What is computational psychiatry?

People with the same diagnosis may have different histories, symptoms and responses to treatment. Even the same symptom can have several plausible causes.

Imagine two people who both report poor concentration. For one, disrupted sleep may be central. For the other, altered responses to reward or uncertainty may matter more. Their experiences are real in both cases. A symptom score alone cannot tell us which account is right, so we need measurements and competing explanations.

Computational psychiatry makes explanations precise enough to simulate and test. The field includes cognitive tasks, reinforcement learning, statistics and machine learning. In CPNS, a major thread is mechanistic modelling of brain dynamics: we combine EEG or MEG with models of circuits, synapses and receptors. We also study how similar inference principles can guide intelligent agents.

Both approaches are useful. A generative parameter is an inference under a particular model, not a direct measurement of an individual receptor or proof of causation.

Sensory data are incomplete and noisy. We bring expectations to them, then update those expectations when evidence arrives.

Suppose you hear a faint tone in a noisy room. Whether a tone is probably present depends on how often tones occur, how often you hear a real one and how often noise sounds like one. Adjust those three ingredients below.

Predictive coding is a useful theoretical framework. Specific mappings from prediction errors to cell populations, oscillations or psychiatric symptoms remain hypotheses that experiments must test.

EEG records electrical potential at the scalp. MEG records magnetic fields produced by neural currents. Both track changes on a millisecond scale, but neither reads out a receptor directly.

We collect brain signals during rest, tasks, sleep or controlled pharmacological studies. Before interpreting them, we check timing, noisy channels and artefacts such as blinks or muscle activity. We can then examine event-related responses, frequency spectra and relationships between regions.

Three meanings of “connection”

One signal can reflect many underlying processes. That is why the link from EEG/MEG to synapses requires a forward model of how proposed neural activity would generate the observed data, and careful checks for alternative explanations.

A forward model turns chosen parameters into predicted observations. Inversion works backwards: which settings make the model's prediction resemble what we recorded?

The curves below are synthetic frequency spectra. Move the sliders until the teal model curve fits the dark observed curve. The toy model has a peak frequency, width and amplitude. These are visual properties of a curve, not receptor estimates or a DCM analysis.

What changes in a real CPNS model?

Instead of directly adjusting a peak, a neural mass model represents interacting populations of cells. Synaptic coupling, receptor-related time constants and transmission delays alter the modelled activity. A measurement model then predicts what EEG or MEG would look like. Dynamic Causal Modelling (DCM) is a framework for fitting and comparing such generative models.

The lab's studies use the same logic on different questions: define a clinical contrast, record a signal, fit a mechanistic model, then ask what the parameter estimates can and cannot support.

The ambition is to connect a person's experience to testable biological hypotheses and, eventually, better choices about intervention. The route from a fitted model to a useful clinical tool is long.

Three checks we keep returning to

• Model dependence. Could different circuits explain the same EEG or MEG pattern? What happens when we compare alternatives?
• Generalisation. Does the finding hold across people, recording sessions, tasks and independent cohorts?
• Clinical meaning. Does it predict an outcome in a way that is reliable and useful for someone outside the original study?

This is why good computational psychiatry needs careful experiments, strong data quality, uncertainty estimates and honest reporting of null or ambiguous results. Models make assumptions visible so we can improve them.

Pick a question you care about, then learn just enough of the toolkit to investigate it. You do not need to master the full maths before collecting or thinking clearly about data.