Method at a glance
- Modality: respiratory
- Measures: Trial-by-trial learning about upcoming interoceptive events — how a person updates their expectation that a cue predicts an inspiratory resistance, expressed as a fitted learning rate and as per-trial prediction certainty and prediction error magnitude usable as fMRI regressors.
- Procedure: Eighty trials. On each, one of two visual cues appears (3.5 s) and the participant presses a button to predict whether a breathing resistance will follow. After a 2.5 s pause, a circle marks a 5 s stimulus window in which either an inspiratory resistance (70% of that person’s maximal inspiratory pressure, delivered by a remotely controlled mechanical circuit) or nothing is applied. The participant then rates how difficult it was to breathe on a VAS (5 s), followed by 7-9 s rest. The two cues are explicitly instructed as a fixed 80/20 pair whose contingencies can only swap; after a stable first 30 trials the pairing reverses four times. Behaviour is fitted with an associative learning model (Rescorla-Wagner, or 2- and 3-level Hierarchical Gaussian Filters) plus a unit-square sigmoid response model, via the TAPAS/HGF toolbox; the fitted trajectories are transformed into stimulus space and split by valence to give positive/negative prediction certainty and prediction error regressors.
- Also called: BLT, breathing associative learning task, interoceptive learning task
Strengths
- The first interoceptive task built to be modelled trial by trial with concurrent imaging. Everything else in the wiki that measures interoception yields a score per person; this yields a trajectory per person, which is what a computational account actually predicts about.
- A safely and precisely manipulable visceral stimulus. Respiration is the one interoceptive channel where an aversive, homeostatically real perturbation can be applied and withdrawn on a five-second schedule without pharmacology — contrast Khalsa’s isoproterenol boluses, which cannot be delivered eighty times.
- Reversals create genuine uncertainty. Four contingency swaps after a stable block mean prediction error is driven by the design rather than by noise, and volatility-sensitive models have something to fit.
- Validated before use: parameter recovery and model identifiability were checked on simulated data at three noise levels, and the fitted trajectory was tested against 15 held-out participants (beta = 3.1, t = 12.1).
- Pre-registered, including the model-selection rule and the fallback that was actually invoked.
Limitations
- Learning is about a cue-outcome contingency, not about the body’s own dynamics. What is learned is that a coloured square predicts a resistance — ordinary associative learning whose outcome happens to be interoceptive. It is therefore not a measure of Forecasting (internal-state prediction) in the sense that page defines (how internal states evolve under their own dynamics), and the wiki should not file it as one.
- No model won. In Harrison et al. (2021) Bayesian model selection gave no candidate a protected exceedance probability above 90% (RW 0.30, HGF2 0.40, HGF3 0.30). The simplest model was used by pre-specified fallback, so every downstream neural result is conditional on a model the data did not select.
- Prediction certainty is not a model parameter. Rescorla-Wagner has no precision or volatility term, so certainty is constructed as the absolute deviation of the predicted probability from 0.5. That is a distance from guessing, not an inferred precision — and the models that do represent volatility explicitly were the ones set aside.
- Predictions are button presses, not beliefs. The dependent variable is a binary yes/no forecast, so the model is fitted to a coarse readout; a continuous expectancy rating would constrain the trajectory far more tightly.
- The resistance is set at 70% of maximal inspiratory pressure — a strong, obviously detectable stimulus. This task measures learning about a suprathreshold event, not perception of a faint one (that is the FDT’s job in the same paper).
- Two of sixty participants had model fits indistinguishable from chance and were excluded, which is a reminder that a fitted learning rate presupposes that the person was learning.
Connections
Introduced in Harrison et al. (2021), and the reason that paper matters more than a well-run anxiety study.
Predict, from a cue, whether the next breath will be hard. Find out. Rate it. Eighty times, with the rule secretly reversing four times.
Why the wiki did not have this
The wiki’s interoceptive methods divide into detection tasks (Heartbeat detection task, Filter Detection Task (FDT), Respiratory Resistance Sensitivity Task (RRST), Heart Rate Discrimination Task (HRDT)) and attention tasks (Interoceptive attention task), plus one control task (Respiratory tracking task with visuomotor perturbation (respyra)). All of them ask about a static capacity: how well you read a signal, or where you point.
None of them produces the thing computational accounts of interoception are theories of. Predictive coding, Interoceptive inference and Active inference are claims about a quantity that changes on every trial — a prediction, updated by an error, weighted by a precision. A per-person accuracy score cannot test any of that. As Petzschner et al. (2021) put it, and computational psychosomatics repeats: “there is little empirical work testing such models’ predictions.”
The BLT is the instrument-shaped part of the answer. Fit a learning model to the button presses; take the model’s own per-trial estimates of prediction certainty and prediction error; use them as regressors against brain activity. The quantities the theory is about become measurements.
What comes out of it
Behaviourally, a learning rate (α ≈ 0.25 in both anxiety groups) and an inverse decision temperature. In Harrison et al. neither differed by anxiety — a null worth holding, because it means the anxiety effect in the brain occurred without any change in how fast people learned.
Computationally, two trajectories per participant, each split by valence:
- prediction certainty — how far from guessing the participant’s expectation was, separately for expectations of resistance (negative) and of no resistance (positive);
- prediction error magnitude — how wrong the expectation turned out, again split by whether the surprise was a resistance appearing or failing to appear.
Neurally, in the paper’s 7T reduced-FOV slab: certainty tracked deactivation of dlPFC, anterior insula, ACC and MFG; error magnitude tracked activation of aIns, ACC, MFG and the PAG. The signs are the ones predictive coding requires — more certainty, less updating, less activity.
The design decision worth copying
The two cues are explicitly instructed as a coupled pair: if one predicts resistance at 80%, the other predicts it at 20%, and they can only swap. Participants are told this.
That instruction is what lets the model be fitted in “contingency space” — a single trajectory rather than two, following Iglesias et al. (2013) — which roughly halves the parameters the data must support. It also means the task measures learning about one volatile rule rather than two independent ones, which is a cleaner target and a narrower one. A participant who did not believe the instruction is being modelled wrongly, and nothing checks whether they did.
What it is not
It is not a measure of interoceptive accuracy. The resistance is set at 70% of maximal inspiratory pressure; everyone feels it. Breathing-difficulty ratings did not differ between anxiety groups (82.6% vs 83.8%). The task deliberately makes perception trivial so that learning is the only thing varying — which is why the same paper runs the FDT separately for the perceptual layer.
It is not Forecasting (internal-state prediction). That page defines the missing arc as prediction of how internal states evolve, including under their own intrinsic dynamics. The BLT’s learning is about an arbitrary visual cue’s relationship to an externally imposed event. The body does nothing on its own here. The gap Forecasting (internal-state prediction) records is narrowed only in the sense that some interoceptive prediction is now modelled and measured; the specific arc remains empty.
It is not conditioning in the Pavlovian defense conditioning sense, though it is close: there is no measured defensive response, the outcome is a graded aversive load rather than a shock, and the dependent variable is an explicit verbal-equivalent prediction rather than a physiological CR.
Open
- The task has been run once, by its authors. No independent replication, no clinical sample, no second channel.
- The model-selection failure is unresolved. Either the models are too similar to distinguish at 80 trials, or the task does not constrain volatility learning enough to separate a fixed learning rate from an adaptive one. Longer sessions or richer readouts (continuous expectancy ratings) would settle which.
- Nothing links BLT parameters to any outcome. Learning rate is a candidate computational biomarker; here it predicted neither group membership nor any questionnaire. The biomarker claim remains unearned by this instrument.