The citalopram question.
Modelled out · confirmed and sized, now subtracted from the signals it touches, and the correction turned out small
One medication ran through these four years, at a dose that changed over time. Before any signal can be read as 'the body', we have to ask what the drug did to the numbers themselves. This is everything we know and have tested about it, in one place.
Why I took it
Citalopram wasn't a shot in the dark. A retired Dutch neuropsychiatrist, Carla Rus, had gone public with a theory that SSRIs might help long covid, not for mood, but by calming the inflammation and the overactivated nervous system that can follow the infection. I was wary at first. I didn't want to change several things at once and lose track of what did what, so I waited. When my CPAP trial came to an end, it freed up the room to try one new thing cleanly, and together with my GP, we decided on citalopram.
So how good is the idea? A literature review for this page weighed it up. The short version: a biologically plausible, recently-energised hypothesis, with much stronger evidence for preventing acute covid than for treating the long covid this actually was.
- The idea plausible, but contested
A 2023 paper proposed that the inflammation after a viral infection lowers serotonin, which in turn dampens vagus-nerve signalling, a chain that could tie together fatigue, brain fog and an over-activated nervous system. Influential, but a published rebuttal argues its core serotonin measurements were unreliable.
- The strongest trial strong, but acute covid, not long covid
The best randomised evidence is for fluvoxamine (a related SSRI) started early in acute covid, where a large trial found it cut hospitalisation, probably through an anti-inflammatory action rather than its effect on serotonin. But that is about stopping the acute illness from worsening, not treating long covid.
- For long covid itself thin, observational
For the chronic use this actually was (an SSRI to treat established long covid) the evidence thins to one large observational study (a modestly lower risk) and reviews calling for proper trials. No adequately-powered randomised trial has reported a clean positive.
- Carla Rus's study exploratory, hypothesis-generating
Her own work is real and peer-reviewed: 95 post-covid patients treated with SSRIs, most of whom improved, brain fog and sensory overload most of all. But it was exploratory by its own description, a questionnaire study with no control group, so it raises the hypothesis rather than proving it.
So this was a reasoned bet on a live, biologically plausible hypothesis, not a settled treatment. That the question was worth asking in one body doesn't mean one body can answer it for anyone else.
The studies behind this, with plain summaries: reading & sources →
How it felt
And it did seem to help. My headaches eased and the brain fog lifted, at least as I felt it from the inside. But it came with a cost I hadn't bargained for: it quietly turned down all my emotions, flattened them. I wasn't anxious or depressed to begin with, so that flattening was pure side effect, nothing I needed. That's what set me on a slow, deliberate taper, easing the dose down over a long stretch and eventually stopping altogether (which happens after this record ends).
Both halves of it are recognised in the literature: emotional blunting is a known effect of serotonergic antidepressants, reported with citalopram specifically, and a slow, flexible taper is the discontinuation approach the evidence supports. That makes my experience a typical one, which still isn't proof of anything in my case.
This is how it felt from the inside, my own report, not a measured result. What the numbers did is a separate question, below.
What it is
Citalopram is an SSRI, an antidepressant. It was taken across the recent years of the record: none in the early era, built up to 30mg through 2024, held there, and tapered down again in 2026. A dose that moves over time is exactly the kind of thing that can masquerade as 'the body changing'.
The question
The watch's stress and body-battery scores are derived from heart-rate variability. SSRIs are sometimes said to blunt heart-rate variability, though, as a literature review for this page found, that evidence is mixed, not settled. Either way the sharp question is empirical: when the dose changed, did these channels measurably move with it, independent of how the body actually felt?
What the data shows
Yes, and cleanly. Each channel was regressed on the citalopram dose. A positive number means the channel rose as the dose rose; the three that matter most cleared significance in the same direction the dose went both up (2024) and down (2026).
- Daytime stress (all-day average) Confirmed
The strongest of the three, measured daytime stress rose with the dose.
+0.57 / mgp < 0.001
- Overnight stress (the HRV stand-in) Confirmed
The same channel the improvised HRV signal is read from; it rose with the dose too.
+0.43 / mgp = 0.001
- The day's lowest Body Battery Confirmed
Moved the other way, as expected: more drug, a lower battery floor.
