A Machine Learning Approach to the Differentiation of Functional Magnetic Resonance Imaging Data of CFS. Baraniuk et al. 2020

[John Mac]

Full Title:
A Machine Learning Approach to the Differentiation of Functional Magnetic Resonance Imaging Data of Chronic Fatigue Syndrome (CFS) From a Sedentary Control.

Chronic Fatigue Syndrome (CFS) is a debilitating condition estimated to impact at least 1 million individuals in the United States, however there persists controversy about its existence.

Machine learning algorithms have become a powerful methodology for evaluating multi-regional areas of fMRI activation that can classify disease phenotype from sedentary control. Uncovering objective biomarkers such as an fMRI pattern is important for lending credibility to diagnosis of CFS.

fMRI scans were evaluated for 69 patients (38 CFS and 31 Control) taken before (Day 1) and after (Day 2) a submaximal exercise test while undergoing the n-back memory paradigm.

A predictive model was created by grouping fMRI voxels into the Automated Anatomical Labeling (AAL) atlas, splitting the data into a training and testing dataset, and feeding these inputs into a logistic regression to evaluate differences between CFS and control.

Model results were cross-validated 10 times to ensure accuracy. Model results were able to differentiate CFS from sedentary controls at a 80% accuracy on Day 1 and 76% accuracy on Day 2

Recursive features selection identified 29 ROI's that significantly distinguished CFS from control on Day 1 and 28 ROI's on Day 2 with 10 regions of overlap shared with Day 1 (Figure 3).

These 10 shared regions included the putamen, inferior frontal gyrus, orbital (F3O), supramarginal gyrus (SMG), temporal pole; superior temporal gyrus (T1P) and caudate ROIs.

This study was able to uncover a pattern of activated neurological regions that differentiated CFS from Control.

This pattern provides a first step toward developing fMRI as a diagnostic biomarker and suggests this methodology could be emulated for other disorders.

We concluded that a logistic regression model performed on fMRI data significantly differentiated CFS from Control.

https://www.frontiersin.org/articles/10.3389/fncom.2020.00002/full

 

[adambeyoncelowe]

Full Title:
A Machine Learning Approach to the Differentiation of Functional Magnetic Resonance Imaging Data of Chronic Fatigue Syndrome (CFS) From a Sedentary Control.
https://www.frontiersin.org/articles/10.3389/fncom.2020.00002/full

Thanks. I'll follow this. It may yield an interesting discussion!

 

[mariovitali]

Very interesting study. They used a validation method which i also don't come across often (shuffling). I am not an expert on this but i found some excerpts suggesting (?) that certain activated areas are associated with panic / anxiety disorder :

Ten AAL regions were significantly different according to the predictive model between SC and CFS before and after exercise and may represent persistent indicators of CFS pathologies. Left and right putamen and right caudate of the basal ganglia may be part of the Affective Network that has been identified by meta-analysis of studies in anxiety (Xu et al., 2019). The primary sensory region (right S1) has been associated with heightened sensory awareness in panic disorder

 

[Hoopoe]

I think this is the part that stood out most

The left Rolandic operculum had the highest coefficient of any region, and was notable for its association with bodily self-consciousness, interoceptive and pain networks

 

[Andy]

I think this is the part that stood out most

And of course we don't know if differences in that area of the brain are a cause or an effect.

 

[Adrian]

. These included a Support Vector Machine (SVM), Random Forest, Decision Tree, and Neural Net. Logistic Regression was the algorithm ultimately selected as it fast to build, repeatedly produced the most accurate and generalizable results, and is easy to implement in practice.

Given their very small data set this doesn't surprise me but if they had more data other classification techniques may start to show better results.

 

[Adrian]

The positive prediction rate which is the proportion of positive classifications (which I assume to be CFS) which are correct is interesting in that it goes up on the second day - even though overall accuracy falls. The negative prediction rate falls. This seems to suggest that on the second day those who were predicted as in the CFS group were more likely to be but the levels of prediction in the CFS group generally went down.

I wonder if this has a relationship to the use of Fukuda where PEM is optional?

 

[Hutan]

Ten AAL regions were selected by the logistic regression models on both days (Table 2), suggesting these ten regions may represent persistent indicators of CFS pathologies. In addition, 19 were significant only on Day 1, and 18 only on Day 2.

Haven't read much of the study. But this all seemed a bit random - 29 areas identified as different between teh CFS cohort and the controls on Day 1 and 28 areas identified as different on Day 2 - but only 10 areas different on both days.

I'll wait to see the study done again with a new and larger set of patients meeting a criteria requiring PEM before thinking much about this one.

 

[Snow Leopard]

I think this is the part that stood out most

The left Rolandic operculum had the highest coefficient of any region, and was notable for its association with bodily self-consciousness, interoceptive and pain networks

Hmm...

"Speech suppression without aphasia after bilateral perisylvian softenings (bilateral rolandic operculum damage)"
https://link.springer.com/article/10.1007/BF02043975

"Interictal discharges have neuropsychological effects in rolandic epilepsy"
https://www.medwirenews.com/epileps...neuropsychological-effects-in-rolandi/7105850
"Indeed, they found that scores for verbal IQ on the Wechsler Intelligence Scale for Children correlated significantly with the dynamic functional connectivity of the left SMG and right rolandic operculum in the phases before and after centrotemporal spikes."

https://www.mdedge.com/psychiatry/a...white-matter-deficit-seen-stuttering-children
"Interestingly, the left rolandic operculum abnormality was not related to ongoing stuttering, because no difference was found in this region between children who recovered and children who continued to stutter. This may indicate that the abnormality indicates a risk for stuttering, not whether there is a chance of recovery, the investigators noted."


Differences within the rolandic operculum before the exercise challenge may simply reflect selection biases with regards to the verbal intelligence of patients versus controls.

Now the differences within patients before and after the challenge may be interesting, but it is important not to confuse cause with effect, as Andy mentioned.