Delay Budgets from the Brain: Statistical Wireless Guarantees for Closed-Loop Pain-and-Mood Neuromodulation

A closed-loop brain stimulator for people with both chronic pain and depression has to listen to several brain regions, decide what state the patient is in, and adjust stimulation. If the sensing and stimulating parts are separate implants that talk wirelessly, every measurement becomes a packet that must arrive before its deadline, sent from a device with almost no energy. Such links are usually designed for the worst case: every stream is treated as equally urgent and provisioned for the deepest radio fade. This proposal argues that the brain gives a better design rule. Pain has a fast sensory component, which tracks where and how much it hurts, and a slower affective component, which tracks suffering and mood. These run through partly separable circuits, so a controller's two decisions have very different tolerance for late or lost data. We propose to turn that tolerance into explicit per-stream delay-violation budgets. The wireless link would then serve each stream only as strictly as the neural decision requires, using opportunistic access to idle medical radio spectrum to fill the gaps.

The tool that makes this precise is effective capacity: the largest steady data rate a fading radio channel can carry while keeping the probability of exceeding a delay bound below a target. Its single knob is a "QoS exponent" θ. A θ near zero means delay-tolerant, and the capacity approaches the ordinary average capacity. A very large θ means delay-intolerant, and the capacity collapses toward the rate that survives the worst fade. Our claim is that sensory-pain streams belong at high θ and affective streams at low θ. We also claim that a patient's preoperative diffusion MRI, which maps white-matter wiring, is a usable prior on which recording sites carry redundant information. Taken together, these should let a distributed implant hold the same decoding accuracy at markedly lower transmit energy than uniform, worst-case provisioning.

Core Claim

Four directions meet here, each with a different role. Pain-mood neurobiology supplies the application and the key asymmetry. Effective capacity is the core allocation mechanism. Spectrum underutilization is the resource being exploited. Diffusion tensor imaging (DTI) is a prior on cross-site redundancy.

The asymmetry is well supported in animals. Blocking ERK1/2 signaling in the anterior cingulate cortex (ACC) removed pain-related anxiety but left the writhing response to pain intact (Zhong et al.). Duloxetine injected into the ACC relieved allodynia without reducing anxiety (Ni et al.). Restoring hippocampal NMDA/ERK signaling eased neuropathic anxiety without rescuing mechanical hypersensitivity (Wang et al.). These three studies motivate treating the sensory and affective read-outs as separable outputs.

Other work constrains the design, because the two components are not independent. An ACC-to-subthalamic projection drives both hyperalgesia and depression-like behavior (Wang et al. 2024). Nucleus reuniens-to-mPFC and ventral pallidal cholinergic circuits do the same (RE-mPFC, Ji et al.). The allocator therefore cannot treat the streams as independent, and the decoder has to model their coupling.

The timescales also differ in humans. Intraoperative recordings of local field potentials (LFPs, the summed electrical activity near an electrode) show that moment-to-moment pain ratings track theta/alpha regularity on a scale of seconds (Ryu et al.). Depression recovery under cingulate stimulation, by contrast, unfolds over weeks (Alagapan et al.).

Formally, for stream $k$ with per-slot service $S_k$ (bits delivered in one slot), the effective capacity is

$$E_c^{k}(\theta_k) = -\frac{1}{\theta_k}\log \mathbb{E}\left[e^{-\theta_k S_k}\right].$$

The delay-violation probability behaves approximately as $P(D_k \ge D_k^{\max}) \approx e^{-\theta_k \mu_k D_k^{\max}}$, where $\mu_k$ is the service rate. This relation lets us map a decoder-derived tolerance $\varepsilon_k$ at deadline $D_k^{\max}$ to a required $\theta_k$.

The allocator then minimizes total implant transmit energy subject to $E_c^{k}(\theta_k) \ge r_k$ for every stream $k$, where $r_k$ is the stream's data rate. Opportunistic access to idle spectrum enters as extra service slots whose availability is random. The contribution is not the optimization, which follows existing effective-capacity power control. The contribution is the procedure that produces $\varepsilon_k$ and $D_k^{\max}$ from the neural decoder and from tractography, together with the evidence that this procedure pays off.

