Atrial fibrillation · one case
I tested this twice. The first attempt used a reasonable proxy and produced almost nothing. The second used a worse-sounding one and produced a three-beat step. Both are below, because the one that failed is the reason the second is believable.
Where this came from
On the nights I went to bed with an empty stomach, my overnight heart rate was clearly lower. Eat late, and it sat higher. Everyone half-knows this — digestion costs work, work costs beats.
What made it worth testing properly was that I had 1,316 nights of continuous measurement and a prediction attached to specific dates: October and November 2022, when I lost a lot of weight, and the main thing I do when I am being careful about weight is not eat before bed.
Four years later I started a beta blocker and my overnight heart rate dropped. The hunch stopped being a curiosity and became the thing standing between me and a wrong conclusion about my medication.
Attempt one
Split the nights into thirds by how my weight had moved over the previous 21 days. Criterion: at least 10 readings in the night, plus a recorded weight within 5 days of both the night itself and of 21 days earlier. 1,134 nights qualify.
| Nights | Weight change | Overnight median | Under 46 | |
|---|---|---|---|---|
| Losing | 378 | −1.50 kg | 50.8 | 5.8 % |
| Stable | 378 | 0.00 kg | 52.0 | 2.4 % |
| Gaining | 378 | +1.35 kg | 52.0 | 1.3 % |
Correlation +0.17. Just over one beat between losing and gaining. A real direction and a useless effect size. If I had stopped here I would have written “weak support, probably nothing” and been wrong about why.
Attempt two
“Losing weight over three weeks” does not tell you whether you ate at eight o’clock last night. It is an average of behaviour over a period, and what I actually proposed is a property of each individual evening.
So I needed a per-night measurement of how much digestion was running at bedtime, and I used the one I already had: my own pulse between 21:00 and 23:00. Not a measure of food directly, but downstream of it, measured every night, and requiring nothing I was not already recording.
Thirds by evening pulse. 1,285 nights have both measurements. Criterion: at least 5 readings between 21:00 and 23:00, on top of the night itself.
This is a step, not a slope. The slow-evening group sits three beats below the other two, and medium and fast are identical at 52. I originally wrote this up as a three-rung ladder going 49, 52, 53, and that was me seeing a gradient that is not there. What the data says is narrower and sturdier: a quiet evening drops the night by three beats. Eating more does not keep pushing it up.
And the share of nights below 46 is the stronger result anyway: 6.1 % for slow evenings, 0.9 % for medium, 1.6 % for fast. A quiet evening makes a sub-46 night about seven times more likely than a medium one.
The signal lives in the specific evening, not in the kilos. Losing weight over weeks does not guarantee a light dinner every night. Eating at eight o’clock does.
The part I find most convincing
I picked October and November 2022 in advance, because I knew what I had been doing then and not what the numbers would say.
| Nights | Overnight median | Nights under 46 | |
|---|---|---|---|
| Oct–Nov 2022 — eating early and light | 59 | 47.0 | 27.1 % |
| Jan–Mar 2023 — right afterwards | 89 | 49.5 | 1.1 % |
Two and a half beats, and 27 % of nights under 46 against one night in 89. Same body, same year, no medication of any kind, no diagnosis yet, nothing to confound it.
One, not none. My first draft said zero and it was wrong by a single night of 89 — the kind of rounding to a rhetorical zero that destroys a document’s credibility the moment somebody checks. And one comparison I wanted and could not make: the months immediately before, June to August 2022, have only 2 nights with enough readings to count. I am reporting that rather than quietly dropping it, because an n of 2 presented as a baseline would have made this look much stronger than it is.
Why this is not a wellness tip
In October 2026 my overnight median was 46, with a typical nightly low of 42. Three weeks into bisoprolol, a beta blocker whose entire job is slowing the heart. The causal story is obvious, complete and satisfying.
October 2022 — no medication, no diagnosis, no arrhythmia on record — came out at 46, with a typical nightly low of 40. Identical, and the untreated year runs lower.
So before anyone, me included, attributes October 2026 to the drug, there is an alternative explanation sitting in my own history with nothing to do with pharmacology.
And the knife cuts both ways, which I have to say here or the companion page and this one contradict each other. In those same weeks my very low nights became far denser: 6 of my 11 lifetime nights at 43 or below land in the 25 nights since I started the drug, which is 1.9 % of my record. I started the beta blocker and I went back to eating early in the same week, and those nights sit inside both explanations. I cannot apportion it, and neither can anyone else with one subject and four weeks.
