The Gutenberg Moment of Dental AI: Local Supercomputing Replaces the Cloud

1. Introduction

Dental implant therapy is a routine procedure in contemporary oral rehabilitation, but clinical complexity and risk can vary substantially between cases. Structured assessment of surgical and prosthetic difficulty, patient‑specific risk factors, and long‑term prognostic considerations is therefore central to safe and predictable treatment planning. The ITI SAC (Straightforward–Advanced–Complex) classification has become an established framework for grading case complexity in implant dentistry, yet its consistent application in daily practice remains challenging and time‑consuming, particularly under routine chairside conditions.

Virtual Patient MedlibreGPT AI‑SAC is a local, large‑language‑model (LLM)–based clinical decision support system designed to operationalise and extend the SAC concept in implantology. The system runs entirely on an on‑premises high‑performance server and uses an open software stack combining LLMs with retrieval‑augmented generation (RAG) over curated clinical literature and guidelines. It aims to provide real‑time, literature‑grounded assessments of case complexity, risk and therapeutic options for all implant patients without transmitting any patient data to external cloud services.

This paper (i) outlines the historical trajectory from early augmented‑reality “virtual patient” systems to the current knowledge‑centric architecture, (ii) summarises a recently published feasibility study in oral oncology that serves as a first empirical validation step for the underlying architecture, (iii) describes the design of the MedlibreGPT AI‑SAC module for implantology, and (iv) discusses data‑governance and regulatory considerations for local, open‑stack clinical AI systems in the context of the European Medical Device Regulation (MDR) and the EU AI Act.

2. Background: From Augmented‑Reality Navigation to Knowledge‑Centric Virtual Patient Systems

2.1 Early Virtual Patient System at Medical University Vienna

In the 1990s, the Department of Oral and Maxillofacial Surgery at the Medical University of Vienna introduced the first clinically deployed „Artma Virtual Patient“ system (ARTMA). This CE Class IIa–certified platform facilitated real-time projection of CT-derived three-dimensional patient data into the surgical field, thereby supporting image-guided interventions through the integration of augmented reality and stereotactic navigation. The system was employed for a diverse range of maxillofacial procedures and served as the foundation for teleconsultations and image-guided remote case reviews across various centers.

These developments demonstrated that digital representations of the individual patient could enhance surgical precision and facilitate knowledge transfer between institutions. The primary focus, however, was on intraoperative visualization and navigation rather than on structured modelling of clinical reasoning or longitudinal treatment pathways.

2.2 From image guidance to clinical reasoning and literature integration

The current Virtual Patient MedlibreGPT AI‑SAC system builds on this historical foundation but shifts the emphasis from spatial guidance to decision support. Instead of projecting tumour margins into the operative field, the system aims to “project” differential diagnoses, guidelines and outcome data into the consultation and treatment planning process in real time.

This transition reflects broader advances in machine learning and natural language processing: large language models can synthesise heterogeneous textual information, while RAG architectures allow direct linkage between model outputs and underlying sources. By embedding SAC‑based criteria and implantology‑specific literature into a local LLM‑driven workflow, Virtual Patient MedlibreGPT AI‑SAC seeks to support not only where and how to operate, but whether, when and under what preconditions implantation should be undertaken in a given case.

3. System Architecture and Local Deployment

Virtual Patient MedlibreGPT AI‑SAC is implemented as a fully local, on‑premises system that leverages high‑performance compute hardware now available at the practice or clinic level. The architecture comprises three main layers:

1. LLM layer (MedlibreGPT)
   • MedlibreGPT is a fork of PrivateGPT tailored for clinical use.
   • It supports open‑weight models (e.g. Mistral, Llama, DeepSeek) deployed via local runtimes such as Ollama.
   • The LLM is used for controlled natural‑language interaction, reasoning over structured case data and synthesising outputs in an explainable format.
2. RAG layer
   • An agentic retrieval‑augmented generation pipeline (e.g. based on Kotaemon, LangChain or Ragflow) integrates structured, locally stored literature into each answer.
   • The document corpus includes clinical guidelines, meta‑analyses and relevant implantology literature.
   • Retrieved passages are explicitly cited in the generated reports, supporting traceability and critical appraisal by the clinician.
3. Data and governance layer
   • A self‑hosted Nextcloud instance is used for storage of patient records, radiological imaging and documents.
   • A PostgreSQL database maintains transactional and longitudinal treatment data at the patient and practice level.
   • A blockchain‑based timestamping component (Virtual Patient Guard) provides tamper‑evident audit trails for prompts, model versions and outputs.