Relation to Prior Work

Effective capacity is mature within wireless research. Cui et al. modeled LTE licensed-assisted access sharing the 5 GHz band with Wi-Fi. They reduced a four-state contention model to an ON-OFF service process and proved that effective capacity is concave in transmit rate, which made joint power and rate control a convex problem. Zhang et al. derived closed forms for spectrum-sharing secondary users with multi-antenna combining under Nakagami fading. Huang et al. showed that energy efficiency and delay bounds can be reconciled in multi-carrier systems. Abdel Khalek et al. came closest to our idea: they tied delay budgets to user-perceived video quality rather than to signal-to-noise ratio, and supported far more users as a result.

We reuse their machinery, especially Cui's concavity result. We differ in where θ comes from. In those papers θ is an application parameter set by hand. Here it is estimated from how a neural state decoder's accuracy degrades under delay and loss, separately for sensory and affective outputs. None of those works considers implants or biological sources, whose traffic statistics change with the patient's state.

On the spectrum side, the scarcity paradox reported by Liu et al. is the resource argument: most licensed sub-6 GHz spectrum is idle at any given place and time. Anwar et al. is a hard constraint on the design. Their power-adaptive underlay access beats fixed-power schemes, but secondary throughput falls sharply as the energy budget tightens from 0.1 J toward $10^{-7}$ J, which is the regime implants live in. Kassab et al. handle event-triggered sharing with multi-agent reinforcement learning. That approach is the natural learned alternative, and we treat it as a baseline.

On the imaging side, Alagapan et al. found that lower fractional anisotropy in the cingulum bundle slowed recovery. Tiruvadi et al. showed that stimulating white matter near the subcallosal cingulate acts through tract-connected distant cortex rather than locally. Both motivate using tract connectivity as a prior on which sites carry shared signal. The closed-loop design in Closed-Loop DBS for Refractory Chronic Pain proposes multi-site sensing across S1, dorsal ACC and orbitofrontal cortex. It also warns that stimulation artifacts corrupt the recordings, which is exactly when a data-free structural prior becomes valuable. It does not address the communication link at all.

Research Program

The work has three parts. The first measures the neural delay tolerance. From multi-site LFP recordings, we train a decoder with two coupled output heads, one for sensory pain intensity and one for affective state. The inputs are band-power and inter-event-regularity features, following Ryu et al.'s auto-mutual-information biomarker. We then replay the feature streams through a simulated link, injecting delay and loss, and trace each head's accuracy against deadline and violation probability. This yields empirical $(D_k^{\max}, \varepsilon_k)$ pairs at a fixed tolerated accuracy loss.

The second part builds the structural prior. From each patient's diffusion MRI, we compute tract-level connectivity and fractional anisotropy between recording sites. We then fit a model mapping structural coupling to observed cross-site redundancy, measured as conditional mutual information between site features. The purpose is to predict, before any data are recorded, which streams can tolerate loss because another site partly covers them. The pipeline is pretrained on TractoInferno-scale data, 284 subjects in the ATM work. Because crossing fibers distort single-tensor estimates, the prior is only allowed to relax budgets, never to tighten them past the data-driven value.

The third part solves the per-stream energy minimization. It uses Cui-style concave effective-capacity constraints with an ON-OFF model for opportunistic medical-band access. We also add a correction for imperfect channel-state information, since existing closed forms assume perfect channel knowledge and implants will not have it.

Baselines

The naive reference is uniform worst-case provisioning. Every stream gets the same large θ, set by the tightest sensory deadline, on a dedicated channel with no opportunistic access. This is how a cautious device engineer would ship a system today. We run it on the same replayed neural traffic, channel traces and energy model. Beating it means equal decoder accuracy on both heads at lower energy per decoded decision.

The strongest analytical competitor is uniform-θ effective-capacity power control in the style of Cui et al., with opportunistic spectrum included. In their LTE/Wi-Fi setting they report a 62.7% effective-capacity gain over water-filling under tight delay limits and a 171.4% energy-efficiency gain. Those numbers come from a different channel and traffic regime, so they calibrate expectations rather than serve as a direct target. For implant-grade energy budgets on neural traffic we found no reported result on Lacuna. Our method must beat this baseline at matched decoder accuracy, and the gain must come specifically from the heterogeneous θ.

The strongest learned competitor is MADSPG from Kassab et al., trained with a reward that penalizes decoder error. It is given the same traffic and a tuning budget in simulator-hours equal to what our method spends estimating tolerance curves. In their four-agent setting they report a 92% average sum event rate, against 60–80% for independent Q-learning. We must match its accuracy-energy tradeoff at a small fraction of its on-device compute and memory.

