A mathematical model can separate two explanations that look similar in everyday life. Someone who rarely chooses a rewarding activity might expect little pleasure from it, or might update their expectations slowly after a good experience. Those are different hypotheses. Computational psychiatry turns them into parameters researchers can compare with observed behavior.

That distinction is the field's main contribution: making an explanation specific enough to test. A model that fits a task doesn't, by itself, establish the cause of a person's disorder.

How computational psychiatry works

Researchers build a model of a process such as learning from rewards, then ask people to complete a task and estimate which parameters best describe their choices. A learning-rate parameter, for example, describes how strongly a new outcome changes an expectation. Reward sensitivity describes how much the reward matters in that model.

The review by Huys, Maia and Frank (2016) distinguishes theory-driven models from data-driven approaches. The former test proposed mechanisms; the latter look for patterns that could help classify participants or predict outcomes. Researchers can combine them, but a useful prediction and a causal explanation are different achievements.

Four common approaches illustrate the range. They overlap: the hierarchical Gaussian filter is itself a Bayesian model, rather than an entirely separate theory of mental illness.

Approach	What the model represents	Example research question
Reinforcement learning	How outcomes change the expected value of an action	Does reduced reward sensitivity help explain anhedonia in a task?
Bayesian perception and free-energy models	How prior expectations combine with sensory evidence	When does a prior expectation outweigh an ambiguous signal?
Symptom networks	How symptoms may interact over time	Could poor sleep and low mood sustain one another?
Hierarchical Gaussian filter	How beliefs about a changing environment are updated	How does someone estimate uncertainty and change?

These are ways to ask questions. None gives a complete account of depression, anxiety or psychosis on its own.

Two examples with measurable predictions

In Huys and colleagues' 2013 meta-analysis, six behavioral datasets covered 392 experimental sessions. Participants completed a probabilistic reward task, and reinforcement-learning models separated reward sensitivity from learning rate.

Major depressive disorder and anhedonia were associated with a larger reduction in reward sensitivity than in learning rate. The distinction matters. It suggests that, in this task, the value assigned to rewards explained more of the observed pattern than a change in how quickly people learned. It doesn't rule out learning differences or establish one mechanism for every form of depression.

Powers, Mathys and Corlett (2017) used a different design. Through a Pavlovian learning task, they induced conditioned auditory hallucinations in four groups distinguished by voice-hearing and treatment-seeking status. People who heard voices were more susceptible. Computational modeling linked the effect to overweighting perceptual expectations.

The experiment makes a proposed mechanism testable under controlled conditions. Extending it to every hallucination, or using it to make an individual diagnosis, would require further evidence.

Symptom networks ask a different kind of question. Borsboom's network theory (2017) proposes that symptoms can sustain each other: poor sleep may increase fatigue, which may make daily activity harder and affect mood. The model encourages researchers to examine those connections rather than assume every symptom comes from one hidden cause.

What this changes for treatment research

If a parameter reliably captures a relevant process, researchers can test whether a treatment changes it and whether that change predicts improvement. That gives treatment research a more precise question than whether a group improved on average.

The difficult part is getting a dependable measurement. Karvelis, Paulus and Diaconescu (2023) reviewed reliability and construct validity in computational psychiatry. Many measures had poor psychometric properties. A parameter that changes substantially when the same person repeats a task is a weak basis for an individual treatment decision, even if a group comparison looks convincing.

This also explains the distance between these models and a chatbot. Our review of CBT chatbot research compares dialogue designs and study methods; it doesn't establish that a chatbot can identify or repair a specific computational mechanism. The AI chatbot meta-analysis addresses symptom outcomes in a different evidence base.

Digital phenotyping supplies another possible source of observations, such as changes in phone use or activity. Combining those observations with a computational model is a research direction. Detecting a pattern, explaining its cause and choosing an effective response each need their own validation.

FAQ

What is computational psychiatry in simple terms?

It's the use of mathematical models to test ideas about mental processes. A model might estimate how someone learns from rewards or combines expectations with sensory evidence, then compare those estimates with behavior and symptoms.

Can a model identify the exact cause of a person's disorder?

A good fit shows that a model can describe the observed data. Several mechanisms can produce similar behavior, so identifying a cause requires additional evidence and reliable measurements.

How is this different from an AI support app?

Computational psychiatry studies proposed mechanisms and predictions. An AI support app provides a conversation or exercises. For Nearby's current purpose and features, see the product facts: emotional support and self-help, rather than diagnosis based on computational models.