Nine Ramadans in Wearable Data — Hassan Dawood

Read the study abstractThe formal Background / Method / Results / Conclusion summary of this study

Background

Ramadan's dawn-to-sunset fast shifts eating, sleep timing and activity for a month, but most physiological evidence on it comes from short clinical studies. This N-of-1 analysis draws on nine consecutive Ramadan cycles (2018 to 2026) of Apple Watch data, three of Oura data, and two of passive phone GPS to characterise within-subject cardiovascular, sleep, behavioural and mobility change relative to matched pre- and post-Ramadan windows.

Method

Each Ramadan was compared against the 30 days before it and the 30 days after Eid using per-year Mann-Whitney U tests. Years were pooled with a sign test on year-level direction and a linear mixed-effects model with a random intercept per year, chosen over naive pooling to avoid pseudoreplication and secular drift. Apple rows were restricted to Watch-native sources; Oura-written and third-party-app rows were removed. Resting heart rate and HRV were decomposed into sleep-window and daytime components. Mobility features (minutes away from home, significant places, trips, last return) were derived from Beiwe GPS with a 150 m home disc and no stored coordinates, with 2024 descriptive and 2025 fully controlled. Wake-time offsets after Eid were tracked as 7-night rolling medians against each year's pre-Ramadan baseline across three sources.

Results

Resting heart rate was lower during Ramadan on both devices in every year (Apple −7.1 bpm, Oura −2.6 bpm, p<0.0001), with a larger drop in the daytime proxy (−8.8 bpm) than the sleep window (−2.0 bpm). An all-day HRV increase (+4.25 ms, p=0.003) disappeared inside the sleep window (−0.3 ms, p=0.91) and matched Oura's null. Daytime HRV was higher (+7.4 ms, p<0.0001, 5 of 5 years), peaking at +17.2 ms in the 16:00 to 20:00 bin. Active energy was lower in the same bin. Bedtime and wake time shifted later (+0.64 h, +0.61 h). Oura recorded more total sleep (+0.58 h, p=0.001) and more REM (+1.9 points, 3 of 3 years); Apple recorded no duration change. Structured workouts fell by about 65% (8 of 9 years). App-logged meditation fell (−4.1 min/day). In 2025 the phone recorded fewer significant places per day (1 vs 3, p=0.001), late-night time out on 7% of days versus 33% (p=0.026), and an earlier last return home (18:27 vs 00:49, p=0.015). Total time away did not change significantly. The wake shift was confined to 2025 and 2026 (+0.7 to +1.6 h across Watch, ring and phone) and absent in earlier years. Pooled post-Eid offsets were +1.0 h, +0.5 h and +0.3 h in the first three weeks.

Conclusion

The findings are hypothesis-generating. Ramadan's lunar drift moves the fasting window through every season over nine years, so season and fasting are confounded by design. The resting heart rate result replicates across devices and windows but has a plausible non-physiological explanation, an activity-linked construction inside Apple's algorithm, that was not ruled out. The HRV correction, the meditation result and the mobility result each contradicted either an earlier analysis or the author's own account. Each was resolved by checking rather than assuming.

I have fasted every Ramadan for the nine years I have worn an Apple Watch. For the last three I have also worn an Oura ring, and for the last two my phone has been quietly logging where I go. None of it was collected with a study in mind. This is how I went back through all of it to ask what a month of fasting actually does to me, and what it took to trust the answers.

The setup

Ramadan is a 29 or 30 day fast from dawn to sunset. No food, no water, and in my case a deliberately slower life. I wake later, exercise much less, protect my sleep, and spend the low-energy afternoons reading or resting. That is the self-report. The question was whether the devices agreed, and where they did not, which one was right.

