A Bilingual On-Premises AI Agent for Clinical Drafting: Implementation Report of Seamless Electronic Health Records Integration in the Y-KNOT Project
Bibliographic record
Abstract
Background: Large language models (LLMs) have shown promise in reducing clinical documentation burden, yet their real-world implementation remains rare. Especially in South Korea, hospitals face several unique challenges, such as strict data sovereignty requirements and operating in environments where English is not the primary language for documentation. Therefore, we initiated the Your-Knowledgeable Navigator of Treatment (Y-KNOT) project, aimed at developing an on-premises bilingual LLM-based artificial intelligence (AI) agent system integrated with electronic health records (EHRs) for automated clinical drafting. Objective: We present the Y-KNOT project and provide insights into implementing AI-assisted clinical drafting tools within constraints of health care system. Methods: This project involved multiple stakeholders and encompassed three simultaneous processes: LLM development, clinical co-development, and EHR integration. We developed a foundation LLM by pretraining Llama3-8B with Korean and English medical corpora. During the clinical co-development phase, the LLM was instruction-tuned for specific documentation tasks through iterative cycles that aligned physicians' clinical requirements, hospital data availability, documentation standards, and technical feasibility. The EHR integration phase focused on seamless AI agent incorporation into clinical workflows, involving document standardization, trigger points definition, and user interaction optimization. Unlabelled: The resulting system processes emergency department discharge summaries and preanesthetic assessments while maintaining existing clinical workflows. The drafting process is automatically triggered by specific events, such as scheduled batch jobs, with medical records automatically fed into the LLM as input. The agent is built on premises, locating all the architecture inside the hospital. Conclusions: The Y-KNOT project demonstrates the first seamless integration of an AI agent into an EHR system for clinical drafting. In collaboration with various clinical and administrative teams, we could promptly implement an LLM while addressing key challenges of data security, bilingual requirements, and workflow integration. Our experience highlights a practical and scalable approach to utilizing LLM-based AI agents for other health care institutions, paving the way for broader adoption of LLM-based solutions.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.012 | 0.016 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.002 | 0.003 |
| Open science | 0.002 | 0.006 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.007 | 0.003 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".