Structuring the Unstructured: How We Use AI to Empower Medical Insight

About the project

One of the ongoing challenges in the healthcare sector is the vast amount of unstructured medical text, such as pathology and lab reports, which are difficult to analyse and transform into structured, usable insights. To address this, we developed an internal solution powered by Large Language Models (LLMs) similar to GPT-4 (via Azure Cognitive Services).

The goal is to enable healthcare professionals to convert complex free-text data into meaningful insights in a manner that is fast, secure, scalable, and highly efficient —ultimately supporting better data-driven decision-making across the sector.

The challenge

Healthcare professionals generate vast amounts of data through clinical documentation, particularly in the form of free-text entries in Electronic Medical Records (EMRs), pathology reports, and laboratory results. These narratives contain valuable insights—ranging from patient symptoms to specific diagnostic test outcomes, such as NTRK mutation tests—but they are unstructured and therefore difficult to analyse at scale.

Traditionally, extracting meaningful information from this type of data has required manual annotation, a process that is labour-intensive, time-consuming, and susceptible to human error.

On top of this, variations in language and coding systems between healthcare institutions—and especially across countries—further complicate the process; not to mention navigating the complex requirements of data privacy, compliance, and security in the context of healthcare regulations.

Finding a way to transform unstructured clinical text into structured, analysable data—quickly, accurately, and in a privacy-compliant manner—has therefore become an essential step towards more efficient, data-driven healthcare.

Our solution

Our team developed a method to automatically identify whether specific tests had been performed and, if so, what the outcomes were. This process is carried out by prompting a Large Language Model (LLM) with targeted questions like: “Has this test been performed?” and “What was the result?”

To ensure this could be done responsibly and securely, LOGEX chose to integrate GPT-4 via Azure Cognitive Services. This setup allows us to harness the power of LLMs in a way that aligns with our contractual, regulatory, and ethical standards. Unlike public tools such as ChatGPT, this configuration guarantees that no data is stored or used for training by external parties. In fact, the AI models are run in a local, isolated environment, ensuring that all data remains within our secure ecosystem throughout the process.

In addition, all data is pseudonymised before processing, meaning that no personal identifiers—such as names or patient IDs—are exposed. From a security perspective, our use of AI is no different from storing and analysing data in an Azure-hosted database. This careful design reflects our commitment to being a trustworthy and responsible partner in healthcare data processing.

The result is a scalable solution that can process thousands of reports in parallel—reducing work that previously took weeks. And because the output is structured and ready for analysis, medical professionals and researchers can focus on generating insights rather than cleaning and preparing data.

“You had to annotate 35,000 reports manually – that takes weeks. Now we can do it in 1.5 hours, and more accurately.”

Maarten van Meeuwen

Data Engineer, LOGEX

Results

Our first project was a collaboration with UMCG, where we focused on analysing over 35,000 reports. This initial project spanned two years and involved experimentation, proof-of-concept (POC), infrastructure set up, and designing a workflow orchestrator that is capable of processing large volumes of data in parallel. Today, this same process can be completed in just 1.5 hours, with increased accuracy.

To ensure reliability of the process, the AI model performs a first pass based on a defined prompt, identifying which tests have been performed and what the outcomes were. These outputs are then reviewed and validated by domain experts, refine the results and create validation sets to ensure accuracy.

As part of the validation process, we typically select a sample of around 500 reports which were annotated manually and also processed by the AI model, allowing us to compare the outputs directly. Interestingly, the AI often catches things human miss, due to the repetitive and routine nature of the task, humans tend to make mistakes. In contrast, the AI demonstrates greater consistency and accuracy, delivering higher-quality insights and enabling doctors to act faster and with more confidence.

This approach also helps bridge the gap between domain expertise and data implementation. Medical professionals, researchers, and analysts can now engage directly with the tools, reducing their dependence on highly technical teams. This fosters a faster route from raw clinical data to meaningful, actionable insights.

Building on the success of the UMCG project, we are now focusing on Spain, where the healthcare system produces a high volume of free-text medical documentation. By scaling the solution to handle different languages and data standards, we are expanding the scope of what can be achieved in both local and international contexts.

“We’re helping customers move from raw data to actionable insights with much less effort. That’s a game-changer.”