All components are open and auditable. The design aims to facilitate compliance with MDR and EU AI Act requirements regarding data quality, transparency, risk management, human oversight and post‑market surveillance.

4. AIDOC Feasibility Study as Foundational Evidence

The underlying architectural approach was first evaluated in a feasibility study in oral oncology using the historical case of Sigmund Freud as an index example of cocaine‑induced midline destructive lesions (CIMDL). In this study (AIDOC), a locally hosted RAG system (Mistral‑7B with a curated CIMDL literature corpus) was compared with a state‑of‑the‑art cloud LLM without retrieval (GPT‑4o).

Across 401 simulated teleconsultations, the local RAG system included CIMDL in the differential diagnosis more than twice as often as the cloud baseline (odds ratio ≈ 2.3, p < 0.01). The study also documented a fully on‑premises data flow, versioned prompts and explicit governance mechanisms as part of a five‑layer clinical decision framework (interaction, patient, solutions, AI, data). The results suggest that, for rare entities requiring deep literature integration, a specialised local RAG system can outperform a general‑purpose cloud LLM without retrieval.

While the comparison does not isolate “local” versus “cloud” per se—since cloud‑based RAG architectures are also feasible—it provides initial empirical support for the Virtual Patient architecture and its suitability for MDR‑conformant deployment. The implantology module described in this paper builds directly on this architecture, extending it from rare oncologic cases to routine implant planning across all patients.

5. SAC Concept and Its Extension in MedlibreGPT AI‑SAC

The SAC (Straightforward–Advanced–Complex) framework was originally developed in oral surgery and subsequently adapted to implant dentistry, culminating in the widely adopted ITI SAC classification and its associated assessment tools. Within this framework, implant cases are graded according to surgical, prosthetic and aesthetic complexity, with the aim of aligning case selection, treatment planning and operator experience.

MedlibreGPT AI‑SAC adopts this conceptual basis and extends it with additional patient‑ and practice‑level parameters. The current schema comprises more than 15 criteria, each mapped to SAC levels (S, A or C), including:

• Aesthetic zone relevance (non‑visible vs highly aesthetic region)
• Bone availability and need for augmentation procedures
• Soft‑tissue situation and requirement for connective tissue grafting
• Systemic comorbidities and medication‑related risk (e.g. diabetes, bisphosphonates)
• Surgical difficulty and anatomical risk (e.g. sinus proximity, neurovascular structures)
• Restorative space, prosthetic span and material requirements
• Occlusal scheme and parafunctional loading (e.g. bruxism, malocclusion)
• Indication for immediate versus delayed implantation and loading
• Soft‑tissue healing demands and intensity of required follow‑up
• Minimal operator experience level considered appropriate for a safe procedure

The overall SAC class is derived from the combination of these parameters with conservative up‑classification in the presence of high‑risk or highly aesthetic features. This extended SAC model is intended to reflect not only the technical difficulty of implant placement but also systemic risk and the anticipated burden of long‑term maintenance.

6. Clinical Workflow for Implant‑Specific Decision Support

In routine practice, MedlibreGPT AI‑SAC is designed to integrate seamlessly into the implant consultation and planning workflow:

1. Data acquisition
   • Collection of medical history, medication, systemic diseases and smoking status.
   • Import of cone‑beam CT/CBCT data, clinical photographs and, where available, intraoral scans into the local Nextcloud instance.
   • Completion of a structured Medlibre SAC questionnaire (approximately 10–15 items) at the chairside via tablet.
2. AI‑assisted assessment
   • Structured inputs, imaging metadata and relevant clinical notes are transmitted via a local API to MedlibreGPT.
   • The RAG layer retrieves pertinent literature (e.g. current systematic reviews and guidelines on augmentation, immediate loading, or peri‑implant disease).
   • The LLM synthesises a case‑specific SAC assessment and explanatory narrative within a latency compatible with chairside use.
3. Output and clinical use
   • A structured report is generated, including:
     • Overall SAC class (S/A/C) with criterion‑level justifications,
     • Evidence‑based therapy pathway suggestions (e.g. need for staged augmentation, choice of timing and loading protocol),
     • Risk and follow‑up recommendations with cited literature.
   • The report is intended as a decision support tool; final treatment decisions remain under the full responsibility of the treating clinician.

In subsequent treatment phases (surgery, osseointegration, prosthetic restoration and long‑term maintenance), the same framework can be used to update SAC classification dynamically as new information becomes available (e.g. implant stability quotient values, peri‑implant bone remodelling, soft‑tissue changes).

