620
edits
Changes
Jump to navigation
Jump to search
mLine 24:
Line 24: − + − + Line 50:
Line 50: + + + + + + + + + + + + + + + + + + + + + + + + +
expand based on SRC discussions
=== Methods ===
=== Methods ===
For all four metrics, I used my existing data exports from Oura and Fitbit to [[Open Humans]], filtering for data that fell within the time period of interest. I marked down the start of the overall stress period as starting on September 13, 2021 as this corresponds to the end of the general summer vacation period in France and my own end-of-summer vacation. For the "end of stress" date, I marked down April 8, 2022 as the date where I got my job offer for leaving. To create a pre- and post-stress baseline view, I additionally exported the 10 weeks of data prior to the stress-onset as well as all remaining data until May 31, 2022 to evaluate the evolution after the stress period. The raw data consists of a single data point per day for each of the variables and I calculated weekly averages to remove some of the noise due to differences in weekdays/weekends etc.
For all four metrics, I used my existing data exports from Oura and Fitbit to [[Open Humans]], filtering for data that fell within the time period of interest. I marked down the start of the overall stress period as starting on September 13, 2021 as this corresponds to the end of the general summer vacation period in France and my own end-of-summer vacation (red dashed line in plots). For the "end of stress" date, I marked down April 8, 2022 as the date where I got my job offer for leaving (red solid line in plots). To create a pre- and post-stress baseline view, I additionally exported the 10 weeks of data prior to the stress-onset as well as all remaining data until May 31, 2022 to evaluate the evolution after the stress period. Additionally, I marked two more points of interest within this time period: November 20, 2021 as the start point of my job application period (blue dashed line in plots) and May 20, 2022 as the last day of the job application period as the last application results came in on this date. Rather than fitting potential dates to the observed data, these four dates were picked before plotting any data, as such the corresponding findings aren't a result of tweaking cut-off points.
Each of the weekly averages was then plotted on a [[timeline]], alongside an annotation for the time periods to evaluate any changes/trends compared to the baselines. The code to reproduce these visualizations using your own data is available as a [[Jupyter]] notebook<ref>https://exploratory.openhumans.org/notebook/168/</ref>.
The raw data consists of a single data point per day for each of the variables and I calculated weekly averages to remove some of the noise due to differences in weekdays/weekends etc. Each of the weekly averages was then plotted on a [[timeline]], alongside an annotation for the time periods to evaluate any changes/trends compared to the baselines. The code to reproduce these visualizations using your own data is available as a [[Jupyter]] notebook<ref>https://exploratory.openhumans.org/notebook/168/</ref>.
== Results ==
== Results ==
Only towards the end of this period I had some anecdotal experience that made me suspect that things might be different: Towards the end of the stress-period I developed a persistent back/shoulder pain, which – while not fully disappearing – strongly subsided when getting the news that I would '''not''' be offered a job at my current workplace, thus making my exit plan to a different institution would happen (blue solid line). Which made me suspect that there might be a strong sense of relief in this outcome, even if I had thought I would have preferred another outcome. As such it is quite interesting for me to see that the retrospective analysis of my data actually ''shows'' that many of those values already went back to the baseline once I had the option to move, despite still nominally being within the job application stress!
Only towards the end of this period I had some anecdotal experience that made me suspect that things might be different: Towards the end of the stress-period I developed a persistent back/shoulder pain, which – while not fully disappearing – strongly subsided when getting the news that I would '''not''' be offered a job at my current workplace, thus making my exit plan to a different institution would happen (blue solid line). Which made me suspect that there might be a strong sense of relief in this outcome, even if I had thought I would have preferred another outcome. As such it is quite interesting for me to see that the retrospective analysis of my data actually ''shows'' that many of those values already went back to the baseline once I had the option to move, despite still nominally being within the job application stress!
== Addendum after self-research chat presentation ==
I presented these findings [[2022-06-16 Self-Research Chat|during the self-research chat of June 16, 2022]]. During this call a number of questions came up, some of which I investigated afterwards.
=== How correlated are the chosen variables? ===
[[File:PCA test variable alignment.png|thumb|How different variables that I measure correlate with each other. ]]
I chose the four variables presented based on their known association to stress generally and occupational stress in particular. But of course it could be the case that they basically all show the same effect, as they could be highly correlated: A rising body weight might be directly related to a reduction in physical activity (as measured by step counts as a proxy). Similarly, an overall lower fitness might explain a higher resting heart rate and lower HRV.
For another recent self-research project – [[What does clustering tell us]] – I had already looked into how different variables correlate by doing a Principal Component Analysis (PCA). Looking at the variable representation in this PCA (see right), it seems that the only two things which are strongly negatively correlated with each other are my heart rate variability (<code>hrv</code>) and my resting heart rate (<code>resting_hr</code>), as indicated by the two arrows more or less pointing in opposite directions. This is not too unexpected to me, as those two quite typically move in opposite directions in response to any factors.
In contrast, the weight measurements (<code>weight_fitbit</code>) are less well aligned with my resting heart rate/heart rate variability, the corresponding arrow being about 45º off. This indicates that there is still a small correlation, but one that is a lot worse than the case of HRV/heart rate. This is even more extreme when looking at the step count metric (<code>oura_steps</code>), which moves in a 90º away from the resting heart rate and heart rate variability metrics, indicating that there is basically no correlation between the step count and those metrics.
With the exception of resting heart rate and heart rate variability it thus appears like there isn't too strong a correlation between the different metrics looked at here.
=== Do you find anything in your sleep metrics? ===
Another question that came up was whether I felt that the stress also impacted my sleep and whether I could see that in any of the sleep metrics. Anecdotally I could say that it definitely impacted my sleep, in particular leading to problems falling asleep due to ruminating. To investigate this further, I also proceeded to plot four different sleep metrics as measured by the Oura Ring using the same approach as outlined above (weekly averages, same four date annotations) in a similar plot.
[[File:Workstress-sleep.png|thumb|Using the same approach to look at four sleep metrics: Total sleep duration, time awake during the night, sleep onset latency, sleep efficiency.]]
As the four metrics I chose:
* Sleep onset latency (time until falling asleep in seconds)
* Sleep efficiency (total time asleep divided by total time asleep plus sleep onset latency and time awake during the night)
* Total sleep amount (in seconds)
* Time awake during the night (in seconds)
These graphs overall seem a lot more noisy, despite being based on weekly averages as well and are also less clearly correlated than the other metrics outlined above. I think partially this is due the inherent problem with detecting sleep that all wearable devices have that do not make use of EEG recordings. Like other wrist/finger-worn wearable devices the Oura Ring uses a combination of accelerometer and heart rate data to predict different sleep stages. Unfortunately, it is not trivial to confidently distinguish between "being asleep in bed" and "laying in bed awake" and as a result a lot of time that is spent being actually awake will be classified as "light sleep" instead. As all of the sleep metrics make explicit or implicit use of identifying sleep-time, a lot of signal here might be lost due to misclassifications. Additionally, as the PCA plot (above) shows all the different sleep metrics seem to be quite correlated between each other one way or another. Despite these limitations, there seems to be a particular period of outlier for in the "time awake" data which is between the onset of the work stress and the start of the job application period, where the awake-time goes up (and the 'sleep efficiency' values go down correspondingly)
== References ==
== References ==