Discriminating Myalgic Encephalomyelitis/Chronic Fatigue Syndrome and comorbid conditions using metabolomics in UK Biobank, 2024, Huang et al

[MelbME]

But the AI doesn’t «understand» anything. You’re really not addressing the issues raised with regards to XAI (explainable AI), you just keep saying you don’t think they are important.

Which wouldn’t be the case for AI because as AI doesn’t «understand» anything. I feel like we’re going in circles..

No the AI at this point doesn't understand. It identifies a pattern. I'm pointing out that people use things that work and end up trusting in the results with their own eyes. Some feel better knowing that somewhere someone understands the biology but that's replaceable with AI pattern recognition for vast majority.

Our understanding of biology is basically an evolving theory that is a simplification of reality. It's different to AI but our health systems are all mostly built on outcome.

 

[Utsikt]

I think very few clinicians actually understand the biology of all that they recommend. It's more a case of, in my experience giving this treatment helps people. There's plenty of things that just work without us knowing all the details.

You’re comparing treatments and diagnostic tests, which are not comparable because for treatments we have evidence that they work through controlled trials.

If you’re handing out a diagnosis, you have an ethical and legal responsibility to be able to explain why you chose that one over another, so that the logic behind the decision can be controlled by others and that you/the healthcare authorities can be held responsible for mistakes if they were very clearly wrong based on the available evidence.

If your answer is «the AI said so», how can that be scrutinised?

No the AI at this point doesn't understand. It identifies a pattern.

So your previous point doesn’t hold up..

 

[MelbME]

You’re comparing treatments and diagnostic tests, which are not comparable because for treatments we have evidence that they work through controlled trials.

If you’re handing out a diagnosis, you have an ethical and legal responsibility to be able to explain why you chose that one over another, so that the logic behind the decision can be controlled by others and that you/the healthcare authorities can be held responsible for mistakes if they were very clearly wrong based on the available evidence.

If your answer is «the AI said so», how can that be scrutinised?

So your previous point doesn’t hold up..

That's what validation on larger and varied cohorts are for, to prove an outcome is reliable that matches to other alternatives of diagnosis.

I never said AI understands. I also added more text to previous reply. I pressed post reply accidentally.

 

[Utsikt]

In our paper we did a sensitivity analysis to rule out the impact. We took the low-comorbidity mecfs group and it reduced signals but many of the lipid signals remained significant.

Is this the relevant section?

Therefore, we created another cohort with 354 ME/CFS individuals with or without hypertension, depression, asthma, IBS, hay fever, hypothyroidism, or migraine and performed association tests (Supplementary Fig. 7) and sensitivity analysis for this subset (Supplementary Data 9).

Thirty-one of the initial 168 ME/CFS biomarker associations remained significant (P < 2.0110−4). SFA% and omega-3 were the only significant associations that produced greater odds ratio in the subset than the full cohort.

The lower odds ratios observed may be attributed to the reduced number of comorbid conditions reported by each individual, rather than the specific condition.

The average number of comorbid conditions was 3.0 for the full cohort and 0.6 for the subset. This suggests that the burden of having several comorbid conditions might exacerbate ME/CFS symptoms (inclusive of symptoms from common comorbid conditions), reflecting a higher disease severity, leading to more pronounced biomarker signals in the full cohort.

 

[Trish]

It's a complex issue in the field because comorbidity is part of mecfs but the accumulating comorbidities have their own impact on biology. No study ever produces a control group that can counter that properly. Probably NIH got the closest and that got a ton of backlash because people rightly highlight that some of that comorbidity may actually be the disease. I've heard some argue that no comorbidities likely means they weren't ME.

People say all sorts of things. That doens't make them true. The ICC criteria require all sorts of things beyond the generally agreed core symptoms of chronic disabling fatigue and PEM, with some other common sympoms such as pain, OI and cognitive dysfunction.

A problem with assuming some comorbidities are part of the disease is that the timing often doesn''t work. For example I have several of the comorbidities you selected as common ones to include in the study, but they were all present for me decades before my ME started. I expect that's true for many pwME. I have never heard of most of the comorbidities you researched having any concurrence in onset with ME/CFS onset or being included in diagnostic criteria.

