Radek goes to Hollywood (Palo Alto)

Way before I got my first computer, I would read and collect PC magazines and imagine what it would be like to play all those wonderful games.

For one reason or another, despite never playing it, the title Frankie goes to Hollywood stuck with me throughout the years and now I get to paraphrase it to title this post :slightly_smiling_face:

In fact, I am not headed to Hollywood but to Palo Alto to meet in person the ingenious people of Curai (a startup scaling the world’s best healthcare for every human being). We started talking a while back and we connected predominantly due to me being active on these forums!

I don’t believe that this is in any sense an outcome to be expected when embracing the fast.ai way of learning. I see it as an auxiliary occurrence to being part of this fantastic community of learners. But it is an inspiring story nonetheless, one that is currently under way, and it would be remiss of me to not report it. This is just one of the many great things that you open yourself up to if you keep learning & shipping, be that posts, projects or kaggle competitions (I’m writing these words in as much for you as a reminder for myself).

I will be in Palo Alto over the next two weeks. There’s a bunch of folks on the forums I interacted with and would love to meet - if you would be around Palo Alto and would be open to grabbing a decaf venti soy latte (or whatever exotic beverages the locals drink!) give me a shout please! Not sure how I will be on availability / mobility but it’s worth a try.

Here’s one to the adventure and the kind souls that extend a helping hand to us along the way! :tropical_drink:



1 Like

Good luck!! But be aware about what they claim as their approach :

“We are using artificial intelligence / machine learning with a user-centric focus to provide instant medical expertise that is accurate, trustworthy, relevant, and actionable.”

is just a commercial message, IMO. Unless they only want to match patient needs with the right specialist expertise.

Once you will go trough the real challenges of this field, you could easily understand how hard is to combine data from different sources in an interpretable model. The failure of IBM watson health is an example of it.

Precise medicine is out of the reach of our current state of AI tools. But of course it is very important to take some small step in that direction. So again good luck!!!

PS: Xavier Amatriain has very strong references, so it will be a super useful experience to share time with him.


Congrats! Sounds like a promising opportunity. We need major disruption in healthcare in order to bring costs down and quality up. Although my degree is in CS/AI I have worked in healthcare since before I graduated in '91. However, I spent most of that time (and still do) applying more of my minor in cognitive science, than AI. Many of the issues with healthcare, especially in the US, are due to the poor design of the healthcare system and all of the perverse incentives that drive behavior. For instance, many poor decisions are made because we don’t have a universal patient health record, or the UI is so bad that the doctor doesn’t even know a result is available and/or cannot see the big picture. We still have results being printed from one facility, faxed, then stored as a digital fax in a different EHR. It is very difficult, for instance, to figure out which of your patients have been pregnant and when, because patients often go to different healthcare systems.

Some of the most positive recent changes have been through value-based pay, which focuses on meeting specific quality metrics, such as giving flu shots, controlling BP, cancer screening, etc. However, it is quite a nightmare to automatically calculate these, because the data is hard to get to (could be in text or a faxed image, etc.) or just plain missing.

Applying AI to imaging and other such critical subtasks seems more doable, since such algorithms can just replace something that is already there, but these are not exactly transformative. They may lower cost if they save expert time, but that depends on how much they cost–healthcare companies are really good at charging a lot for new tech.

Predictive algorithms are also clearly useful, but only if causal knowledge is available to act on the predictions and only if you can show that they can transfer to new sites–something that has been an issue in almost all cases. Some of the companies in this space tune their models for each site. Then again we have the missing, bad data problem. Even structured data is wrong about 30% of the time (for diagnosis codes).

I’m personally more interested in causal predictive models, since those should show better out of distribution performance. However, 80% of my work is geared toward dashboards that show how we are doing on our metrics, including gap lists of patients for use at the point of care. I’d love to do more AI here, but I have to work where the funding is–at least for my day job. This is why we need companies outside of healthcare working to transform it–those of us working within healthcare have so many urgent problems to solve that it is hard to step back and make major changes.