Atrial fibrillation · one case
Apple’s Afib History gave me 26%, then 8%, then 2% three weeks running. I pulled the raw beat-to-beat intervals behind those same five weeks and counted it myself. The fall was real. The flat 2% was not.
Turn on Afib History and once a week your watch hands you a percentage: the share of the time it believes you were in atrial fibrillation. Mine, for the five weeks after I was diagnosed:
| Week | Apple | Counted from the raw beats | Episode hours |
|---|---|---|---|
| 21–28 Sep | 26% | 19.3% | 32.5 h |
| 28 Sep–5 Oct | 8% | 6.6% | 11.1 h |
| 5–12 Oct | 2% | 1.2% | 2.0 h |
| 12–19 Oct | 2% | 3.9% | 6.6 h |
| 19–26 Oct | 2% | 3.2% | 5.3 h |
The shape agrees, and that matters: both methods say the same thing about the big move, a fall from a quarter of the time to a few percent in a fortnight. I started flecainide and bisoprolol on 28 September, in the middle of it. What happened after is where the two numbers part company.
Apple printed 2% three times. The beats underneath say 1.2%, then 3.9%, then 3.2% — the middle week carried three times the arrhythmia of the one before it, and the watch showed me the same digit for both.
That is not Apple being wrong. It is Apple rounding a small number to one digit, which is what a weekly summary is for. But it means the thing you are most likely to do with that screen — read three 2s in a row as a flat line, as the condition settling — is the one reading the data does not support.
A burden that reads flat at the bottom of its scale is not telling you it is flat. It is telling you it has run out of digits.
This is the part I had no intuition for, and it is the reason I wrote this page. Across those three 2% weeks, in 1,188 recorded sessions:
| Episodes detected | 19 |
| Nights with at least one, out of 21 | 7 |
| Total time in episode | 13.9 h |
| Longest single episode | 4.5 h |
| Sessions flagged, of 1,188 | 61 |
Two percent was not a stray minute. It was nineteen separate runs of atrial fibrillation across a third of the nights, one of them four and a half hours long, and a second of three and a quarter. I felt none of them.
They are also, overwhelmingly, at night: 73% of the flagged sessions in those five weeks fall between 22:00 and 08:00. Overnight, 10.8% of sessions are flagged; during the day, 5.5%.
On 25 September, the day I was diagnosed and turned Afib History on, my watch changed how often it records beat-to-beat data: from about 11 sessions a day to about 58. Nothing about my heart changed that morning. The sampling did.
So the week Apple labels 26% is two different weeks stitched together. Before the switch, 21 to 24 September: 38 sessions, 1 flagged. After it, 25 to 27 September: 141 sessions, 61 flagged, 43%.
If you compare your own burden across a date when your device changed — a new watch, a software update, a feature turned on — you are measuring the device, not yourself.
It cuts the other way too, and in my favour this time. October is the month with the lowest share of flagged sessions in my whole year, 4.1%, and it is also the month the watch measured me the most. A number that falls while the measuring gets denser is a number you can believe.
Over those five weeks my watch recorded, on average, 50 minutes of actual beat-to-beat data a day. That is 3.5% of the day. It is spread thin rather than bunched — 18.5 of the 24 hours have at least one reading in them, and the typical gap is 15 minutes — but it is still 50 minutes of evidence about 1,440.
And the watch is blind exactly where I most want to see: it does not sample beat-to-beat intervals during exercise at all. Not once, in any of my training sessions.
Whatever percentage your watch shows you, the true figure is that or higher. Never lower. It is a floor, not a measurement.
Apple’s weekly figure is stored on the phone like any other health sample, and comes straight out of an encrypted iPhone backup. The second column is my own count: every beat-to-beat recording in the same weeks, with a session flagged as fibrillation when more than 52% of consecutive intervals differ by more than 5% from each other and the spread of intervals exceeds 13%.
That threshold is not a guess, and it is not Apple’s. It was fixed against the years before any of this started: run over 2021 to 2024, when I had no diagnosis and no symptoms, it fires on 0.75% of sessions. That is the false-positive floor of the method. The 5.1% inside the 2% weeks sits well above it.
The full extraction, including why three of the five obvious routes to this data do not work, is written up in reading your own heart data. The device’s own ECG verdicts, and how often it declines to give one, are in fifty-four without a verdict.
One year of this, for scale: 5,845 sessions since November 2025, 1,213 of them flagged, 146 episodes, 565 hours. Monthly share of flagged sessions, November through October: 13.4, 27.8, 27.6, 16.0, 16.1, 36.0, 31.1, 33.2, 21.7, 20.5, 34.5, 4.1. The hours are not comparable month to month — before late September the watch sampled too sparsely for short episodes to show their length — which is why the share of flagged sessions is the figure to read, not the hours.
If your watch shows you a burden figure, you can do exactly what I did to it. Paste the block below into whatever model you use and attach your own export. It will not diagnose you, and it is written to tell you when your data cannot answer the question.
SOURCE: https://afibcases.com/afib-burden-two-percent You are helping me read the atrial fibrillation burden figure my own device gives me. Work from my data only. Do not reassure me, do not diagnose me, and say plainly when my data cannot answer a question. MY DATA: [attach your export. Ideally beat-to-beat intervals, in Apple Health these are the heart rate variability sessions; otherwise timestamped heart rate. Plus the weekly burden figures your device reported, if it reports any.] DO THIS, IN ORDER, AND SHOW ME THE NUMBERS AT EACH STEP: 1. BEFORE ANYTHING ELSE, count how much data there is per day: how many recording sessions, how many minutes of actual recording, and how many hours of the day contain at least one reading. Tell me what share of the day is actually measured. Every later percentage is a floor, not a measurement, and you must repeat that whenever you give me one. 2. Plot sessions per day over the whole period and find any date where the sampling rate changes by more than half. A new device, a software update or a feature being switched on all do this. Name those dates. Do not compare any two periods that straddle one without telling me you are doing it. 3. For each recording with beat-to-beat intervals, compute two things: the share of consecutive intervals that differ from each other by more than 5%, and the standard deviation of the intervals divided by their mean. Flag a session when the first is above 52% AND the second above 13%. 4. CALIBRATE BEFORE YOU CONCLUDE. Find the longest stretch in my record from before any diagnosis or symptoms and run step 3 on it. Whatever share of sessions fires there is my false-positive floor. Report it. Any later figure that is not well above it means nothing. 5. Now give me, per week: sessions recorded, sessions flagged, the share, the total hours those flagged sessions span, and that as a share of the week. Put my device's own burden figure next to it where I have one. 6. Group consecutive flagged sessions into episodes, breaking an episode when a clean session falls in the middle. For each week give me the number of episodes, how many distinct nights they fall on, and the longest one. This is the part the percentage hides. 7. Split flagged sessions by time of day, night against day, and tell me whether mine cluster. TELL ME PLAINLY: - Where my device's figure and your count disagree, and by how much. - Whether any week where my device shows the same rounded number as another week actually differs underneath. - What my lowest week actually contained, in episodes and nights. - What you could not determine from my data, as a list, before any conclusion. WHAT THIS IS: a reading of one person's own recordings by that person. Not a diagnosis, not evidence about anyone else, and not a substitute for the device's own classification or for a doctor's.
What this is not. One person’s self-recorded data, re-read by that person. A single case, not evidence about any population, and not medical advice. Burden is one of the things a cardiologist weighs when deciding about treatment, which is exactly why I wanted to know what mine was made of — but what to do about it is a conversation with a doctor, not with a watch and not with this page.