 

[forestglip]

In our paper we did a sensitivity analysis to rule out the impact. We took the low-comorbidity mecfs group and it reduced signals but many of the lipid signals remained significant.

From the sensitivty analysis paragraph:

We recognise that the other 265 comorbid conditions not analysed in this study may influence the biomarker associations. Therefore, we created another cohort with 354 ME/CFS individuals with or without hypertension, depression, asthma, IBS, hay fever, hypothyroidism, or migraine and performed association tests (Supplementary Fig. 7) and sensitivity analysis for this subset.

Is this saying an ME/CFS group was created which excluded only those 7 conditions, and compared that group to C2? So could the ME/CFS group still have included individuals with type 2 diabetes or other conditions?

I guess the concern could be put simply, as I stated above: if the study was redone with other conditions, such as migraine or asthma, replacing the ME/CFS group as the "heterogenous" condition, would we see increased triglycerides, glucose, cholesterol, etc, in these conditions? If most conditions studied this way would show similar findings, it's hard to imagine these lipid findings are showing something specific to ME/CFS. It would make sense to me that this is what would happen if, for example, migraine cases could also have type 2 diabetes, but the other groups could not.

 

[MelbME]

Is this the relevant section?

Yep exactly. From memory the general population has 1 to 2 disease diagnoses and me/cfs cohort has 6 or more. That's on average.

 

[Utsikt]

I would say that many people (not all) often don't know why a decision is made but still trust it is medically sound if there is at least one human somewhere that understands it. In this case, instead of it being that one human that understands it, it's AI.

No the AI at this point doesn't understand. It identifies a pattern. I'm pointing out that people use things that work and end up trusting in the results with their own eyes. Some feel better knowing that somewhere someone understands the biology but that's replaceable with AI pattern recognition for vast majority.

It’s irrelevant what people «feel». And you’re dodging the issue of declining trust in AI despite increased use and capabilities.

Our understanding of biology is basically an evolving theory that is a simplification of reality. It's different to AI but our health systems are all mostly built on outcome.

As multiple people have been pointing out, the system also requires justification. If the justification is «this trial demonstrated that X works for Y, you have Y, so I’m giving you X», that’s fine.

If it’s «the AI said it», it’s not.

That's what validation on larger and varied cohorts are for, to prove an outcome is reliable that matches to other alternatives of diagnosis.

Then we’re back to the «why» because it can’t get better accuracy than just applying the criteria, and you’d still need to do the other tests to rule out alternative diagnoses, because this test wouldn’t demonstrate that you only have ME/CFS and nothing else.

I never said AI understands. I also added more text to previous reply. I pressed post reply accidentally.

See the first quote. I guess it’s ambiguous, I read it as «it’s AI (that understands it)». But agree that AI doesn’t understand so at least that’s cleared up.

 

[MelbME]

It’s irrelevant what people «feel». And you’re dodging the issue of declining trust in AI despite increased use and capabilities.

As multiple people have been pointing out, the system also requires justification. If the justification is «this trial demonstrated that X works for Y, you have Y, so I’m giving you X», that’s fine.

If it’s «the AI said it», it’s not.

Then we’re back to the «why» because it can’t get better accuracy than just applying the criteria, and you’d still need to do the other tests to rule out alternative diagnoses, because this test wouldn’t demonstrate that you only have ME/CFS and nothing else.

See the first quote. I guess it’s ambiguous, I read it as «it’s AI (that understands it)». But agree that AI doesn’t understand so at least that’s cleared up.

Right I see the confusion. That first quote was meant to mean, in some cases it's a person who understands and instead of that person it's AI (who doesn't understand). I was more trying to make a point of the network of trust built on the trust that a few people actually understand what is happening (we don't know they understand, we trust they do).

The validation process is: we have clinicians diagnose these patients as mecfs and these others as not having mecfs, the algorithm correctly gets to the same answer with a certain accuracy repeatedly. People then trust the algorithm as a substitute for the clinician diagnosis.

