Phase-Locked Control of Pain's Affective Channel

Pain is not one signal. A growing body of work shows that the physical registration of a noxious stimulus — where it is, how strong it is, and the reflex that pulls your hand away — runs through different circuits than the suffering it causes. In mice, silencing a small ensemble of basolateral amygdala neurons abolishes pain-motivated avoidance while leaving withdrawal reflexes and sensory detection fully intact (An amygdalar neural ensemble that encodes the unpleasantness of pain). In humans, scalp EEG separates an early, phase-locked gamma response that tracks stimulus intensity from a late, induced gamma response that tracks subjective unpleasantness and is shifted by emotional context (Gamma Band Oscillations Reflect Sensory and Affective Dimensions of Pain). Meanwhile, a separate literature on phase-locked neuromodulation has shown that delivering stimulation at the right moment of an ongoing brain rhythm — rather than continuously — can causally gate perception and suppress pathological rhythms, as in the 21.5% tremor reduction achieved only when alternating-current stimulation was locked to the tremor rhythm's phase (Non-invasive suppression of essential tremor via phase-locked disruption of its temporal coherence).

These two literatures have never been joined. No study has delivered stimulation phase-locked to the endogenous rhythm of an affective pain hub with the goal of selectively switching off suffering while leaving sensory detection untouched. The closest works stop short on every axis: Shirvalkar et al. propose closed-loop deep brain stimulation across sensory, affective, and cognitive pain nodes but only as an untested conceptual protocol, with no sub-cycle phase targeting and no data (Closed-Loop Deep Brain Stimulation for Refractory Chronic Pain); Park et al. deliver a rhythm-mimicking ultrasound waveform to the spinal cord but entirely open-loop and without isolating the affective dimension (Harnessing Theta-Gamma Coupled Brainwave Using Ultrasound for Spinal Astrocyte Revitalization and Sustained Neuropathic Pain Relief); and Ryu et al. record intracranial signals in awake chronic pain patients but never stimulate (Spontaneous pain dynamics characterized by stochasticity in awake human LFP with chronic pain). This proposal fills that gap: it asks whether pain unpleasantness is a phase-controllable quantity.

Core Claim

The claim is that the temporally regular theta/alpha (4–12 Hz) activity that Ryu et al. found tracks spontaneous pain in human affective hubs — regularity measured by inter-event-interval auto-mutual information correlated with momentary pain ratings at $\tau = 0.50$, $p < 0.01$ — is not merely a correlate but a control variable: stimulation delivered at a calibrated phase of that rhythm can disrupt the affective channel of pain while sparing the sensory channel. Two facts make this plausible and one fact makes it dangerous. It is plausible because phase-locked stimulation has already produced causal, phase-specific effects in perception and motor pathology: identical magnetic pulses at the peak versus trough of an alpha cycle produce opposite perceptual outcomes (The Phase of Ongoing Oscillations Mediates the Causal Relation between Brain Excitation and Visual Perception; The Phase of Ongoing EEG Oscillations Predicts Visual Perception). It is dangerous because Mirkhani et al. proved that in vivo pathological networks sit at only 5–10% of theoretical maximum synchrony, and that under such conditions mistimed pulses amplify rather than suppress the rhythm (Response of Neuronal Populations to Phase-Locked Stimulation: Model-Based Predictions and Validation) — and Taylor et al. showed computationally that an improperly timed single pulse can extend a seizure rather than abort it (Abatement of epileptic spike-wave discharges through single pulse stimulation). So the proposal is calibration-first: before stimulating anyone, we estimate each subject's own phase-response properties and use them to select the provably suppressive phase.

The design principle comes from Mirkhani et al., who validated in Parkinsonian rats that the amplitude response to a pulse at phase $\theta$ is predicted by the slope of the phase response curve:

$$\text{ARC}(\theta) \propto -\frac{d}{d\theta}\,\text{PRC}(\theta), \qquad \theta^{*} = \arg\max_{\theta}\,|\text{ARC}(\theta)| \text{ with } \text{ARC}(\theta) < 0$$

where PRC is the measured shift in rhythm phase caused by a pulse, ARC is the resulting amplitude change, and $\theta^{*}$ is the phase at which a pulse maximally shrinks the rhythm's amplitude. Our controller selects $\theta^{*}$ per subject, per session, and re-estimates it continuously as the rhythm drifts — directly attacking the phase-drift problem Schreglmann et al. documented but did not solve.

