Does Treating Refractory Neuropathic Pain Raise Dementia Risk? A Federated Mediation Study

People whose nerve pain does not respond to treatment are usually moved up a ladder of drugs: gabapentinoids, opioids, tricyclic antidepressants, often several at once. These drugs act on healthy brain tissue as well as on the damaged nerves, and they are known to cause cognitive blunting. Patients with refractory pain are also often older and carry vascular disease, which is itself one of the earliest signals of late-onset Alzheimer's disease. Two questions are therefore tangled together. Is chronic refractory pain associated with later Alzheimer's or dementia diagnosis? If so, how much of that link runs through the drug escalation rather than through the pain or its underlying cause? No single hospital sees enough of these patients to answer either question. The records cannot be pooled across hospitals because they are protected health information.

We propose a federated mediation analysis. Each hospital keeps its records and sends only summary statistics. The analysis then splits the pain-to-dementia association into a part that runs through cumulative drug exposure and a part that does not. The methodological core follows a simple rule we call "shared biology, local practice." How strongly a drug affects dementia risk should be roughly the same everywhere. How readily doctors escalate drugs for refractory pain differs widely across sites and countries. Our estimator pools the drug-to-dementia effect across sites and lets the pain-to-prescribing effect vary by site, using only per-site score statistics. Existing federated survival estimators either force every parameter to be common or treat every site as unrelated. We claim that neither choice gives a valid indirect-effect estimate when prescribing practice differs across sites.

What the three directions contribute

The neuropathic pain literature supplies the motivation and the main confounding structure. Conventional channel blockers cannot separate pathological high-frequency firing from normal firing. C2230, a preferential use- and state-dependent CaV2.2 channel blocker makes this argument while designing a more selective blocker, and the neurofeedback meta-analysis describes cognitive side effects that drive patients to stop therapy. Both papers motivate the drug-mediation hypothesis. Neither measures long-term cognitive outcomes. The rTMS work constrains our design. In a crossover trial of high-frequency motor-cortex stimulation, Hosomi et al. found sustained relief in patients with spinal or peripheral lesions but only transient relief in patients with stroke lesions. This shows that central post-stroke pain differs biologically from other refractory pain. Separately, Iturria-Medina et al. report that vascular dysregulation is the earliest detectable event in late-onset Alzheimer's. Taken together, these two results mean that any pain-dementia association in stroke-origin pain is heavily confounded by vascular disease. We turn this into a diagnostic. If the drug pathway is real, its indirect effect should appear in peripheral-origin pain (postherpetic neuralgia, diabetic neuropathy, spinal cord injury) at a similar size to central pain. If vascular confounding drives the association, the effect should concentrate in post-stroke pain.

The Alzheimer's literature defines the outcome and cautions against claiming a mechanism. Post-amyloid accounts emphasize inflammatory and metabolic pathways. Examples are neurotoxic microglia driven by interferon-gamma suppressing the antioxidant regulator Nrf2 (Kang et al.) and inflammation-driven damage to cholinergic circuits (Hu). The cholinergic finding makes anticholinergic tricyclics a plausible culprit. Our study measures diagnoses and prescriptions, not microglia, so these papers supply hypotheses about which drug classes to separate, not evidence we can test directly.

The federated statistics literature supplies the methods. CAFTA fits accelerated failure time models from center-level score functions and sensitivity matrices. An accelerated failure time model is a survival regression in which covariates stretch or shrink the time until the event. CAFTA proves equivalence to pooled estimation under common parameters. Distributed mediation analysis shows that exact global Sobel and MaxP mediation tests can be rebuilt from per-site linear-regression t-statistics, provided the number of sites $K$ stays well below $\sqrt{N}$. That paper handles neither censored outcomes nor site-varying paths. Privacy-protected causal survival analysis uses per-site efficient influence functions and an $\ell_1$-penalized weighting to drop sites whose distribution has shifted. It targets total effects, not mediation. Causal discovery with missing values warns that deleting records with missing values that are not missing at random creates collider bias. This applies to pain scores and cognitive screens in health records, which are recorded mainly when a clinician is already worried.

Core Claim

The model works on log survival time. For patient $i$ at site $k$, let $A_i$ indicate refractory neuropathic pain, $M_i$ the standardized cumulative drug burden in a fixed exposure window, $X_i$ baseline covariates, and $T_i$ the time to dementia diagnosis, possibly censored. The model is

$$M_i = \alpha_k + \beta_k A_i + X_i^\top \theta_k + e_i, \qquad \log T_i = \mu_k + \tau A_i + \gamma M_i + X_i^\top \phi + \sigma \varepsilon_i ,$$

where $\beta_k$ is the site-specific effect of refractory pain on prescribing, $\gamma$ is the shared effect of drug burden on time to dementia, $\tau$ is the direct effect, and $\varepsilon_i$ follows a Weibull, log-normal or log-logistic error. The site-$k$ indirect effect is $\beta_k\gamma$. On the log-time scale, products of coefficients keep their usual mediation meaning, which is why we use accelerated failure time models rather than Cox models.

