Extended amygdala function is governed by threat probability, not uncertainty

TL;DR: For 15 years, the dominant model of anxiety neuroscience has held that the central amygdala (Ce) handles certain threats while the bed nucleus of the stria terminalis (BST) handles uncertain ones — a framing baked into NIMH's RDoC framework. This paper argues the whole edifice rests on a confound: in essentially every classic experiment, "uncertain" threats were also low-probability threats. Using a task that mathematically separates probability from two distinct flavors of uncertainty, plus monkey lesions, the authors show that self-report, skin conductance, and both EA regions track threat probability, not uncertainty. The twist that saves the useful part of the old model: BST is specifically tuned to low-probability threat, Ce to high.

The old model and why it was so appealing

Picture the extended amygdala as two switchboards. The canonical account, descended from Davis and Walker's rodent startle work in the 1990s, assigns them crisp, complementary jobs. The central amygdala (Ce) fires for threats that are certain, acute, imminent — the tone that always precedes a shock. The bed nucleus of the stria terminalis (BST) fires for threats that are uncertain, diffuse, sustained — the brightly lit room a nocturnal rodent dreads, where something bad might happen but you don't know when.

This was a beautiful model. It unified a sprawling literature — contextual conditioning at short vs. long intervals, cued vs. uncued shocks, forward vs. backward conditioning — under one axis: certainty. It mapped "fear" onto Ce and "anxiety" onto BST. And it gave clinical neuroscience a tidy target: if anxiety disorders are about pathological responses to uncertainty, BST is your handle.

The trouble, which had been accumulating quietly, is that the predictions kept failing. In rodents, manipulating either region affects both imminent and diffuse threats. In human fMRI, certain and uncertain threats light up a heavily overlapping circuit including both Ce and BST. The field reached an impasse, blamed partly on the fact that "uncertainty" had never been defined with any rigor.

The key move: uncertainty is not one thing

Here is the conceptual pivot. "Uncertainty" in everyday threat language secretly bundles three mathematically distinct quantities. The authors borrow from behavioral economics to pull them apart.

Let p be the probability of an aversive outcome.

• Probability is just p — how likely the bad thing is. A certain-threat cue (p=1) isn't uncertain at all; it's simply bad.
• Risk is the variance of the outcome, \mathrm{Var}[p]. This peaks at p=0.5 (maximal coin-flip unpredictability) and vanishes at both p=0 and p=1. Crucially, risk is an inverted-U function of probability, which makes it orthogonal to probability — you can vary one without the other.
• Ambiguity is when the probability itself is partly unknowable — you know p lies in some range but not where. Quantified as \mathrm{range}[p].

Once you see this decomposition, the historical confound jumps out. A "diffuse" threat (lights-on in a cage) means threat is possible but unlikely at any given moment — low p. A temporally uncertain threat spread over a 10-second window is ~10% per second instead of 100% at a known instant — low p. A distal predator has many chances to withdraw — low effective p. In every classic paradigm, the "uncertain" condition was also the low-probability condition. The field had been reading a probability axis and labeling it uncertainty.

The experiment: M-DART

The human task (N=104, deliberately recruited across a wide anxiety range, racially diverse) is a thermometer-style cue indicating threat likelihood from 0–100% in 10% steps, ending in a compound aversive outcome (unpleasant image + sound + wrist shock) or a benign one. On ~37% of trials the bar is partially occluded by a mask, injecting ambiguity (you know p only within a range). They collected trial-by-trial distress ratings, skin conductance (SCR), and fMRI throughout, then used parametric regressors to estimate the unique contribution of Probability, Risk, and Ambiguity to each measure.

The orthogonalization is the clever part. Risk enters as the inverse-quadratic of probability, so it's separable by construction. Ambiguity is estimated alongside a "rational probability" covariate (visible p + half the hidden range), ensuring ambiguity effects can't be an artifact of shifted expected probability.

What they found

Probability dominates everything, and it isn't close.

Whole-brain, the probability regressor lit up the classic fear/anxiety network — Ce, BST, hypothalamus, periaqueductal gray, midcingulate, anterior insula — while Risk and Ambiguity produced comparatively sparse activation. A timecourse analysis confirmed the probability signal appears during the anticipatory cue period, dissociable from the shock response itself, so it isn't an outcome artifact.

Now the direct test of the canonical model (Fig. 2a). Both Ce and BST were significantly modulated by Probability (Ce z=7.00, BST z=6.22, both p<0.0001) and by neither Risk nor Ambiguity (all non-significant). And there was no Region × Facet interaction — the two regions did not divide the labor the canonical model demands. That is a clean refutation of the central claim.

