Whining about obstacles facing medical AI

In the category of "things that will seem really absurd in a couple decades," I thought I'd write down a few recollections of a consulting job Novamente had for a medical data management company, which I'll here call ABCD.

I started the ABCD project with high hopes of using AI to help hospitals optimize their processes and cure people better. It hasn't worked out that way yet.

As a couple examples

1)
We wanted to try to predict which medical therapies tend to lead to best outcomes for patients, contingent on various properties of the patient and their diagnosis. Oops. Hospitals don't really keep data on the quality of patient outcome. They measure outcome by LOS, length of stay in the hospital after a procedure is done ... but this tends to be determined by how long the patient's insurance will pay for them to stay in the hospital. There are also typically some handwritten notes by the doctor reporting something about the patient's condition, but in an unsystematic way...

2)
The different hospitals serviced by ABCD store data about the products they use (scalpels, gloves, EEGs, whatever) in their databases in unsystematic ways, without indexing them by any kind of universal product ID. So ABCD wound up using our AI tech to **guess the product ID** for a product, based on looking at its corresponding database record in a hospital database. This is a really irritating AI application because the AI is doing a lot of work to make a 90% accurate guess of something that some human really should have just typed into the DB in the first place (the product ID).

Eventually these sorts of problems will go away, I'm sure ... but right now they mean that real-world medical AI is bloody hard to do.

If there were a hospital that were committed to collecting data in a manner conducive to medical AI (or to serious statistical analysis, for that matter ... the requirements are largely the same), then a worthwhile partnership could be undertaken. They would need to actually record information in a database in a manner useful for AI analysis using current tools, e.g.

• identify products by their universal product IDs
• have each doctor make some kind of assessment of a patient's outcome in machine-readable format

etc.

This seems elementary but based on our experience with ABCD (which serves many major hospitals) it is not standard practice.

In a couple decades all this stuff will be systematized and these issues will seem absurd ... like Babbage having trouble building his Analytical Engine because the screws were all slightly different sizes. But in the actual moment these sorts of "minor technical issues" can loom really large.

An analogous issue is the current lack of any technology combining virtual worlds and robot simulators. This is obviously a "minor issue" in the big picture: both of these technologies exist, so it's "just" a matter of software integration. But in real life, in the thick of doing real work, 5 man-years of software integration work becomes a big deal and prevents great things from happening, or causes them to happen 5-10 years later than would otherwise be the case...