SoylentNews is people
SoylentNews
Google is training machines to predict when a patient will die
A woman with late-stage breast cancer went to a hospital, fluids flooding her lungs. She saw two doctors and got a radiology scan. The hospital's computers read her vital signs and estimated a 9.3% chance she would die during her stay.
Then came Google's turn. An new type of algorithm created by the company read up on the woman — 175,639 data points — and rendered its assessment of her death risk: 19.9%. She died in a matter of days.
The harrowing account of the unidentified woman's death was published by Google in May in research highlighting the healthcare potential of neural networks, a form of artificial intelligence software that is particularly good at using data to automatically learn and improve. Google had created a tool that could forecast a host of patient outcomes, including how long people may stay in hospitals, their odds of readmission and chances they will soon die.
What impressed medical experts most was Google's ability to sift through data previously out of reach: notes buried in PDFs or scribbled on old charts. The neural net gobbled up all this unruly information then spat out predictions. And it did so far faster and more accurately than existing techniques. Google's system even showed which records led it to conclusions.
Scalable and accurate deep learning with electronic health records (open, DOI: 10.1038/s41746-018-0029-1) (DX)
(Score: 3, Insightful) by c0lo on Tuesday June 19 2018, @07:23AM (5 children)
So, both algoes estimated the chances of the woman to leave the hospital alive (that is not die in the hospital) to over 80% - that is, 4 chances out of 5. And that woman actually died.
I know a single case doesn't make a statistically representative data set, so... why exactly should be impressed:
- because the woman died with 80% estimated chances to live (so there is a chance both algos are shit, with the non-Google one twice as shitty)? *or*
- the fact that the PR spin use a single data point to glorify a Google algo?
https://www.youtube.com/watch?v=aoFiw2jMy-0 https://soylentnews.org/~MichaelDavidCrawford
(Score: 0) by Anonymous Coward on Tuesday June 19 2018, @08:01AM (1 child)
In 2010 there were 35.1 million inpaitent stays in the U.S., with 715,000 deaths (source [cdc.gov]), so the chance of dying in an average visit was 2.25%. Hence both algorithms predicted an elevated chance of death for this patient.
Think of it as a weather forecast. If the forecast calls for a 20% chance of rain, do you bring an umbrella? Or only when it says 50% or higher?
(Score: 2) by takyon on Tuesday June 19 2018, @08:13AM
If you have some simple details about patients, then the predictions may seem a lot less impressive.
For example, the age of the patient. The older, the more likely they are to die. Has a history of heart failure: simple true/false with no additional details. If true, boost that death percentage. Oxygen saturation. If it is dipping below 90%, that is not good news for the patient. And of course, the condition the patient is being treated for. Some conditions are more dire than others.
Google does it with 175,639 data points. Can it be done almost as well with just 5 or 10? Well, the study is open access.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 2) by PiMuNu on Tuesday June 19 2018, @10:22AM (2 children)
> why exactly should be impressed:
Because... "AI" ... and Google.
(Score: 2) by c0lo on Tuesday June 19 2018, @10:36AM
~* shudders *~
https://www.youtube.com/watch?v=aoFiw2jMy-0 https://soylentnews.org/~MichaelDavidCrawford
(Score: 2) by Gaaark on Tuesday June 19 2018, @11:58AM
"> why exactly should be
impressedscared:Because... "AI" ... and Google."
FTFY.
You welcome (I'm scared too)
--- Please remind me if I haven't been civil to you: I'm channeling MDC. ---Gaaark 2.0 ---