Author: Ossi Koskinen, Head of Key Accounts at Siemens Healthineers. He is responsible for strategic client relationships in Finland and Sweden.
In healthcare, the question is no longer whether artificial intelligence can be used. The much harder question is how a solution that has proven effective can be integrated into daily work in a way that genuinely benefits healthcare professionals and patients.
This question comes up repeatedly in my discussions with hospital leaders, clinical experts and IT professionals. There are many solutions and promising pilots, but the journey from a pilot to established use is often surprisingly long.
This is not a new phenomenon.
I have spent much of my career working with healthcare transformation and technology implementation. Over the years, I have been involved in projects such as developing a pain assessment tool for what was then Windows Phone and paving the way for remote consultations at a time when there was still relatively little interest in them.
In hindsight, investing in Windows Phone was not one of the most successful technology choices of my career. However, the experience taught me something that still holds true: a solution that works technically does not necessarily make for a service that works in practice.
The situation with remote consultations was somewhat different. The necessary technology had largely existed for years before the surrounding environment was ready for widespread adoption. The COVID-19 pandemic changed the situation rapidly. Today, it seems quite natural that a patient in Finnish primary care can also be seen by a healthcare professional working from another country. This is currently the case, for example, with an acquaintance of mine who lives in France.
Technology can therefore be ready long before the surrounding operating practices, structures and people are. With AI, this is exactly where many organisations currently find themselves.
Many solutions perform well from a technical perspective. Yet the promised benefits are not always visible in everyday work. The reason often lies in how the solution connects with hospital systems, workflows and decision-making.
I have seen situations where the objective was to save healthcare professionals’ time. In practice, the user first reviews the original information, then the AI-generated assessment and finally checks the matter again in another system. The new solution was supposed to eliminate work, but instead introduced an additional verification step.
In such a situation, the algorithm may be working exactly as intended. Yet the healthcare professional’s work does not become any easier.
AI-generated information needs to be available in the system and at the point in the workflow where the healthcare professional is already making the relevant decision. The information must be presented in an understandable format and must enable action. Otherwise, AI can easily become just another tool alongside the existing ones.
In medical imaging, the opportunities already extend far beyond the analysis of an individual image. AI can support patient preparation, image acquisition, image interpretation, reporting and the integration of information into clinical decision-making. What matters is not an individual application, but the entire pathway from the patient to the image, from the image to the report and from the report to treatment.
This is why, in my view, an AI initiative should not begin by asking where we could add AI.
I would start with much more practical questions. Where does work come to a halt? Where do delays occur? Where is the same information documented more than once? Where do healthcare professionals have to switch between systems? Which tasks consume time even though they do not require clinical expertise?
Once the actual problem has been identified, we can assess whether AI is the right solution. Sometimes it is. In other cases, changing the operating model or adopting a somewhat simpler technical solution may be more helpful.
The same practical approach is needed when measuring success. The accuracy of an algorithm is important, but on its own it does not demonstrate the value of a solution. In healthcare, the impact should be visible in patient care, the work of healthcare professionals or the organisation’s operations. Preferably in more than one of these.
The effects must also be monitored after implementation. Systems and imaging equipment are updated, patient populations change and new versions of AI solutions are released. Implementation therefore does not end on the day the solution is made available to users.
This may sound self-evident, but in practice, responsibilities can easily remain unclear. Who monitors the solution’s performance? Who handles user feedback? When is local validation required? How are the effects of software updates assessed? And who makes the decision if the solution no longer performs as expected?
This work requires clinical users, IT professionals, technical experts, leadership and vendors. People who understand both clinical work and the opportunities and limitations of technology are especially valuable. They are often able to identify problems that may be overlooked when the issue is examined from only one perspective.
Managing the overall environment becomes more difficult as the number of AI solutions increases. Ten separate solutions do not yet constitute a functioning environment. Each of them requires integrations, access rights, maintenance, monitoring and support. If the overall environment is not designed in a coordinated way, users can easily find themselves faced with a new set of digital work steps.
In the best-case scenario, healthcare professionals do not need to think about which AI solution they are using at any given moment. The necessary information appears at the right point as part of their normal work. The healthcare professional assesses the information, makes the decision and remains accountable.
The next interesting step is to look beyond individual workflows and examine hospital operations as a broader system. In this context, I find the concept of an operational digital twin particularly interesting.
The term may sound technical, but the basic idea is highly practical. If we can create a sufficiently up-to-date view of patient flows, staffing, facilities, equipment, schedules and the interdependencies between care processes, we can assess the effects of decisions before implementing them.
For example, what happens if demand for diagnostic examinations increases at one hospital? Should patients be directed to another unit, should staff shifts be adjusted, or should appointment schedules be reorganised? How does servicing one imaging system affect patient waiting lists, staff workload and follow-up care?
In a real hospital, it is difficult to test these alternatives. Operations continue at all times, patient care cannot be put on hold, and the impact of an individual change is not always easy to distinguish from everything else that is happening. In a digital model, different alternatives can be compared before changes are implemented in practice.
This could also help target AI investments more effectively. Instead of starting with an existing solution and looking for a use case for it, organisations could analyse their operations to identify where waiting, manual work, breaks in information flow or uneven capacity utilisation are causing the greatest harm.
Only then should we consider whether AI is needed.
However, even a digital model does not solve everything automatically. The utilisation rate of one system may improve while the patient’s total time in the care pathway becomes longer or the workload of staff increases. Objectives cannot therefore be defined solely in terms of efficiency, speed or cost.
There must be agreement on what constitutes good performance. This includes the quality and safety of care, the work of healthcare professionals and the patient’s overall experience. Ultimately, these choices are matters of leadership. A digital twin can help make the alternatives and their potential consequences visible, but people decide what is acceptable and what should be pursued.
A broader opportunity arises when the scope is expanded from an individual department to the entire hospital, and later to an interconnected system consisting of multiple services or hospitals. This would make it possible to examine staff availability, capacity utilisation and patient transitions between different stages of care within the same model.
There is also a word of caution. The broader the system being modelled, the greater the need for reliable and up-to-date data. A shared understanding of the intended objectives is also required. In practice, it therefore makes sense to begin with well-defined problems whose effects can be identified and measured.
Finland has a strong foundation for this. We have digital infrastructure, health data and extensive expertise. There is also considerable collaboration between hospitals, research organisations, public-sector stakeholders and companies. Despite these strengths, progress will not come from launching new pilots alone.
The more difficult phase begins when a solution that has proven effective needs to be integrated into the right workflow, its impact needs to be demonstrated and its use needs to be scaled beyond the first unit. Succeeding in this phase usually requires more persistence than launching the initial pilot.
The purpose of AI is not to remove the human being from the relationship between patients and their care. It should give healthcare professionals more time for tasks that require experience, judgement and empathy. The need for these qualities will not disappear as technology develops.
When I assess a new solution, I often return to a few simple questions:
- What problem are we trying to solve?
- How will the healthcare professional’s work change in practice?
- How will the effects be monitored during use?
- How does the solution connect with other systems and processes?
An operational digital twin adds one more question:
What happens elsewhere in the system if we change this part?
The answers to these questions usually reveal more about the likelihood of success than a technical demonstration or a product claim.
AI is developing rapidly, but the most important question in healthcare remains highly practical: does the solution help healthcare professionals do their work better, more smoothly and more safely?
If the answer is visible only on presentation slides and not in daily work, the implementation is not yet complete.
Ultimately, success is determined by workflows.
