Development of an expert system for the classification of myalgic encephalomyelitis/chronic fatigue syndrome 2024 Yagin et al

Abstract

Myalgic encephalomyelitis/chronic fatigue syndrome (ME/CFS) is a severe condition with an uncertain origin and a dismal prognosis. There is presently no precise diagnostic test for ME/CFS, and the diagnosis is determined primarily by the presence of certain symptoms. The current study presents an explainable artificial intelligence (XAI) integrated machine learning (ML) framework that identifies and classifies potential metabolic biomarkers of ME/CFS.

Metabolomic data from blood samples from 19 controls and 32 ME/CFS patients, all female, who were between age and body mass index (BMI) frequency-matched groups, were used to develop the XAI-based model. The dataset contained 832 metabolites, and after feature selection, the model was developed using only 50 metabolites, meaning less medical knowledge is required, thus reducing diagnostic costs and improving prognostic time. The computational method was developed using six different ML algorithms before and after feature selection. The final classification model was explained using the XAI approach, SHAP. The best-performing classification model (XGBoost) achieved an area under the receiver operating characteristic curve (AUCROC) value of 98.85%. SHAP results showed that decreased levels of alpha-CEHC sulfate, hypoxanthine, and phenylacetylglutamine, as well as increased levels of N-delta-acetylornithine and oleoyl-linoloyl-glycerol (18:1/18:2)[2], increased the risk of ME/CFS. Besides the robustness of the methodology used, the results showed that the combination of ML and XAI could explain the biomarker prediction of ME/CFS and provided a first step toward establishing prognostic models for ME/CFS.

Open access, https://peerj.com/articles/cs-1857/

 

I don't see if they explain how what criteria they use to select participants.

 

They appear to have used the open access metabolomics data from the following paper which used 1994 Fukuda definition for ME/CFS.

"Prospective Biomarkers from Plasma Metabolomics of Myalgic Encephalomyelitis/Chronic Fatigue Syndrome Implicate Redox Imbalance in Disease Symptomatology"
https://www.mdpi.com/2218-1989/8/4/90
by Arnaud Germain, David Ruppert, Susan M. Levine and Maureen R. Hanson.

Hanson has a good rep imho but its a tiny cohort size. 32 PWME and 19 controls, its about the same size as Kerrs affymetrix study, too small. These wont have been randomly picked is my guess and are essentially cherry picked. But I think really this AI paper is proof of concept and they just needed a data set and chose this one for reasons of realpolitik.

Logically we already know epidemiologically ME can occur in clusters.
see https://me-pedia.org/wiki/List_of_myalgic_encephalomyelitis_and_chronic_fatigue_syndrome_outbreaks

This suggests a common cause for clustered cases per location but replicability is poor globally, suggesting it may be unsafe to assume the same causes in different locations, or between these and cases arising outside clusters. This is simply a huge unknown. The search for one virus ended after Mikovits/XMRV debunking. So the possibility remains open that multiple pathogens can cause ME each with its own unique signature.

We have only just discovered obelisk viroids so goodness know what else is out there.
https://www.scientificamerican.com/...-found-in-human-gut-may-be-virus-like-entity/

To me clustering suggests a pattern of infection but it is strangely self limiting. I wonder if it may be something weird like Dengue where you are more likely to get hemorrhagic fever if you were previously infected with a different strain of Dengue, by which I mean ME may be the result of two or more subclinical epidemics colliding. We dont know.

The question is which signatures are unique to a cluster and which generally applicable to all PWME and we just dont know with such a limited sample. All of this decreases my confidence that the results for this small localised cohort, will be applicable to ME patients around the world.

IMHO its the right idea to use AI this way and the data it is working on just needs to be much x10^4 bigger and from all around the world.

Just sayin'... brave attempt, needs bigger data. Also should be looking for subtypes.

 

So, they derived the model with the dataset from 51 people, and then showed that the model achieved the predictive value of 98.5% when applied to the same 51 people? I don't know why anybody would fall for this classic statistical fallacy.

 

Do you think they'd get similar results if they used letters in the subject's names rather than blood factors?

 

These new discoveries about our bodies is kind of frightening, in terms of "what else have they missed that is really important?" Maybe ME is caused by a microbe (still undiscovered) that affects certain cells (still overlooked as different from similar-looking cells) in an organ (still undiscovered)?

 

They needed 50 out of 832 available metabolites for a model to separate patients and controls. To me that suggest they had a really hard time separating the two groups.

 

Presumablely this also suggests that verification in another lab (other labs) and different populations of people with ME/CFS is desirable/necessary.