That's outcome driven. It's the same way that we don't understand the biology of how some drugs work but they work in large trials and we use them.

 

[MelbME]

People say all sorts of things. That doens't make them true. The ICC criteria require all sorts of things beyond the generally agreed core symptoms of chronic disabling fatigue and PEM, with some other common sympoms such as pain, OI and cognitive dysfunction.

A problem with assuming some comorbidities are part of the disease is that the timing often doesn''t work. For example I have several of the comorbidities you selected as common ones to include in the study, but they were all present for me decades before my ME started. I expect that's true for many pwME. I have never heard of most of the comorbidities you researched having any concurrence in onset with ME/CFS onset or being included in diagnostic criteria.

No I know. Just saying you get a lot of conflicting information on this topic. Which comorbidities would you allow?

We selected most common comorbidities in the UK Biobank cohort of ME/CFS.

 

[MelbME]

From the sensitivty analysis paragraph:

Is this saying an ME/CFS group was created which excluded only those 7 conditions, and compared that group to C2? So could the ME/CFS group still have included individuals with type 2 diabetes or other conditions?

I guess the concern could be put simply, as I stated above: if the study was redone with other conditions, such as migraine or asthma, replacing the ME/CFS group as the "heterogenous" condition, would we see increased triglycerides, glucose, cholesterol, etc, in these conditions? If most conditions studied this way would show similar findings, it's hard to imagine these lipid findings are showing something specific to ME/CFS. It would make sense to me that this is what would happen if, for example, migraine cases could also have type 2 diabetes, but the other groups could not.

Yes that's what it's saying. We reduced the shared comorbidities down as far as we could before the number of patients were dramatically impacting the power. We'd have liked to go to 0 but then we wouldn't have enough data, so we did sensitivity analysis and pleiotropy analysis.


No not unless altered blood lipids were part of the condition of the patient then you would not see these results based on how we did this analysis. Whether lipids are part of the mechanism of the disease remains to be seen, deconditioning could explain the result simply because severity is higher and people are less active because of it in the patient cohort. We were sensitive to not harp on about that in the manuscript.

I think to get even more accuracy we would need to identify control pair matches that matches every comorbidity of every patient and remove the patients that have no close match. It would be an exploratory paper and not one that aimed towards a differential diagnostic like this one. But it's something we have considered doing with a couple new PhD students in our lab.

 

[forestglip]

No not unless altered blood lipids were part of the condition of the patient then you would not see these results based on how we did this analysis.

I guess I don't see why not. Maybe I'm not understanding something about the methods.

There will probably be a portion of migraine sufferers, or asthma sufferers, who have type 2 diabetes or any of dozens of other conditions, just due to some people having more than one condition due to random chance.

If this group of migraine sufferers is compared to a group of, say, depression sufferers in which all of the individuals with type 2 diabetes and any other conditions were excluded, would you agree that we would expect to see a higher proportion with type 2 diabetes in the migraine group, basically by definition, with this not necessarily indicating anything about migraines? If this is the case, then we may also see differences between the groups that are related to diabetes and unrelated to migraine, like maybe high glucose.

 

[MelbME]

I guess I don't see why not. Maybe I'm not understanding something about the methods.

There will probably be a portion of migraine sufferers, or asthma sufferers, who have type 2 diabetes or any of dozens of other conditions, just due to some people having more than one condition due to random chance.

If this group of migraine sufferers is compared to a group of, say, depression sufferers in which all of the individuals with type 2 diabetes and any other conditions were excluded, would you agree that we would expect to see a higher proportion with type 2 diabetes in the migraine group, basically by definition, with this not necessarily indicating anything about migraines? If this is the case, then we may also see differences between the groups that are related to diabetes and unrelated to migraine, like maybe high glucose.

Yes you would have more diabetes but less than 5%, it generate noise but not a strong signal because the other more than 95% don't have diabetes. Multimorbidity is a feature of me/CFS, you don't see the level of multimorbidity in any of these other conditions. That's why I highlight that while the lipid profiles look to be part of the condition from this analysis, doesn't mean it's mechanism.

