What are the necessary conditions and criteria for a theoretical model of ME/CFS?

We have a paper submitted to a Nature journal that you will probably be interested in. We're hoping it gets published soon. We looked at common co-morbid conditions in an ME cohort and then compared ME to cohorts of that co-morbidity to see if we could find differences that had more to do with an ME signature than the co-morbidity. We also looked for a diagnostic signature that would separate ME from these comorbid diseases. We got one that was ~75% accurate. So we think there is potential here.

 

Thanks a lot for sharing! I'm very much looking forward to reading it once it is published.

With AI classifiers seemingly becoming the norm, are you aware of any work going into something along the lines of "instead of trying to distinguish as many LC or ME patients from HC as possible, even if the differences in actual measurement values might be very marginal and might require combinations of things that apparently aren't pathological connected (say:cortisol+EBV), we are looking at trying to find a diagnostic signature that yield relatively consistent values in HCs and can explain pathology but who's measurement values differ drastically in LC or ME from HC, but only does so for a subset of patients".

The danger of this approach is of course that one could be identifying differences attributable to co-morbidities rather than "true" ME or LC, but it still seems worthy exploring for me, especially if one then can then look at it in a follow-up study in a similar fashion to the one you submitted to a Nature journal.

In the past, I have been rather perplexed when even the most brilliant researchers have made claims surrounding their diagnostics, for example claims surrounding "cortisol being a biomarker for LC" when they should be well aware that AI classifiers cannot distinguish signal from noise but are simply very good at finding structures independently of whether those structure represent signal or noise and have been wondering whether there shouldn't be far more focus on the explainability power of a diagnostic signature rather than purely being focusing on its differentiating capacity and gobbling up those p-values that some statistican throws at you.

 

Well one does "sell" the impact of work to get published in a good paper from time to time and to help get future grants/funding, that's probably most of the reason. And maybe sometimes they're just not as experienced in the diagnostic space. In 2015 we did the first metabolomics profile of ME vs controls, decent numbers for the time. We could get 100% separation of groups using 6 metabolites. We did explore the idea of a biomarker but in that process quickly realized how far away we were (had no mechanism we were building off and hadn't validated it nor compared it to disease cohorts). I was mid-PhD but it was a good learning process to at least speak to people in the diagnostic area about how realistic the chances were.

POT (postural orthostatic tachycardia) occurs in healthy controls. POTS is when it is produces symptoms chronically and comes with an array of symptoms that aren't obviously related. So is that what you refer to as a pathology. I mean why not separate patients by comorbidities?

 

A prospective study would start with healthy people and test them every time they get sick, or at least a blood sample. No waiting. At even six months the key biochemical processes may be much less apparent. This should probably be a much larger study looking at a much larger range of diseases. No point in just tracking ME or LC, especially since funding might be easier.

 

From my understanding that is essentially what LIINC has done for LC.

It was done for EBV and MS and took somewhere around 40 years and more than 10 million participants (and yet one tiny wrong test result in that study completely changes the risk factor they had calculated). Allegedly those authors thought about ME at the same time but decided against it due to the necessary efforts and unknown results. With Covid known to cause ME/CFS as well you'd likely have to control for Covid infections as well and I think that is extremely hard.

It's just very hard to get sensible data on something that affects almost everyone (EBV & Covid) but where only a much smaller minority develop symptoms afterwards (ME/CFS), especially when you don't even know whether negative controls are actually negative controls.

It seems more sensible to me to just start off with those that are already affected. Possibly someone has a clever way around those problems but I haven't seen anyone suggest anything sensible.

 

Well I think characterising 500-1000 ME/CFS would give a great overview of the heterogeneity and it would be a target for developing severity markers.

Obviously decode targeted 20k+ people, this wouldn't be to that level but could be a start. Need more of these large-scale data projects.