AI pioneer warns of fragmentation

The deployment of AI in hospitals is in danger of falling victim to fragmentation. This is what founder Wouter Kroese of AI pioneer Pacmed warns. "In the future, healthcare providers will start using dozens and perhaps hundreds of AI-based insights and predictions. If those are all point solutions, it will become unaffordable, un-implementable and unmanageable."

Kroes's cry for help does not come out of the blue. In 2014, Pacmed developed the first plans to improve healthcare using digital data and artificial intelligence (AI). Ten years later, Kroese and his colleagues can look back on some high-profile collaborations, including with Amsterdam UMC and Santeon, but also on some--as Kroese puts it--"painful statistics." "We have been involved in dozens of different development projects. The results were actually always beautiful and those involved enthusiastic, but implementation and use rarely came about. A promising first algorithm is only a fraction of the work to get to a used and scalable product."

Painful conclusion

The laborious adoption of AI is closely related to the fragmented organization of healthcare, according to Kroese. Partly because of this, there is a strong tendency for hospitals and departments to seek solutions on their own. But point solutions in the field of medical AI are dead in the water as far as Kroese is concerned. "We have managed to bring AI to the patient twice in recent years. In both cases, our software proved to be able to generate predictions and insights that help improve specific medical decisions. Yet in both cases it did not prove to be a viable and scalable product. It's time for a hard and, for Pacmed, painful conclusion: Medical AI decision support is not viable as a point solution."

EHR integration

Scale, according to Kroese, is indispensable to address some of the fundamental issues surrounding the application of AI in healthcare. Among these, he thinks of data quality and the limited accessibility of data sources, such as the electronic health record (EHR). "We started rather naively at the time to process, harmonize, validate, enrich and standardize all ICU data, but it takes an awful lot of time to link. Not to mention integration into all the disparate and differently configured EHR systems."

Mortal sin

EHR integration is just one of the technical hurdles. Added to that are issues such as data extraction, CE certification, continuous data monitoring and cloud hosting. Together, they make the cost "far too high for a point solution," Kroese said. "We've paid learning money in this regard over the past decade. We obviously had to start somewhere. And then a point solution is the logical starting point. But the fact that ten years later in the Netherlands we keep starting this over and over again in hundreds of places is a mortal sin. In the future, healthcare providers will be using dozens and perhaps hundreds of AI-based insights and predictions. If those are all point solutions, it becomes unaffordable, un-implementable and unmanageable."

AI hype

Still, this is the trend Kroese does see. "All over the country there are new initiatives. It is a shame that these are so fragmented and that those involved usually start all over again. We spend millions of euros and thousands of scarce hours on AI in the Netherlands. Then it is crucial that these actually contribute to future-proofing the healthcare sector, which is under enormous pressure. Otherwise we run the risk of blowing up an AI hype now. If that soon snaps, we lose time and attention that AI is going to desperately need in the coming years."

Reuse data

Pacmed aims to lead by example by consistently linking further activities to benchmarks such as value, impact, accountability and scalability. "Together with hospitals, we are further developing Pacmed Critical, our software for timely and safe discharge from the ICU, into an integrated and ICU-wide product that brings together all valuable information from the needs of caregivers," Kroese said. "That means, for example, using the same data for different purposes. We will additionally facilitate the development of AI for hospitals and help ICUs bring their AI ambitions, research and projects to the bedside."

That this is possible is proven by the ongoing activities in OLVG, according to Kroese. "We have now been able to develop and roll out a new functionality for better capacity utilization in a few months, whereas it took us seven years to develop our discharge software."