The Apple Watch gave me resting heart rate, HRV, active energy, sleep episodes, workouts and mindfulness sessions from 2018 onward. Different metrics start in different years. Oura gave me resting heart rate, sleep HRV, sleep stages and its own activity number from 2024. My phone runs Beiwe, the research platform I work on. Its passive GPS gave me a mobility record from March 2024. And I had the Ramadan calendar, nine start dates and nine Eids.

One thing the calendar does that matters for everything below. Ramadan follows the lunar year. It moves about eleven days earlier every year. In 2018 it ran mid-May to mid-June. In 2026 it ran mid-February to mid-March. Over nine years the fasting month has crossed every season.

What I was actually asking

Not "does fasting change my heart rate." That question has one answer and it is obvious. The questions I could actually test were narrower. Does the change replicate across two devices with different sensors and different definitions of resting? Does it hold in every year, or is the average hiding a few big years? When the two devices disagree, is one of them wrong or are they measuring different things? And does the record confirm my own account of the month, or contradict it?

The last question ended up producing the most interesting results. Two of my own beliefs did not survive.

Getting one Ramadan to compare to another

Every year became a set of three windows. Thirty days before Ramadan, Ramadan itself, and thirty days after Eid. Each metric was compared across those windows with a Mann-Whitney test, one year at a time. That gave me eight or nine independent results per metric instead of one.

Pooling those years was the first place I could have cheated. Throwing all the days into one big test would have been the flattering choice. It also would have been wrong. Days within a year are not independent of each other. My baseline resting heart rate has also drifted across the decade. A big pooled test would count that drift as a Ramadan effect. Instead I used two things. A sign test on the direction of each year's result, which only asks "how many years agree." And a mixed-effects model with a random intercept per year, which lets each year have its own baseline. When those two disagree, that is information.

I also tried to build a seasonal control. The idea was to match each Ramadan to the same calendar dates in a non-fasting year. With nine years of dates it was not feasible. I dropped it rather than forcing a weaker version. That means nothing in this study is season-controlled. It is the biggest limitation of the whole thing. I say so wherever a finding could be confounded by it.

The clean-up that ate most of the time

Before any of that, the Apple data had to be made trustworthy. Three problems surfaced, all by checking rather than assuming.

Apple Health merges data from every source. Oura writes into it. Third-party workout apps write into it. So every Apple metric was restricted to rows the Watch itself produced. That was where the first bug turned up. Apple's source field contained a non-breaking space. A plain text match on the device name silently dropped 2018 to 2021. Four years of data were missing until I noticed the row counts were wrong.

The second bug was in sleep. Records that straddle midnight were being split and mislabelled for about 13 percent of nights. The fix was small. My first attempt at the fix was not. It assumed a night-of convention instead of checking Oura's. It changed 80 percent of nights. A correction that touches 80 percent of the data is not a small correction. That mismatch was the signal to stop and verify.

The third was in the sign test itself. Ties were being counted as trials without counting as evidence. Harmless for continuous metrics, which never tie at the year level. It would have quietly understated the workout result, where a third of days are zero. Caught before reporting.

The result that was right and still wrong

Resting heart rate was the headline. It drops about 7 beats per minute on Apple and about 2.6 on Oura, in every year on both devices. That was the easy part.

HRV was where I learned something about method. The first pooled pass said HRV rises during Ramadan by 4.25 milliseconds, p equal to 0.003. Six of eight years agreed. The number was computed correctly. It was also answering a different question than the one I had asked.

The catch came from another project. In an earlier comparison of the two devices I had established that Apple's HRV is an all-day average while Oura's is measured only during sleep. Ramadan moves my bedtime and wake time about 40 minutes later. So the readings that make up Apple's daily average land at different hours during Ramadan. Change when the readings happen and the average moves, with no physiological change at all.

I rebuilt the comparison inside the scored sleep window. The effect vanished. Minus 0.3 milliseconds, p equal to 0.91. Apple and Oura now agreed there was no sleep HRV effect.