The paradox, and the rescue

Here's where it gets interesting rather than merely negative. The authors could still replicate the old finding when they ran the old analysis: contrasting 50%-threat vs. safety, BST lit up more than Ce (z=−2.95, p=0.003). So prior results were real — they were just misinterpreted.

The resolution: model each region's BOLD as a logistic function of probability and compare the curves. BST's sigmoid is left-shifted and steeper — it ramps up at low probabilities. Ce's is shifted right — it responds to moderate/high probabilities. Both track probability; they're tuned to different parts of the range.

Bootstrapped sigmoid fits confirmed BST has a significantly lower midpoint x_0 (U=193,918, p<0.0001) and steeper slope k (p=0.0007). The probability-bin transition analysis agreed, though more weakly (omnibus F(3,400)=2.61, p=0.051 — marginal).

The causal test in monkeys

fMRI is correlational, and you can't lesion a human BST. So they turned to rhesus macaques and the Human Intruder Paradigm. The key design: two "uncertain" contexts that differ in probability. In No-Eye-Contact (NEC), an intruder stands in profile — threat is uncertain but moderate/high. In Alone-following-NEC (the intruder has left), threat is uncertain but low — nothing is currently there, but something might return.

The canonical model (both are uncertain) predicts BST lesions reduce freezing in both. The probability model predicts a reduction only in the low-probability ALN context. Seven monkeys got MRI-guided ibotenic-acid BST lesions, each paired with a cage-mate control.

Result: a significant lesion × context interaction (z=2.15, p=0.031). Lesioned animals froze less in ALN (z=−3.14, p=0.002) but were indistinguishable from controls in NEC (z=−0.10, p=0.920). Restricting to animals with confirmed bilateral BST overlap strengthened the ALN effect (z=−6.16, p<0.0001). Locomotion showed no such interaction, ruling out a gross motor deficit. And in a novel-conspecific social encounter — a naturally low-probability threat — lesioned animals also froze less (z=−4.14, p<0.0001), hinting that this low-probability system may underlie shyness and social anxiety.

How strong is this, really?

The disambiguation itself is the paper's best contribution, and it is genuinely clean: orthogonalizing Risk against Probability is mathematically forced, and the Ambiguity-with-rational-probability-covariate is a reasonable design. The parametric human result is convincing — large z-values, convergence across self-report, SCR, and BOLD, and a timecourse control against outcome contamination.

Be skeptical on several fronts, though. First, the null for uncertainty is partly an absence of evidence — Risk and Ambiguity simply didn't reach significance, and N=104 may be underpowered to detect genuine but smaller uncertainty effects, especially in small subcortical ROIs. The abstract's "refute" is stronger than "we found probability dominant and uncertainty undetectable." Second, the refined BST-low/Ce-high story leans on the bin-transition analysis that is only marginal (p=0.051); the bootstrap sigmoid comparison is the real support, and bootstrapped p-values from pooled resampling can overstate certainty.

Third, and most important for the causal claim: the monkey lesions were not BST-specific. The methods are admirably honest — "most lesions were not specific to the BST region," with damage extending into ventral anterior caudate, and only three animals had confirmed bilateral BST overlap. The effects survive in that subset, but with n=3 one should hold the causal conclusion loosely. The probability difference between NEC and ALN is also asserted rather than independently measured.

Finally, the authors flag the right next question themselves: is it probability specifically, or expected value (p \times intensity), or a threat-response policy? They didn't vary intensity, so they can't yet separate these. A hungry animal shouldn't freeze for a high-probability low-stakes threat, which pure probability can't explain.

Why it matters

If this holds, it redraws a map that an entire translational field — and NIMH's RDoC framework — has been navigating by. "Anxiety = aberrant uncertainty processing" becomes "anxiety = exaggerated response to low-probability threat," which is both more ethologically grounded (it connects naturally to the Predator Imminence Continuum) and more computationally tractable: once you anchor on probability, you can import hazard rates, temporal-difference prediction, and the whole machinery of reward learning into threat anticipation. It also sharpens the clinical hypothesis — that miscalibrated tuning of EA neurons to low threat probability, especially in BST, drives pathological anxiety — into something testable with cell-type-resolution tools.

The section most worth your time is the Results around Figure 3 (the confound argument and the sigmoid-fitting), where the intellectual payload lives: it's where a clean negative result turns into a constructive, falsifiable replacement model rather than just tearing down the old one.