Sites send score vectors and information matrices for the shared block $(\tau, \gamma, \phi, \sigma)$ in CAFTA's renewable, non-iterative style. They also send local estimates and standard errors for $\beta_k$. The coordinator then tests the joint null "$\gamma = 0$ or all $\beta_k = 0$" with a MaxP-type statistic. Following Distributed mediation analysis, we will derive the asymptotic calibration under the $K = o(\sqrt N)$ regime. The coordinator also checks whether a single $\gamma$ is defensible by comparing each site's influence-function contribution to $\gamma$ against the pooled value. It down-weights sites that disagree using the $\ell_1$ scheme of privacy-protected causal survival analysis.

The new element is this partially pooled structure, together with the theory showing when it recovers the indirect effect that a pooled analysis would give.

Research Program

The first component is methodological. We will derive the distributed estimator and its variance, prove asymptotic equivalence to a centralized analysis that pools $\gamma$ and keeps $\beta$ site-specific, and characterize the bias of full pooling. That bias is a weighted average of $\beta_k$ that does not match any site's or the population's indirect effect when exposure prevalence differs across sites.

Missing pain-severity and cognitive-screening data are handled in two ways. First, inverse-probability-of-observation weights are estimated locally and folded into the scores. Second, a local missing-value causal-discovery step, following MVPC, audits whether the assumed confounder graph acquires spurious edges under deletion.

Reverse causation is the main threat to validity. Early, undiagnosed dementia can itself change how pain is reported and how drugs are prescribed. We handle this with three design choices. The exposure window ends two to three years before outcome follow-up begins. We use a negative-control outcome with no plausible drug pathway, such as cataract surgery. We use a negative-control mediator: topical lidocaine or capsaicin, which treat the same pain without acting on the brain.

The second component is empirical, using sites that store records in the OMOP common data model of the OHDSI network (Hripcsak et al.). LEGEND-T2DM shows that this network can already run harmonized, distributed treatment-effect studies at scale. Refractory neuropathic pain is defined computably as a neuropathic pain diagnosis plus failure of at least two first-line drug classes. Etiology is split into central (post-stroke) and peripheral or spinal groups. Drug burden is separated by class into gabapentinoid, opioid, and anticholinergic tricyclic exposure.

Baselines

The naive reference point is a local-only analysis with random-effects meta-analysis. Each site fits its own mediator and outcome models and computes its own indirect effect, and the estimates are combined with inverse-variance weights. This is what a consortium would do by default. Run on the same sites, covariates and lag windows, it should be badly underpowered when each site has few refractory pain patients. The evidence for this expectation comes from two papers. In CAFTA, fixed- and random-effects meta-analysis on the transplant registry missed a donor-ethnicity effect that the pooled estimator detected; the file gives no coefficient values for this comparison. In Distributed mediation analysis, single-site tests found no mediation ($p > 0.05$) while the distributed MaxP test reported $p = 0.0007$.

The strongest existing method is full-pooling CAFTA with a Sobel test. Both regressions are fitted by CAFTA with all coefficients common, and the indirect effect is the product of pooled $\hat\beta$ and $\hat\gamma$. It is the strongest competitor because it is already proven equivalent to pooled estimation whenever the common-parameter assumption holds. It uses exactly the same summary statistics and communication rounds as our estimator, so the comparison is fair by construction. We beat it if, in simulations where $\beta_k$ varies across sites, our estimator has lower root-mean-square error for site-specific and population indirect effects and correct Type I error while full pooling does not. We must not lose to it when $\beta_k$ is in fact constant. No reported result was found on Lacuna for full-pooling CAFTA on a mediation estimand.

The influence-function survival estimator of privacy-protected causal survival analysis is a third comparator, for the total effect only. That paper reports up to 59% lower root-mean-square error than local-only models in its own simulations. Our total effect $\tau + \beta\gamma$ should match its total-effect estimate on shared simulations. A mismatch would indicate that the accelerated failure time model is misspecified.

Two ablations isolate the key ingredients. Removing partial pooling turns the method into full-pooling CAFTA. Removing the influence-function site down-weighting tests whether shared-$\gamma$ violations matter in practice. In simulation, a centralized oracle on pooled data sets the ceiling. All methods use the same lag windows, covariate sets and model-family selection procedure.

Why not just…?

