Discussion in 'BioMedical ME/CFS Research' started by Andy, Apr 2, 2019.
Paywalled at https://www.clinicaltherapeutics.com/article/S0149-2918(19)30112-2/fulltext
FYI - This paper is not paywalled now.
For everyone's benefit this is a paper describing the exercise testing that Klimas has been doing in ME/CFS, taking blood at 8 timepoints, and measuring hormones and cytokines, and how they fitted a computer model to it. This is the one she has raved about in her videos that requires supercomputers to run.
Some quotes from the paper describing what they did.
Here is what they measured
They extended their previous 2013/2014 model by using data generated from a machine learning approach. Don't know if you have seen this @mariovitali
And yes @mariovitali they used machine learning yet again
Then I get a bit lost (understatement of the year) so I wanted to look at the data they measured shown in Supplementary Figure A3. From what I can see there is huge overlap between patients and controls, and the mean differences are pretty small...........
They then did their magic and came up with 26 models to explain what is happening and tested two drugs "in silicon" models. You can read the details in the paper (watch that your brain doesn't explode trying to understand )
The question I have is, if you create a model that is selected/chosen using patient data (along with the literature), don't you need a second cohort of new data to test your model against to see that your model is real? Especially when there is so much overlap between healthy and patient experimental data and the mean differences between the data is pretty minimal. Otherwise, how can you trust the model?
EDIT: Posting an image of the experimental data didn't turn out so well. Here is a link to it
I will have to read this very carefully , it appears that they use a number of techniques which also includes elements of network analysis (betweeness score) which i also use in my methodology.
The standard way of assessing a machine learning model is to use three subsets of data : Train, Validation and Test set. The Train set is used to have the algorithm(s) "learn" about the problem and you get to tune the algorithm(s) by looking how it performs on the validation set. The actual performance of the classifier though takes place on the test set (recall that the algorithms are being trained and optimized using the train and validation sets so the validation set is not a representative performance evaluation since it is used for the optimization).
As discussed , this paper describes a very complex work but i must say it is interesting because i can confirm the -possible- involvement of IL-6 and TNF-α from the methodology i use. The -possible- involvement of HPA and HPG axis is also there and -interestingly- the tool suggests that cases of ME patients with hypothyroidism may be a cause of this HPA dysregulation. Of course this is a hypothesis and i do not know if this stands medically speaking.
EDIT for correction : The hypothesis suggests that HPA Dysregulation may be the cause of hypothyroidism in some ME patients, not the other way around as it is suggested above.
Here is an example run , executed just today. Observe how il-6 and tnf-α are being selected along with hypoperfusion, hypoxia and vasoconstriction. Note that these results have to do with the symptoms of ME/CFS and not the cause.
I do not know why angiotensin, albumin are selected as being important (the same for norepinephrine - not shown). ckd = chronic kidney disease. Glutamate -as discussed- is considered a key topic according to Machine Learning AND Network Analysis.
BDNF was recently being mentioned as result of a research effort whose title i do not recall.
Also , important is caloric restriction. What this suggests (as a hypothesis) is that caloric restriction may either have a positive or a negative effect to the symptoms of ME/CFS. Unfortunately we do not know the context suggested (i.e if the certain feature is associated with symptom or non-symptom state)
This could be why some of them are linked - to do with blood pressure/blood vessel control
Albumin seems related to electrolyte balance & water in blood control and also blood volume.
Cytokines are the subject that always seems to come back and I could understand mecfs as the side effect of cytokines,even if these individual cytokines aren’t very elevated taken as a whole a little bit on each cytokine could add up to a lot
Cort's just published an article about this on Health Rising:
It seems like this team's work on hormones in ME/CFS could prove to be extremely important.
But that's just my amateur impression - what do other S4ME members think?
If you believe their model then yes. But this is the raw data from actual measurements. Did they really only measure 2 hormones and model the rest based on their literature search???? Here is Supplementary figure 5.
Here are some quotes from the paper on ACTUAL measurements :-
Individual t tests at each time point does not support a conclusion that there is a difference in progesterone. Estrogen was elevated at all time points but again, refer to the actual data - there is a huuuuge overlap of datapoints between the two groups, so while the mean may be different, it is possible that could just be by chance.
This confirms that measurements for important hormones were unavailable..........
And others don't notice a difference in their data in ME.
So if I read correctly, hormone differences exist in the model, but they do not have data to confidently back that up yet.
Separate names with a comma.