IBM's AI Can Predict How Parkinson's Disease May Progress In Individuals (engadget.com) 7
An anonymous reader quotes a report from Engadget: [R]esearchers from IBM and Michael J. Fox Foundation (MJFF) say they've developed a program that can predict how the symptoms of a Parkinson's disease patient will progress in terms of both timing and severity. In The Lancet Digital Health journal, they claim the software could transform how doctors help patients manage their symptoms by allowing them to better predict how the disease will progress. The breakthrough wouldn't have been possible without the Parkinson's Progression Markers Initiative, a study the Michael J. Fox Foundation sponsored. IBM describes the dataset, which includes information on more than 1,400 individuals, as the "largest and most robust volume of longitudinal Parkinson's patient data to date" and says it allowed its AI model to map out complex symptom and progression patterns.
Well duh. (Score:2)
Anyone can predict how something may progress.
It can predict, but... (Score:2)
...it can't email you the results.
That's nice..... (Score:2)
I mean being able to predict what's going to happen with the disease might have an application I guess, but much more useful, I think, would be having tactics for mitigation, prevention, or ideally reversal.
In mythology Cassandra's prophecies were flawless, but they never made an iota of difference.
Thank-you Michael (Score:2)
Here, let me fix that for you (Score:2)
"Researchers from (...) developed a program that _might_ predict how the symptoms (...)"
To wit, I can write a program that "could" predict the outcome of a coin toss.
Re: (Score:2)
Just get alternate dimensions involved and your ability to write that program goes up!
https://gaming.stackexchange.c... [stackexchange.com]
Yo Grark
Is this actually useful? (Score:3)
"Median time to state 8 starting from state 5 was 225 years (95% CI 125–not reached). 464 (76%) of the 610 PDBP participants had consistent baseline state assignment whether using only cross-sectional data or all available patient data. Again, the most common discrepancy was between states 1 and 4 (37 inconsistent assignments). For the binary terminal state prediction, 496 (81%) of the 610 participants had consistent predictions."
Maybe this is better info than anyone had before, but it does not seem particularly useful to know that progression from state 5 to state 8 will take 2 to 25 years or that the model has inconsistent predictions 19% of the time for the binary terminal state. 1400 people as data does not seem like a lot either. We've learned that biology is individual and what works for the most number of people often does not work for individuals, so maybe this model will make better predictions (not necessarily better life for the patients but better predictions) for a majority of people but will fail all the rest of them. In general when I read about machine learning approaches to solving problems and they report things like "86% accuracy!" I always wonder about the other 14%. 90% accuracy means it is wrong one out ten times. The results here are also heavily weighted towards males (67%) and they say "because the majority of the cohort identify as White, our results might not be generalisable to the entire Parkinson's disease population."
A lot of medical studies seem to be underpowered, biased, and have large design flaws, yet people keep doing them. Then they say "this shows promise" but they rarely seem to follow up with much larger, unbiased, well designed studies. From the outside it looks like they are always taking the easiest, cheapest way to get publishing fame without actually advancing science much.
Maybe I'm missing the point of this study and they actually discovered something useful about the probability of transitioning between a few specific states that was unknown and that is what is useful here?