Separating how much it hurts from how people use the pain scale

When a person rates pain as "7 out of 10," that number mixes two things. One is how strongly their nervous system responds to the stimulus. The other is how they use the rating scale: some people reserve high numbers for the worst imaginable pain, others reach for them readily, and some want to seem tough or to please the investigator. Quantitative sensory testing (QST) is the set of lab methods that apply precisely controlled heat, cold, or pressure and record where sensation turns painful and how intense it feels. It is usually described as the objective counterpart to questionnaires. But human QST still ends in a self-report: the participant says "now it hurts," or gives a number. Scale-use bias therefore sits inside the very measurement that is supposed to be objective, and a stoic participant can look like someone with low sensitivity.

We propose a short calibration protocol and a latent-variable model that pull these two quantities apart for each person. Each participant rates calibrated painful stimuli, rates a set of non-painful anchor stimuli (sound loudness and light brightness) whose felt intensity varies little across healthy people, and performs forced-choice discrimination trials that estimate sensitivity without using any rating scale. The anchors pin down the person's reporting style. The forced-choice trials pin down their sensitivity. The painful ratings must be explained by the two combined. The resulting per-person reporting-style parameter is then put to two tests: whether it corrects between-person comparisons of pain reports, and whether it can personalize automated facial pain estimators for a new patient without first collecting any pain ratings from that patient.

Core Claim

Scale-use bias and pain sensitivity can each be identified at the individual level if the protocol adds two channels that ratings alone lack. The first is a rating channel whose input is roughly the same for everyone, which exposes scale use. The second is a scale-free sensory channel, which exposes sensitivity. Once identified, the reporting-style estimate should do three things. It should stay stable across sessions. It should move when reporting is manipulated but not when sensation is manipulated, and sensitivity should show the reverse pattern. And it should replace the labeled calibration sequences that current personalized facial pain models require.

We write the model compactly. Let $r_{ij}$ be person $i$'s rating of painful stimulus $j$ with physical intensity $s_j$. Let $a_k$ be a non-painful anchor, and let $f$ be a population-level function giving its felt intensity:

$$r_{ij} = g_i\big(h_i(s_j)\big) + \epsilon_{ij}, \qquad r^{a}_{ik} = g_i\big(f(a_k)\big) + \epsilon_{ik}.$$

Here $h_i$ is the person's sensitivity function, which maps stimulus to felt pain, and $g_i$ is their reporting function, which maps felt intensity to a number. We take $g_i(x) = \alpha_i + \beta_i x$, an offset plus a gain, with a saturation term at the scale ends. The forced-choice accuracy $P_i(\text{correct} \mid s, s')$ depends only on $h_i$. The model is fit as a hierarchical Bayesian model over people. The noise term is allowed to vary with the input, following the heteroscedastic modeling in Rudovic et al., because rating noise itself differs between people.

Why this is different from existing work

Several lines of prior work motivate the proposal.

Facial pain estimation research has documented the gap between expression and report. Personalized Automatic Estimation of Self-Reported Pain Intensity from Facial Expressions showed that an unpersonalized model predicts self-reported pain from facial landmarks with an intraclass correlation (ICC) of only 0.04 against the 0–10 visual analog scale (VAS). It repaired this with an expressiveness score, defined as the ratio of an observer's rating to the patient's VAS, which requires labeled sequences from each new patient. DeepFaceLIFT avoids per-patient labels by feeding in age, gender, and complexion. Those are coarse proxies that say nothing about a given person's scale use. Neither paper touches a controlled stimulus. Our proposal replaces both the label-based expressiveness score and the demographic proxies with a quantity measured directly against known physical inputs.

Dualé et al.'s French validation of the Pain Sensitivity Questionnaire plays a dual role: it both motivates and constrains the work. The questionnaire asks people to imagine everyday painful events. Its scores correlated with above-threshold heat and cold-pressor ratings but not with pain thresholds. In surgical patients they also tracked anxiety and catastrophizing. One reading is that the questionnaire partly measures reporting style rather than sensitivity. That makes it a direct and cheap competitor to our anchor-based estimate, and we treat it as one.