7. Data Governance and Local vs Cloud Deployment

A central design choice of Virtual Patient MedlibreGPT AI‑SAC is the exclusive use of local, on‑premises deployment for both computation and data storage. This decision is motivated by several considerations:

• Data protection and regulatory alignment

  Local processing avoids cross‑border data transfers and reduces the complexity of demonstrating compliance with the General Data Protection Regulation (GDPR), particularly with regard to processing of identifiable health data by third‑party processors outside the EU/EEA.

• Longitudinal and practice‑specific learning

  Because patient data remain within the practice environment, it is possible, in principle, to analyse longitudinal outcomes at the level of individual patients and compare treatment strategies across similar cases within a single practice or clinic. Such analyses can support practice‑specific quality improvement and hypothesis generation.

• Transparency and independence

  An open‑stack, local architecture limits dependency on external cloud providers, their evolving business models and potential constraints on access or pricing. All core components, including data schemas and model configurations, are under the control of the institution.

At the same time, it is important to recognise that cloud‑based solutions can also be designed to comply with GDPR and medical‑device regulations, for example through EU‑based processing, robust contractual frameworks and appropriate pseudonymisation strategies. The present work does not aim to provide a comprehensive legal comparison between deployment models. Rather, it argues that for many small and medium‑sized practices, a fully local implementation offers a favourable balance of regulatory simplicity, data sovereignty and technical feasibility, particularly in view of the increasing availability of high‑performance compute hardware at the practice level.

8. Regulatory Considerations and MDR Pathway

Given its intended use as a clinical decision support system in implant dentistry, Virtual Patient MedlibreGPT AI‑SAC falls under the scope of the European Medical Device Regulation as software as a medical device (SaMD). Its functionality—supporting assessment of case complexity, risk and treatment planning in a high‑risk domain—suggests classification in the higher MDR risk classes (e.g. Class IIa or above), though the precise classification must be determined in consultation with competent authorities and a notified body.

The AIDOC feasibility study, published in Oral Oncology Reports, represents an initial step on this regulatory pathway by:

• Demonstrating clinical performance of the underlying architecture in a delimited diagnostic setting,
• Documenting a transparent, on‑premises data flow compatible with GDPR, and
• Introducing versioned prompts and governance mechanisms that support traceability.

For the implantology module, a dedicated validation study is planned to evaluate agreement between AI‑generated SAC classifications and expert assessments, as well as the system’s impact on treatment decisions and workflow. The subsequent MDR‑conformant technical documentation is expected to include:

• A detailed risk‑management file,
• Usability and human‑factors evaluations,
• Validation reports for SAC‑related outputs, and
• A post‑market surveillance plan addressing model performance monitoring, model and content updates, and mechanisms for handling adverse events or clinically relevant errors.

The open and auditable nature of the software stack, together with local data storage and blockchain‑based audit trails, is intended to facilitate compliance with MDR requirements concerning explainability, auditability and human oversight.

9. Outlook

Advances in hardware, including compact systems with petaflop‑scale performance and emerging on‑device LLM capabilities in tablets and smartphones, are likely to further reduce the technical barriers to local deployment in clinical environments. In the near term, Virtual Patient MedlibreGPT AI‑SAC is implemented on an on‑premises server accessed via secure VPN connections from chairside devices. In the longer term, parts of the functionality—particularly inference with compact models and access to compressed guideline embeddings—may be executed directly on mobile devices without any network dependency.

Future work will focus on rigorous clinical validation in implantology, refinement of SAC‑related criteria and aggregation rules, systematic evaluation of user interaction and explainability, and comparative assessments of local versus hybrid deployment models. Beyond implant dentistry, the general architecture may be transferable to other domains where structured risk stratification, longitudinal outcome analysis and tight integration with up‑to‑date literature are central to clinical decision making.

Literature

Michael Truppe, Kurt Schicho, Michael Figl, Simone Holawe, and Christos Perisanidis. “Artificial Intelligence in Oral Cancer: A Feasibility Study Informed by Freud’s Case.” Oral Oncology Reports, 2025.

FWF Research Project P 12464 The optical interface for augmented reality in computer-assisted navigation and 3D visualization in surgery; Michael Truppe, Markus Eckholt, Rolf Ewers, Klaus Ehrenberger, Helmar Bergmann, Franz Watzinger, Monika Cartellieri, DOI: 10.13140/RG.2.1.3117.8403