Sana: what a health AI shouldn't do

On May 11 I went to a hackathon at Universidad Torcuato Di Tella. The challenge: there are places in Argentina where the nearest specialist is six hours away. How can AI help there? We had two hours, in teams of up to five, to come up with an answer and pitch it.

One condition was clear: an appointment app didn't count. The problem starts when no doctor is available and someone needs to know what to do right now. The example we used: a kid with a high fever at night, hours away from the nearest hospital.

Our answer: Sana

A chat that asks narrowing questions, more like Akinator than a form: symptoms, age, how long it's been going on, medical history. At the end it gives guidance: what it could be, how urgent it is and what to do now.

What we decided it wouldn't do

• —It doesn't diagnose. It guides and prioritizes: do I need to go to the hospital now, or can it wait?
• —It doesn't hold on to serious cases. Chest pain, trouble breathing, fainting or seizures go straight to a person, or the user is told to seek care right away.
• —It doesn't depend on a good connection. Messages are light, and questions are saved offline and sent when the signal comes back. It can be used from a health post.
• —It doesn't pretend to be a doctor. Every answer says it's preliminary guidance, based on protocols supervised by professionals.

What we built in two hours

The team put together the concept and the pitch, and Eitan and I built most of the interface prototype, so you could see what using it felt like. It wasn't a working medical system: it was a way to show the idea.

People liked the idea. And that's where it stayed: we didn't take it further.

What I take from it

• —In healthcare, what the product doesn't do matters as much as what it does. Half the pitch was about limits.
• —A pitch people like doesn't mean a validated need. For that we would have had to talk to doctors and to the people who'd use it.
• —Two hours is enough to show an idea, not to prove it.