This document summarizes the Health eHeart Study, which links Fitbit step data to health measures. It notes that health warning signs are often missed, and that real-world walking data is hard to obtain, messy to analyze, and usually not linked to health metrics. The study analyzes daylong and yearly step time-series with features like maximum pace and steps. It finds links between walking patterns and cholesterol levels and decodes daily breathing scores from step data with 61-64% accuracy. Challenges include small clinical groups and correlated features.
5. Problem
[1] Health warning signs often go
unnoticed (or ignored).
[2] “Real life” walking data is:
(a) hard to obtain
6. Problem
[1] Health warning signs often go
unnoticed (or ignored).
[2] “Real life” walking data is:
(a) hard to obtain
(b) messy to analyze
7. Problem
[1] Health warning signs often go
unnoticed (or ignored).
[2] “Real life” walking data is:
(a) hard to obtain
(b) messy to analyze
(c) often unlinked to health measures
8. Problem
[1] Health warning signs often go
unnoticed (or ignored).
[2] “Real life” walking data is:
(a) hard to obtain
(b) messy to analyze
(c) often unlinked to health measures
31. Challenges
[1] Dealing with dramatically unbalanced groups (i.e. the
most interesting clinical populations are tiny).
32. Challenges
[1] Dealing with dramatically unbalanced groups (i.e. the
most interesting clinical populations are tiny).
[2] Engineered features are highly correlated (despite my
best creative efforts).
33. Challenges
[1] Dealing with dramatically unbalanced groups (i.e. the
most interesting clinical populations are tiny).
[2] Engineered features are highly correlated (despite my
best creative efforts).
With more time:
34. Challenges
[1] Dealing with dramatically unbalanced groups (i.e. the
most interesting clinical populations are tiny).
[2] Engineered features are highly correlated (despite my
best creative efforts).
With more time:
— Create more and better features
35. Challenges
[1] Dealing with dramatically unbalanced groups (i.e. the
most interesting clinical populations are tiny).
[2] Engineered features are highly correlated (despite my
best creative efforts).
With more time:
— Create more and better features
— Link step data with additional clinical/demographic info