Research Program

The program has three linked components, moving from offline design to causal tests in humans and rodents.

First, offline controller design and simulation on existing data. We will refit the awake-human LFP recordings of Ryu et al. (five patients, sensory thalamus, subgenual cingulate, and periaqueductal gray, with continuous pain ratings) with Kuramoto-type oscillator models, estimate each subject's PRC from spontaneous event timing, and verify the ARC–PRC-derivative relationship that Mirkhani et al. confirmed with a mean correlation of $R = -0.84$ across 12 of 13 rats. We will then simulate three controllers on these fitted models — phase-locked, power-thresholded (stimulate when band power is high, at uncontrolled phase), and asynchronous open-loop — and quantify the fraction of simulated pulses that amplify rather than suppress the pathological regularity. This component also builds the machine-learning pieces: an adaptive phase tracker (the endpoint-corrected Hilbert transform of Schreglmann et al. reduced phase estimation error from $47^\circ$ to $9^\circ$, but has no drift auto-tuning; we add online recalibration), and a responder-prediction model that classifies which subjects' rhythms are regular enough to be phase-trackable, following the approach Schreglmann et al. used to predict tremor-treatment responders with 97% accuracy from baseline regularity features.

Second, an interventional human experiment in acute experimental pain. We extend the paradigm of Lyu et al. — emotional picture priming followed by painful cutaneous stimulation, with separate intensity and unpleasantness ratings — by adding real-time phase-locked transcranial stimulation timed to the pre-stimulus theta/alpha phase at midline frontal and centroparietal sites overlying the medial pain network. The critical readout is the double dissociation they established: early gamma amplitude correlates with perceived intensity at Spearman $\rho = 0.608$ and is invariant to emotional priming, while the late gamma difference between negative and positive primes correlates with unpleasantness at $\rho = 0.511$. Our prediction is that suppressive-phase stimulation selectively reduces the late, induced gamma response and unpleasantness ratings, leaving the early phase-locked gamma response and intensity ratings unchanged. We deliberately phase-lock to the slow 4–12 Hz carrier rather than to gamma itself, because a single 80 Hz gamma cycle lasts 12.5 ms while sensor-processing latencies run 5–10 ms; Tan et al. showed that affective hubs coordinate fast gamma activity through the slower phases of beta rhythms in exactly this carrier sense (Intracranial EEG signals disentangle multi-areal neural dynamics of vicarious pain perception), and embedded hardware such as the NeuralTree chip, with a 7.8 μs front-end settling time at 0.227 μJ per classification, makes the real-time loop feasible (NeuralTree: A 256-Channel 0.227μJ/class Versatile Neural Activity Classification and Closed-Loop Neuromodulation SoC).

Third, a rodent causal test in chronic neuropathic pain. In spared-nerve-injury mice, we record anterior cingulate cortex LFP and deliver closed-loop, phase-locked stimulation of the ACC — either electrical, or optogenetic targeting of the monosynaptic ACC-to-subthalamic-nucleus projection that Wang et al. showed drives both mechanical allodynia and depressive-like behavior when hyperactive (The anterior cingulate cortex controls the hyperactivity in subthalamic neurons in male mice with comorbid chronic pain and depression). Readouts follow the dissociation logic of Corder et al.: von Frey mechanical thresholds, operant avoidance along a thermal gradient, reflex withdrawal, and depressive assays. The target profile is theirs — avoidance and affective readouts normalized, reflex withdrawal and sensory detection intact — but achieved with a reversible, moment-bounded electrophysiological intervention rather than chronic chemogenetic silencing. Importantly, allodynia in this model normally recruits the amygdala unpleasantness ensemble (Corder et al.), so we will additionally image BLA ensemble activity to test whether phase-locked ACC control propagates downstream to the ground-truth affective encoder.

