Source proof for Cliniques universitaires Saint-Luc deploys an on-premises Qwen3-235B assistant inside its Epic EHR

…his burden, yet real-world deployments in live hospital systems remain limited. We implemented a secure, GDPR-compliant, on-premises LLM assistant integrated into the Epic EHR at a European university hospital. The system uses Qwen3-235B with Retrieval Augmented Generation to deliver context-aware answers drawing on structured patient data, internal a…

…eval, and note drafting, which together accounted for over 70% of interactions. Following the pilot, the system was deployed hospital-wide and adopted by 1,028 users who generated 14,910 conversations over five months, with more than half of clinicians using it at least weekly. Usage remained concentrated on information access and documentation support, in…

…ese tools useful and safe in day-to-day workflow. To support our GenAI efforts, we invested in a server with 8 NVIDIA H200 GPUs, allowing us to host the largest available models, such as DeepSeek V3. In addi…

…of existing workflows and should only appear on demand Integration Since 2020, the Cliniques universitaires Saint-Luc hospital has utilized the Epic EHR system, which offers various mechanisms for integrating third-party applications. We s…

…supply its title together with the signature date. External clinical documents. Belgium’s eHealth platform employs a decentralized architecture for sharing healthcare documents. Because our hospital is located in Brussels, we use the Réseau Santé Bruxellois (RSB), one of the platform’s four regional…

…secure, GDPR-compliant, on-premises LLM assistant integrated into the Epic EHR at a European university hospital. The system uses Qwen3-235B with Retrieval Augmented Generation to deliver cont…

…e are credited. PMC Copyright notice PMCID: PMC12716761 PMID: 41417848 Abstract Electronic Health Records (EHRs) have improved access to patient information but substantially increased clinicians’ documentation workload. Large Language Models (LLMs) offer a potential means to reduce this burden, yet…

…n, often in structured fields. In addition to this double documentation burden, the availability of clinical documents also increases the volume of information available that physicians have to process when preparing a consultation, on average, 20 minutes per patient for emergency physicians at our hospital. T…

…s critical for downstream systems that rely on predictable, well-formed inputs. We use LiteLLM to manage access control and system reliability by enforcing usage limits and distributing requests across multiple vLLM instances. This allows us to maintain stable performance and fair resource allocation, eve…

…minimalist chat-based interface accompanied by a list of previous interactions. We built on the open-source OpenWebUI interface, simplifying it for clinical users while preserving its core chat functionality…

…stem, which offers various mechanisms for integrating third-party applications. We selected web-based integration, as Epic can embed web apps directly in its interface and provide secure access via the SMART on FHIR protocol. This study is designed as a deployment feasibility and adoption study rather th…

…phasizes transparency throughout the retrieval and response generation process. Every answer generated by the model is accompanied by a structured list of the sources consulted, as shown in Fig 4A and 4B . These sources include both documents within the EH…

…patient data, internal and regional clinical documents, and medical literature. A one-month pilot with 28 physicians across nine specialties demonstrated high engagement, with 64% of participants using the assistant daily and generating 482 multi-turn conversations. The most common tasks were summari…

…s over five months, with more than half of clinicians using it at least weekly. Usage remained concentrated on information access and documentation support, indicating stable incorporation into everyday clinical workflows. Feedback vol…

…chnical implementation, measure user adoption, and characterize usage patterns. We deliberately did not measure clinical outcomes or time savings, as this would require a controlled study design inappropriate for an initial d…

…his burden, yet real-world deployments in live hospital systems remain limited. We implemented a secure, GDPR-compliant, on-premises LLM assistant integrated into the Epic EHR at a European university hospital. The system uses Qwen3-235B with Retrieval Augmented Generation to deliver conte…

…nature of the documents allows for more accurate searches, especially semantic. We therefore opted for full-text semantic search using Apache Solr instead of vector databases. The documents are chunked based on level 1 and 2 titles present in the document…

…heterogeneous and often stored as PDFs, whose reliable parsing is non-trivial. We therefore rely on Docling [ 37 ], an open-source, fully on-premise tool that converts numerous formats (including PDF) into machine-readable representa…

…ese tools useful and safe in day-to-day workflow. To support our GenAI efforts, we invested in a server with 8 NVIDIA H200 GPUs, allowing us to host the largest available models, such as DeepSeek V3. In addi…

…apabilities, especially to support thousands of users in a university hospital. The GenAI strategy of our institution is to host medical applications on-site and not to rely on vendor services that do not allow for the level of customiza…

…apabilities, especially to support thousands of users in a university hospital. The GenAI strategy of our institution is to host medical applications on-site and not to rely on vendor services that do not allow for the level of customiza…

…onfirmed that no patient data would leave the hospital infrastructure, and that the application met the standards of the EU General Data Protection Regulation (GDPR) [ 40 ]. Access control and data minimization. Access control is enforced throug…

…several obligations for high-risk AI systems being phased in over time [ 42 ]. Given the clinical context of use, we proactively aligned the system’s development and governance with the requirements applicable to high-risk medical AI systems. In particular, we implemented: Clear documentation of intended use, model behav…

…ital’s network of physician “EHR champions,” representing multiple specialties. These clinicians participated throughout the development lifecycle, from use-case definition to interface testing. Before go-live, the application was demonstrated in a test environment during a…

…ghout the development lifecycle, from use-case definition to interface testing. Before go-live, the application was demonstrated in a test environment during a dedicated one-hour review session, which served to surface potential limitations and incorporate clinician feedba…

…e system is to integrate seamlessly into the day-to-day workflow of physicians. We held early workshops to familiarize physicians with LLMs and to collect their envisioned use cases [ 31 ]. Through this process, we obtained a set of use cases prioritized by the…

…w. During the last week of the pilot, the complete system was enabled. Adoption After completing the brief training session, physicians received production access to the model. Over the 30-day observation window, 18 physicians interacted with the system da…

…eval, and note drafting, which together accounted for over 70% of interactions. Following the pilot, the system was deployed hospital-wide and adopted by 1,028 users who generated 14,910 conversations over five months, with more than half of clinicians using it at least weekly. Usage remained concentrated on information access and documentation support, in…

…his burden, yet real-world deployments in live hospital systems remain limited. We implemented a secure, GDPR-compliant, on-premises LLM assistant integrated into the Epic EHR at a European university hospital. The system uses Qwen3-235B with Retrieval Augmented Generation to deliver conte…