I Repatriated My Heart Rate

I have worn some form of fitness tracker since 2020. Resting heart rate, sleep stages, steps, workouts, weight from the connected scale: my body has been emitting a telemetry stream for six years. Until this spring, every byte of it lived exclusively in the clouds of three vendors, which means the most intimate dataset I generate was the one I held the least. A measurer who does not hold his own measurements is a sensor, not a scientist.

So in April I repatriated it: every extractable byte, into one open-format database on my own hardware, with an automated daily pipeline going forward. Here is the report, including the part where the export buttons got graded.

The extraction audit

Three vendors, six years, one legally mandated export apiece (GDPR and its cousins oblige them, which is the only reason some of these buttons exist). Graded:

Vendor	Request to delivery	Format	Completeness	Grade
Watch vendor	26 hours	Mostly JSON, documented	98% (heart rate at full resolution)	B+
Scale vendor	9 days	CSV, undocumented columns	~85% (body composition downsampled)	C-
Old tracker (2020-2022)	31 days, two support tickets	ZIP of oddities	61% by my count	F

That F deserves its paragraph. The old tracker's export contained daily summaries but not the intraday data the device demonstrably recorded and displayed for years. The high-resolution stream, an estimated 4.1 million samples of my own heart, exists only in their warehouse, where it presumably still trains someone's models. The data about my body is complete somewhere; just not anywhere I can reach. File that sentence under everything this site is about.

Total recovered: 6.2 million data points, 1.9 GB, now sitting in SQLite files (a format with a documented commitment to readability until at least 2050) on the NAS, in the 3-2-1 backup rotation with the photos. The daily sync pipeline runs at 03:00 and has failed twice in 90 days, both self-healed on retry.

What owning the merged dataset immediately taught me

The repatriation paid for itself inside a month, because the vendors' walls had been hiding cross-correlations. Each app knew its silo; none could join the tables. With everything in one database, three queries produced findings six years of dashboards never surfaced:

Alcohol and resting heart rate. A single evening drink raises my sleeping heart rate by a median 4.2 bpm and cuts measured deep sleep by 19 minutes. I knew the literature said this. There is a difference between the literature and a chart of your own sinus node, and the difference is behavioral: my drinking has dropped by roughly half since I plotted it, with no willpower expended.

The scale and the watch disagree about progress. During strength-training blocks, weight is flat while resting heart rate falls. Either metric alone misleads; the join tells the truth. No vendor dashboard could perform this join because the join key was me, and I was not for sale to each other.

Sleep predicts tomorrow's discipline. Nights under 6.5 hours are followed by a 31% drop in the probability I complete the next day's planned workout. My fitness limiter was never fitness. It was bedtime, and the proof required owning both tables.

The principle, generalized

Health data is where the stakes of data ownership stop being abstract. Financial records get statements; photos get sympathy when lost. But the longitudinal record of your heart is irreplaceable in the strictest sense (you cannot re-measure 2021) and it is precisely the dataset vendors guard most jealously, because it is worth the most.

My rule going forward, and my recommendation: any device that measures your body must pass the export test before purchase, not after. Request the sample export, check the resolution, time the process. A device whose data you cannot fully extract is not measuring you for you.

Confidence that in ten years I will consider this repatriation among the five most valuable weekends of the decade: 0.89. The heart generates about 100,000 data points a day. They should beat somewhere you own.