Physiologically Coupled Augmentation for Nociception Monitoring from Biosignals

Machines that read pain from the body's signals (heart rhythm, sweat-gland activity, muscle electrical activity) are what closed-loop, opioid-sparing anesthesia would need. These models are starved of data exactly where it matters: strong, rare pain responses. The usual fix for scarce data is augmentation, meaning the creation of modified copies of training examples. Existing multimodal augmentation was built for images paired with text or audio, and it ignores the fact that a single painful stimulus drives several body signals at once, each with its own delay. We propose an augmentation that separates each recording into a stimulus-locked response and a subject-specific background. It then creates new training examples by moving a pain response, intact and time-aligned across all channels, onto another person's background activity, and by jointly shifting its timing and strength within bounds the body actually allows.

The clinical motivation is specific. Opioid-free and opioid-sparing regimens that combine ketamine, dexmedetomidine and regional blocks reduce opioid use by roughly 40% during surgery and 55% after it (Opioid sparing anesthesia in patients with liver cirrhosis). Even so, doses are fixed or adjusted by heart rate and blood pressure, with no direct reading of whether nociceptive input is getting through. Reviews of opioid-free anesthesia name the missing real-time titration signal as a barrier to adoption (Research Progress on the Clinical Application of Opioid-Free Anesthesia). The case for pre-emptive dosing, which aims to block the spinal "wind-up" that turns surgical pain into lasting hypersensitivity, has the same gap. Nothing in current practice checks whether the pre-emptive block actually held. Under general anesthesia the face is uninformative and muscles may be pharmacologically paralyzed, so any such monitor must work mainly from autonomic signals. That is the regime where data scarcity is most severe.

Core Claim

Our claim is that, for physiological pain data, the right unit of augmentation is the stimulus-evoked response shared across channels, not the individual channel and not the whole modality. We model a trial in channel $c$ as

$$x_c(t) = a_c\, r(t - \tau_c) + b_c(t),$$

where $r(t)$ is a shared evoked response waveform, $a_c$ and $\tau_c$ are that channel's gain and delay (for example, skin conductance responds seconds after heart rate), and $b_c(t)$ is ongoing background activity unrelated to the stimulus. The response is estimated per trial by aligning to stimulus onset. The background is estimated from the pre-stimulus baseline and from baseline trials. Augmentation then has three moves. First, response transplant pastes a high-intensity trial's response onto another subject's background at the same onset. Second, coupled warping rescales the shared $r$ in time and amplitude for all channels together. Third, bounded lag jitter perturbs each $\tau_c$ only within the range of delays observed in real data for that channel pair. The label follows the response, so class semantics are preserved. Cross-subject variation, which dominates leave-one-subject-out error, is added deliberately.

This differs from prior augmentation in a way we can test. MIDAS (MIDAS) swaps entire modalities between samples to create deliberate conflict and relies on soft labels to handle it. We swap only the part of the signal that carries no label, so no conflict is created. LeMDA (Learning Multimodal Data Augmentation in Feature Space) perturbs fused embeddings with an adversarially trained autoencoder and has no notion of time or delay. Joint mixup (Wang et al.) averages two inputs with one shared mixing ratio. For evoked responses with different onsets and lags, that averaging blurs the responses into a signature no real body produces. Geometric synchronization methods (Zang et al.; Ma et al.) supply the principle we borrow: whatever changes in one stream must be reflected consistently in its partner streams. Here the consistency is temporal and physiological rather than spatial.

Research Program

The first stage builds the decomposition and validates it as a signal model before using it for augmentation. On BioVid Part A, which has ECG, facial EMG, skin conductance and video recorded during graded heat stimuli across 87 subjects, we estimate $r$, $a_c$, $\tau_c$ and $b_c$ per trial. We check two things: that transplanted trials reproduce the empirical cross-channel lag and amplitude statistics of real trials, and that a held-out real-versus-synthetic discriminator performs near chance. A failure here sends us back to the model before any downstream claim.

The second stage trains pain-intensity classifiers with and without the augmentation under leave-one-subject-out evaluation. We run two input regimes. The full regime includes video. The physiology-only regime keeps ECG and skin conductance, with EMG optional, and stands in for the anesthetized patient. The analgesia setting enters as a stress test rather than a claim of clinical validation. We build test sets in which a single channel group is blunted in the way a drug class would blunt it: sympathetic blunting of heart rate and sweat response for an α2 agonist, loss of muscle signal for paralysis. We then ask whether models trained with coupled augmentation degrade more gracefully than models trained with channel dropout. These perturbations are never used in training, so the test is not circular.

