The Corpus That Reads Your Medicine Cabinet
An online survey already exists that asks thousands of ME/CFS and long-COVID patients which treatments they felt helped. It reports over 150 treatments ranked by perceived benefit (Eckey et al. 2025). The survey is genuinely useful, but it cannot tell you which biological pathway is broken, because it asks whether a treatment helped — not at what dose, for how long, or by what mechanism it acted.
The difference between a popularity ranking and a mechanism ranking is almost entirely a matter of what the report contains. This post is about that difference.
1 Why “LDN helped me” pools to nothing
The central proposal of the primary document’s dose-response work is that in ME/CFS, where on the dose scale a drug works may itself be the information. In that analysis, low-dose naltrexone helps some patients at 0.5–1.5 mg and then stops helping as the dose rises, because its dominant mechanism at low doses (a partial blockade of TLR4 that triggers the cell’s own anti-inflammatory Nrf2 programme) is proposed to switch off when the blockade becomes too complete (Younger, Parkitny, and McLain 2014; Kučić et al. 2021; Calabrese and Kozumbo 2021). A patient whose benefit disappears as the dose climbs may have told you something about their biology. A patient who reports only “LDN helped” has not.
This is why the corpus needs the dose. A report of “LDN did not help” is not poolable, because it does not distinguish a patient who failed at a dose where the drug could work from a patient who failed at the top of a range, where the mechanism was never engaged. The dose-response series makes this point for one patient’s own titration; at population scale the same principle decides whether the entire corpus is useful or inert.
The minimum field set, and why each field earns its place:
- Drug — the intervention itself.
- Dose and dose band — the load-bearing variable. A null inside a tested window is information; a null at the ceiling is not.
- Duration — because a benefit that appears at three weeks and declines at three months is a different finding from one that never appears. An adequate trial has to have lasted long enough to show its effect.
- Response domain — which symptom changed. A drug that fixes brain fog while leaving pain untouched is telling you those two symptoms run on different pathways. A single “better/worse” score destroys that signal.
- Side-effect type — because a side effect is a readout that the target was engaged, independent of whether correcting it helped. A patient who stopped a drug for a specific side effect has still contributed mechanistic data.
- Stop reason — the closest thing to a free-text adjudication of what actually happened.
- Community discussion volume — an unusual field, and a necessary one, explained below.
2 How hundreds of one-patient trials become one instrument
Once reports are poolable, they form a matrix. Each drug is a row; each mechanistic hypothesis (TLR4 neuroinflammation, TRPM3 channel dysfunction, mitochondrial substrate limitation, mast-cell activation, autoimmune signalling, and so on) is a column. Each cell asks: if this hypothesis were the patient’s rate-limiting problem, what would this drug have done?
A positive response leaves the hypotheses the drug should have addressed standing, but only after the placebo floor below is subtracted, since an unblinded positive is partly expectation. A null response, recorded inside the drug’s window, removes them from the candidate set — unless the hypotheses are “spectrally identical” (two mechanisms that would predict the same response to every drug we can safely test, so no response can tell them apart) or the target has adapted. A specific side effect identifies a system the drug touched. The corpus would be designed to resolve, by constraint, to the small set of hypotheses compatible with everything observed. That is the same logic the primary document applies to a single patient’s drug history, run at population scale.
Two refinements matter.
First, the corpus should look for responder subgroups, not an average. Heterogeneous response to a drug is itself diagnostic. When a large ARDS trial split patients by a simple biological marker, the two groups responded differently to the same drug, so a trial that was “null” overall contained a real, mechanism-revealing split (Calfee et al. 2018). ME/CFS shows substantial, partially reproducible within-diagnosis heterogeneity of the same kind (Vaes et al. 2023). A corpus that reports “the average patient improved a little” has thrown away its most valuable finding; one that reports who responded has found a subtype.
Second, the corpus needs its own placebo floor. If you pool every report, you capture not only drug effects but the community’s baseline expectation, the natural fluctuation of the illness, and the recall bias of looking back. The way to measure that baseline is to include interventions with no plausible mechanism at all: the supplements and fads that patients try and that no hypothesis predicts should work. Their pooled response is the floor. Subtract it, and what remains is closer to a mechanism signal. Without it, the corpus cannot tell a weak real signal from a strong expectation effect. The floor has a limitation worth stating: its validity depends on the hypothesis space being complete, and a floor measured on inert supplements need not equal the expectation attached to a specific popular drug, so it corrects the community-average expectation, not drug-specific inflation.
3 The community-discussion field, and why it is there
The seventh field is uncomfortable but essential. If reports are recruited from a community, they carry that community’s beliefs with them, and those beliefs shape both which drugs are tried and how their effects are reported. A widely discussed treatment will accumulate enthusiastic reports whether or not it works. If the corpus cannot distinguish a drug’s pharmacology from its popularity, readers risk taking the community’s prior consensus as independent evidence.
So the corpus must record, per drug, some measure of how much it is discussed online or prescribed at the recruiting clinics. The top-ranked pathways are then tested against it: does the signal track dose and molecular target, or does it track how much the drug is being talked about? If the latter, the proposal fails. Without that test the proposal cannot separate signal from popularity; it is the condition under which the idea may exist at all.
4 What this would take
Most of the raw material already exists. Patients have run these trials; registries and surveys already collect some of the fields; the statistical machinery for pooling single-patient experiments is mature (Bourke et al. 2020; Schmid and Yang 2022; Punja et al. 2016). What is missing is a shared schema (an agreement on what a report must contain) and the discipline to collect failures as carefully as successes.
That is a small thing to ask, and a large thing to change: it would turn a scattered archive of personal experience into the closest thing a biomarker-less disease has to a population-level ranking. The ranking would reflect the patients well enough to contribute (mostly mild-to-moderate), and it would take dedicated channels to reach severe and very-severe patients, whose reports are the most informative and the hardest to collect.
Next: What a Pile of Drug Reports Can and Cannot Tell You — the limits, including the one that would kill the whole idea.
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. Any dose figure here is an illustration from the primary document’s analysis, not a prescribing guide. Discuss any medication decision with your clinician.
This post draws on the pharmacodiagnostic corpus specification and minimum data schema in Loth 2026. Heterogeneous-response subtyping follows (Calfee et al. 2018); the emerging-practice substrate follows (Bourke et al. 2020).