HÄVG tests Privatemode to bring AI and doctor-patient confidentiality together.

General practitioner practices struggle with a shortage of skilled workers and high documentation overhead, which ties up up to 40 percent of working time. Therefore, the demand for AI-assisted solutions that meaningfully relieve these tasks is increasing. Falling back on solutions such as ChatGPT is not an option, as this would violate medical confidentiality (§ 203 StGB) and the GDPR. Operating one's own AI infrastructure with GPUs is not feasible for most practices, economically, administratively, or in terms of the necessary know-how. Standard cloud AI solutions carry the risk that data may be visible to providers or used for training. 

Furthermore, standard cloud AI solutions do not offer general practitioner practices the full potential of AI in a medical context. The use of specialized LLMs or enrichment with domain knowledge in the HZV context is not possible.

HÄVG is evaluating secure and high-performance AI approaches closely aligned with the needs of general practitioner activity, in order to provide a specialized AI solution that supports practices in documentation, in the administrative area, and in quality-assured knowledge queries, without incurring data protection risks or causing high costs.  

In a feasibility study, HÄVG is testing the Confidential Computing approach for the deployment of AI. The approach is oriented toward the "Healthcare Confidential Computing" specification of gematik. In order to be able to use AI functions securely from the cloud, Privatemode AI was integrated in initial use cases. Privatemode, the Confidential Computing-based AI service by Edgeless Systems, ensures that data remains protected at all times — during storage, transmission, and also during processing. On this basis, HÄVG can provide AI functions quickly and scalably from the cloud, while user data is demonstrably protected on a cryptographic level against access by model, cloud, and AI service providers. 

The combination of Privatemode with HÄVG's own HZV interface into the practices also enables the specialization of applications to the needs of general practitioner practices in the medical and HZV context. 

Implementation of the feasibility study 

On the basis of Privatemode, HÄVG has developed an "HZV AI Chat" that can be used as a secure AI chat by all staff members of the general practitioner practice. An option for internet integration was also implemented in a data-protection-compliant manner by means ofverifiable human-in-the-loop search queries. 

At the same time, the HZV AI Chat offers the possibility of being enriched with domain knowledge. In the feasibility study, enrichment with medical domain knowledge in the form of selected medical guidelines was tested, with the provision and use of this knowledge taking place under strict quality assurance processes. Looking ahead, enrichment with HZV domain knowledge as well as the use of specialized medical LLMs are planned. 

Architecture 

For the implementation of the feasibility study, a direct connection was developed between the "HZV AI Chat" application in the general practitioner practice and the backend of Privatemode AI:

By means of Remote Attestation, the Privatemode Proxy verifies whether the backend of Privatemode AI has not been tampered with. After successful verification, the proxy forwards the AI request to the backend.  

1. Within Privatemode AI, only application layer access exists. The client can submit a request for processing; the cloud provider has no access.  
2. In the Privatemode backend (Trusted Execution Environment), the AI request is passed to the LLM and processed. 

With this architecture, HÄVG has created a practice-ready concept for the secure use of cloud AI in general practitioner practices. Various open-source AI models, such as GPT-OSS-120B or Whisper, can be connected via the existing HZV interface without dedicated GPU hardware or additional AI infrastructure in the practices, and support use cases such as documentation assistance, quality-assured knowledge queries, and transcription. 

In addition to the models, further AI-feeding systems for data preparation are also provided in the same secure Confidential Computing environment, so that upstream and downstream processing steps remain continuously protected. 

The result is a genuinely sovereign cloud solution, whose trustworthiness is underpinned by strong cryptography and technical protection during storage, transmission, and processing. General practitioner practices can thereby be meaningfully relieved - particularly with regard to the extensive documentation burden - without incurring risks in the protection of health data. 

Outlook:  

Looking ahead, the use of specialized medical LLMs is planned, among other things in collaboration with partners such as the Fraunhofer IAIS.