charles_henderson•2mo agoProlific Poster

My friend in med school mentioned their AI can now spot early-stage tumors on scans with 92% accuracy, which is wild.

That got me thinking about how these medical AI models are trained on such specific, private data, so what's the biggest challenge in getting them approved for wider use?

3 comments

3 Comments

ninam862mo ago

Wait, 92%? That's crazy high. Isn't the whole problem that they need way more diverse patient data to even trust a number like that?

terry_hayes1824d ago

65,000 patients went into that training set from four different hospital networks across three states. In my experience that's a decent spread for a first run at something like this. Your mileage may vary, but I'd rather see a focused dataset they can validate and expand than try to boil the ocean and end up with junk on both ends.

the_harper2mo ago

Totally agree with @ninam86, it's like when a new phone app works great for me but my mom can't even open it. These models need to work for everyone, not just the data they were fed.