If you are interested, I think the only ME/CFS paper that has ever matched all other comorbidities between ME and controls is the NIH paper (of course multi-morbidity is removed which could be a feature). We could attempt it in UK biobank but we didn't in this paper. This paper was about trying to differentiate a signature of ME vs 7 common comorbid conditions.

I think the multimorbidity discussion is an important one. I think good hypotheses account for it. An example would be the Edwards, Cambridge, Cliff hypothesis. That mechanism would result in presentation of a multi-morbidity disease given how we diagnose conditions. Low-grade macrophage activation could explain the lipid and inflammation markers we saw in this study.

 

[Hutan]

That is the strength of our paper and why it is well regarded. I'd say less than 5 papers in all of ME/CFS research have actually tried to tackle the comorbidity you talk of. Probably this paper and a couple of Pontings papers and I guess the NIH paper. I think for this reason you would find all of these papers in a top 10 ranking of publications in ME/CFS if you asked your favourite AI.

There have been many well ranked papers that actually aren't accurate. Saying 'everyone important thinks the paper is great' doesn't cut much ice here, because we know that there are papers and ideas that many important people have backed that are nonsense.

In our paper we did a sensitivity analysis to rule out the impact. We took the low-comorbidity mecfs group and it reduced signals but many of the lipid signals remained significant. This is all in the paper that Hutan has read through yet they keep attacking our work. I've discuss in PMs as well. It doesn't seem like Hutan is actually listening to what I'm saying but I wish they would. The things they are trying to attack this paper about are actually the strengths it has against just about every other paper in the field. I'm only not replying because every time I try explain they come back more aggressively, now calling for retractions. This has been going on for awhile and not specific to this paper.

I'm sad that you characterise my engagement with your papers in that way. I have tried to explain the problem I see, in multiple ways and giving detail, and you have not yet given adequate explanations. The questions I have raised deserve answers better than appeals to authority and suggestions that the person raising the questions is needlessly attacking. I think it is reasonable, after the effort I have made to get the answers and then not getting them, to begin to assume that there really is a problem.

I think it really matters whether the claims in this paper are true, and whether they are widely seen as true. Indeed, if the paper is well regarded and highly ranked but is actually wrong, then the regard and ranking make the problem bigger. So, I think it is worth trying to continue to see if my concern is well-founded. That is what I am trying to do here.

The prevalence of obesity in the UK Biobank population is around 25%. The prevalence of hypothyroidism in the UK Biobank population is around 4%. Obesity is a very common health condition in the UK biobank population, hypothyroidism is much less common. Obesity is strongly associated with perturbations of blood lipids. In this paper, there is a hypothyroidism cohort, but no obesity cohort. In fact, as far as I can see, people who are obese are actively excluded from all of the cohorts except for the ME/CFS cohort.


So

A. How do you know the differences in blood lipids you identified are not related to things like obesity and Type 2 diabetes rather than ME/CFS?
1. What are the percentages of people with obesity and Type 2 diabetes in the ME/CFS and in the C2 and the 7 disease cohorts used in this paper?
2. Do these percentages reflect the actual percentages of people with obesity and Type 2 diabetes in the whole Biobank population and in the Biobank populations with each disease (along with any number of comorbidities)?
3. How do you know that the differences in blood lipids shown in Figure 1 (between ME/CFS and the C2 No Diseases cohort) are not due to differences in the prevalence of obesity and Type 2 diabetes, given that the ME/CFS group members were allowed to have obesity and Type 2 diabetes, but the C2 No disease members could not?



I have a new question, about how the adjustment for the cholesterol-lowering medication was made.

Biomarker associations and multiple testing correction​

Logistics regression was used to estimate the odds ratio for biomarker associations with ME/CFS and comorbid cohorts against the C2 cohort. Odds ratios were adjusted for sex, age, cholesterol-lowering medication and fish oil supplements.

We know that the rate of use of cholesterol-lowering medication in people in the ME/CFS cohort is normal for the age group (around 15.7%). We also know that the use of cholesterol-lowering medication is almost nothing in the various disease cohorts in this paper (generally less than 1%), because the members of those cohorts were highly selected and could not have any co-morbidities. So, the situation is that there's a significant percentage of people using cholesterol-lowering medication in the ME/CFS group, and almost no one using cholesterol-lowering medication in the disease groups (and the C2 group).

