Thread by @POhukainen

Our latest paper fresh out in @ATHjournal! We asked: "If we used an #AI algorithm to group individuals based on similar blood profile of 44 #lipid measurements, could we improve coronary #heartdisease risk prediction?" Turns out: no.

Thread! 🚨 (1/13)

atherosclerosis-journal.com/article/S0021-…

We looked to see if we could detect subgroups in the population that would have a) similar blood profiles across 14 lipoprotein subclasses and b) possible differences in CHD risk. The algorithm we use is called a self-organising map (SOM)

See: academic.oup.com/ije/article-ab… (2/13)

SOM is a tool for multivariate subgrouping. Essentially the user puts in variables and the algorithm groups individuals based on similarity of these inputs. Here's a rough emoji-schematic of how this applies to population cohorts (3/13)

In practice, SOM is viewed through colourings that display the distribution of of each variable on the map. For example, if we had used "darkness of skin" as an input variable and viewed that variable, we would see a cluster of darker skinned people show up (4/13):

But a SOM can also reveal information that's not in the input. For example, hair colour is correlated with age so if hair colour was an input variable, we would see clustering of age, as well. We could then view the same map but just colour the "age" variable (5/13):

So how does this look like in our lipid data? Like this. We can see that for example, all non-HDL-lipids cluster and correlate with each other, like they're supposed to. But interestingly, we also see that there are roughly 4 subgroups, as indicated by lines on the plots (6/13):

If we look at the actual lipoprotein particle concentration profiles of these four subgroups, here's how they look (7/13):

Notice the red bar in histograms? It's apoB concentration. All the groups seem to have an ascending mean concentration of apoB, and since the SOM doesn't use artificial cut-off-values, the actual groups contain overlapping distributions of apoB (8/13):

So the next question then was obviously if this unsupervised lipid data-driven subgrouping would provide useful information about CHD risk.. And by useful, we mean would it provide better risk stratification than apoB, which is an established risk marker. Turns out: no! (9/13)

There's a lot of excitement and straight-up hype about AI-based risk prediction algorithms applied to #BigData. However, our data should be taken as caution that more data and machine learning may actually NOT be as awesome as some people think (10/13)

Clustering people based on an advanced lipid profile can be interesting as such but it doesn't seem to improve risk prediction. This makes sense because apoB is the main pathological component so adding information about other lipid metrics may even worsen risk resolution (11/13)

Of course, we were only looking at lipid-mediated risk and didn't do a deep dive on all other risk factors. Nor did we specifically look for the *best* lipid-predictor. Those things were beyond the scope of this study and subject for separate studies (12/13).

Huge thanks @mak_sysepi, @Sanna_Kuusisto, @MakinenGroup, @johanneskettune for the collaboration and mentoring! This is my first 1st author paper since moving to systems #epidemiology from the wet lab where I did my PhD. I've learned a LOT and now gotten up to good speed 😊🙏🏼