AI in healthcare: a reality check at Karolinska

AI moves fast. Healthcare takes time.

It was definitely a reality check for me, because AI evolves so fast, but regulation and bureaucracy slow down adoption like crazy. Imagine AI identifies a promising new treatment today. There is still a long way from that discovery to knowing whether it works safely in people and making it available to patients.

I get it, we need to make sure treatments are safe and effective. Testing takes time, and so does getting something to work in a real healthcare setting. But I left with the feeling that the current regulatory process will break in our AI world. If discovery keeps accelerating, how will everything that comes afterwards keep up?

Virtual patients you can practise with

However, one example that Karolinska presented really stuck with me because it shows how you can be innovative with AI despite all the regulations: virtual patients that you can practise difficult conversations with.

The AI plays a patient with a particular background and responds to how you ask questions. You have the conversation, then get feedback. So you get practice handling hard topics, including situations where a patient’s background and cultural experiences may affect the conversation.

The platform is called Virtuella patienter and was developed at the Centre for Psychiatry Research at Karolinska Institutet. It is designed for clinicians and students. You can find it at vp.ki.se.

Visit Virtuella patienter at KI

Asking children about violence at home

They described how this started with helping healthcare staff practise asking children about violence at home. That’s a conversation you really want to handle carefully.

Karolinska’s account of the original project describes a chatbot with child avatars, each with a different backstory, and the option of digital coaching and evaluation. Staff can repeat the exercise as they become more comfortable asking difficult questions.

I really liked this example because I can totally see how practice like this could improve how patients feel when talking to doctors. Having somewhere to work through an awkward question, or realise that your wording could be misunderstood, seems very useful before you are sitting opposite a real person.

Read about the original child-protection project at Karolinska (Swedish)

The people who understand the problem

I also liked it from the development side. You need researchers and clinicians who understand the problems in the field, people who understand the technology, and people who can bring them together.

Experienced clinicians helped shape these scenarios. They know what makes a conversation difficult, what staff need to practise and what useful feedback should look like. Those are things you need to understand when you build a tool like this.

That part is very familiar to me from working with AI in other fields. You need to spend time with the people doing the work to understand where the technology could actually help. This was a really interesting example of what can come out of that.

Further reading

• About Karolinska Institutet
• Centre for Psychiatry Research: virtual patients and clinical conversation training (Swedish)