The Threat Your Parents Made

Keep making that face and it will stay that way.

It’s a threat about the relationship between states and traits. Hold an acute expression long enough, the warning goes, and it calcifies into something permanent — a disposition you can no longer put down. Nobody’s face has ever actually frozen (as far as I know). But the underlying intuition is worth taking seriously: that repeated states leave a residue, and that residue eventually is you.

Our newest paper runs the intuition in the other direction. Rather than asking whether repeated states sculpt traits, we asked whether the trait is already legible in the brain when nothing is happening at all — whether six minutes of someone staring at a fixation cross contains enough structure to predict how much empathy they subsequently report on a questionnaire.

Where We Left Off

In our 2020 study, we showed that resting-state functional connectivity within and between two theory-driven networks predicted trait empathic concern better than the canonical intrinsic networks did. The two networks came from a specific conceptual model: a bottom-up resonance network (inferior frontal gyrus, inferior parietal lobule, superior temporal sulcus, insula, amygdala, somatomotor cortex) that simulates others’ internal states, and a top-down control network (DLPFC, TPJ, lateral OFC, medial prefrontal and paracingulate cortex) that regulates how much of that simulation reaches behavior.

That result had a familiar vulnerability. It came from a single cohort, validated with leave-one-out cross-validation — a procedure that holds out one participant at a time but never leaves the room. Cross-validation inside one sample tells you a model is internally consistent. It does not tell you the model has learned something about empathy rather than something about the fifty-one people you happened to scan.

Training on Strangers

So we replaced the pseudo-holdout with a real one. Two independently recruited cohorts (N = 96 total), scanned at UCLA’s Ahmanson-Lovelace Brain Mapping Center on the same 3T scanner but years apart, were used to train and test in both directions: fit the model on Cohort 1, test it on Cohort 2, then reverse it, and average.

The features were (relatively) straightforward. For each participant we extracted mean BOLD time courses from 198 spherical ROIs, computed pairwise Pearson correlations within each network, regressed out sex, and fed the resulting edges to a LASSO regression — a regularized model that shrinks most coefficients to zero and keeps only the connections that carry weight. Empathy was measured with the four subscales of the Interpersonal Reactivity Index: empathic concern, personal distress, fantasizing, and perspective-taking. 

The stress test was harsher than we designed it to be. The two cohorts differed significantly in their empathic concern scores — Cohort 1 simply reported more of it. The model generalized across that distributional gap anyway.

The Control Network Never Missed

The control network predicted all four subdimensions, with accuracies between r = 0.31 and r = 0.36, explaining roughly 11–13% of the variance in each. Classical resting-state networks managed only empathic concern and personal distress.

Empathic concern was the most broadly represented by a wide margin, predicted significantly by seven separate networks or combinations, with accuracies from r = 0.28 to r = 0.45. The salience network was the strongest single predictor. Combined resonance-and-control connectivity outperformed either network alone — consistent with the idea that empathic concern is not a thing one network does, but something that emerges from the traffic between simulation and regulation.

Personal distress was the only subdimension the whole-brain model captured, which fits its profile as the more primitive, arousal-driven relative of empathic concern — the response that shows up earlier in development and stays closer to raw vicarious activation.

Why Fewer Features Won

The most consequential result is a methodological one. The whole-brain model had access to 34,716 possible connections. The combined resonance-and-control model had 1,540. The smaller model won.

The obvious explanation — that LASSO simply prunes harder when given more — turned out to be wrong. Across every significant model, regardless of input size, LASSO converged on a comparable handful of non-zero coefficients. The advantage of a theory-driven network therefore isn’t regularization behavior. It’s that a greater proportion of the edges you hand the model carry signal in the first place.

That matters at a moment when the field has an increasing sense that we have, as Finn put it, “squeezed the lemon enough” with resting-state scans. Perhaps the lemon isn’t spent.

What This Opens Up

The clinical motivation is direct. Self-report questionnaires and in-scanner empathy tasks both require a participant who can understand instructions and respond — which excludes many of the populations where empathic function matters most to measure. A model that reads trait empathy from six minutes of passive rest doesn’t have that requirement.

We’re releasing the model and datasets openly (OSF; GitHub) so other groups can test them against their own cohorts. The immediate methodological frontier is sex interactions, which our cohort sizes couldn’t support — splitting by sex would leave roughly 25 participants per training group, too few to preserve the generalizability the whole design was built for.

Two caveats worth stating plainly: these are accuracies in the r ≈ 0.3–0.45 range, meaningful at the group level but not yet individually diagnostic, and both cohorts were healthy young adults scanned on one scanner. Generalization to clinical populations is the hypothesis, not the finding.