(2016-12-14) Alexander Mental Disorders As Networks

Scott Alexander SSC Journal Club: Mental Disorders As Networks. Suppose you have sniffles, fatigue, muscle aches, and headache. You go to the doctor, who diagnoses you with influenza and gives you some Tamiflu.

assuming that the causal structure of her field matches the causal structure of better-understood diseases like influenza.

Generations of psychiatrists have noticed that different symptoms all tend to show up together and follow a similar pattern, suggesting some kind of deep connection between them. So psychiatrists follow the influenza model and attribute this collection of linked symptoms to a latent variable called “depression”.

Psychiatric disorders are diagnosed through clusters of symptoms, but we don’t expect every person to have every symptom in the cluster. For example, we diagnose depression when a patient has five out of nine symptoms on a list including fatigue, guilt, sleep disturbance, suicidality, et cetera. Each of these symptoms is often but not always present in a patient who has most of the others – for example, 75% of depressed patients have sleep disturbances, but 25% don’t.

But all psychiatric disorders are hopelessly comorbid with each other

If someone meets criteria for one DSM disorder, there’s a 50% chance they’ll have another one too. 60% of people with major depression also have an anxiety disorder. This is awkward when compared to eg the 75% sleep disturbance rate. Why are we calling sleep disturbance a “symptom” of depression, but anxiety a “comorbid condition” with depression?

The problem isn’t that nothing ever clusters together – depression, for example, is a very natural category. But so are various subtypes of depression. And so are various supertypes of depression, like depression + anxiety, or depression + psychosis, or depression + anxiety + psychosis

When people try to use sophisticated clustering algorithms on psychiatric disorders, they usually come up with something like this, where there are only three supercategories instead of the 297 different diagnoses in the DSM. And even three supercategories are pushing it – people with psychosis are far more likely to have depression too! Having any number of categories starts seeming arbitrary and fuzzy.

So Nuijten, Deserno, Cramer, and Borsboom (from here on: NDCB) ask: what if that’s wrong? What if there isn’t a latent variable like “influenza”? What if it’s symptoms all the way down?

Consider a network in which each symptom is a node, connected to all the others by pathways with certain weights on each direction. So for example, “sleep disturbance” might be connected to “fatigue” by a strong path – people with disturbed sleep are much more likely to be tired. These might both be connected to “low mood

A sample image from the paper, showing two possible simple networks of depression symptoms

Also from the paper. This shows a more complicated (and apparently empirically validated) network of symptoms. MD is major depression. GAD is generalized anxiety disorder. The nodes are all different symptoms – for example, “inte” is “loss of interest in activities” and “musc” is “muscle tension”.

Each node might affect the others with a certain delay.

This type of loopy network might be stable and self-reinforcing. Maybe your boss yells at you at work, which makes you have a bad mood. Then even if the direct effect of your boss would go away quickly, if it causes suicidal thoughts which cause guilt which cause more low mood, then the cycle can stick around forever.

In NDCB’s model, all possible psychiatric symptoms are connected like this in a loose network. Particularly tight-knit symptom clusters that often active together and reinforce each other correspond to the well-known and well-delineated psychiatric diseases, like depression and schizophrenia. But there are no natural boundaries in the network; low mood and poor sleep may be closely connected to each other, but they’ll also be more distantly connected to anxiety, and even more distantly connected to psychosis

There are lots of good things about thinking about psychiatric problems this way:

1. It helps explain how life stressors can cause depression.

2. It explains how treating depression symptoms can treat the depression.

3. It explains how therapy can treat depression.

4. It explains the polygenic structure of mental illnesses. If a mental illness were one specific thing, we would expect it to have one specific cause, or at least be limited to genes active in one specific area or process. In fact, it’s hard to come up with anything that genes involved in these illnesses have in common other than “they’re mostly expressed in the brain” – and sometimes not even that

5. It helps explain why so many brilliant people searching for The One True Cause Of Depression have come up empty.

II.

Actually, this last one deserves more explanation. NDCB think of these symptoms as visible patient complaints (“poor sleep”, “feelings of worthlessness”), and treat the connections between them as common sense (“if you don’t sleep, you’ll probably be fatigued”, “if you feel very guilty, you might attempt suicide because you think you deserve to die”). But their theory also works for networks of biological dysfunctions, or networks that combine biological dysfunctions with common-sense observed symptoms.

For example, we know that there’s a link between depression and inflammation. But it’s not a very good link; not all depressed people have increased inflammation, not all people with increased inflammation get depressed, and drugs that decrease inflammation don’t always cure depression. There’s similarly good evidence linking depression to folate metabolism, serotonergic neurotransmission, BDNF levels, and so on. Suppose we made a graph like the ones above, except that instead of putting things like “poor sleep” and “feelings of guilt” on it, we used “inflammatory dysfunction”, “folate metabolism dysfunction”, “serotonin dysfunction”, and “BDNF dysfunction”. There are a lot of reasons to expect these things to interconnect – for example, folate helps produce a cofactor necessary for serotonin synthesis, so any dysfunction in folate metabolism could make a problem with serotonergic neurotransmission more likely.

common-sense concepts mostly don’t have simple well-delineated biological equivalents. But what I’m saying is that the model where all of these things are observable symptoms, and the model where they’re all disturbances in brain chemicals and metabolism, aren’t necessarily in conflict.

III. Are there any problems with this theory?

There are a couple of disorders that really don’t fit this model. Bipolar disorder, for example, doesn’t quite work as a collection of self-reinforcing symptoms.

The impressive success of ketamine also counts as a point against. NDCB imagine psychiatric disorders like depression as gradually fading out on a symptom-by-symptom basis, eventually reaching a point where enough symptoms are gone that the rest of them aren’t self-reinforcing and just sputter out. This matches the course of eg SSRI treatment, where the medications will gradually improve a few symptoms at at time over the space of a month or so and maybe cause a full remission if you’re lucky. It doesn’t really match ketamine, where every aspect of depression vanishes instantly, then returns after a week or so without treatment.

I don’t think it’s possible to cure depression by blasting every symptom simultaneously. That is, suppose somebody is depressed with symptoms of poor sleep, poor appetite, low energy, suicidality, and low mood. Ambien can make them sleep. Pot can make them eat. Adderall can give them energy. Clozaril can make them stop wanting to kill themselves. And heroin can perk up mood. So if you gave someone Ambien, pot, Adderall, Clozaril, and heroin at the same time, would that cure their depression? I’m pretty sure no one has ever tried this, but I don’t think anyone’s reported exceptional results from less extreme cocktails like Adderall + trazodone + pot, which I’m sure a bunch of people end up taking.

Despite these issues, I feel like something along these lines has to be true. There are too many things that sort of kind of cause psychiatric problems, and too few things that look like One True Causes.

I guess I’ve felt for a long time that some kind of weird change in attractor states of biological systems is the best way to explain these kinds of things, but I was never able to express what I meant