Qualitative feedback highlights reduced cognitive load and workflow fit. Findings demonstrate the feasibility and clinical value of agentic, multimodal AI interfaces and inform design principles for next-generation decision-support tools in healthcare settings worldwide.
Recommended Citation
Roosan, Don; Khan, Rubayat; Essien-Aleksi, Inyene; Nirzhor, Saif; and Hai, Fahmida, "Empowering Clinicians with an Agentic AI for Voice-Driven EHR Exploration" (2025). PACIS 2025 Proceedings. 11.
https://aisel.aisnet.org/pacis2025/general_topic/general_topic/11
Empowering Clinicians with an Agentic AI for Voice-Driven EHR Exploration
Clinicians increasingly struggle to locate critical insights hidden within sprawling electronic health records (EHRs). We design, implement, and evaluate an agentic artificial-intelligence prototype that enables voice-driven, conversational exploration of EHRs. Built on a domain-adapted BART transformer and micro-services architecture, the system parses natural-language queries, retrieves structured laboratory values and unstructured note excerpts, and proactively suggests complementary information. Using a de-identified multi-institutional corpus we fine-tune medical vocabulary embeddings and implement a query decomposition engine that links narrative text with temporal data trajectories. Simulated user sessions involving practising physicians achieved 97 % retrieval accuracy, median 2.5-second response time, and a 40 % reduction in information-seeking time compared to conventional keyword search. Qualitative feedback highlights reduced cognitive load and workflow fit. Findings demonstrate the feasibility and clinical value of agentic, multimodal AI interfaces and inform design principles for next-generation decision-support tools in healthcare settings worldwide.
Comments
General