I fed 28 days of health data into an LLM. The patterns feel obvious in hindsight, but I couldn’t see them in the charts.
Restorative sleep predicts my next-day stress. When deep and REM sleep drop below 30% of total sleep, my average stress rises to 38. Above 45%, it falls to 24—a 37% reduction.
The stress triggers are blunt:
- Low restorative sleep (<1h) → +12 stress points next day
- Missed sleep data → +15 stress points next day
- Short total sleep (<6h) → +8 stress points next day
Missed sleep data is the worst offender. The nights I forget to charge my watch or fall asleep on the couch are usually the nights I need tracking most.
Patterns I Couldn’t See
My best recovery day is Sunday, with an average of 2h15m of deep sleep. Friday is worst, at 1h05m. The week’s accumulated fatigue is visible in the data.
Stress peaks at 11am-1pm and 6-8pm. The first is probably work pressure hitting its stride. The second might be the commute, or just the body’s natural circadian dip meeting whatever chaos the day delivered.
Stress is lowest from 7-9am, when my morning routine is working as designed.
The Prediction Problem
Based on last night’s 50 restorative minutes, or 22% of total sleep, today’s predicted stress is 35-45 and likely to pass 40 by evening. The current reading of 23 is still below that range.
The forecast turns a retrospective chart into a decision: aim for at least 7.5 hours tonight, avoid alcohol and screens, and sleep before 11pm.
My deep sleep improves 40% when I sleep before 11pm. Knowing that number makes the tradeoff concrete. An extra hour of Netflix costs me almost half my recovery.
From Metrics to Recommendations
None of the five features is revolutionary: restorative sleep ratio, sleep-to-stress correlations, day-of-week patterns, time-of-day patterns, and predictive alerts. Their value comes from combining them into a recommendation instead of scattering them across a dozen charts.
A dashboard showed me each metric. The LLM connected last night’s sleep to today’s stress.
Take care of yourselves out there. Get some good rest tonight.
Related: When AI Replaces Dashboards - The earlier post about reverse-engineering Garmin’s API to understand recovery telemetry.