Therefore, how the adjustment for use of cholesterol-lowering medication was done could have a significant impact on the assumed blood lipid levels. I assume that the levels of cholesterol-related lipids in people taking cholesterol-lowering medications were increased?

B. How were the odds ratios adjusted for cholesterol-lowering medication use?
What lipids levels were increased? What percentage of people had their lipid levels increased in each cohort?

(To be clear, the act of adjusting lipid levels due to medication use is not necessarily wrong. It's just that, if groups have been selected so as to mostly exclude people who qualify for the adjustment, that biases the odds ratios.)

edited to fix the use of cholesterol-lowering medication percentages

 

[MelbME]

There have been many well ranked papers that actually aren't accurate. Saying 'everyone important thinks the paper is great' doesn't cut much ice here, because we know that there are papers and ideas that many important people have backed that are nonsense.

I'm sad that you characterise my engagement with your papers in that way. I have tried to explain the problem I see, in multiple ways and giving detail, and you have not yet given adequate explanations. The questions I have raised deserve answers better than appeals to authority and suggestions that the person raising the questions is needlessly attacking. I think it is reasonable, after the effort I have made to get the answers and then not getting them, to begin to assume that there really is a problem.

I think it really matters whether the claims in this paper are true, and whether they are widely seen as true. Indeed, if the paper is well regarded and highly ranked but is actually wrong, then the regard and ranking make the problem bigger. So, I think it is worth trying to continue to see if my concern is well-founded. That is what I am trying to do here.

The prevalence of obesity in the UK Biobank population is around 25%. The prevalence of hypothyroidism in the UK Biobank population is around 4%. Obesity is a very common health condition in the UK biobank population, hypothyroidism is much less common. Obesity is strongly associated with perturbations of blood lipids. In this paper, there is a hypothyroidism cohort, but no obesity cohort. In fact, as far as I can see, people who are obese are actively excluded from all of the cohorts except for the ME/CFS cohort.


So

A. How do you know the differences in blood lipids you identified are not related to things like obesity and Type 2 diabetes rather than ME/CFS?
1. What are the percentages of people with obesity and Type 2 diabetes in the ME/CFS and in the C2 and the 7 disease cohorts used in this paper?
2. Do these percentages reflect the actual percentages of people with obesity and Type 2 diabetes in the whole Biobank population and in the Biobank populations with each disease (along with any number of comorbidities)?
3. How do you know that the differences in blood lipids shown in Figure 1 (between ME/CFS and the C2 No Diseases cohort) are not due to differences in the prevalence of obesity and Type 2 diabetes, given that the ME/CFS group members were allowed to have obesity and Type 2 diabetes, but the C2 No disease members could not?



I have a new question, about how the adjustment for the cholesterol-lowering medication was made.


We know that the rate of use of cholesterol-lowering medication in people in the ME/CFS cohort completely normal (around 10%), if not slightly lower than normal, compared to the whole UK Biobank population (12%). We also know that the use of cholesterol-lowering medication is almost nothing in the various disease cohorts in this paper (generally less than 1%), because the members of those cohorts were highly selected and could not have any co-morbidities. So, the situation is that there's a significant percentage of people using cholesterol-lowering medication in the ME/CFS group, and almost no one using cholesterol-lowering medication in the disease groups (and the C2 group).

Therefore, how the adjustment for use of cholesterol-lowering medication was done could have a significant impact on the assumed blood lipid levels. I assume that the levels of cholesterol-related lipids in people taking cholesterol-lowering medications were increased?

B. How were the odds ratios adjusted for cholesterol-lowering medication use?
What lipids levels were increased? What percentage of people had their lipid levels increased in each cohort?

(To be clear, the act of adjusting lipid levels due to medication use is not necessarily wrong. It's just that, if groups have been selected so as to mostly exclude people who qualify for the adjustment, that biases the odds ratios.)