Proposed Experiments

The human experiment uses the IAPS picture-priming paradigm with painful cutaneous stimulation, in a within-subject design crossing prime valence (negative, neutral, positive) with stimulation timing (suppressive phase $\theta^{*}$, amplifying antiphase, phase-scrambled, no stimulation). Primary metrics are late gamma amplitude and unpleasantness ratings; secondary metrics are early gamma amplitude, intensity ratings, and inter-trial phase coherence. The decisive check is an interaction: a reduction of late gamma and unpleasantness at $\theta^{*}$ with early gamma and intensity statistically unchanged, and no such pattern under phase-scrambled stimulation at matched energy. The rodent experiment's decisive check is the Corder-profile dissociation, plus the downstream BLA imaging result. A pre-registered negative control matters equally: if stimulation at the amplifying phase worsens avoidance behavior, the phase-specificity claim is confirmed in both directions; if neither phase does anything, the rhythm is a correlate, not a mechanism.

For decoding-based readouts in the rodent and any human pilot with intracranial access, we will use the feature-ensemble approach of Tan et al., whose support-vector-machine decoding of vicarious pain reached 76.28% accuracy and required multi-regional coupling features, not single-node power — a reminder that the affective channel is distributed and the controller should be evaluated against network-level features.

Baselines

The do-nothing reference point is spontaneous fluctuation of the pain state, anchored by Ryu et al.'s observational finding that regularity tracks subjective pain at $\tau = 0.50$, $p < 0.01$ in their five-patient cohort; our intervention beats this baseline only if it actively moves pain ratings or regularity in the predicted direction within a session, not merely predicts them. The strongest existing method baseline is open-loop rhythmic stimulation: continuous, non-phase-locked delivery at the same carrier frequency and the same total energy per session. Its best-documented performance in neuropathic pain is Park et al.'s open-loop ultrasound waveform, which raised mechanical thresholds from 14% to 67% of baseline with relief persisting five days — but that was spinal, reflex-based, and did not isolate the affective dimension, so for the human affective readout no reported result found on Lacuna for open-loop stimulation effect on late pain gamma or unpleasantness ratings; we will generate that number ourselves under matched energy. The second method baseline is power-thresholded closed-loop control in the spirit of Shirvalkar et al.'s conceptual framework — stimulate when band power exceeds a threshold, with phase uncontrolled — run with the same detection latency and tuning budget. The key ablation isolates our ingredient: the identical hardware, tracker, and energy budget, but with pulse phases randomly scrambled; any effect that survives scrambling belongs to the rhythm per se, not to phase-locking. Finally, a pharmacological reference profile comes from Ni et al., who showed that duloxetine microinjected into the ACC relieved sensory hypersensitivity without improving comorbid anxiety (The anterior cingulate cortex contributes to the analgesic rather than the anxiolytic effects of duloxetine in chronic pain-induced anxiety); our target is precisely the opposite profile, which makes it a useful directional contrast. All baselines run on the same subjects, same sessions where possible, same per-subject tuning budget (phase and threshold parameters cross-validated on calibration blocks), and same delivered energy.

Why not just…?

Why not just stimulate continuously, as clinical DBS already does? Because the literature gives specific evidence that unphased pulses waste and sometimes invert their effects: Schreglmann et al. found non-phase-locked stimulation completely ineffective for tremor while phase-locked stimulation achieved a 21.5% amplitude reduction; Taylor et al. showed a mistimed single pulse can reinforce a pathological rhythm; and Mirkhani et al. showed asynchronous pulses land at random phases whose effects cancel. That said, Askari et al. did report clinical pain improvement from conventional continuous subthalamic DBS in Parkinson's patients (Dorsal subthalamic deep brain stimulation improves pain in Parkinson's disease), so this is a real threat, not a straw man. We settle it empirically: the open-loop matched-energy arm is a first-class baseline, and if it matches the phase-locked arm on the affective readouts, the honest conclusion is that phase-locking is unnecessary for pain — a finding worth publishing either way.

Why not just threshold on band power and trigger stimulation, skipping sub-cycle timing? Because the pain signal that matters is temporal regularity, not amplitude: Ryu et al. showed their regularity measure outperformed power spectral density for tracking subjective pain and opioid relief, meaning a power threshold fires at the wrong moments. And Dugué et al. established that excitability swings between opposite states within a single cycle, so a power-triggered burst still lands at an arbitrary phase with an unpredictable sign of effect. Our power-thresholded baseline, run at matched detection latency, tests this directly; the simulation study additionally predicts its amplification rate before any human is stimulated.

