Have a personal or library account? Click to login
Computation in Psychotherapy, or How Computational Psychiatry Can Aid Learning-Based Psychological Therapies Cover

Computation in Psychotherapy, or How Computational Psychiatry Can Aid Learning-Based Psychological Therapies

Computational Psychiatry
Volume 2 (2018): Issue 0
By: Michael Moutoussis,  Nitzan Shahar,  Tobias U. Hauser and  Raymond J. Dolan  
Open Access
|Feb 2018

Figures & Tables

00014f01c.png
Figure 1.

Correspondence between CBT and computational conceptualizations of a paradigmatic case of depression (Louie in text). Left: CBT diagram as used in textbooks, in self-help books, and in working with patients with depression. Several aspects of it, such as the postulated vicious cycle including the bold arrow, are still inadequately validated. Right: the postulated CBT process is related to the language of computational psychiatry. Key questions for research are posed in terms of probabilistic reasoning, computational models, and parameters characterizing individuals.

Box 1.

Therapy concepts relevant to inference (alphabetical order).

• Attribution: Inference that a particular cause was responsible for an observed event.
• Attributional bias: A cognitive bias (see below) affecting attribution (above).
• Avoidance: Emission or withholding of activity so as not to enter a feared external situation, as would be ordinarily expected. For example, ‘School avoidance’ = not going to school (Lovibond et al., 2009)
• Catastrophizing: A cognitive bias, postulating unwarranted inference of unaffordable and/or severe and irreversible loss.
• Cognitive (or reasoning) bias: Propensity to draw inferences unwarranted by the totality of information available to a healthy person in the particular context. Many use it to imply that no qualitative information-processing or organic deficit is present. Hails from learn ing theory and the psychosocial therapies, in opposition to the medical model.
• Cognitive deficit: Lack of an adaptive information-processing function. Many use it to imply that a reasonably well-defined brain problem (gene, lesion, mis-development, etc.) adversely affects information processing. It hails from neuropsychology / medical model. Computational psychiatry elegantly reconciles ‘bias’ and ‘deficit’.
• Dysfunctional assumption: Belief such as ‘unless I always succeed, I’m a loser’, used as a rule guiding behavior (‘... so I must always succeed’). It can lead to unhelpful inference (‘I failed in this, what a loser I am!’) and hence psychopathology.
• Ex-consequentia emotional reasoning: Treating the presence of an emotion as evidence for congruent inferences-e.g., ‘I am anxious, therefore there is danger around’. Tradi tionally considered a reasoning error, we argue that it is in the first instance an adaptive process of inference on the basis of interoception—though of course only in the first instance.
• Exposure: Entering the feared situation, whether withholding or emitting safety behaviors. See also ‘response prevention’.
• Mentalizing: Inference of mental states as causes of the intentional behavior of self and others, relevant to the self.
• Mentalizing breakdown: Stereotyped inference about mental states, usually resulting in unwarranted inferences that take place under strong emotions. Closely related to mind- reading (below), but more elaborated, as it’s used in conditions where such biased theory- of-mind is of central importance.
• Mind-reading: A cognitive bias, postulating an overconfident inference that others hold negative views about the self. Note that in CBT this term is more narrowly defined than in common parlance.
• Psychotherapy: A formal treatment aiming to change mental function and perception by using existing sensory inputs in order to improve mental health. Here we include at least the psychological therapies of all modalities (behavioral, analytic, systemic, etc.) and the rehabilitation therapies.
• Response-prevention: Withholding safety behaviors (see below) upon entering the feared situation (what happens in Exposure with Response Prevention, or ERP).
• Reinstatement: The re-appearance of an older pattern of behavior in response to a stimulus, well after new behaviors have been successfully learnt, as a result of non- specific factors such as the passage of time or the occurrence of non-specific stress.
• Schema: Constellation of properties pertaining to a state, and appropriate actions to be performed in response. A person’s mini-model of an aspect of a situation. Some therapy theorists write about emotion as part of schema, but here we follow the more traditional analysis of emotional feelings as interacting with schemas which include beliefs about self, world and appropriate action.
• Safety-behavior: In general, any behavior carried out in order to reduce or avert a serious feared outcome (Salkovskis, 1991). In that sense avoidance is a safety behavior that removes the patient from the anxiogenic situation. However ‘safety behavior’ usually means usually means protective behaviours emitted within the anxiogenic situation.
00014f02c.png
Figure 2.