Why not just run a standard OHDSI cohort study with Cox models and meta-analysis? This answers whether refractory pain predicts dementia, but not whether drug escalation carries the effect. Product-of-coefficient mediation is not well defined on the hazard scale when the outcome is common. Site-by-site meta-analysis also loses the power that mediation tests depend on, which is the failure mode shown in Distributed mediation analysis. We include the Cox total-effect analysis as a descriptive companion, not as a competitor.

Why not just pool everything with CAFTA and compute Sobel? This is a real threat. If prescribing escalation turns out to be similar across sites, it is all that is needed, and our extra structure buys nothing. The experiments settle this directly. We will estimate the spread of $\beta_k$ across real sites from the exchanged local estimates and report it. If the spread is small, we will say full pooling suffices. In that case the contribution is the clinical finding, not the estimator. The large regional differences in a related neuromodulation literature suggest the spread will not be small. In the rTMS-for-Alzheimer's meta-analysis of Zhang et al., Asian trials reported SMD 0.40 against a pooled SMD of 0.27. This does not prove that prescribing differs by region, but it makes uniform practice implausible.

Why not just train a federated deep classifier to predict dementia from pain history? Methods such as UG-FedDA reach 90.54% normal-versus-Alzheimer's accuracy on ADNI in a federated setting. However, accurate prediction does not show which pathway carries the risk. A drug-mediated and a vascular-confounded association would predict dementia equally well. Only the mediation decomposition, with the etiology split and the negative-control mediator, separates them.

Why not just recruit a prospective cohort? The pain cohorts behind the mechanistic findings contain 7 to 14 patients (Mori et al., Hosomi et al.). Dementia takes years to develop, so a prospective study powered for mediation is not feasible before the observational evidence exists.

Proposed Experiments

Simulation comes first. We will build a semi-synthetic generator that extends the clinically grounded graph of the Neuropathic Pain Diagnosis Simulator with drug-burden and dementia-time nodes. The generator varies the number of sites (up to the $K < \sqrt N/2$ boundary), the spread of $\beta_k$, censoring, and the missing-not-at-random mechanism. Metrics are bias, root-mean-square error and confidence-interval coverage for indirect effects, along with Type I error and power of the MaxP test.

Second, we will check the machinery on public data. We will reproduce CAFTA's oracle match on the transplant registry it used, and run the missing-data audit on CogUSA, where MVPC has already been shown to violate fewer expert constraints than deletion.

Third, we will run the federated study across OMOP sites. Three checks decide whether a positive result can be trusted. The negative-control mediator (topical agents) must show a null indirect effect. The negative-control outcome must show a null total effect. Lengthening the lag must leave the indirect effect stable. If these checks pass, the comparison between central and peripheral etiologies says whether the risk comes from the drugs or from the vascular disease that often accompanies post-stroke pain.

References

  1. C2230, a preferential use- and state-dependent CaV2.2 channel blocker, mitigates pain behaviors across multiple pain models
  2. The analgesic effect of electroencephalographic neurofeedback for people with chronic pain: A systematic review and meta‐analysis
  3. &lt;b&gt;Treatment of intractable neuropathic pain by high-frequency repetitive transcranial magnetic stimulation (rTMS) of the primary motor cortex &lt;/b&gt;
  4. Early role of vascular dysregulation on late-onset Alzheimer’s disease based on multifactorial data-driven analysis
  5. Neurotoxic Microglial Activation via IFNγ‐Induced Nrf2 Reduction Exacerbating Alzheimer's Disease
  6. Ethanol extract of Evodia lepta Merr. ameliorates cognitive impairment through inhibiting NLRP3 inflammasome in scopolamine-treated mice
  7. Collaborative Inference for Accelerated Failure Time Model Using Clinical Center‐Level Summary Statistics
  8. Distributed mediation analysis with communication efficiency
  9. Privacy-Protected Causal Survival Analysis Under Distribution Shift
  10. Causal Discovery in the Presence of Missing Values for Neuropathic Pain Diagnosis
  11. Large-scale evidence generation and evaluation across a network of databases for type 2 diabetes mellitus (LEGEND-T2DM): a protocol for a series of multinational, real-world comparative cardiovascular effectiveness and safety studies
  12. Comparative efficacy of rTMS on different targets in Alzheimer’s disease: a systematic review and meta-analysis
  13. UG-FedDA: Uncertainty-Guided Federated Domain Adaptation for Multi-Center Alzheimer's Disease Detection
  14. Study of cortical motor representation and cortical excitability in refractory neuropathic pain
  15. &lt;b&gt;The mechanism of high-frequency rTMS &lt;/b&gt;&lt;b&gt;for intractable pain, a TMS study &lt;/b&gt;
  16. Neuropathic Pain Diagnosis Simulator for Causal Discovery Algorithm Evaluation