23.07 Algorütm: Kuidas andmeteadlane enda une ja taastumise lahti häkib?

Key wellness strategies & self-care / recovery techniques

• Use a wearable to collect sleep/recovery-related signals (e.g., wrist/hand device like Oura/Whoop) so you can look for patterns rather than guesses.
• Track more than one recovery signal:
  • Resting heart rate (RHR)
  • Heart rate variability (HRV)
  • Respiratory rate (breaths per minute)
• Separate “sleep quality” from “body recovery.”
  • Instead of defining recovery purely by sleep (to avoid “data leak”), redefine recovery as how rested your body is, regardless of how much sleep you got.
• Use context and exclude abnormal days
  • Remove days marked as “sick” or with unusually low recovery to avoid breaking normal-rhythm assumptions.
• Use moving windows / recalculation vs long-term averages
  • Recompute daily metrics relative to the average of the last ~30 days to understand what’s happening today/tomorrow rather than mixing in long-term trends.
• Adjust training/today’s activity based on morning metrics, not “push through” regardless of how you feel.”
• Consider the psychological effect of being measured
  • Monitoring can nudge better choices even if the measured data never becomes part of a perfect model.

Key productivity / analytics habits that support better self-care

• Build a personal “data pipeline”
  • Export wearable data → compute intermediate metrics → visualize → produce a simple monthly review (often in PDF form).
• Talk to the data (review as a dialogue)
  • Not just “the graph looks nice,” but using it to decide what to do next.
• Don’t outsource thinking to AI/automation
  • AI can help produce graphs, but interpretation depends on the questions you care about.
• Use statistical testing to avoid false conclusions
  • Correlation ≠ causation.
  • Test whether observed differences could be due to chance.
• Avoid common analysis pitfalls
  • “Linearity trap”: correlations may only hold in part of the dataset—plot to check.
  • Misaligned definitions/time windows: understand when “a day” starts/ends in the wearable.
• Validate methods with time-aware testing
  • For time series, use approaches consistent with training on the past and testing on the future (cross-validation is trickier for time series).

Example approach to analyze sleep → recovery (as described)

1. Collect raw signals automatically from the wearable + add subjective logs (caffeine, rest days, feelings).
2. Run correlation/association analysis.
3. Redefine “recovery” to reduce data leakage (don’t let “sleep” trivially define recovery).
4. Combine multiple recovery metrics into one index (dimensionality reduction), after checking reasonable value ranges.
5. Use moving-window normalization (e.g., compare to last 30 days).
6. Keep iterating over months/years because some effects (like day-of-week) need more data to detect reliably.

List of presenters/sources

• Presenters/hosts: Priit Liivak; Eerik Jõgi (host)
• Guest: Eriks Ven (data scientist at Pipedrive)
• Wearables / data sources mentioned: Oura; Whoop; (also “Exo” mentioned for data export); Pipedrive (employer of the guest)
• Misc. / platform: Cloud Code (AI tool used for analysis)