21
Computational Psychiatry
*

Current psychiatric categorizations of patients are based on a taxonomy of symptoms. While this categorization seems to be reliable (in that different practitioners tend to categorize similar patients in similar ways), there seems to be a disconnect between the categorization and the underlying mechanisms. The detailed understanding of decision-making systems presented in Chapters 615 and the engineer’s view on the brain looking for failure modes, vulnerabilities, and dysfunction (Chapters 1820) provide a new view on psychiatry, which is now emerging as a field called “computational psychiatry.”

Our discussions of addiction, problem gambling, and PTSD suggest that the neuroscientific view of the decision-making system has implications for psychiatric disorders. Since the introduction of the DSM-III in 1980,A psychiatry has been developing a taxonomy of categories for identifying patients, presumably under the assumption that patients who fall in the same category should require similar treatment. Although these categories seem to be reliable in that different psychiatrists tend to categorize example patients into similar categories (that is, they agree on what categories look like), these categories have not been as good at identifying treatments for patients.2

In part, this arises because many of these categories are disjunctive; they are made up of subcomponents linked by an “or” relationship—if you have symptom A or symptom B then you are in category C.3 If the two symptoms included in the disjunctive category are both reliable, then the disjunctive combination of them will also be reliable. This disjunctive categorization can be seen by the predominance of DSM categories based on adding up symptoms. In the DSM-IV-TR, the diagnosis of problem gambling can be made if the patient has five or more of ten criteria. This means that two people can both be identified as problem gamblers even though they have no symptoms in common. The fact that the categories are reliable does not imply that they are valid, that they say anything about shared mechanism.4

The other problem with the categories identified in the DSM is that they are based on symptoms. As everyone knows, a given symptom could be caused by any of a number of potential underlying problems. Imagine a patient who arrives at a hospital with chest pain. Imagine if the doctor tells the patient, “Some proportion of chest-pain patients require open-heart surgery and some proportion of chest-pain patients respond well to an antacid, so we’ll try both.” Treatment would obviously succeed better if we reliably differentiate patients who were having heart attacks from patients who had indigestion.B

Sometimes the only option available is to cure the symptom. For a doctor looking down at a patient who has just lost a leg from a landmine explosion, there is nothing to do at the time but treat the wound and eventually provide a mechanical, replacement leg. Similarly, the inability to initiate movement (akinesia) and the slowness of movement initiation (bradykinesia) seen in Parkinson’s disease are due to the loss of dopamine cells in the brain.6 Just as it would be great to address the underlying causes that placed the landmine in the field7 (Ain’t gonna study war no more), it would be great to prevent the loss of dopamine neurons in the first place. However, the only treatment option for a patient presenting with Parkinson’s disease is treating the symptoms, either by increasing the dopamine available to the surviving cells (such as through dopamine precursors like levodopa or through attempts to replace the lost cells) or by allowing the brain to sidestep whatever it was that the dopamine cells were providing (as is done in deep-brain stimulation).8

Similarly, sometimes the right treatment is to prevent the symptoms from doing too much damage to the body and allowing the body to heal itself. It is the symptoms caused by cholera that kill, not the disease itself. Rehydration is a reliable treatment that effectively cures the disease by treating the symptoms.9

At this point, there is a recognition that there is a problem in psychiatry, but practitioners are still debating how to restructure the taxonomy so as to get at the mechanisms we do know, while allowing for the fact that there are a lot of mechanisms we don’t know, all while continuing to do our best to treat patients, many of whom desperately need help.10 A new field of computational psychiatry seems to be emerging to make these connections.11 The term is meant to invoke the use of theoretical constructs (like those we’ve been exploring in this book) to redefine the categories and proposed treatment for psychiatric disorders.

The question at hand is thus What are the mechanisms and causes that underlie psychiatric illness? What is depression? What drives it? What is schizophrenia? How do we define it? What is an anxiety disorder? Psychiatry has identified lots of different psychiatric illnesses: the DSM-IV-TR has over a thousand different diagnoses. How many of these are the same dysfunction manifesting itself in different ways? Is “social phobia” (fear of being out in society) the same or a different disorder than “agoraphobia” (fear of places where escape or help might be difficult)? How many of these are different dysfunctions manifesting as a single symptom? (See our discussion of addiction in Chapter 18.)

Anxiety disorders, for example, entail a pathological fear of taking actions in specific situations. In humans, anxiety manifests itself in terms of fear and worry about future outcomes. In animals, anxiety manifests itself in observable avoidance and fear behaviors.12 The circuits and brain structures involved are similar in both animal and human anxiety experiments.13 They involve the structures we’ve examined as part of the Pavlovian action-selection system (Chapter 8) and structures we’ve examined as part of the Deliberative action-selection system (Chapter 9). Current models suggest that there are two components that drive anxiety. First, direct negative expectations of outcomes can drive fear-related actions through the periaqueductal gray (PAG) and the amygdala, which we saw earlier to be a key part of the Pavlovian fear circuitry (the rustle in the grass that signals the lion; the tone that came before the shock). Second, indirect negative expectations of outcomes can be generated by hippocampal–prefrontal interactions that are then evaluated as likely leading to a negative outcome14 (don’t go down that dark alley!).

