A wearable measures what my body did. It rarely knows what I was doing. HRV, resting heart rate and sleep stages are outcomes. The behavior behind them lives in other places.
My texts show who I was talking to and how late. Screen time shows when I was up on my phone. Location shows whether I left the house. A mood log shows how the day felt. A gym log shows the load my body was recovering from. Put next to the ring's numbers, that context turns a low readiness score into a question I can answer.
I started doing this in 2020 with a Google Sheet and a Google Data Studio dashboard. By 2021 it was a Tableau workbook fed by manual exports. Today it's a set of Python scripts, a few scheduled jobs and one local dashboard. Below is the current stack, grouped by what each source tells me. Each entry says how the data reaches my laptop.
The body
Each device measures these signals its own way. I keep every source separate and compare them side by side.
- Apple Watch and iPhone. Data reaches me two ways. The Health Auto Export app writes daily CSVs into an iCloud Drive folder on a schedule. A script copies each new file, untouched, into a dated raw folder. It then merges every file into one daily table and one workout table. When two files disagree about the same day, the newer file wins. That makes the script safe to rerun.
- The full Apple Health history. For everything back to 2015, I use Export All Health Data in the Health app and AirDrop the zip to my Mac. Each export gets its own dated folder, so older pulls stay around for comparison. A parser turns the XML into one CSV per metric.
- Oura. Oura retired personal access tokens in December 2025. So I registered an OAuth app, approved it once in the browser and saved the tokens locally. A script pulls sleep, readiness, SpO2, stress, sessions and 5-minute heart rate. Oura's refresh tokens work only once. The script saves the new pair after every refresh.
- WHOOP. Same pattern through WHOOP's developer API. WHOOP rescores some records after the fact. So each pull starts a week before the last record I already have. Newer versions replace older ones.
One thing in Apple's export is easy to miss. Every timestamp carries the phone's UTC offset at the moment of the export, not the moment of the reading. An August export stamps January heart rate with a summer offset. The instant is still right. The wall-clock hour is off by one for half the year. My parser now converts every timestamp to a real time zone. Sleep records carry their own zone, so they stay correct even on trips.
The training
The body sources tell me how I recovered. These two tell me what I was recovering from.
- Strong. Every set, rep and weight from the gym. Strong exports a CSV from the app. I log single-leg exercises as separate left and right entries. That makes side-to-side strength comparable over time.
- Slopes. Every ski run, with speed, vertical and a GPS track. The app keeps its own export files in iCloud. A script reads them from there.
The mood
- Daylio. A daily mood rating with tags and notes. Daylio backs up its full history to iCloud about twice a day. Each backup holds every entry. My script reads only the newest file and rewrites the table.
- Surveys. Standardized questionnaires like PROMIS-10, the general health measure, filled out through Google Forms. The responses land in a Google Sheet.
The places
- Beiwe. Beiwe is the open-source research platform I work on at Harvard. I enroll myself in its test studies on my own phone. It records GPS, phone motion and survey answers the way a study participant's phone would. A script downloads my data through Beiwe's API in monthly or daily chunks. It skips any chunk it already has.
- Google location history. From Google's Timeline export.
These two look interchangeable. They aren't. Google's current export holds only visits that its own algorithm inferred. The raw trace underneath is gone before the file is made. Movement through a park collapses into a path with no place attached. Once it placed me overnight inside a city park in February. The real spot was almost certainly a building next to the park. My older Google export, covering 2011 to 2021, did contain raw GPS fixes. That level of access is gone now. Beiwe still gives me the raw fixes. With those I can measure time outdoors myself.
The people
- Calls and texts. A script reads the Mac's own call history and Messages databases. iOS doesn't offer this data to third-party apps. The Mac keeps a synced copy anyway. Newer messages store their text in a binary format, which the script also decodes.
- A hashed version. A second script runs the same pipeline but hashes every contact identifier with a salt that stays on my machine. It never writes message text to disk. Digital phenotyping studies use call and text metadata like this as a proxy for social connection.
The phone
- Screen Time for my iPhone, Mac and Apple TV. The Mac's own usage comes from its local Screen Time database. With Share Across Devices turned on, iCloud also syncs the iPhone and Apple TV usage to the Mac. That copy lands in Apple's undocumented binary log format. A script decodes it with a schema adapted from the open-source aw-import-screentime project. It then rebuilds app sessions from the raw foreground and background events.
Turning on Share Across Devices for the iPhone alone wasn't enough. Nothing arrived for over a day. Sync started only after I turned it on for the Mac as well.
The clinic
- Clinical records. My health system sends diagnoses, medications, procedures and vitals to the Health app as FHIR files. They come along inside the full Apple Health export.
- Visit notes. The patient portal lets me view each visit's documents. I saved the ones I needed, like physical therapy, orthopedics and imaging notes, into one file.
How it runs
The calls, texts and screen time exports run every Sunday morning as two scheduled macOS jobs. That schedule matters. The Mac's own app-usage database held about 26 days when I checked. The call history may have a limit too. Anything older than that can't be backfilled.
Getting the jobs to run took one non-obvious step. The scheduler starts processes on its own, not from Terminal. They don't inherit Terminal's Full Disk Access. I had to grant that access to bash and Python directly.
Most sources follow the same layout. Raw files go into a dated folder and are never edited. Scripts build processed tables from them. Each analysis lives in its own project folder with notes on what it found. None of the raw data or processed tables are in a public repository.
On top sits one dashboard, a single HTML file that makes no network requests. Six sources feed it so far. They are the Apple Watch, Oura, WHOOP, Daylio, calls and texts, and screen time. It shows sleep, heart and recovery, fitness, mood, communication and screen time over the same week, month or year. Metrics with more than one source can show them all on one scale. Beiwe, location, the gym and ski logs and the clinical records still live in their own project folders.
Why the context matters
The clearest case so far was my own torn ACL. My iPhone's walking asymmetry jumped to 93% the morning after the fall, days before my watch reacted. The full story is in the ACL write-up.
The questions I'm working on now need two or more of these sources at once. Do late-night texting and phone use show up in the next night's sleep? Does time outdoors track with mood? And what happens to my sleep, movement and social contact during Ramadan?