Why not just silence the affective encoder with drugs or chemogenetics? Corder et al. proved this works superbly in mice — avoidance abolished, reflexes intact — and Zhong et al. and Wang et al. each found molecular switches that alter pain-related anxiety without shifting sensory thresholds (Activation of Anterior Cingulate Cortex Extracellular Signal-Regulated Kinase-1 and -2 (ERK1/2) Regulates Acetic Acid-Induced, Pain-Related Anxiety in Adult Female Mice; Differential roles of hippocampal glutamatergic receptors in neuropathic anxiety-like behavior after partial sciatic nerve ligation in rats). But these tools are not human-translatable on any near horizon, they act over minutes to weeks rather than within a single cycle, and pharmacology can cut the wrong way — Ni et al.'s duloxetine relieved the sensory channel while leaving the anxiety untouched. The scientific question the rodent experiment answers is whether the dissociation Corder achieved chemogenetically can be reproduced by reversible phase-locked control of the upstream rhythm. If it cannot, that is strong evidence the 4–12 Hz regularity is epiphenomenal, and the field should stop treating it as a control target.

Decisive Checks

Three results would make or break the claim. First, the simulation must show that a per-subject calibrated phase-locked controller suppresses the fitted pathological rhythm with a materially lower amplification rate than asynchronous or power-thresholded control — otherwise the Mirkhani synchrony bound (5–10% baseline synchrony) makes the whole enterprise too risky to take to humans. Second, the human experiment must show the double dissociation on Lyu et al.'s readouts: late gamma and unpleasantness down at $\theta^{*}$, early gamma and intensity unchanged. Third, the rodent experiment must reproduce the Corder profile with BLA ensemble activity as the downstream witness. One methodological caution we take from the generalization literature: cross-participant EEG models collapse without individualized calibration (a within-participant accuracy of 0.9864 for a shallow classifier dropped to 0.57004 cross-participant in Towards Generalizable Learning Models for EEG-Based Identification of Pain Perception), so every controller here, including the phase tracker and responder model, is calibrated per subject on blocked initial data and evaluated on held-out stimulation blocks only.

References

  1. An amygdalar neural ensemble that encodes the unpleasantness of pain
  2. Gamma Band Oscillations Reflect Sensory and Affective Dimensions of Pain
  3. Non-invasive suppression of essential tremor via phase-locked disruption of its temporal coherence
  4. Closed-Loop Deep Brain Stimulation for Refractory Chronic Pain
  5. Harnessing Theta-Gamma Coupled Brainwave Using Ultrasound for Spinal Astrocyte Revitalization and Sustained Neuropathic Pain Relief
  6. Spontaneous pain dynamics characterized by stochasticity in awake human LFP with chronic pain
  7. The Phase of Ongoing Oscillations Mediates the Causal Relation between Brain Excitation and Visual Perception
  8. The Phase of Ongoing EEG Oscillations Predicts Visual Perception
  9. Response of Neuronal Populations to Phase-Locked Stimulation: Model-Based Predictions and Validation
  10. Abatement of epileptic spike-wave discharges through single pulse stimulation
  11. Intracranial EEG signals disentangle multi-areal neural dynamics of vicarious pain perception
  12. NeuralTree: A 256-Channel 0.227-μJ/Class Versatile Neural Activity Classification and Closed-Loop Neuromodulation SoC
  13. The anterior cingulate cortex controls the hyperactivity in subthalamic neurons in male mice with comorbid chronic pain and depression
  14. The anterior cingulate cortex contributes to the analgesic rather than the anxiolytic effects of duloxetine in chronic pain-induced anxiety
  15. Dorsal subthalamic deep brain stimulation improves pain in Parkinson's disease
  16. Activation of Anterior Cingulate Cortex Extracellular Signal-Regulated Kinase-1 and -2 (ERK1/2) Regulates Acetic Acid-Induced, Pain-Related Anxiety in Adult Female Mice
  17. Differential roles of hippocampal glutamatergic receptors in neuropathic anxiety-like behavior after partial sciatic nerve ligation in rats
  18. Towards Generalizable Learning Models for EEG-Based Identification of Pain Perception