The third stage transfers the method to EmoPain (EMOPAIN Challenge 2020), which has surface EMG, motion capture and video. There the "stimulus" is a movement onset rather than a heat pulse, and the rare class is protective behavior. This tests whether the response-plus-background decomposition generalizes beyond laboratory-evoked pain. As an exploratory analysis connected to sensitization, we will use BioVid's repeated stimuli to check whether estimated response gains drift upward within a session, which would be a crude behavioral analogue of temporal summation. We will report it only descriptively.

Baselines

The naive reference is the same encoder trained with no augmentation and with inverse-frequency class weighting. Class weighting is the cheapest imbalance fix, so any gain from our method must exceed it.

Independent per-channel augmentation applies jitter, scaling and time-warping to each channel separately. This is standard practice for biosignals and the default a reviewer would reach for. Joint mixup, following Wang et al., applies one ratio to all channels. Channel or modality dropout is included because it is the simplest robustness method for missing channels, and it is the most dangerous competitor on the physiology-only and blunted-channel tests. MIDAS and LeMDA are the strongest published multimodal augmentation methods. On Kinetics-Sounds, MIDAS reports 74.88% accuracy against 63.92% for joint training (MIDAS). LeMDA reports about a 6% accuracy gain on Hateful Memes and Petfinder (LeMDA). Neither has been run on biosignals, and we will port both.

For pain-specific reference points, the file offers the following. PKDOT reports a 4.6% absolute gain over a visual-only lower bound on BioVid Part B, where it distills physiological signals into a video student (Distilling Privileged Multimodal Information). Progressive multi-source domain adaptation reports 0.88 average accuracy, against 0.70 for standard multi-source adaptation (Progressive Multi-Source Domain Adaptation). EmoPain's Stacked-LSTM reaches an average F1 of 0.57 but only 0.24 on the protective class. We found no reported result on Lacuna for a physiology-only (ECG plus skin conductance) leave-one-subject-out classifier on BioVid. Our no-augmentation model therefore defines that reference, and we report it.

Fairness conditions apply to every method. All methods share the same encoder, the same subject splits, the same number of training epochs, and an equal hyperparameter search budget of 20 configurations each. Augmentation methods also get matched counts of synthetic samples.

Two ablations isolate the key ingredient. In the first, transplanted responses get independent random onsets per channel, which breaks synchrony while keeping everything else. In the second, the response is kept but the background is not swapped, which removes the cross-subject variation.

Our criteria for beating the baselines are stated in advance. On BioVid no-pain versus highest-intensity classification, we require a gain over the best baseline whose confidence interval across subjects excludes zero. On EmoPain, we require a clear improvement over 0.24 protective-class F1 relative to our own reimplemented baseline. On the blunted-channel stress test, we require smaller degradation than channel dropout.

Why not just…?

Why not just augment each channel independently? LeMDA's authors found that uncoordinated augmentation of paired inputs can hurt multimodal models (LeMDA). For biosignals, the specific risk is that independent time-warps destroy the heart-rate-to-skin-conductance lag, which is itself evidence of an evoked response. The synchrony-breaking ablation tests this directly. If independent augmentation matches our method, then coupling does not matter and the claim fails.

Why not just use mixup or LeMDA? Mixup averages misaligned responses, and LeMDA's latent perturbations cannot represent a delay change. We expect both to help on average accuracy but not on the rare high-intensity class. That expectation is testable, and LeMDA's roughly 6% gains on static benchmarks mean it could surprise us. If it matches our method at lower engineering cost, we will say so.

Why not just use modality dropout for robustness to anesthesia? This is the real threat for the stress test, because dropping channels at random directly teaches the model to cope when a channel is missing. Our bet is that drugs attenuate channels rather than delete them. A model that has seen only missing-or-intact channels may misread a half-blunted skin conductance signal as low pain. If dropout matches us on the blunted tests, we will narrow our claim to data scarcity and rare-class performance, where dropout adds no new response examples.

Why not just oversample rare classes? Oversampling repeats the same few high-pain responses from the same few subjects. The background-swap ablation separates "more copies of the response" from "responses seen against new people." Oversampling amounts to the no-swap condition without warping.

Decisive Checks

The idea stands or falls on three results. First, under leave-one-subject-out evaluation on BioVid in the physiology-only regime, coupled augmentation must beat channel dropout, MIDAS and LeMDA on the highest-intensity class, and the synchrony-breaking ablation must lose most of that gain. Second, the transplanted samples must preserve real cross-channel lag statistics and fool a real-versus-synthetic discriminator. Third, the gains must carry over to EmoPain's protective class, where the evoked event is movement rather than heat. If the first result holds but the synchrony ablation does not hurt, the contribution is only cross-subject background swapping, and we will report it as that. A clinical, closed-loop evaluation during real multimodal analgesia is deliberately outside this proposal. What this work supplies is the evidence needed to justify building one.