Applied Case: The Category Goes In the Brain

Off to the Frontier.

I am racing to the frontier.

A new paper in Frontiers in Psychiatry examines the “aberrant network topology of cortical-subcortical circuits in schizophrenia and bipolar disorder.”

Excellent.

• We have schizophrenia.
• We have bipolar disorder.
• We have a brain scanner.
  • Cool.
• We have graph theory.

Finally, maybe, somebody can show me where these things actually are.

Go ahead.

Show me the schizophrenia.

So the researchers recruited 89 people diagnosed with schizophrenia, 57 people diagnosed with bipolar I disorder, and 45 healthy controls.

The schizophrenia and bipolar diagnoses were assigned by certified psychiatrists according to DSM-5 criteria.

Symptom severity was assessed with the Positive and Negative Syndrome Scale.

And so then everybody went into the scanner.

Wait. Go back.

The schizophrenia was already there.

Not in the brain.

In the grouping variable.

Before just one functional network was constructed, before one correlation was calculated, before one single graph had an edge, the people entering the study had already been divided into:

• schizophrenic,
• bipolar,
• healthy control.

That is completely normal study design.

It is also going to matter a whole hell of a lot.

Because this experiment can now tell us whether populations selected using those categories differ along the measurements being taken.

It cannot independently discover that those categories are the natural objects into which those measurements divide.

The category just went in first.

Keep your eye on that nasty thing.

It is about to take a long, wet trip.

What the Fuck Are We Measuring?

• The participants lie inside an MRI scanner with their eyes closed.
• The researchers collect resting-state functional MRI data.

Fine.

Where is the schizophrenia?

Well, first, the continuous activity recorded across the brain is processed and anatomically divided using the AAL-90 atlas.

• Ninety regions.

The time series for each region becomes a node.

Okay.

Where is schizophrenia?

Next, the researchers calculate Pearson correlations between every pair of those regional time series.

Now we have a 90-by-90 functional-connectivity matrix.

Ok.

Where schizophrenia?

Only positive correlations are retained.

Negative correlations are set to zero.

Alright. Noted.

Not even going to ask.

And then the matrices are thresholded across a range from 0.10 to 0.40.

• Above the relevant threshold, an edge survives.
• Below it, the edge does not.

And then the matrices are converted into binary graphs.

• 1.
• 0.
  • Connection.
  • No connection.

The strength that survived the threshold has now become a yes.

And so from those graphs the researchers calculate synchronization, degree, betweenness centrality, efficiency, path length, assortativity, hierarchy, modular connectivity, and other topological properties.

Okay. Fuck it.

It's why I came all the way to the frontier.

I’m asking again.

Where the fuck is schizophrenia?

Nothing I have just described here is fake.

Nothing here requires fraud.

Nothing here even strikes me as obviously bad methodology at all.

• Every operation is an instrument.
• And every instrument makes a cut.
  • The atlas cuts a continuous organ into ninety regions.
  • Regional averaging cuts away variation inside those regions.
  • Correlation selects one kind of statistical relation among the resulting signals.
  • Removing negative correlations cuts again.
  • Thresholding cuts again.
  • Binarization cuts again.
    • Graph theory then gives us an extraordinarily useful language for asking what the resulting relations look like.

This is science. This is also why the sentence “the brain scan shows schizophrenia” should make your hand move toward the emergency brake and mouth make a noise like “oeuea”.

Because, oh man, oh God, just fucking look at everything that just happened between

• brain
  • and
    • graph.

We have not uncovered a little schizophrenia object hiding behind the hippocampus.

We have constructed an instrument for examining relations among signals.

And we are not done yet.

Please, Choose One of Five Brains.

The paper next organizes those ninety regions into five fun & functional modules.

• Module I covers somatosensory, motor, and auditory functions.
• Module II covers visual processing.
• Module III covers frontal-parietal attention.
• Module IV includes the default-mode network.
• Module V contains limbic, paralimbic, and subcortical systems associated with memory and emotion.

Question:

Were those five modules discovered from the organization of these participants' data?

Answer:

No.

They were defined a priori from prior literature.

Again:

That is allowed.

Prior knowledge is allowed. Scientific instruments do not have to awaken each morning with total amnesia.

But the distinction has to survive.

• You selected an anatomical atlas.
• You selected a connectivity measure.
• You selected which correlations could exist.
• You selected a threshold procedure.
• You selected a graph representation.
• You selected five inherited modules.
  • Then, you asked reality what happened inside that representational system.

And here, wonderfully, fantastically, the authors themselves eventually say just this.

The AAL-90 template is relatively coarse. It reduces anatomical specificity.