In fact, when rats or mice come to an open space, they tend to hesitate, leaning out and then returning back to shelter, as if they are nervous about whether to go or not.15 This “stretch-attend posture” is very similar to the vicarious trial and error behavior that we saw in Chapter 6. Like vicarious trial and error behavior, the stretch-attend posture behavior depends on the hippocampus, the prefrontal cortex, and their interaction.16 This suggests that this hesitation likely entails predictions of future dangers. Of course, rats and mice are prey animals and dangers really do lurk around every corner. Humans are generally at the top of the food chain and tend not to worry as much about being eaten by cats or owls.C Generally humans are more concerned with social dangers, but the same stress–response mechanisms seem to be involved.18 Anxiety attacks can come from a variety of sources, including social interactions (such as public speaking), financial futures, personal danger, danger to one’s friends or family, availability of escape and rescue (in claustrophobia and agoraphobia), and even rare circumstances (such as earthquakes or buildings collapsing). The DSM-IV-TR considers PTSD a form of anxiety disorder; certainly, the description of soldier’s heart sounds like a panic attack.19 As another example, anxiety is now considered to play an important role in obsessive-compulsive disorder (OCD). Interestingly, OCD seems to be a combination of an anxiety disorder about future dangers combined with an uncertainty about the memory of completed actions.20

Computational models are beginning to be built to explain fear and anxiety. While they haven’t yet led to changes in treatment, there is consensus building as to the underlying mechanisms. We’ve already discussed addiction, gambling, and PTSD in the previous chapters. Similarly, other psychiatric diseases are beginning to be addressed by computational models—for example, depression, schizophrenia, and other disorders.21

There are a number of major psychiatric disorders that are related not to individual action-selection systems, but rather to social interaction systems—for example, borderline personality disorder and sociopathy or psychopathy.22 These interactive dysfunctions entail people who do not show normal social interactions. We will leave these interactive issues aside until we can address neuroeconomics, game theory, the propensity for humans to organize into groups, and the mechanisms through which cooperation within those groups is maintained (in Chapters 22 and 23).

With all of these examples (as with all psychiatric disorders), there is an important issue that it is extremely difficult to define the line between normality and dysfunction.23 There has been a lot of discussion in recent years about the overdiagnosis of attention-deficit/hyperactive disorder (ADHD).24 Where is the line between a rambunctious boy who is not getting time to run every day and ADHD? How much is ADHD being used as a diagnosis to control overcrowded classrooms? Similarly, is it really true that one American in ten is depressed? (In 2008, 10% of Americans were taking antidepression medication, presumably mostly selective serotonin reuptake inhibitors [SSRIs].25) How often are these SSRIs being used to increase confidence, and how much are they being used to treat normal grief?26

On the other hand, it is important to recognize that there are people with severe psychiatric dysfunction who desperately do need help. True depression and normal grief are very different things.27 Suicide is the tenth leading cause of death in the United States.28 Similarly, while some rambunctious children may be incorrectly diagnosed with ADHD and treated with Ritalin, there are others who really do have a problem and for whom Ritalin is the difference between getting an education or not. Trichotillomania (hair pulling) is not about the 10-year-old who twirls her hair nervously while shuffling her feet and smiling cutely up at you, nor is it about pulling that first gray hair. Trichotillomania is about people who have pulled their head bald and their scalp bloody.29 There are people with severe dysfunction who do need help.

At this point, we don’t have clear answers. Although this is not a new problem and has been recognized as an issue within psychiatry for many years,30 there does seem to be a change coming where the connections are being identified between neuroscientific mechanisms and psychiatric dysfunction.31 What these mechanisms are, where the severe dysfunctions are, and how to treat them are all open questions being actively explored within the neuroscience, psychological, and psychiatric communities.

Books and papers for further reading

• Paul R. McHugh and Phillip R. Slavney (1998). The Perspectives of Psychiatry. Baltimore, MD: Johns Hopkins.

• P. Read Montague, Raymond J. Dolan, Karl J. Friston, and Peter Dayan (2012). Computational psychiatry. Trends in Cognitive Sciences, 16, 72–80.

• Steven E. Hyman (2007). Can neuroscience be integrated into the DSM-V? Nature Reviews Neuroscience, 8, 725–732.

• Thomas Insel et al. (2010). Research Domain Criteria (RDoC): Toward a new classification framework for research on mental disorders. American Journal of Psychiatry, 167, 748–751.

• Matthijs A. A. van der Meer, Zeb Kurth-Nelson, and A. David Redish (2012). Information processing in decision-making systems. The Neuroscientist, 18, 342–359.