Pain: A Statistical Account models felt pain as a Bayesian estimate of threat and cites a red-versus-blue cue experiment in which an identical cold stimulus was rated as roughly twice as painful under a danger cue. That framework raises a question our design can answer and theirs cannot: does contextual threat change sensation (a shift in $h_i$) or only reporting (a shift in $g_i$)?

The domain-adaptation and distillation work on thermal-pain video, P-MSDA and PKDOT, treats differences between people as shifts in visual feature distributions, and uses stimulus level as the label. Neither separates the report from the stimulus. Self-report bias research outside pain supplies the manipulations we borrow. Examples include inflated well-being reports in both arms of a psychotherapy trial (ACT vs. MBSR trial) and image-protective underreporting (Bergen scale invariance study).

Research Program

We will collect a new dataset, because no dataset in the file contains calibrated stimuli, non-pain anchors, scale-free sensory tasks, facial video, and self-reports together. UNBC-McMaster has VAS ratings but no controlled stimuli. BioVid has calibrated heat but no per-trial self-report in the form we need.

The sample will be about 120 healthy adults, larger than Dualé's 85-person volunteer cohort. Each participant will complete two sessions one to two weeks apart. Each session takes about 40 minutes. It includes contact-thermode heat stimuli at six intensities spanning threshold to tolerance, punctate pressure with calibrated von Frey filaments, and loudness and brightness anchors each at six levels. All ratings use the same 0–10 scale. Each session also includes a two-interval forced-choice heat discrimination block and continuous facial video with skin conductance. Participants also complete the Pain Sensitivity Questionnaire and standard anxiety and catastrophizing scales.

In a subset of sessions we apply two manipulations. The reporting manipulation varies observation and scale framing: an investigator is present and states that "most people rate this around 3," versus an anonymous, unobserved condition. The sensory manipulation is topical capsaicin sensitization of the test site, which works through the same heat-sensing channel (TRPV1) that the spinal TRPV1 study links to heat hypersensitivity. A third manipulation, the danger-versus-safety cue from the Tabor account, is included as an open probe.

The model is fit per person with pooling across people. From the fits we derive a compact calibration protocol: the minimum number of anchor and forced-choice trials that recovers $\alpha_i$ and $\beta_i$ to within a set tolerance. The target is a protocol of under ten minutes. Finally, we train facial pain estimators and condition them on the estimated reporting-style parameters. The estimators are pretrained on UNBC-McMaster and BioVid and fine-tuned on our data with leave-one-subject-out splits.

Baselines

The naive reference is raw ratings taken at face value. For between-person analyses, sensitivity is simply each person's mean painful rating or rating-defined threshold. For facial estimation, it is an unpersonalized landmark regressor. On UNBC-McMaster, Martinez et al. report MAE 3.67 and ICC 0.04 for their unpersonalized model. We will re-establish this reference on our own data with the same splits.

The strongest existing personalization method with labels is the expressiveness-score approach of Martinez et al.. They report MAE 2.46 and ICC 0.34 on UNBC-McMaster. We give it its best case: one and two labeled sequences per test participant, taken from the same session as our calibration. The strongest label-free methods are DeepFaceLIFT, which reports MAE 2.18 and ICC 0.35 on UNBC-McMaster, and the Riemannian motion-dynamics regressor of Automatic Estimation of Self-Reported Pain by Interpretable Representations of Motion Dynamics, which reports MAE 2.31 under leave-one-subject-out. We reimplement both on our data with the same backbone and the same hyperparameter search budget, which is equal trials per method on validation subjects only.

Beating the baselines means two things. First, zero-shot calibration must reduce MAE and raise ICC relative to DeepFaceLIFT on our held-out participants. Second, it must at least match the one-sequence expressiveness score without using any pain label from the test participant.

