I am trying to reinvent how we interact with medication instructions, using new technology to make everyday clinical work feel more natural.
Tired of entering medication instructions one field at a time? The dose goes into one box, the frequency into another, the duration somewhere else. An instruction that takes a single sentence to express becomes a sequence of clicks.
I want healthcare software to work more naturally. A clinician should be able to write an instruction, see how the software interprets it, and check the result.
That is the idea behind Posologie: a small, specialised AI model that translates free-text medication instructions into structure. A temporary demo is available at posologie.stighellemans.com. I am sharing it to learn where the model needs to improve and where this approach could make a practical difference.
Write the instruction, review the schedule
The video below shows a fictional instruction being built in three steps: a tablet with breakfast, a duration of seven days, and then a second phase with a tablet at breakfast and at the evening meal.
The result is a visual schedule that the healthcare professional can check and adjust. The aim is simple: spend less time navigating forms while keeping a clear view of what the instruction means.
Let people write naturally
We need a shift in how we ask healthcare professionals to document their work. A sentence is often the easiest way to explain what someone should take, when they should take it, and how the treatment changes over time.
I want that freedom to remain. Software should help turn our words into information that other systems can use, while letting us check that it understood us correctly.
This becomes even more relevant as we start talking to our devices. AI scribes can turn a conversation or dictated note into written text. The next opportunity is to make those words useful in the tools we use every day, without having to enter the same information again.
A specialist for a repetitive task
Large generative AI models are incredibly capable, and I see them as essential tools. Their ability to handle many different problems is a major strength.
But when the same clearly defined task comes up thousands of times, a smaller model trained specifically for that job deserves a closer look. It can run on modest hardware, respond quickly and become inexpensive to use at scale.
With enough good training examples, I believe a specialist can match or even outperform a much larger generalist on its particular task. That is the potential I want to explore and test.
Promising early results
In an early comparison, an experimental version of our small model ran on a Raspberry Pi, a compact, inexpensive computer. It returned a result in about 2.4 seconds, compared with 22.5 seconds for the larger-model workflow we tested. The two approaches also produced similar scores on a small set of reference examples.
The chart assigns the full monthly cost of a dedicated cloud server to this one model. In that scenario, running it costs about €1.10 per 1,000 instructions at 10,000 instructions a month, as the hosting bill is spread over more uses.
But a small model does not need a server all to itself. It can share a server with other applications, or run on hardware an organisation already owns. That can make the additional cost much lower, even at modest usage. The useful comparison is what it costs to add the model to your own setting, including the computing capacity, electricity and maintenance it needs. Flexible deployment is part of what makes small models attractive for healthcare.
Help me find the useful applications
The temporary demo is a way to learn from people who recognise this problem in their own work. I would like to understand which instructions are difficult to capture, which mistakes the model makes, and where a proposed schedule would save time or still create extra work.
Please try Posologie using fictional examples, without patient information. It is a research demo; its output needs checking. If you encounter a failure, have an idea for another application, or want to explore collaboration, email stig.hellemans@uantwerpen.be.
My goal is to bring this capability into healthcare and make the greatest useful impact with the limited resources we have. Giving healthcare professionals a more natural way to work is a good place to start.
Developed in collaboration with Dirk Broeckx and Robert Vander Stichele.