Then I asked the follow-up. If not in sleep, what about the day? Daytime HRV really is higher during Ramadan, by 7.4 milliseconds, in five of five years. Broken into four-hour bins it peaks in the late afternoon before iftar, at 17 milliseconds. Active energy in those same bins is lower. Higher HRV where there is less movement is what you would expect from sitting still. That does not rule out a fasting effect on the nervous system. It means the simpler explanation was tested and is still standing.

Nothing in a p-value flags that a metric is responding to the clock instead of the body. What caught it was re-applying a confound I had already found somewhere else.

Where the devices disagreed

Sleep duration is the one finding I could not settle. Oura says I sleep about an hour more during Ramadan, with more REM, in all three of its years. Apple says nothing changes. The Apple sleep record in those years is Watch-only and covers 17 or 18 nights per window. It may just lack the power to see what Oura sees. My own account, that I deliberately protect sleep and cut other obligations to do it, leans toward Oura. That is a lean, not a result.

Exercise was cleaner. Workouts fall by about two thirds, in eight of nine years. Oura's active energy falls in every year. Apple's active energy is pooled-significant. Seven of nine years are lower, and the sign test on that count is not significant. Two years went the other way. That is exactly the case the two-test approach was built to catch. A pooled p-value alone would have overstated it.

Where the record contradicted me

I would have told you I meditate more during Ramadan. The app log says the opposite, down by half in 2024 and by 94 percent in 2026. I checked whether a 2025 sync bug that produced one-minute fragments was driving it. It was confined to 2025. The likely explanation is that what goes up during Ramadan are practices no app logs. The proxy is narrower than the behaviour. But I only know that because I looked.

The second contradiction came from the phone. I assumed Ramadan evenings were busier, with iftar dinners and night prayers pulling me out of the house later. The GPS record for 2025 says the month was more home-bound. Fewer distinct places per day, one instead of three. Late nights out on 7 percent of days instead of 33. Last return home at 18:27 instead of 00:49. Total time away was not significantly different. One recurring Friday-evening outing survived and every other weekday's evening time fell. GPS only starts in March 2024. So this is one controlled year, a single data point. It sits consistently next to the companion communication study. That one found more people reached by message during Ramadan. More checking in, less going out.

The mobility work was not part of the original study. It was the sibling I planned and never started. I added it a month later. It also fixed a shared piece of code. The rank-biserial effect size in the period-comparison script had its sign inverted. I caught it because a metric with a lower Ramadan median came out positive. Magnitudes and p-values were unaffected. It is still worth recording.

A habit I did not know was new

The phone project answered one more question I had not thought to ask. The later wake time that shows up in the pooled Apple result is not present in every year. 2021, 2023 and 2024 show no shift at all. 2025 shows about an hour, on the Watch, the ring and the phone's accelerometer independently. 2026 shows an hour and a half. Sleeping in during Ramadan is a habit I picked up two years ago. Three sensors agree on when.

After Eid it takes about three weeks to fade, pooled across the shifted years. Week one is still a full hour late, week two half an hour, week three twenty minutes. Per year and per device it is messier than that. In two cases the offset never fully returned within 75 days. Another device on the same nights said it had. Two sources disagreeing about the same mornings means their baselines differ. I treat three weeks as the central estimate and not a fact.

What is not in here

Diet and mood were never collected. That was a scoping decision, not a gap. Season is not controlled. There is one subject. Dozens of tests were run without formal correction. Year-over-year agreement is the substitute. Apple's resting heart rate algorithm is proprietary. If it leans on low-activity daytime periods, less daytime activity during Ramadan could lower the number without the heart changing. That is the single largest threat to the headline result. It was not ruled out.

Where the numbers are

Every chart, per-year comparison, pooled test and table lives on the results page. It runs from the resting heart rate result through HRV, sleep, exercise, meditation, mobility and the wake-time follow-up. A timeline at the top lets you put any metric against all nine Ramadans at once. The resting heart rate line is the one to look at first. Nine amber bands, nine dips.