Ten years ago, I did not deliberately begin a health project. I simply kept my phone in my pocket, later put on a watch, and went on with life.
When writing the original article, I exported all my iPhone Health data for the first time. Unzipped, it was a 1.3 GB XML file containing more than three million records. Those steps, heart-rate readings, and workout routes spanned years of entrepreneurship, transitions, and starting again.
My thought then was simple: the data had kept accumulating, but I had no system for looking back and making sense of it.
I still want to connect those scattered records. But connecting them means more than placing two charts side by side. First, I need to understand what each chart records, what it misses, and which questions it cannot answer.
Ten years of data is not a complete ten-year record of my body
In the export described in the original article, records began in August 2015. Apple Watch joined in 2016 and Fitbit in 2023. Different metrics covered different periods; detailed sleep records and some health metrics in that earlier summary were concentrated in the later years.
The inventory then included more than 1.13 million heart-rate records, 29,532 sleep records, and 966 GPS workout tracks. These are historical figures retained from the original article, not results recalculated for this revision.
The word “record” is especially easy to misread. A single sleep episode may be split into many segments; multiple sources may also record the same period. Clearly, 29,532 sleep records do not mean 29,532 nights.
I need a small information card for each metric: which device produced it, its date range, how many days are usable, which periods are missing, and whether the algorithm changed. Without that card, a ten-year span can make the data seem more complete than it is.
The break in weight records is a good reminder. In my earlier analysis, weight records stopped after 2022. The reason was ordinary: my smart scale broke, and I never replaced it.
No record does not mean no change. The gap should remain visible on the chart, rather than allowing AI to fill it with an apparently smooth line.
Two watches cannot certify each other’s accuracy
I once compared the two devices on the same night: total sleep differed by only two minutes, deep sleep by 21 minutes, and REM sleep by 15 minutes.
That example shows different classifications, not which device was accurate. Without an appropriate reference assessment on the same night, two similar readings can both be wrong. Disagreement also cannot be resolved simply by choosing a brand or relying on an impression of its sensors.
I therefore no longer retain the original recommendation to prioritize Apple Watch for REM because it is more accurate. My comparison of a single night cannot support that conclusion.
HRV is even easier to mix up. Apple’s HealthKit documentation defines the relevant field as SDNN; Fitbit’s API provides RMSSD. Both are expressed in milliseconds, but they are calculated differently. Apple field definition; Fitbit API definition
The same “HRV” label does not make them a continuous series. Fixing the source and calculation method, then checking sampling periods, usable data, and device changes, provides a basis for comparison. Fitbit also has HRV data; my earlier claim that it was exclusive to Apple Watch was incorrect.
Before adding up sleep, decide what counts as a night
The original article listed 883 nights from Fitbit and 537 from Apple Watch. Their periods overlap. They cannot simply be added together as 1,420 distinct nights.
For me, rebuilding a sleep summary starts with these definitions:
| Question | My approach |
|---|---|
| Which date contains sleep that crosses midnight? | Fix the date convention and time zone, retaining the original start and end times. |
| What if two devices record the same night? | State the primary source; identify backup data separately rather than counting both. |
| What about time in bed, wakefulness, and sleep stages? | Handle them separately so time in bed is not added to sleep duration. |
| What about naps or a missing night? | List naps separately; do not fill missing data with zero hours. |
| Are Apple Core and Fitbit Light equivalent? | Retain the source names rather than assuming identical algorithms. |
Only after these definitions are settled does it make sense to recalculate monthly averages. I have removed the claim that “33 consecutive months fell below the target” and the conversion of shortfalls into over a thousand hours of sleep debt. Those statements moved too quickly past coverage, deduplication, and health interpretation.
