One of the problems that really prevents the change is that analysis of patient data is so highly protected that it's virtually impossible to do proper studies. We assume that people with such access are performing them, but I know from second-hand experience that actual time spent on real treatment studies is very small and most of the time is spent dealing with policy and patient privacy issues.

Large insurance and government-based (VA, CMS, etc.) providers have the most data, but either what they're looking for isn't exactly aligned with what patients need, or the organizations are so fubar (e.g. the VA) that real studies are unlikely to happen.

There really needs to be some kind of research exception/feedback loop on massive scale data analytics that allows some group of objective researchers to mine patient data and put in place actionable recommendations that work their way into medicine rapidly -- perhaps with some penalizing stick insurance providers can leverage against practitioners to make sure they adopt newer better practices.

I'm a GP, and I see dozens of patients every year with cases of Foozle disease. I was taught in medical school to give them orange pills, but I follow developments in general medicine and I know the new purple pill has seen good results for Foozle too. Nobody has published even a small study about which is better.

The law says since a responsible doctor wouldn't know one is better I can choose either. I could give all my female Foozle patients the orange pills and prescribe the purple for everybody else, or I could just pick whichever is cheapest, or I could even pick randomly for each patient. All legal.

But, if I pick randomly AND record the results, thus doing science to find out which was best, that's an unauthorised medical experiment and I could lose my job or even go to jail.

I am trying to make the survival data of a clinical trial combined with sequencing data of the patients publicly available. Nearly all the patients have died by now, as the patients in this cohort patients have a very poor prognosis (median survival of about 24 months). It is still almost impossible to release clinical data (diagnosis, number and location of metastasis, age of patients, etc) combined with sequencing data without violating privacy laws.

In the meantime, there is almost no career incentive for me to make the data available. I do not have, and am unlikely to get a permanent contract. In my next job interview I will be judged by the number of publications on my CV.

When it's the other way around and someone survives 10 years on a "2 years left to live" diagnosis, we seem quick to credit their strength, determination, spirituality, etc., rather than looking at how inaccurate the original prediction might have been.

What's the other 10%?

I'm serious - I've got a better-than-average understanding of statistics, and everything I've seen referred to as "AI" seems to really be just a statical model with a thin layer of code over the top to make decisions based upon it on the fly. Is there more to the state of the art that isn't apparent from the outside?

In fact, the only thing that comes to mind that doesn't fit that description would be genetic algorithms, but I don't hear much about them these days.

"...the novelty of the approach does not lie simply in incremental model performance improvements. Rather, this predictive performance was achieved without hand-selection of variables deemed important by an expert, similar to other applications of deep learning to EHR data. Instead, our model had access to tens of thousands of predictors for each patient, including free-text notes, and identified which data were important for a particular prediction."

So it sounds like the advance here is actually in the following: "a generic data processing pipeline that can take raw EHR data as input, and produce FHIR outputs without manual feature harmonization".

The article doesn't explain this very clearly. Yay!

I think you want to say Skoda or Seat. Both owned by VW and the cheap cars.

It's better if this becomes an ethical/moral problem about when to care for patients and to what degree, rather than a guessing game problem while anybody's guess is as good as anybody else's.

at a conference, i once heard that the best single predictor of ICU patient mortality was reduction of doctor attentiveness. in other words, patients in ICUs whom doctors began to visit less, treat less, basically start to ignore -- those patients were the ones for whom death was imminent.

the challenge is that it's hard to untangle cause and effect. was it a self-fulfilling prophecy? etc.

https://en.wikipedia.org/wiki/Harold_Shipman

Have you ever asked her if she wants to do the job of predicting when a patient will die? How many hours would she need to spend reading the patient's history to come up with a prediction she feels satisfied to use to decide what service the patient will get? I'm not a medical professional but my gut reaction is I don't want to do anything with predicting or sentencing if some machine can do it almost as well as I can.

I remember this story of a radiologist who told me he thinks we spend too much money for too little during the last about six months of a patient's life. If we had better information on when the last six months starts, maybe we could reduce the cost of healthcare? Apparently, we spend cost to 18% of GDP on healthcare in the US? Iirc, most of Europe is closer to 12?

Just think about the machines/algorithms your wife already (indirectly) uses to predict outcome for patients.

- An MRI scan, which is created from physical laws (Maxwell's law together with a quantum mechanical understanding of hydrogen atoms) and reconstructs a 3D-image of the different tumors. Knowing if and where metastasis are present have an enormous impact on the prognosis.

- The algorithms used to align sequencing reads from biopsies to determine the mutation status, which are critical to determine the prognosis of some tumors.

In fact, I'd wager that if your wife did not have access to these algorithms, she would already perform worse. And, in my opinion, the distinction between machine learning and the algorithms used to reconstruct 3D-images or align sequences is somewhat arbitrary.

The problem isn't that insurers are horrible people, it's that insurance is a bad fit for healthcare.

Medicine has advanced to a point where some serious illnesses can be kept at bay for decades. People with some formerly mortal diseases can get medicines that cost $100k/year and/or professional help every day, and live well for thirty years.

This means that each of rich societies has to choose:

1. Raise health insurance rates until the budget covers everything that's possible medically. However, in some rich countries, people's after-tax income is only a few doublings above the current health care plus pension costs...

2. Decide that some treatments aren't worthwhile, ie. everyone with Hyperthis or Abnormalthat Syndrome gets cheap palliative treatment and a peaceful, gentle death.

3. Decide on a per-patient basis, often involving numbers such as "cost of treatment" and "years left of productive life".

Option 1 is the humane way, but slightly impossible. Options 2 and 3 involve treating people like dollar signs, one way or another. It's an unpleasant choice, not an avoidable one.