The modules were imposed from prior literature, and the authors explicitly acknowledge that their modular findings may partly reflect assumptions imposed by the framework rather than network organization discovered directly from the data.

Oh. Thank fucking God.

They know.

They know. So keep that in mind too.

We will be returning to these very fine people after I finish yelling at their very good paper like a moron.

The Relations: They Just Keep Coming Back.

So, what did they find?

Quite a bit.

At the global level, both the schizophrenia and bipolar groups showed lower network synchronization than the healthy controls.

At the nodal level, both clinical groups showed altered network properties across regions including the hippocampus, temporal areas, parietal areas, occipital areas, and thalamus. Some effects were stronger or more extensive in the schizophrenia group.

For several regions, nodal degree followed an especially interesting pattern:

SCH > BD > HC.

At the modular level, the schizophrenia group had fewer surviving connections inside the somatosensory-motor module, more inside the limbic/subcortical module, and fewer between the somatosensory-motor and default-mode modules.

Both clinical groups had denser connectivity between the frontal-parietal attention module and the limbic/subcortical module.

This is real information. Do not flatten it behind my ranting.

There are shared findings. There are findings more pronounced in schizophrenia.

There are findings apparently more specific to schizophrenia.

There are gradients. There are overlaps.

There are different relations at different scales.

Woah. Wait.

Look at what kind of nouns we are using now.

• Synchronization.
• Degree.
• Centrality.
• Connectivity.
• Density.
• Integration.
• Segregation.
• Modules.
• Relations between modules.

So, hold on.

You started with two psychiatric objects.

And the closer the instrument gets to the activity of the brain, 

the more the nouns begin dissolving into relations.

That is interesting. Actually, that is extremely interesting.

Because these results do not arrive saying:

• HERE IS SCHIZOPHRENIA,
  • AIDAN,
    • WE KNOW YOU CAME ALL THE WAY TO THE FRONTIER JUST TO SEE IT,
      • LOOK AT THIS RADICAL SHIT.

They arrive saying instead:

• Here are differences in the topology of relations among selected signals, under this parcellation, under these preprocessing choices, under these thresholds, among populations selected using these diagnostic categories.

That sentence is so much uglier.

THEY TESTED THE BRIDGE.

Now we get to my favorite part. These researchers did something important.

• They did not stop at finding group differences.
• They asked whether the altered network measurements were related to clinical symptom severity.

That is the bridge.

If these topological abnormalities are helping explain the clinical phenomena by which the diagnostic categories become meaningful in the first place, then we should want to know how the network measurements relate to those phenomena.

So, the researchers tested associations between the significantly altered network parameters and PANSS scores while controlling for several covariates.

And, after false-discovery-rate correction:

Nothing.

No significant correlations.

Okay.

Listen.

Everybody just remain calm.

Hear them out.

This does not prove that these network features have nothing to do with symptoms.

• The sample may be too small.
• PANSS may be too coarse for the relevant relations.
• Medication may interfere.
• Clinical heterogeneity may wash out more specific effects.
• The study is cross-sectional.

There are many possible reasons the bridge failed to appear.

Right. Still.

Those possibilities all point in the same epistemic direction:

You did not establish the bridge.

But then, we enter the Discussion section.

And suddenly, the network differences may contribute to poor control of imagination.

They may relate to difficulty distinguishing imagination from reality.

They may help produce hallucinations and bizarre ideation.

Increased connectivity involving limbic systems may bias attention, destabilize emotion, and contribute to delusions.

Guys.

What?

What happened?

• You tested the bridge.
  • The bridge did not survive correction.

So why the hell are we driving the Discussion section across it?

You can propose mechanisms. Please, propose mechanisms. That is part of what Discussion sections are for.

But speculation has to remain visibly downstream of evidence.

“Could contribute to” is not the same state as “was associated with in this study.”

The paper does know that. The language occasionally tries to forget.

But Why Am I Yelling at You?

Then, something very inconvenient happens to my article.

• The authors start listing their limitations.
  • And they are good limitations.

The sample is relatively small.

The design cannot establish causality.

The diagnostic groups contain heterogeneous patients.

The atlas is coarse.

The modular structure was imposed in advance and may partly reflect those assumptions.

All of the patients were medicated, creating a potential confound.

And, yes, the authors specifically flag the absence of significant correlations between their network measures and PANSS symptoms.

You tested the relationship, reported the null result, acknowledged that the representation itself imposes assumptions, admitted that causality cannot be inferred, preserved medication as a possible confound, and called for longitudinal and more homogeneous follow-up work.

So what the hell am I yelling at you for?

You just wrote the problem down.

This paper is doing something very useful.