Two ablations isolate our ingredients. The first removes the sensory-affective split: one pooled tolerance curve is set from a single-output decoder. The second removes the DTI prior and learns redundancy from recorded data alone, and we plot performance against the number of days of recording.

Why not just…?

Why not just compute the biomarker on the implant and transmit a few bits? This is the most serious threat, and it narrows our claim rather than killing it. If each site sends only a scalar feature, throughput stops mattering. Energy per delivered bit and tail delay still matter, because every packet must still clear a deadline over a fading, shared channel on a budget where Anwar et al. show throughput collapsing. Our design therefore assumes on-implant feature extraction from the start. The question is how strictly each feature stream must be served, and we frame the comparison in energy-per-decision at low rates, not in capacity.

Why not just learn cross-site redundancy from recorded LFPs and drop DTI? Once weeks of clean data exist, the data-driven estimate should win, and we expect it to. The structural prior matters at cold start, immediately after implantation, and during stimulation, when artifacts corrupt the recordings (Closed-Loop DBS). The ablation that plots performance against recording days tests this directly. If the DTI prior gives no advantage even on day one, we drop it and report that result.

Why not just use reinforcement learning for access and skip the analytical model? MADSPG needs four neural networks per agent, which is heavy for an implant. It also offers no explicit delay guarantee that a regulator or clinician could check. Effective capacity gives a verifiable bound and a convex allocation. The MADSPG baseline measures whether the learned policy's accuracy advantage, if it has one, justifies that cost.

Proposed Experiments

No public multi-site pain-and-mood LFP benchmark appears in our file. The cohorts cited, with 5, 6 and 10 patients, are small and not stated to be released. The first experiment therefore uses rodent spared-nerve-injury recordings from ACC, subthalamic nucleus and somatosensory sites, where the sensory and affective read-outs are established behavioral assays. It is replicated in both sexes, to address the male-only and female-only splits in Wang et al. 2024 and Zhong et al.. The second experiment uses de-identified human deep brain stimulation (DBS) recordings obtained through collaboration, with preoperative diffusion MRI. Channels are simulated with Nakagami fading and a measured medical-band occupancy model.

The first decisive check is whether the two decoder heads really have different tolerance curves. If the affective head is no more delay-tolerant than the sensory head, the central premise fails. The second is whether heterogeneous-θ allocation reduces energy per decision relative to uniform-θ effective-capacity control at equal accuracy on both heads. The third is whether the DTI prior shortens the time to reach the data-driven energy level after implantation.

References

  1. 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
  2. The anterior cingulate cortex contributes to the analgesic rather than the anxiolytic effects of duloxetine in chronic pain-induced anxiety
  3. Differential roles of hippocampal glutamatergic receptors in neuropathic anxiety-like behavior after partial sciatic nerve ligation in rats
  4. The anterior cingulate cortex controls the hyperactivity in subthalamic neurons in male mice with comorbid chronic pain and depression
  5. Excitatory projections from the nucleus reuniens to the medial prefrontal cortex modulate pain and depression-like behaviors in mice
  6. Plasticity in ventral pallidal cholinergic neuron-derived circuits contributes to comorbid chronic pain-like and depression-like behaviour in male mice
  7. Spontaneous pain dynamics characterized by stochasticity in awake human LFP with chronic pain
  8. Cingulate dynamics track depression recovery with deep brain stimulation
  9. Effective Capacity of Licensed-Assisted Access in Unlicensed Spectrum for 5G: From Theory to Application
  10. Capacity of Spectrum Sharing Cognitive Radio with MRC Diversity under Delay Quality-of-Service Constraints in Nakagami Fading Environments
  11. Energy-Efficiency Optimization for MIMO-OFDM Mobile Multimedia Communication Systems With QoS Constraints
  12. Delay-Constrained Video Transmission: Quality-driven Resource Allocation and Scheduling
  13. Intelligent and efficient development of wireless networks: A review of cognitive radio networks
  14. Interference-based optimal power-efficient access scheme for cognitive radio networks
  15. Multi-Agent Deep Stochastic Policy Gradient for Event Based Dynamic Spectrum Access
  16. Network Action of Subcallosal Cingulate White Matter Deep Brain Stimulation
  17. Closed-Loop Deep Brain Stimulation for Refractory Chronic Pain
  18. Anatomy-to-tract mapping infers white matter pathways without diffusion streamline propagation