Wearable records can help me understand sleep habits, but a chart alone cannot diagnose insomnia or another condition. The AASM’s 2018 position statement similarly advises interpreting consumer sleep data within an appropriate clinical evaluation, rather than using it as a substitute. AASM statement
What I want to recover is exercise’s place in my life
In 2018, I began entering triathlons and joined 日日三鐵, a group whose members choose swimming, cycling, or running each day. I usually chose running: put on shoes, and I could head out.
One reason was to set an example for my son, showing him how his father cared for his body and maintained discipline. Another came from entrepreneurship: when the body cannot keep up, decisions and execution can suffer too.
The original GPS inventory fell from 340 tracks in 2022 to 39 in 2025. It prompted me to think back on changes in work, growing responsibilities, and exercise gradually being squeezed out of my schedule.
But fewer tracks do not mean exercise fell by the same proportion, much less that it reached zero. Indoor exercise, leaving the device behind, or a synchronization gap may leave no GPS track. Nor does a track represent a fixed duration or intensity.
The original article also combined falling HRV and rising resting heart rate into a conclusion that my autonomic nervous system was “under chronic stress.” I would now withdraw that conclusion. I have not recalculated the raw data for this revision; even if the curves are correct, they cannot independently identify the cause or establish a diagnosis.
The records can help me ask more specific questions: How much have I slept lately? Is exercise still on my calendar? Am I experiencing persistent discomfort? The data does not have to diagnose a disease before it becomes worth changing my habits or discussing my concerns with a physician.
Alongside the three pipelines, draw the data flow
The original design had three parts: Fitbit queries through an API and MCP, longer-term processing of Apple Health XML, and selected summaries published on the website.
Fitbit’s official documentation also announced that the legacy Web API would shut down on October 30, 2026, with subsequent access requiring migration to the Google Health API. I have not verified whether my integration has migrated. This section therefore records the architecture at the time, rather than claiming it remains directly usable today.
XML was the export method I used, not Apple’s only route for data access. Apple also lets users authorize apps to read and write health data; suitability depends on permissions and purpose. Apple guidance
A large XML file lends itself to streaming, parsing it in sections, indexing types, and then preparing daily and monthly summaries. I want AI to receive the necessary aggregates and definitions, rather than handing over my entire health history simply because the file is accessible.
The original claim that “everything runs locally, with no data passing through third-party servers” also needs correction. Fitbit’s API connects to a provider service in the first place. Running an MCP server on a Mac describes only the location of one part of the system. The MCP architecture documentation defines a protocol for providing context; it does not guarantee local model inference.
If a cloud model is used, tool results and context supplied to it may still leave the computer. I therefore need to check separately where data comes from, where it is processed, which fields are sent to the model, and which summaries become public. Without checking the actual configuration, I cannot claim the entire pipeline keeps data local.
Website updates require the same distinction. The original article described a design that updated every ten minutes and deployed automatically after a push. Current site rules have disabled automatic deployment on push: updating data and publishing it are separate steps. A script’s presence does not prove a schedule keeps succeeding; a successful scheduled run does not prove the public page has updated either.
I want each summary to show its data period, most recent successful update, and known gaps. Failures should be visible, rather than leaving a “live” label in place.
Next, make the system worthy of trust
I still want to add lifestyle annotations and have considered integrating more devices. Those are ideas for the next stage, not functions accepted as complete in this article.
Anomaly alerts are similar. “HRV below baseline for three consecutive days” could be a rule to test, not an established medical threshold, and certainly not a reason for AI to prescribe an intervention automatically. First establish whether the data is valid; then decide what to flag and who will assess it.
Before publication, I would also narrow the data shared. Detailed routes, precise timestamps, raw health records, and identifying information should not be published merely because they make an attractive chart. Model queries should likewise retrieve only what is needed for the question.
The most valuable part of these ten years of records is the chance to look back: I once made exercise part of daily life, and I also allowed important things to be gradually crowded out.
The system I want should help me see those changes while honestly preserving what it does not know. When data can be checked and interpretations corrected, I have a better chance of carrying my next choice into everyday life.


💬 Comments
Loading...