• The researchers found patterns.
• They reported them.
• They used graph theory to expose relationships that cannot be reduced comfortably to one localized brain lesion or one clean pathological switch.
  • They preserved several reasons their own interpretation could be wrong.

That is what real-deal scientific correction looks like. Excellent shit.

The problem is larger. These very good researchers are working inside a field whose ordinary grammar makes a very particular little move feel natural:

1. Begin with schizophrenia.
2. Begin with bipolar disorder.
3. Compare their brains.
4. Find different distributions of some measurements.
5. Call the measurements abnormalities of schizophrenia and bipolar disorder.
6. Ask whether those abnormalities can become biomarkers of the disorders with which the populations were selected.

And there is valuable research hidden inside that pipeline.

There is also a category trying very hard to become an ontology.

These researchers are not hiding that problem.

In places, their own paper exposes it.

Okay, I Did What You Asked, Now Please, Put the Biomarker Down.

The abstract concludes that the distinct topological characteristics found here could become useful biomarkers for differential diagnosis, treatment guidance, and response evaluation.

The conditional matters. They say could.

Good move.

Because right now, you do not have that biomarker.

You have group-level statistical differences.

There is no diagnostic classifier here that receives an unknown patient and reliably returns schizophrenia or bipolar disorder.

There is no held-out prediction establishing individual diagnostic performance.

There is no sensitivity and specificity showing that this graph can sort clinical cases at useful rates.

There is no prospective demonstration that these measurements guide treatment.

There is no demonstrated prediction of treatment response.

The paper itself closes by calling for further work to identify promising network biomarkers and therapeutic targets.

Exactly. Yes.

Candidate signal. Research program. Interesting result. Keep going.

But, watch what happens to language when it leaves the paper.

• A group-level neuroimaging difference can become a potential biomarker.
• A potential biomarker becomes a biological marker.
  • A biological marker becomes evidence that schizophrenia is “visible in the brain.”
    • And five transitions later somebody is waving around a picture of a cortex explaining that science has finally shown us where the schizophrenia is.

No. They haven't.

I already drove all the way down to the frontier and checked. 

Wasted my whole Sunday.

A Category Can Work Without Owning Reality.

None of this requires anyone to declare schizophrenia imaginary, Thomas.

Failed Field Analysts: Thomas Szasz and the Conceptual Knife

He saw the prison inside the hospital and tried to free the patient by abolishing the hospital’s language.

Modal Path EthicsAidan Lawson

That would be satisfying for approximately twelve seconds and then immediately recreate the exact same problem from the opposite direction, Thomas.

Diagnostic categories can do work.

“Schizophrenia” can gather recurring experiences into a shared clinical vocabulary. It can help clinicians communicate. It can make some forms of care reachable, organize research, identify populations at elevated risk for particular outcomes, help somebody find other people who recognize what they are going through, carry legal protections, and preserve knowledge that would become harder to coordinate if every person had to begin description from zero.

An instrument does not become illegitimate because it simplifies.

All instruments simplify.

The question is whether the simplification remembers that it simplified.

A category becomes dangerous when successful coordination gets promoted into ownership of the thing being coordinated.

Then:

• The diagnosis organizes the patient.
• The diagnosis explains the patient.
• The diagnosis predicts the patient.
• The diagnosis interprets disagreement from the patient.
• The diagnosis tells us what sort of experience the patient is having.
  • The diagnosis eventually becomes more stable than the human being it was built to help describe.

That is when the category stops serving contact. It starts demanding that contact arrive in category-shaped pieces.

This paper is interesting precisely because the measurement does not cooperate cleanly.

Reality just keeps leaking across the edge.

I generally recommend you let it.

The Ruling.

The category went in first.

• Eighty-nine people entered one group because they had already been diagnosed with schizophrenia.
• Fifty-seven entered another because they had already been diagnosed with bipolar I disorder.
  • Then the scanner collected resting-state functional data.
    • The brain became ninety regions.
    • The regions became time series.
    • The time series became correlations.
      • Some correlations disappeared.
      • The others became thresholded edges.
      • The edges became graphs.
      • The graphs became topological measurements.
        • The measurements became group differences.

And after all of that, the original categories returned to collect the findings under their legal names.

Fine. That is definitely one legitimate way to organize research.

Just keep the direction of inference visible.

Do not confuse the category that selected the cohort with an object discovered inside the cohort.

This paper went looking for the neural topology of schizophrenia and bipolar disorder. What came back was stranger and more useful:

• relations,
• overlaps,
• gradients,
• different severities,
• different connections,
• different cuts through one living system,
• and a pile of reasons to keep investigating.

The category can stay. Its jurisdiction gets smaller.

• Put schizophrenia into the grouping variable.
  • Nature gives you a graph.