Use Cases

Countless documents are generated and stored in many industries. The number of available documents makes it difficult for staff to extract, prepare, and make quick, informed decisions from the multitude of data.

In clinics, extensive medical letters are produced daily. Due to time constraints, it is challenging for medical staff to get a quick assessment of the situation at hand.

Large Language Models (LLMs) offer promising possibilities for extracting relevant data from clinical documentation. The extracted data can be used in a structured form for further analysis or decision support, leading to significant relief for medical staff and an improvement in the quality of documentation in daily work.

Challenges of Health Data Analysis

One's own health and the responsibility for patients' health is a highly sensitive topic. To secure the trust of doctors and patients in AI-supported healthcare in the long term, several aspects need to be considered.

Data Protection

The GDPR classifies health data as a "special category of personal data" that may only be processed under certain conditions. The Medical Device Regulation (MDR) also demands compliance with data protection, particularly in clinical trials. For this reason, it is necessary to use locally hosted LLMs to maintain control over personal data. Access by third parties is thus excluded.

Explainability

Because the medical field is a critical domain directly linked to people's health, the results of AI-supported solutions must be transparent and explainable. The challenge in implementing corresponding products is the selection of suitable models. There are fundamentally two types of LLMs:

1. Proprietary Models: Proprietary LLMs, such as GPT or Gemini, are like a black box – there is no insight into the structure and processes within the models. The exact model architectures are therefore unknown, which leads to hard-to-follow decision processes.

   
   

2. Open Source Models: For open source models, the source code is publicly accessible. This results in numerous advantages over proprietary solutions:

   • Cost efficiency

   • Flexibility

   • Public source code

   • Trustworthiness

   • Potential explainability

Open source models thus provide an optimal combination of security and control.

Implementation of an AI-Supported App

CarByte has developed an app based on local open source models that supports both doctors and patients through AI. The functionalities include:

• Reading medical documents

• Automatic detection and categorization of lab values & free text

• Analysis of documents

• Creation of structured summaries

• Understandable explanations of medical terms for patients

• Translation function

• Clear presentation of the extracted text and sound data based on a dashboard

• Speech-to-text functionality, for example, for capturing the medical history

• Creation of medical letters

• Integrated feedback loop for fine-tuning the models based on user feedback

It should be noted that the app will only serve as support for medical staff, and the medical professional remains the primary authority for any decisions.

Outlook

It is expected that LLM-based systems will eventually be adopted in all German medical specialties to ensure high-quality medical care. This supports both the work of specialists and the understanding of patients. The local application contributes to protecting the personal data of each individual and fostering trust in AI-supported solutions in healthcare.