Your Failed Treatments Are Data
You have tried things. Everyone with ME/CFS has. A low dose of naltrexone that helped the fog for two months and then stopped. A stimulant that made the mornings possible and the afternoons worse. Antihistamines that fixed a symptom you did not know was mast-cell related. A rack of supplements of which at most one did anything. A drug a specialist suggested that did nothing at all.
The standard reading of this history is: some things worked, most did not, and that is just how it goes. This post makes a different claim. Each of those trials was a small experiment on one biological node — and the results, read together, could be evidence rather than noise, under the model this series sets out.
1 A drug trial is a constraint, not a story
The primary document develops this idea under the name pharmacodiagnostics: using a drug response to infer which mechanism is active, not just whether the drug is worth continuing. The logic is simple.
Every drug in the ME/CFS toolkit works on a specific target. Low-dose naltrexone engages TLR4 and the Nrf2 pathway (mechanistic evidence is animal and cell-line only) (Kučić et al. 2021; Kwilasz et al. 2021), TRPM3 channels (Cabanas et al. 2018), and — by inference, never directly measured in ME/CFS — orexin neurons (Grossberg et al. 2011), with the dose needed for each still uncertain. Low-dose aripiprazole is a partial agonist at dopamine D2/D3 receptors (Burris et al. 2002; Shapiro et al. 2003), and clinical reports of benefit at low doses exist (Crosby, Kalantar, and DeRisi 2021), though its proposed action on microglia is mechanistically plausible and not yet validated in humans. Ketotifen stabilises mast cells (Afrin 2013; Molderings et al. 2016). CoQ10 supports the mitochondrial electron transport chain (Castro-Marrero et al. 2013; Maes et al. 2009). None of these is a “treatment for ME/CFS” in general. Each is a probe of one node.
When a probe works, it suggests that node was active, and that correcting it mattered enough to change your symptoms. When a probe fails within its proper dose range, it suggests the opposite: that node was not the thing limiting you. The null is information, under the assumptions of the model (a point the dose-response series makes for one patient at a time).
The paper does not propose a future technology; it just makes explicit how a clinician already reasons, in an informal way: “We tried an antihistamine and it didn’t help, so let’s look at the autonomic side.” It turns that intuition into a systematic method.
2 The pool is what makes it a signal
One patient’s drug history is a handful of constraints. Hundreds of patients’ histories, pooled, become something no single trial can be: a population-level picture of which nodes are consistently worth probing and which are consistently dead ends.
Pooling single-patient trials into a group estimate is an established statistical method — used in pediatric ADHD, in physiotherapy, in hypertension — and multi-level models can combine many single-patient experiments while handling the complications (trend, carryover effects, autocorrelation) that make naive averaging misleading (Punja et al. 2016; Nikles et al. 2022; Samuel et al. 2023; Schmid and Yang 2022; Zucker, Ruthazer, and Schmid 2010). The method exists. What does not exist is a corpus of ME/CFS drug reports captured in a form the method can use.
The statistics are the least interesting part. The interesting part is what the pooled pattern could show that no laboratory currently can: a ranked list of candidate rate-limiting pathways — derived not from theory but from the collective response of the people who have it. The pool over-represents patients well enough to report — mostly mild-to-moderate — and cannot speak for severe or very-severe disease (Richiardi et al. 2007; Loth 2026).
3 The counter-intuitive part: failures may be the better data
Everyone assumes the valuable reports are the successes. This helped me. But the two kinds of report are not symmetrically trustworthy.
Positive reports are inflated. Anyone who tries a treatment they believe in can improve for reasons that have nothing to do with the drug’s mechanism — expectation moves real symptoms, especially subjective ones like pain and fatigue, even when a placebo is given openly with full disclosure (Flávio-Reis et al. 2025; Kleine-Borgmann and Bingel 2025). Communities reinforce each other: a plausible treatment gets tried enthusiastically, early good experiences get reported, and a consensus forms that may not survive a controlled test (Loth 2026). The PACE trial is the paper’s standing example of what happens when a treatment’s reputation outruns its evidence (White et al. 2011; Wilshire et al. 2018).
Failed trials are less easy to inflate. A patient cannot expect themselves into no effect. A null report is harder to manufacture than a positive one, and it is systematically missing from the published record — negative results are generally the least likely to be written up and the easiest to forget — though an unblinded corpus also introduces its own reporting pressures (Cutroneo et al. 2023). This means the reports a corpus that actively solicits failures would capture are precisely the ones the literature discards. In a disease where most treatments do not work for most people, that missing stratum may be the most informative.
This is a hypothesis, not a fact. It has a test — if null reports replicate better across independent groups of patients than positive reports do, the asymmetry is real — and it has a caveat that this series will not drop: nulls are not magically protected either. A trial that happens to coincide with a spontaneous worsening can record a failure that was never the drug’s, and an unrecorded change in something else — a new co-prescription, a seasonal shift, a stretch of unusually good sleep — can suppress a real effect and manufacture a null. Whether failures are net more trustworthy is a question to be measured, not assumed.
4 Significance
If the idea holds, it changes the status of something you already have. Your medication history — every trial, every dose you titrated, every drug you abandoned because it did nothing — becomes a contribution to the only kind of evidence a biomarker-less disease can currently produce. It is more than a private medical footnote. This does not mean trialling treatments to produce data: drug trials need a prescriber, and this series does not encourage self-experimentation. The corpus aggregates what patients have already done under clinical care.
The catch is that almost none of it is recorded in a form that can be pooled. That is the subject of the next article.
This is not medical advice. It describes a research proposal. No one should start, stop, or change a treatment on the basis of a pooled community ranking.
Next: The Corpus That Reads Your Medicine Cabinet — what a report has to contain, and why “LDN helped me” pools to nothing.
This post draws on the pharmacodiagnostic framework and the population-corpus section of Loth 2026. The poolability of single-patient trials rests on (Punja et al. 2016; Schmid and Yang 2022; Zucker, Ruthazer, and Schmid 2010); the expectation-inflation caveat on (Flávio-Reis et al. 2025; Kleine-Borgmann and Bingel 2025).