Virtual Patient AI-SAC

Virtual Patient AI-SAC is a pioneering clinical decision support system for implant dentistry that signifies a „Gutenberg moment“ in medical AI. It operates the open-source LLM MedlibreGPT entirely on a local AI supercomputer (AI on the edge, ASUS Ascent GX10 NVIDIA GB10) using an open-source software stack. This enables real-time, literature-based assessment of the complexity, risk, and treatment options for all implant patients, without transmitting a single byte to cloud services.

The SAC classification system, introduced by Sailer and Pajarola in 1999 for oral surgery, has evolved over 26 years from a simple risk stratification tool to an AI-enhanced clinical decision support system. The journey started as physician-led innovation through its standardisation by the ITI, to its current transformation into Virtual Patient AI-SAC—a return to physician-driven development. We demonstrate how local AI supercomputing, developed and controlled by clinicians, represents not merely a technological advancement but a restoration of medical autonomy in the digital age.

Core Innovation: AI-Enhanced SAC Classification

At the heart of the system is the AI-supported SAC classification (Straightforward / Advanced / Complex), which extends the classical ITI criteria by incorporating:

• Patient-specific factors (comorbidities, medications)
• Soft tissue status and occlusion
• ISQ values (implant stability)
• Imaging data (DVT/CBCT)
• Real-time linkage to a curated literature database

Historical Context

The system builds on pioneering work from the 1990s at MedUni Vienna’s Department of Oral and Maxillofacial Surgery with the ARTMA Virtual Patient System (CE Class IIa), which used augmented reality and stereotaxy for image-guided surgery. Virtual Patient AI-SAC shifts the focus from pure imaging to clinical reasoning and literature knowledge, bringing differential diagnoses, guidelines, and outcome data to the chairside in real-time.

Open-Source Architecture & Data Sovereignty

The system is built entirely on open-source components:

• LLM Layer: MedlibreGPT (fork of PrivateGPT) with open models (Mistral, Llama, DeepSeek) via Ollama
• RAG Layer: Agentive Retrieval-Augmented Generation (Kotaemon/LangChain/Ragflow) embedding structured, locally-stored literature
• Data Layer: Nextcloud (HIPAA/GDPR-compliant) for patient records, images, documents
• PostgreSQL: Single Source of Truth for transactional data, Blockchain Timestamping (Virtual Patient Guard) for immutable audit trails

All components are open-source and auditable—critical for meeting MDR (traceability, risk management) and EU AI Act requirements (transparency, data quality, human oversight).

The Feasibility Study (published in Oral Oncology Reports) demonstrated that a locally-hosted, RAG-based system (Mistral-7B with CIMDL literature) significantly outperformed cloud LLMs without retrieval in recognizing cocaine-induced midline destructive lesions (CIMDL). This publication marks the first formal step on the MDR certification path for the Virtual Patient AI system.

Next steps:

1. Classification as SaMD High-Risk / Class IIa/b-III
2. Technical documentation (risk management, usability studies, validation)
3. Audit by notified body and MDR conformity assessment

Virtual Patient MedlibreGPT AI-SAC demonstrates that local, open AI supercomputing solutions are:

• Clinically superior (personalized, context-sensitive intelligence)
• Economically sensible (one-time hardware investment vs. unlimited subscription)
• Regulatorily compliant (GDPR, MDR, AI Act by design)

This transition from cloud monopoly to decentralized, physician-controlled intelligence represents the Gutenberg moment of medical AI: Knowledge leaves the closed cathedrals of the cloud and returns—in the form of an open, verifiable system—to every practice.

Sample Telekonsultation Interaction

https://medlibre.craft.me/szMBmMnlZzQFGo

Links

https://doi.org/10.1016/j.oor.2025.100773

Michael Truppe, Kurt Schicho, Michael Figl, Simone Holawe, Christos Perisanidis, Artificial intelligence in oral cancer: a feasibility study informed by Freud’s case, Oral Oncology Reports, Volume 17, 2026, 100773, ISSN 2772-9060,

Augmented Reality Image Guided Surgery in the 1990’s

https://www.artma.com

Artma.com is an archived site from the late 1990s presenting the “Artma Virtual Patient®”, an early augmented‑reality and image‑guided surgery system for head, spine, ENT, and maxillofacial procedures, used in international research, telemedicine projects, and conference demonstrations.

Google Scholar

https://scholar.google.com/citations?hl=de&user=ABDceFgAAAAJ

Michael Truppe is a researcher with about 2,900 citations and an h‑index of 21 (14 since 2020), known for work on computer‑assisted navigation, image guidance, teleconsultation, and augmented reality in cranio‑, maxillofacial‑, ENT‑ and implant surgery, and he is now working extramurally in private practice.

Open Source Software

https://github.com/eurodoc-telemedizin

EURODOC Telemedizin’s GitHub presents “Virtual Patient AI” as an open, physician-centered telemedicine and clinical decision support platform built around a specialized medical foundation model (“MedlibreGPT”), combining local (on‑prem) and cloud LLMs, strict data protection with GDPR/HIPAA-aligned on-site storage, blockchain-based timestamping for tamper-evident documentation, and integration with tools like Nextcloud and hybrid EHR systems; the organization hosts several public and private repositories (e.g., medlibreGPT, medlibre-ragflow, aidoc, medlibre-docker-services) using mainly Python, Shell, Java, TypeScript, and SCSS, and traces its roots to early AR-guided surgery and teleconsultation work originating in Vienna in the 1990s.