Would you trust a machine learning system to calculate the best treatment for you?
This question might sound a little futuristic, or even like science fiction, to some. The use of machine learning in healthcare is, however, the core of the IMEDALytics research project, in which we’re joining forces with Philips GmbH (Innovative Technologies department), the intensive care department at the Uniklinik Aachen, RWTH Aachen and Trier University.
Starting situation
In order to explain what specifically is being researched under the IMEDALytics project, let us briefly explain the current treatment situation. On intensive care wards, care staff must ensure that patients’ fluid balance is stable alongside their many other treatment tasks (such as administering medication). Any bodily fluids lost by the patient must be compensated for quickly. This treatment is known as “volume substitution”.
The aim of this therapy is to restore a normal blood volume to improve circulation. The hereby increased oxygen supply in the tissue means better overall organ function.
Establishing the correct dose isn’t easy: doctors and care staff must work out the optimal indication, a suitable infusion solution and the tailored application of the right dose. If they don’t manage this balancing act, the result can be undesired long-term consequences such as more serious care or long-term ventilation.
What is IMEDALytics?
The aim of the research project is to develop a digital system that improves medical staffing and the quality of care received by intensive care patients. The system should also support ambulance staff in estimating each individual patient’s risk level and further treatment.
Current guidelines are to be combined with innovative machine data analysis technology. This process consists of multiple steps:
Medically relevant guidelines that are only available in text format are semantically collected using natural language processing approaches and integrated into the recommendation.
We combine individual patient status data with treatment path templates based on huge patient data sets.
The IMEDALytics system learns and constantly generates new knowledge from collected data fed to it.
In future, this newly attained knowledge can be used to develop new, data-supported processes to create evidence-based guidelines.
Specifically, this means that the care staff using the IMEDALytics system will be provided with qualified suggestions for the patient’s optimal treatment. The final decision is always made by a human.
Doctors and nursing staff have highly specialized knowledge, lots of experience and precisely learned routines. Machine intelligence can evaluate huge amounts of data, identify complex interactions and provide precise recommendations. Using this kind of statistical analyses in healthcare is a tempting prospect, and the potential to identify hidden patterns and support the treatment of patients is great.
What is our role in the project?
Our task? The human-oriented design of an efficient, effective, satisfactory interface to comply with ISO norm 9241-210 and 62366.
Specifically, this means that we are conceptualizing and designing a graphical user interface (GUI) that will allow the user to use the whole of the IMEDALytics system - including algorithmic recommendations.
The medical context comes with particular challenges for technical systems and their user interfaces - and, therefore, us. It is essential that the system seamlessly integrates into everyday life at the clinic and the processes of medical staff so that it is actually seen as helpful and used in their busy work days. A high degree of usability and a positive user experience are essential.
Users are also under pressure and stress. Mistakes in treatment can have grave consequences but staff are left with very little time to think things through.
Essential for the system: use context
In order to do these design requirements justice, it’s essential to glean a deep understanding of all user groups. Through intensive interviews and job shadowing, we are creating a multifaceted view of staff’s everyday lives and the various treatment options for volume substitution.
One of the first challenges arose right at the beginning: current guidelines and algorithmic recommendations have to be visualized so that they accommodate individual user preferences. Then these need information that varies in terms of complexity and depth:
For diagnostics (retrospective) and
Suggestions for treatment, in parallel to the risk assessment (prognostic)
Easy-to-understand information on current monitoring
In addition, it must be considered where interaction with the system can take place naturally and which user group the system takes into consideration. Again, the key here is that doctors have the opportunity to enter feedback into the system manually.
Summary
It is our aim to optimally complement human skills with the possibilities offered by a machine learning system. Over the three-year period, we will be working with our project partners to create an IT-based decision-making support system for volume substitution in intensive care patients.