I am not trying to rule the drug out. That is a cardiologist’s call in December, with three months of data instead of three weeks. What I can do is make sure the alternative is on the table when the question is asked, with the numbers attached.
The difference between arriving with “the beta blocker has slowed me down” and arriving with “my nights are at 46, and the same month four years ago with no medication was also 46 — what do you make of that?” is the difference between a consultation about a side effect and a consultation about my actual heart.
This is now a standing rule in how I read my own data: check the dinner explanation before attributing anything to the medication.
The version that was wrong
Early in October the numbers said late dinners triggered my atrial fibrillation. Nights where my 21:00–23:00 pulse was in the top third had arrhythmia 33 % of the time. The calmest third: 6 %. Five-fold, clean mechanism, actionable conclusion.
It was backwards. The fibrillation was already running, and the fibrillation was what raised my evening pulse.
I had measured the arrhythmia at nine in the evening, called it digestion, and then been impressed that it predicted the arrhythmia at two in the morning. The correction is one filter: ask only about episodes that start after bedtime, and exclude every night already in arrhythmia at 21:00. The three groups then give 9 %, 6 % and 12 %. Nothing at all.
Note what this does and does not touch. The heart-rate finding above survives intact — a quiet evening really does predict a quiet night. What evaporated was the much bigger claim, that dinner timing causes fibrillation. I have no evidence for that, and the data I thought supported it was an artefact.
In self-tracking everything you measure is measured on the same body at the same time, so your candidate cause is very often downstream of the outcome you are trying to explain. Before believing any association in your own data, ask whether the arrow could point the other way. It took me one day to catch this. If I had not caught it, it would have gone to a cardiologist as a finding.
Run it rather than read it
In prose, the procedure is short: take each night’s median heart rate between 01:00 and 06:00, take the median between 21:00 and 23:00 as a stand-in for how much digestion is running at bedtime, sort the nights into thirds by that evening figure, and compare. Report the share of nights below a fixed threshold as well as the median, because in my data that column carried the result and the median barely moved. Exclude any night when something else was already happening, name any period you want to test before you look at it, and compare the same month across years because the seasonal cycle is larger than most of what you are hunting.
You do not have to do any of that by hand. Paste the block below into whatever you use, attach your export, and it will do the work and argue with you about it.
SOURCE: https://afibcases.com/dinner-predicts-the-night You are helping me run an analysis on my own wearable heart-rate data. Work from my data only. Do not reassure me and do not diagnose me. MY DATA: [attach or paste your export: timestamped heart-rate readings. Apple Health, Garmin, Oura, Fitbit or a CSV of timestamp + bpm.] DO THIS, IN ORDER, AND SHOW ME THE NUMBERS AT EACH STEP: 1. For each night, compute the MEDIAN heart rate between 01:00 and 06:00 local time. Median, not mean. Discard any night with fewer than 10 readings in that window and tell me how many nights you discarded and how many remain. 2. For each night, compute the median heart rate between 21:00 and 23:00 of the preceding evening. Require at least 5 readings or drop the night. This is a stand-in for how much digestion is running at bedtime - say so, and say it is a proxy, not a measurement of food. 3. Sort the nights into equal thirds by that evening figure. For each third report: how many nights, the median evening pulse, the median overnight heart rate, and the PERCENTAGE OF NIGHTS BELOW A FIXED THRESHOLD that you pick from my own distribution. 4. Tell me whether what you see is a step or a slope. If two of the three groups are identical, say so plainly instead of describing a gradient that is not there. 5. Ask me whether there is any night when something else was going on - alcohol, illness, a hard session, an arrhythmia. Exclude those and re-run step 3. Show me both versions. This step changes answers and is not optional. 6. Ask me whether there is a period in my history where I know I behaved differently. Only after I have named the dates, compute that period against the one immediately after it. 7. Compare the same calendar month across different years, never adjacent months. Resting heart rate has a real seasonal cycle and it is larger than most of the effects worth finding. THEN CHALLENGE ME. Specifically: - For any association you find, ask whether the arrow could point the other way: could my candidate cause be an effect of the outcome? On this subject's data that single question destroyed a five-fold result. Apply it before you tell me anything is a finding. - Tell me which groups have too few nights to say anything. - Give me every criterion you used next to every number, so the figures can be reproduced and so I can see which ones move if the criterion changes. - Round nothing to a rhetorical zero. If it is one night in 89, say one night in 89. FOR COMPARISON, one person's published result using exactly this procedure (n=1, 1,285 nights, endurance swimmer, 55): slow evening, median 56 -> night median 49.0, 6.1% under 46 bpm medium evening, median 61 -> night median 52.0, 0.9% fast evening, median 68 -> night median 52.0, 1.6% A step, not a slope. A weight-trend proxy on 1,134 nights gave only +0.17 correlation and about one beat, which is why the per-night measure is the one that works. Do not expect my data to match. Tell me where it differs.