Three ablations isolate the key ingredient. The first removes the anchors, so style is inferred from painful ratings and forced choice alone. The second removes forced choice. The third replaces our style estimate with the Pain Sensitivity Questionnaire score, anxiety, and catastrophizing entered as covariates.

For between-person correction, the reference methods are within-person z-scoring and the questionnaire-as-covariate model. The target is agreement with scale-free sensitivity. Corrected pain ratings should correlate with forced-choice discrimination thresholds more strongly than raw ratings or questionnaire-adjusted ratings do, and the comparison is tested on held-out stimuli.

Why not just…?

Why not just z-score each person's ratings? Within-person standardization removes offset and gain for free, and for tracking one person's change over time it may be all that is needed. We say so plainly and will not claim an advantage there. It fails for between-person questions, because it erases true sensitivity differences along with scale use. Everyone ends up with mean zero, including a person with nerve-driven hypersensitivity. The decisive comparison is agreement with forced-choice thresholds across people. There, z-scoring should do no better than chance, while the anchor-corrected ratings should do better.

Why not just use the Pain Sensitivity Questionnaire as the reporting-style measure? This is the real threat. The questionnaire takes minutes, involves no pain, has test-retest ICC of about 0.63, and in Dualé et al. it tracked above-threshold ratings and distress rather than thresholds, which is exactly what a reporting-style measure would do. If the questionnaire score explains as much variance in $\alpha_i$ and $\beta_i$ as the anchors do, and matches them in facial personalization, the right conclusion is that the questionnaire is the cheap calibration tool. The protocol would then shrink to it plus a forced-choice block. The ablation that swaps our style estimate for the questionnaire settles this directly. Our concern is that imagined scenarios mix imagined sensitivity with scale use, so we expect the questionnaire to capture only part of $g_i$. That expectation is exactly what the ablation tests.

Why not just collect one labeled pain sequence per patient, as the expressiveness-score method does? A single VAS rating is cheap in an alert adult. It is unavailable at the moments that matter, for example in a sedated or non-verbal patient, and it carries the same reporting bias it is meant to correct. The comparison in the Baselines section gives this method its best case. If one labeled sequence beats our calibration by a wide margin even in healthy adults, the zero-shot argument reduces to a feasibility argument for non-verbal settings, and we will report it that way.

Why not just use skin conductance or other physiological signals as the objective readout? Autonomic responses have their own person-specific gain, and they are driven by arousal and anxiety as well as nociception. They substitute one unknown per-person transform for another. We record them anyway and test whether they add information beyond the reporting-style estimate in the facial model.

Decisive Checks

The central test is a double dissociation. The reporting manipulation should shift $\alpha_i$ and $\beta_i$, estimated from both anchors and painful ratings, while leaving forced-choice heat thresholds unchanged. Capsaicin should lower forced-choice thresholds and shift $h_i$ while leaving anchor ratings unchanged. If the observation manipulation moves forced-choice thresholds, or capsaicin moves loudness ratings, the separation fails and the model is wrong.

Three further checks support the separation. The first is style estimated from the loudness anchor and from the brightness anchor should agree, which tests the assumption that anchors are felt alike across people. The second is style estimated from heat stimuli should predict ratings of pressure stimuli, because scale use should be shared across pain types while sensitivity need not be. The third is style parameters should be more stable across the two sessions than the questionnaire's 0.63.

The threat-cue probe comes last. Whether a danger cue moves $h_i$ or only $g_i$ is a result either way, and it would give the Bayesian threat account a measurement it currently lacks.

The applied payoff is the zero-shot facial personalization comparison against DeepFaceLIFT and the expressiveness score described above. A useful secondary analysis tests whether our style estimate is related to demographics. If reporting style differs across groups, that would bear directly on whether sensitivity-referenced measurements could reduce the treatment disparities that Q-Pain found in language-model analgesia recommendations: a 7% higher denial rate for Black women. We leave that application for later work.