New paper using AI Model to Identify CFS

[BrightCandle]

I still think there is value in a diagnostic test that does not reach biomarker standards because it can be used clinically for diagnosis and as an input into research for finding patients. These are very much still needed for prevalence and recognition which will lead to proper funding.

 

[hapl808]

Yeah, not going to respond in detail because I think we wholeheartedly disagree on almost every point. There's just no such thing as a 'coin flip' meaningless marker with 99% predictive value and perfect experimental rigor. Not really how coins work.

As for the study (which I guess is the point), I touched lightly on my problems with it at first glance. They compare ME/CFS to healthy controls only. I see this a lot. Besides a very small sample size (50 patients), you really need a cohort of people with non-ME/CFS illnesses so you don't find you're classifying just the difference between sick and healthy. Ideally you'd need a significant sample size of healthy and some samples of different diseases - neurological, autoimmune, etc.

It seems they're trying to avoid overfitting, but when I see a 99% AUC with such a small sample set and a poorly chosen control, it seems very likely that it's still overfitting the data. The testing of 768 metabolites is quite interesting and the data aren't meaningless - just that I think the cohort design is a much bigger problem than the AI methodology (which at a quick glance seems pretty solid). I have no way to know if they are identifying metabolites of illness, or metabolites of ME/CFS.

That said, I think it's only appropriate to cite ChatGPT about the specific metabolites highlighted.

Insights into Disease Mechanisms​

1. Metabolic Dysfunction: The involvement of these specific metabolites suggests that metabolic dysfunction might play a role in ME/CFS. Each metabolite is involved in different biochemical pathways, indicating potentially complex and multifaceted metabolic irregularities in ME/CFS patients.
   • Oleoylcholine: This compound could indicate alterations in lipid metabolism or signaling.
   • Cortisone: As a glucocorticoid, cortisone is involved in stress response and immune function, hinting at potential dysregulation in these systems.
   • 3-Hydroxydecanoate: This is a medium-chain fatty acid, potentially pointing to abnormalities in fatty acid metabolism.
   • C-glycosyltryptophan: This is a modified amino acid, which could indicate changes in protein synthesis or degradation, or alterations in gut microbiota (as gut bacteria can modify tryptophan).
2. Interconnected Pathways: These metabolites could be part of interconnected metabolic pathways that are disrupted in ME/CFS, suggesting a complex interplay of factors rather than a single causative agent.
3. Potential Therapeutic Targets: Understanding the roles these metabolites play in ME/CFS could lead to targeted therapies that address these specific metabolic dysfunctions.

Limitations and Further Research​

1. Correlation vs. Causation: While these metabolites are correlated with ME/CFS, it's not clear whether they are a cause or an effect of the disease.
2. Individual Variability: ME/CFS is a heterogeneous condition, and these metabolites might not be relevant for all patients.
3. Need for More Research: Further studies are required to understand the exact roles of these metabolites in ME/CFS and how they interact with other factors like genetics, environment, and immune function.

 

[hapl808]

In a lot of respects, it's a well done study. I really wish they had chosen even 26 healthy, 26 ME/CFS, and 26 mix of various diseases. Even with such a small sample size, that would be a huge improvement.

If the correlations held fast, that would imply much more robust information about ME/CFS. If that weakened all the correlations, then that might imply that it was classifying healthy vs diseased states, not a specific illness.

 

[Forummember9922]

The study doesn't really have to be discussed in detail, because we can all easily agree upon it not being a meaningful result.

Well I am grateful we can all have separate views; I believe that it is moreso a step forward than meaningless and that you do not speak for everyone.

And no you don’t have to provide evidence for your arguments, even when you are 100% right, but it may have the undesired effect of watering down your argument to people who do not know what you know.

 

[Forummember9922]

FWIW Key figures from earlier study (2020) that also observed metabolite abnormalities in 'steroids' as well as cholines

Figure 1. Box plot distribution of logged values for the metabolites that are part of the acyl choline pathway. Controls (CTRL) are shown in red and patients (myalgic encephalomyelitis/chronic fatigue syndrome (ME/CFS)) in blue

Figure 3. Display of the fold change (controls/patients) of the median, averaged for each of the 94 sub-pathways. Roman numbers at the bottom of the figure are assigned as follow for each of the nine super-pathways: I = amino acids, II = carbohydrates, III = cofactors and vitamins, IV = energy, V = lipids, VI = nucleotides, VII = partially characterized molecules, VIII = peptides, and IX = xenobiotics. Labelled sub-pathways are discussed in the manuscript; brown ones are over-abundant in controls compared to patients while purple ones are the opposite. The number associated with each sub-pathway reflects the number of metabolites included.

 

[hapl808]

Here's an example of AI being used not just to find spurious correlations, but to investigate potential new drugs and then try to unravel the basis. I think AI is much more explainable in many forms than people realize - the black box neural net can be opened up and examined, plus various boosting models and other architectures that allow one to examine correlations, etc.

https://www.scientificamerican.com/article/new-class-of-antibiotics-discovered-using-ai/

 

[Osaca]

Here's an example of AI being used not just to find spurious correlations, but to investigate potential new drugs and then try to unravel the basis. I think AI is much more explainable in many forms than people realize - the black box neural net can be opened up and examined, plus various boosting models and other architectures that allow one to examine correlations, etc.

https://www.scientificamerican.com/article/new-class-of-antibiotics-discovered-using-ai/

Yes, this has for quite some time already been one of the extremely useful applications of AI in medicine. My debate was never about AI not being extremely useful in medicine, even if one only looks at image processing one can see that it is the case, my debate was about what we've previously discussed.