On your own export
Take it with you
Paste this into ChatGPT, Claude, Gemini or whatever you use, add your own numbers, and have the conversation. It carries the source link and a line telling the model what it is looking at, so nothing is lost when it leaves this page.
SOURCE: https://afibcases.com/dinner-predicts-the-night
WHAT THIS IS: one person's self-recorded heart data. A single case, not
evidence about any population, and not medical advice.
I want to analyse my own overnight heart rate data and understand what
drives it. I have an Apple Watch / Garmin / Oura export.
REFERENCE CASE (n=1):
- Endurance swimmer, 55, paroxysmal atrial fibrillation.
- 1,316 nights of continuous data, 24 Feb 2021 to 22 Oct 2026.
Criterion: >=10 heart-rate readings between 00:00 and 07:00.
- Overnight median by year: 54, 52, 50, 51, 51, 52. No trend.
ATTEMPT ONE - WEIGHT TREND AS A PROXY - WEAK:
1,134 nights split by 21-day weight change. Criterion: >=10 night
readings, plus a weight within 5 days of the night and of 21 days
before.
losing -1.50 kg -> night median 50.8, 5.8% of nights under 46
stable 0.00 kg -> night median 52.0, 2.4%
gaining +1.35 kg -> night median 52.0, 1.3%
correlation +0.17
A real direction and a useless effect size.
ATTEMPT TWO - EVENING PULSE (21:00-23:00) AS A PROXY - STRONG:
1,285 nights with both measurements. Criterion: >=5 readings
between 21:00 and 23:00 as well as the night.
slow evening, median 56 -> night median 49.0, 6.1% under 46
medium evening, median 61 -> night median 52.0, 0.9% under 46
fast evening, median 68 -> night median 52.0, 1.6% under 46
NOTE: this is a STEP, not a slope. Medium and fast are identical at
52. Only the slow-evening group separates. A quiet evening makes a
sub-46 night about seven times more likely than a medium one.
Why attempt one failed: weight change over three weeks is an average
of behaviour; the hypothesis is a property of each single evening.
PRE-REGISTERED PERIOD (dates named before looking):
Oct-Nov 2022, deliberate weight loss, eating early:
59 nights, median 47.0, 27.1% of nights under 46
Jan-Mar 2023, right afterwards:
89 nights, median 49.5, 1.1% under 46 - that is ONE night of 89,
not zero.
No medication in either period.
Not available: Jun-Aug 2022 has only 2 qualifying nights, too few
to use as a baseline.
WHY IT MATTERS:
Oct 2026, 3 weeks on bisoprolol: 22 nights, median 46.0, typical
nightly low 42, 45.5% of nights under 46
Oct 2022, no medication at all: 29 nights, median 46.0, typical
nightly low 40, 37.9% of nights under 46
IDENTICAL medians, and the untreated year runs LOWER on the
nightly low. The drug is not the only candidate explanation.
BUT, both ways: 6 of his 11 lifetime nights at 43 bpm or below fall
in the 25 nights since starting the drug (1.9% of the record). He
started bisoprolol AND went back to eating early in the same week,
so the two cannot be separated with one subject.
CAUSALITY TRAP HE FELL INTO AND CORRECTED:
Late dinner appeared to trigger AF: 33% of nights vs 6%. Wrong
direction - the AF was already running and was itself raising the
evening pulse. Excluding nights already in AF at 21:00 and asking
only about NEW episodes: 9%, 6%, 12%. Nothing. The heart-rate
finding survives this; the claim that dinner timing causes
fibrillation does not, and he withdrew it.
MY DATA:
[paste your own night medians, evening pulses, any period where you
know you behaved differently, any medication and when it started]
Please help me run the same analysis. Be strict with me about:
(a) whether any association I find could be running backwards,
(b) whether I have enough nights in each group to say anything,
(c) seasonality - compare same month to same month.
One body, 1,285 nights. I cannot tell you whether the same step appears in someone who is not endurance-trained, how big it is at a normal resting heart rate, or whether any of it interacts with atrial fibrillation the way I assumed before I caught the arrow pointing the wrong way. Those need hundreds of people.
That is what the rest of this site is for.