Dutch
French
BulgarianHospital Case: Local AI on Lenovo Hardware
Faster documentation, reduced administrative burden and full data control with on-device LLMs (Llama.cpp).
Doctors spend up to 2 hours less per day on administration thanks to real-time transcription and summaries
during consultations, with full control through human-in-the-loop. No cloud tokens or variable costs
for AI inference, and sensitive patient data remains within the hospital network.
Summary of Results (Pilot)
- Up to 2 hours less administration per doctor per day (measured in the pilot, depending on specialty).
- Real-time transcription and summaries during consultations; doctor remains in control via human-in-the-loop.
- No cloud tokens or variable costs for AI inference; predictable TCO.
- Sensitive patient data remains within the hospital network (on-device and on-prem).
Challenge
Doctors and nurses spend a substantial portion of their time on documentation and administrative tasks. This reduces direct patient time and increases workload. Existing cloud AI services partially address this, but raise questions regarding GDPR, data sharing outside the EU and unpredictable costs. The hospital sought a solution that is privacy-secure, fast and scalable—without disrupting care processes.
Solution Overview
- On-device LLMs via Llama.cpp on Lenovo ThinkStation P workstations and Lenovo ThinkPad P laptops in consultation rooms.
- RAG setup: models receive context from protocols, guidelines and internal documents (securely indexed).
- Batch and heavy workloads on on-prem Lenovo ThinkSystem GPU servers with NVIDIA AI Enterprise for management and security.
- Integration with the EHR (HIX/Epic): summaries and transcripts are stored immediately after doctor approval.
- Human-in-the-loop: the doctor reviews and approves, with clear disclaimers ("AI-assisted, not a diagnosis").
Architecture & Operation
Edge (consultation room): microphone → on-device speech-to-text → summary/action list → doctor approval → EHR.
Datacenter (on-prem): larger models and batch processing, model management, monitoring and audit log.
Security: all data remains within the hospital network; encryption at rest and in transit.
Resilience: offline mode in consultation rooms; synchronization once a connection is available.
Integration with the EHR
- Support for HIX/Epic integrations (HL7/FHIR) for storing summaries and codings.
- Single Sign-On (SSO) and role-based access; logs are included in the EHR event log/audit trail.
- Automatic highlights (medication, allergies, ICPC/SNOMED suggestions) to support record quality.
Privacy & Compliance (GDPR)
DPIA completed and data flows documented (on-device, on-prem; no data sent to public clouds).
Compliant with NEN 7510 and ISO 27001 frameworks: encryption at-rest/in-transit; retention periods per policy.
Strict role and rights structure: periodic audits and pentests; automated PII masking where appropriate.
Implementation & Adoption
- Phase 1 (pilot, 4–6 consultation rooms): setup, data connections, measurement plan and training key users.
- Phase 2 (scaling): rollout per clinic/department, KPI monitoring and continuous model tuning.
- Training & change: short, role-based sessions; clear guidelines for AI use and limitations.
- Support: 24/7 monitoring of services, clear escalations and fallback scenarios.
Risks & Mitigations
- Hallucinations: human-in-the-loop and source references (context) by default; no autonomous writing.
- Bias and safety: periodic output evaluations and dataset updates; clinical validation where required.
- Continuity: rollback plan per department; clear fallback to manual workflow.
Results & Measurement Plan
The measurement plan follows before/after metrics per department. The KPI table below serves as a reporting template:
| KPI | Baseline | After 12 Weeks | Notes |
|---|---|---|---|
| Admin time per doctor per day | — | — | Goal: up to −2 hours |
| Documentation turnaround time | — | — | Minutes/consult |
| Patient satisfaction (NPS/CSAT) | — | — | Per clinic/department |
| EHR record quality | — | — | Regular sampling |
| Cloud AI costs/month | — | — | Goal: €0 for inference |
Experiences from Practice
"Thanks to local AI I can keep my attention on the patient; the summary is ready in the EHR by the time I finish the consultation."
— Cardiologist
About the Hardware
Consultation rooms: Lenovo ThinkPad P laptops or ThinkStation P workstations with NPU support for on-device AI.
Datacenter: Lenovo ThinkSystem GPU servers with NVIDIA AI Enterprise for scalable, managed inference.
Management: central model management, version control and monitoring; updates via change windows.
Why Local AI in Healthcare (Netherlands)
- Data minimization and sovereignty: patient data remains within Dutch/EU jurisdiction.
- Lower latency and higher availability at bedside or in consultation rooms.
- Predictable cost structure and no dependence on external token prices.
Next Step
Want to know how this works within your hospital? Request the reference architecture (PDF) and the tailored pilot plan. We are happy to share lessons learned and the measurement framework.
