What would even count as a mood biomarker?
An affective-state biomarker is a feature of ongoing neural activity that tracks how a patient feels right now — a state marker, to be distinguished from a trait marker that merely correlates with having the diagnosis. For a closed loop you need a state marker: something that rises when symptoms rise and falls when they remit, fast enough to act on. Candidate features are the familiar ones from earlier volumes — band power in specific frequency bands, cross-frequency coupling, and connectivity between limbic sites — but now regressed against felt mood rather than a cursor.
Decoding mood as a hidden state
The most principled approach treats mood as a low-dimensional latent state that you never observe directly but that drives both the neural recordings and the occasional self-report. This is exactly the state-space modelling you met in Volume II, now with two observation channels: dense neural features and sparse mood labels. Human intracranial studies have used this framing to show that multi-site limbic activity carries decodable information about mood variation over days.
Mood as a hidden state driving two observations. The latent xₜ evolves with dynamics A; neural features yₜ are available every moment, but mood labels mₜ arrive only occasionally. A method like preferential subspace identification learns the sub-space of xₜ that is shared with mood — not merely the loudest neural variance.
The subtlety that method addresses is crucial: the largest source of variance in a limbic recording is usually not mood. A naive decoder locks onto whatever varies most — arousal, movement, artefact. Techniques such as preferential subspace identification explicitly extract the part of the dynamics that is behaviourally relevant to mood, which is why intracranial mood decoding is a modelling problem, not just a filtering one.
A personalised biomarker, one patient at a time
The landmark human demonstrations of closed-loop psychiatric neuromodulation have so far been N-of-1: in a single patient, investigators recorded from multiple sites during a period of symptom monitoring, searched for a feature whose fluctuations tracked that patient's symptom state, and then used it to trigger stimulation. The biomarker that emerged was individual — a specific band-limited signal at a specific site — and would not be assumed to transfer to the next person. This is the concrete face of personalised psychiatric neuromodulation.
A simple, interpretable form the discovered biomarker often takes: a logistic read-out over log band-powers P_k at chosen sites/bands. Its transparency matters — a clinician can inspect which band at which site is driving the decision, which a black-box decoder hides.
The ground-truth problem, stated plainly
Every result above is limited by its labels. You are fitting a decoder to a handful of subjective ratings per day, over a recording window of days to weeks, in one person — a regime where it is dangerously easy to find a feature that correlates by chance. The honest practitioner treats a discovered mood biomarker as a hypothesis to be prospectively validated, not a fact, and designs the search to control that risk. We return to exactly this failure mode — over-fitting an individual biomarker — in the final guide.