Simply put. If you care about comorbidity control then in every thread with every article, why aren't you questioning controls? None of them control for comorbidities.

Why aren't you asking every paper to be retracted? I think if you think about this question then you will probably understand why I feel a strong bias in my direction.id love to be wrong but I constantly deal in probability and I notice outliers.

We show the BMI numbers in one of the tables.

Cholesterol-lowering medication use was a binary covariate (yes/no) in the model. The model estimates the odds ratio for each biomarker independent of whether or not the person was taking a cholesterol-lowering drug. It does not change any individual person’s measured biomarker values. It simply removes the confounding effect of the medication from the group-level association.

 

[Hutan]

TLDR: Simply put, you actively selected your controls to be different from the disease group on aspects unrelated to the disease labels but highly material to the question you were asking.


The problem is that you are claiming that this study tells us that blood lipids in ME/CFS are dysregulated. But, you have excluded most of the people in the UK Biobank who are likely to have unhealthy blood lipid levels from the comparator groups. The only group that was allowed to have people with those conditions was the ME/CFS group. Actually, the blood lipid levels in people with ME/CFS might be very normal. We know the prevalence of use of cholesterol lowering medications is normal in the UK biobank ME/CFS group.

You haven't controlled for the conditions that are most likely to explain unhealthy blood lipids levels. In fact, it's worse, you actively selected your comparator groups to exclude people with the conditions most likely to have unhealthy blood lipid levels. So, the Hypothyroidism group excludes people who are obese and/or have Type 2 Diabetes. The ME/CFS group does not. You are not comparing like for like. The problem is in comparing a heterogeneous cohort with homogeneous ones.

All the study tells us is that people with ME/CFS have, on average, less healthy blood lipid levels than groups of people specifically selected to be likely to have healthier blood lipid levels than the whole population with the disease label.

It's like the M&M example I gave above. Two jars of M&Ms, that might have different percentages of various colours. You want to find out if the percentages of colours are in fact different. Then you take out nearly all the yellow ones out of one jar. And then you say 'look, this other jar has more yellow M&Ms!'.

Your team essentially took out nearly all of the people with co-morbidities that can reasonably affect blood lipid levels out of comparison jars, but not out of the ME/CFS jar. You actively selected comparator groups to make them less like the disease group on the issue that was being investigated.

 

[Eddie]

Simply put. If you care about comorbidity control then in every thread with every article, why aren't you questioning controls? None of them control for comorbidities.

Why aren't you asking every paper to be retracted?

If it is true that that most studies do not control for comorbidities then that might explain why so many results can not be replicated. If for example a handful of pwME in the intermural study had also had diabetes; that would have been major issue and would likely have produced some erroneous results. If comorbid conditions can not be excluded entirely, then at least matching the number of comorbid between the groups seems almost necessary to make sure the difference is not a result of the comorbid disease.

If we knew what causes ME/CFS then this would not be such an issue. But since we are searching for clues anywhere, then we have to particularly careful to ensure that differences are related to ME not some other cause like another condition.

 

[Eddie]

I would also guess that studies that collect a large number of different data points are also much more vulnerable to differences in comorbidities between groups. It is unlikely that the comorbidities impact on any one measure but they probably have some impact somewhere. So I think its is fair to be stricter with studies that are trawling for interesting results.

 

[Sean]

1. Not up to following this debate. That said:

2. Comorbidities do have to be factored in somehow. Almost all humans have them, and getting a clean comorbidity-free sample from any group is difficult at the best of times, let alone from a group who have been very sick for many years, some for decades.

I didn't have comorbidities at the time of onset, or for many years after (that I know of), but I do have some now. Not surprising at my age (early 60s), and after more than 4 decades of moderate-severe ME/CFS.

3. This is a difficult but important debate and I appreciate the main players continuing to engage. We need to keep in mind that we all ultimately have the same goal of understanding what is going on and how to treat it.

 

[Hutan]

Why aren't you asking every paper to be retracted? I think if you think about this question then you will probably understand why I feel a strong bias in my direction.id love to be wrong but I constantly deal in probability and I notice outliers.