−1.13 / mgp < 0.001
- Overnight breathing rate No effect
A clean negative, and an honest one: not everything the watch records responds to the drug. The test discriminates.
≈ 0p = 0.86
Two more channels point the right way without clearing the bar on their own: resting heart rate is weakly consistent, and overnight Body-Battery gain is only partial (the 2024 build-up has no usable data).
Three independent reads agree: the dose going UP in 2024, the dose coming DOWN in 2026, and a clean 2025 control year where the dose held steady and nothing moved. That convergence is why this is called confirmed, not merely suggestive. And a later re-audit put the obvious rivals to the test (weight gain, lost fitness, aging, the seasons) and the dose effect survived all of them: a slow trend that only moves one way can't fake an effect that tracks the dose both up and down.
Which way it points
More citalopram nudged the watch's stress numbers up, not down. That isn't a contradiction once you know what the watch measures. Its 'stress' score was never tracking mood, or even how I felt; it's a read of how activated my nervous system was, worked out from heart-rate variability, and blind to the cause. So the drug that seemed to settle my head could, at the same time, leave a fingerprint the watch reads as more arousal. Both can be true at once: a calmer person, a busier-looking number.
what the evidence saysWhy the number rose, exactly, is less settled than it sounds. The usual story (SSRIs blunt heart-rate variability, which a Garmin-style score then reads as higher stress) is only partly borne out: the best meta-analysis finds the effect mixed and design-dependent, citalopram's own effect is mild, and no study has ever validated a consumer device's variability-to-'stress' transform under an SSRI. So I hold the measurement, not the mechanism: in this one body, across four years, the number moved with the dose; that part is solid; the reason is a plausible explanation, not a proven one.
Why it matters for the findings
Stress and Body Battery are the very channels the crash-precursor signals are read from. The dose changed across the four years, so any difference these channels show 'early versus late' is part body and part drug; they can't be cleanly separated by looking. That's what makes citalopram a highly likely confounder of the whole change over time, and not a side-note: it sits inside the main signals, not beside them.
What it still can't settle
With one person, the drug can't be fully untangled from the recovery it ran alongside, the taper and the body getting better move together in time. The honest reading is 'consistent with a dose-graded response to citalopram, after the recovery slope is absorbed', not proof the drug acted alone. A between-person study could settle it; this design cannot.
where it stands
Modelled out, corrected, and the correction is small
On the driver ledger this has moved from 'bounded' to 'modelled out'. We didn't just size the dose effect; we subtracted it, re-reading each signal it touches with the drug removed. That's the step this page used to call pending; it's now done.
The result: The result is deliberately undramatic. Most of the scorecard's signals are built from night-to-night changes, and a slowly-moving dose largely cancels out of a change; so subtracting it barely moved them, and nothing grew stronger once the drug was out. The one signal that stands, the collapse in overnight HRV variability, still stands net of the medication: +19.7 points becomes +18.7. So citalopram is real, confirmed, and now accounted for. It turns out it was never doing the scorecard's work.
On the ledger of everything that could explain the patterns: the driver ledger →
what this case can add
What this one case can add
As a test of whether the drug works, this is one humble anecdote and no more: a single person who happened to get better tells you nothing reliable about citalopram for long COVID. But there is one place it may genuinely add something, and it is not about the drug at all. It is about measurement. What I have is a rare thing: a within-person, dose-graded, four-year record of an SSRI meeting a consumer wearable, the dose climbing through 2024, tapering through 2026, with a flatter 2025 in between as a kind of natural control. And it lands on exactly the gap the literature turned up: no published study has ever checked what a Garmin-style 'stress' or Body Battery number does under an SSRI. So the honest contribution is a caution, not a discovery: if you run a wearable study on people taking these drugs, the drug can move the numbers on its own, and a design that ignores that will misread them.
Measurement, not efficacy. This says nothing about whether citalopram helps anyone; it flags a confound that future wearable studies should design around. Still one body, and 'stress' here is the watch's arousal reading, not mood.
The full analysis, in the research repo: Citalopram dose-response (methodology, v3) ↗ · Citalopram phase stratification ↗ · Intervention effects (descriptive) ↗ · Net-of-driver correction (R20/R16) ↗