Longevity Research Initiative · Bayesian Causal Atlas
Every effect size on this page is extracted from a named, published systematic review or randomised trial. Nothing is simulated. Where the evidence conflicts, the conflict is shown rather than averaged away.
Toggle interventions to assemble a candidate regimen. Each card shows the published effect, its uncertainty, and an E-value.
Benefit vs. tolerability for interventions reporting both NNT and NNH.
Headline effect as reported in the source. These are separate trial results.
Lower NNT = more responders. Higher NNH = better tolerated. Upper-left is the efficient frontier.
Move the cutoff. Interventions whose relative benefit (the risk ratio of that agent’s own favourable binary outcome versus its comparator) meets or exceeds the threshold are highlighted as the effective set; the rest dim into the long tail. Outcomes differ across agents (≥50% pain relief · ≥1-grade disability improvement · staying relapse-free), so read this as a triage ranking, not a combinable joint effect.
Causal framing, honestly bounded.
Experiment. The question is causal: for a person with peripheral neuropathy, what is the effect of intervention X on a defined endpoint (incidence of neuropathy, ≥50% pain relief, or change in a validated symptom score)?
Method. The strongest evidence here is from randomised controlled trials and Cochrane/network meta-analyses of them. Randomisation is the identification strategy: it closes the backdoor paths between treatment and outcome by design, so for these estimates no post-hoc confounder adjustment is required. For the one observational signal included — the cardiovascular risk associated with gabapentinoids — randomisation is absent, so residual confounding by indication is the live threat, and that is exactly what the E-value below quantifies.
E-value. For each risk-ratio-type estimate the table reports an E-value (VanderWeele & Ding): the minimum strength of association, on the risk-ratio scale, that an unmeasured confounder would need with both treatment and outcome to fully explain away the observed effect. It is computed transparently as E = RR + √(RR·(RR−1)). A larger E-value means the finding is harder to dismiss as confounding.
What this Oracle deliberately does not do: it does not report point estimates for the probability of necessity / sufficiency (PN, PS, PNS). Those Tian–Pearl counterfactual quantities require the full joint distribution of treatment and outcome plus a monotonicity assumption, and that information is not recoverable from published trial summaries. Inventing them would be the opposite of a grounded analysis. Where a quantity cannot be identified from the available data, it is left out and said so.
Result. The interventions split cleanly into three tiers: a small group with robust, replicated benefit (glycaemic control in type 1 diabetes; duloxetine; capsaicin 8% patch); a contested middle (gabapentin, alpha-lipoic acid, pregabalin); and a low-certainty tail (amitriptyline despite first-line status, acetyl-L-carnitine, B12). See §05 for the conflicts.
| Element | Stance taken in this build |
|---|---|
| Identification | RCT randomisation closes backdoor confounding for efficacy estimates; no synthetic adjustment applied. Observational HRs flagged separately. |
| Effect sizes | Taken verbatim from the cited source (RR, OR, SMD, NNT, NNH, risk difference). No pooling beyond what each source already performed. |
| E-value | Computed only for risk-ratio-type estimates, using the published point estimate. ORs treated as RR-approximations and flagged. |
| SMD effects | Reported with their heterogeneity (I²). No E-value (not a risk ratio). Large I² is shown, not hidden. |
| Combination / cross-correlation | Not modelled. No trial measured the joint effect of these agents, so additivity is treated as an unsupported assumption, not a result. |
| Counterfactual probabilities (PN/PS/PNS) | Not estimable from summary data → omitted by design. |
| Certainty grade | Mirrors the source's GRADE / quality assessment where stated (e.g., Cochrane "very low" for amitriptyline and acetyl-L-carnitine). |
Every cell sourced.
| Intervention | Population | Endpoint | Effect (95% CI) | NNT | E-value | Certainty | Source |
|---|
The parts a marketing page would hide.
Alpha-lipoic acid. Earlier meta-analyses reported a very large symptom benefit (pooled SMD ≈ −2.26), but with severe heterogeneity (I² ≈ 74%) and an intravenous-vs-oral split. The 2024 Cochrane review of ALA as a disease-modifying agent concluded it probably has little or no effect on neuropathy symptoms at six months, with all studies at high risk of attrition bias. The honest read: short-term symptomatic signals exist; durable disease modification is not established.
Gabapentin. Cochrane finds a real responder benefit in diabetic neuropathy (RR ≈ 1.7; NNT ≈ 6.6), yet a 2021 network meta-analysis found no significant difference from placebo on pain score (SMD −0.25, p = 0.09). Both can be true: a minority respond substantially while the average shift is small. Separately, observational data link gabapentinoids to modestly elevated cardiovascular and thromboembolic risk — plausibly confounded by indication, which the E-values quantify.
Amitriptyline. It is first-line in many guidelines, yet the Cochrane review rates the evidence very low quality — only two of seven trials beat placebo, and more than half of patients had an adverse event. Guideline status here reflects long use and low cost more than trial strength.
Acetyl-L-carnitine. One meta-analysis shows a modest pain benefit (SMD −0.45) but with I² ≈ 85% and several manufacturer-funded trials; the 2019 Cochrane review is "very uncertain." Treat as hypothesis, not conclusion.
The type-2 diabetes problem. Intensive glycaemic control robustly prevents neuropathy in type 1 diabetes (annualised risk difference −1.84%, p < 0.00001) but the effect in type 2 is small and not statistically significant (−0.58%, p = 0.06). The single most evidence-backed prevention strategy does not transfer cleanly between the two diseases — a fact often blurred in general "control your sugar" messaging.
Limitations. This is an evidence-synthesis tool, not medical advice and not a substitute for a clinician. Effect sizes come from heterogeneous trials with differing populations, durations, comparators and risk of bias; they are not directly comparable head-to-head except where a network meta-analysis did so. Endpoints differ (incidence vs. pain response vs. symptom score) and cannot be summed. Decisions about any of these interventions — especially prescription drugs and anything during cancer treatment — belong with a qualified clinician.
Primary sources. Callaghan et al., Cochrane (enhanced glucose control, CD007543); Lunn et al., Cochrane 2014 (duloxetine, CD007115); Derry et al., Cochrane 2019 (pregabalin, CD007076); Wiffen/Moore et al., Cochrane (gabapentin, CD007938); Moore et al., Cochrane 2015 (amitriptyline, CD008242); van Nooten et al. 2017 (capsaicin 8% network meta-analysis); Baicus et al., Cochrane 2024 (alpha-lipoic acid, CD012967) & Mijnhout et al. 2012 meta-analysis; Rolim et al., Cochrane 2019 (acetyl-L-carnitine, CD011265) & Li et al. 2017; Smith et al., JAMA 2013 (duloxetine for CIPN); observational gabapentinoid cardiovascular cohort (PMC9438165). Full URLs in each card.
Longevity Research Initiative · Bayesian Causal Atlas · grounded edition. Generated .