Avoidance learning and spontaneous extinction. Each trial starts on spatiotemporal state 1, signaled to the agent by a light. All possible states are enumerated. Time moves the agent up and to the right. There are two spatial states, left and right of the gray barrier. A small cost is needed to perform avoidance, that is, to jump the barrier toward the safe states 7–11. A) Punishment. Before learning avoidance, the agent does nothing, and time alone lands it in State 6, where it receives the adverse outcome: a shock (gray arrow). B) Acquisition of negative state value. Earlier states quickly acquire negative value by association, but avoidance has not yet been learned. C) Avoidance. Jumping is acquired. D) Extinction. The shock at State 6 has been turned off, and avoidance wanes after a number of unshocked trials. E) Acquisition and loss of avoidance in extinction. Before Time 0, the avoidance action is not available, and shocks are received ˜10 s after the light (as if States 7–11 are unavailable in A–B). At Trial 0, the avoidance response becomes available (7–11 in A–D), and the shocks are switched off. Rodents show vigorous reduction in the latency of avoidance during Trials 1–20; that is, they behave as in C despite State 6 now being innocuous. Avoidance is maintained for many trials in extinction (i.e., without any adverse outcomes) but eventually decays. After 270 trials, the rodents remain for longer than the previously shocked latency on the previously shocked compartment. Adapted with permission from “A Temporal Difference Account of Avoidance Learning,” by M. Moutoussis, R. P. Bentall, J. Williams, and P. Dayan, 2008, Network, 19, p. 140.

00014f03c.png
Figure 3.

Modeling avoidance learning, exposure-with-response-prevention, and spontaneous extinction. A) State space labeled as in Figure 2. The white arrows are the action “stay,” while the black curved arrows indicate “jump to safe state.” B) Simulation of associative, model-free learning: I–III (dark gray), exposure to harm while avoidance is inhibited by the experimenter; III–IV, rapid establishment of effective avoidance when it is no longer inhibited, in extinction (this is entirely analogous to Trials 0–20 in Figure 2E); IV–V, as in human psychopathology, here vigorous avoidance dramatically impairs its own extinction (this mimics behavior around Trials 20–100 in Figure 2E); light gray (to VI), therapy-like ERP, where the mouse is prevented from carrying out avoidance behavior and thus learns that State 6 is no longer dangerous; VI–VII, further gradual extinction of avoidance. The reader is referred to Moutoussis et al. (2008) for further technical details, including the relevant neuropharmacology. Adapted with permission from “A Temporal Difference Account of Avoidance Learning,” by M. Moutoussis, R. P. Bentall, J. Williams, and P. Dayan, 2008, Network, 19, pp. 144, 150.

00014f04c.png
Figure 4.

Translating questions about therapy mechanisms in computational terms helps form testable predictions. Left: important questions about the onset and maintenance of disorders. Right: predictions, in computational terms, that can be addressed in further research.

00014f05c.png
Figure 5.

Highly simplified examples of inference in therapy, either prone (top row) or resistant (bottom row) to failure owing to overaccommodation. The arrows denote the strength of responsibility that the patient attributes to causes. In each row, before therapy, the patient attributes negative experiences to being inadequate. In the case of overaccommodation, positive experiences in therapy are mostly attributed to the therapist (equivalent to creating a new schema), without changing beliefs about the self. In assimilation (bottom row), the new cause (therapist) does not account for the new (success) experiences, so that beliefs about the self change (equivalent to adapting existing schema). The last column shows inference upon encountering new stressors outside therapy.

DOI: https://doi.org/10.1162/CPSY_a_00014 | Journal eISSN: 2379-6227
Journal RSS Feed
Language: English
Submitted on: Dec 19, 2016
|
Accepted on: Sep 16, 2017
|
Published on: Feb 1, 2018
Published by: MIT Press
In partnership with: Paradigm Publishing Services
Publication frequency: 1 issue per year
Keywords:
computational psychiatry,
belief updating,
Bayesian inference,
cognitive-behavioral therapy,
mentalization-based therapy,
near-miss disaster,
avoidance,
therapy failure,
reinforcement learning,
exposure-with-response-prevention

© 2018 Michael Moutoussis, Nitzan Shahar, Tobias U. Hauser, Raymond J. Dolan, published by MIT Press
This work is licensed under the Creative Commons Attribution 4.0 License.

Previous articleVolume 2 (2018): Issue 0Next article