Not every paper is wrong, not every paper is likely to materially misdirect research funding if it is wrong. Of course I'm not asking every paper to be retracted.

I want your team to be doing brilliant work that finds the answer to ME/CFS, so that I and others with ME/CFS and the people yet to get ME/CFS can be in a world where there is more understanding of our illness. So, I don't want you and others wasting time on research questions that are not well founded. To be honest, I'm well over this, I'm tired. This is not fun and I have many other things I'd rather do. But, if I'm right, then this paper is misleading in a substantial way and will misdirect research.

Chris, it might be most comfortable to assume that I am singling your team out for criticism on some basis other than the quality of the paper and its potential to impact on the success of future research. I assure you, I am not. I think it's a shame if the focus turns from the facts of the discussion to the nature of the person challenging the findings.

We show the BMI numbers in one of the tables.

The table gives the median BMIs. You don't tell us the percentage of people who are obese in each cohort. The ME/CFS group has the highest median BMI of all the groups, with the median well into the overweight range. It would be useful to know the percentages of people with obesity in each of the groups in this study; I think this is a key piece of information that should have been presented given its relevance to blood lipids. Obesity is a health condition, and so I assume the homogeneous cohorts (i.e. cohorts where people could only have one health condition) excluded people who are obese.

It would be good to know either way, if people who are obese were allowed to be part of the disease and C2 cohorts. Chris?

Cholesterol-lowering medication use was a binary covariate (yes/no) in the model. The model estimates the odds ratio for each biomarker independent of whether or not the person was taking a cholesterol-lowering drug. It does not change any individual person’s measured biomarker values. It simply removes the confounding effect of the medication from the group-level association.

The paper says that odds ratios were adjusted for cholesterol-lowering medication use. With 15.7% of the ME/CFS group using these medicines, but e.g. only 0.5% of the people in the IBS group using the medicines, only 0.6% of the people in the migraine group using the medicines and even in the hypertension group, only 9.7% using the medicines, it seems likely that the adjustment had a material effect on the odds ratios for the biomarkers for each group. Your answer doesn't tell me anything about the effect of the adjustment.

Those cholesterol-raising medication use rates in the disease groups are surely not indicative of the rates of use in the UK Biobank populations who have those diagnoses. The use of these medicines in the ME/CFS group is about the same as the rate of use by the whole Biobank population.


The key question surely is 'are the blood lipids of people with ME/CFS, on average, different to people of the same age and sex?'. I've gone back to read the abstract of the paper, and found it very interesting that it actually makes no claims about blood lipids being different. Perhaps the authors or the peer reviewers realised that the paper could not actually provide any evidence that blood lipids levels characterise ME/CFS with the comparison groups it used?

Nevertheless, the study did aim to find differences:

Therefore, by using a large heterogenous ME/CFS cohort and various homogenous negative and positive control groups based on common comorbidities of ME/CFS, we sought to (1) identify discriminatory and shared blood metabolomic biomarkers for ME/CFS and comorbid conditions, (2) distinguish ME/CFS and individuals with overlapping comorbid conditions using machine learning and (3) characterise the altered biological pathways that underlie the ME/CFS nuclear magnetic resonance (NMR) metabolomics profile.

And the Discussion claims to have found differences:

This metabolomics analysis presents a lipoprotein profile for ME/CFS, highlighting significant associations of the disease with VLDL subclasses and size. These findings pinpoint a triglyceride and cholesterol transport problem, potentially arising from enzyme dysregulation, such as lipoprotein lipase (LPL).

Again, we can't know if the identified differences are due to ME/CFS or the fact that the ME/CFS group included comorbidities known to impact on blood lipid profiles while the comparison groups did not. I have not looked in detail, but I think its very possible that the reported differences could be explained by the presence of people with obesity and Type 2 diabetes and other conditions in the ME/CFS group, while those people were actively excluded from the comparison groups.

I think I've set out the problems I see as well as I can and am starting to repeat myself. From the information currently available, I don't think we can rely on the findings of this paper. I think it builds an edifice of clever, detailed computations on the foundation of a flawed study design.