Found the study about machine learning predicting cardiovascular risk

(Bob M) #1

This is from Malcolm Kendrick’s book, The Clot Thickens. It’s the last study he uses.

They applied various machine learning techniques, and the best one was a neural network:

image
image.png734×273 17.4 KB

The top 10 risk factors are in Table 3, which I show in part below:

image
image.png778×566 48 KB

To get all the rankings, you have to go here:

https://pmc.ncbi.nlm.nih.gov/articles/instance/5380334/bin/pone.0174944.s001.docx

image
image.png765×417 26.4 KB

LDL is way down on the list. For the neural network, it was 46th. BMI was 43rd.

Also, I think the #1 risk factor in the “ML: Neural Networks” is wrong. It’s not “atrial fibrillation” but is instead “COPD” according to the downloaded Docx document.

Edit: One of the factors that is bad in every system is…age. Getting old apparently isn’t good. :wink:

(Central Florida Bob ) #2

Life is an invariably fatal, sexually transmitted disease.

Some may prefer the word “condition” to disease.

(Edith) #3

I wonder if one of the reasons the ML/AI was better was because it probably didn’t have any preconceived ideas about what CVD indicators were the most important.

(Bob M) #4

I thinks that’s most likely. Also, the Acc/Aha algorithms only use a few variables, whereas the other machine-learning systems used 48 risk factors I believe.

It’s like Dr. Kendrick – he basically started trying to figure out what caused heart disease. Things like stress, pollution, smoking, etc., these are associated with very high levels of heart disease, independent of LDL. He wasn’t trapped looking only at LDL.

Edit: I also think this is why COPD is highest in the neural-network version: this probably means you’ve smoked a lot or been subjected to pollution, which means you have problems with your lungs and endothelium. This is what’s causing the heart disease, not LDL.

(Doug) #5

Bob, lots of great discussions lately. For whatever problems we have, it’s a cool time to be alive.

(Bob M) #6

Doug, it is cool, though everyone wants us to integrate AI into our work somehow. I haven’t found anything useful yet, but we’re still testing.

(Central Florida Bob ) #7

There’s a math blog, which means you have to be pretty pathetic to read it. Which means I’ll be the guy who points it out. The blog is called William M. Briggs - Statistician to the Stars. In my first programming class back in (I’m not even sure) 1973, there was a limerick printed out on the wall that burned into my brain. The long version was:
I really hate this damned machine
I wish that they would sell it.
It never does do what I want
Only what I tell it.

The tie in to Briggs is he always reminds us that all models and that includes AI only do what we tell them. For @ctviggen, I really liked this discussion of COPD, stress, pollution, smoking and how all of those seem to outweigh LDL. I personally think cholesterol (yes, and LDL) as an important cause died (as a model) probably a generation ago.

But then Briggs and that “Only what I tell it” tapped me on the shoulder.

(Edit to remove some dumb typos that some leprechaun inserted after I first posted)