Details for compare_fitbit_healthkit.ipynb

Published by gedankenstuecke


This is a simple example notebook that takes data from both Fitbit and Apple HealthKit and compares how the step counts between the two data sources correlate.


Tags & Data Sources

activity tracking data evaluation Open Humans Healthkit Integration Fitbit Connection


Please log in to comment.

Last updated 2 months, 3 weeks ago

An example of using Open Humans notebooks for activity data analysis

Getting started

In this notebook we want to compare our personal fitbit and Apple HealthKit data. If you want to run this notebook on your personal data in the Open Humans Notebooks you need to have both Fitbit as well as the HealthKit Integration data in your Open Humans account.

Initialize packages

We'll need some Python packages to get our personal data out of Open Humans and to play around with it.

In [1]:
%matplotlib inline

import os, requests, json
import datetime
import pandas as pd
import matplotlib
import numpy as np
import matplotlib.pyplot as plt
from ohapi import api

Let's now start by getting a list of all the data we have on Open Humans. To do this we need our access_token. Using the os library we can access it from the environment variables which are set in our notebooks using os.environ.get('OH_ACCESS_TOKEN')

In [2]:
user = api.exchange_oauth2_member(os.environ.get('OH_ACCESS_TOKEN'))

We now have a json object that contains all the details about our user account. We can look at it:

In [3]:
#user #commented out to not leak all download links ;)
print("not for the published notebook ;-)")
not for the published notebook ;-)

Amongst the keys we see that user["data"] gives us a list with all data files. Iterating over it gives us the following:

In [4]:
for i,f in enumerate(user['data']):
0	fitbit-data.json
1	Runkeeper-activity-data-2017.json
3	ssr_4471.taxonomy.json
5	ssr_4471.taxonomy.json
6	BG85_2.vcf
7	fitbit-data.json
9	jawbone-moves-data.json
10	Runkeeper-activity-data-2018.json
11	Runkeeper-activity-data-2017.json
12	openhumans-bosc2018.pdf
13	Runkeeper-activity-data-2018.json
14	1.json
15	healthkit-export_2015-11-01_2015-11-30_1511288864.58225.json
16	healthkit-export_2015-12-01_2015-12-31_1511288864.58225.json
17	healthkit-export_2015-12-31_2016-01-31_1511297617.26078.json
18	healthkit-export_2016-02-01_2016-02-29_1511297617.26078.json
19	healthkit-export_2016-03-01_2016-03-31_1511297617.26078.json
20	healthkit-export_2016-04-01_2016-04-30_1511297617.26078.json
21	healthkit-export_2016-05-01_2016-05-31_1511297617.26078.json
22	healthkit-export_2016-06-01_2016-06-30_1511297617.26078.json
23	healthkit-export_2016-07-01_2016-07-31_1511297617.26078.json
24	healthkit-export_2016-08-01_2016-08-31_1511297617.26078.json
25	healthkit-export_2016-09-01_2016-09-30_1511297617.26078.json
26	healthkit-export_2016-10-01_2016-10-31_1511297617.26078.json
27	healthkit-export_2016-11-01_2016-11-30_1511297617.26078.json
28	healthkit-export_2016-12-01_2016-12-31_1511297617.26078.json
29	healthkit-export_2017-01-01_2017-01-31_1511297617.26078.json
30	healthkit-export_2017-02-01_2017-02-28_1511297617.26078.json
31	healthkit-export_2017-03-01_2017-03-31_1511297617.26078.json
32	healthkit-export_2017-04-01_2017-04-30_1511297617.26078.json
33	healthkit-export_2017-05-01_2017-05-31_1511297617.26078.json
34	healthkit-export_2017-06-01_2017-06-30_1511297617.26078.json
35	healthkit-export_2017-07-01_2017-07-31_1511297617.26078.json
36	healthkit-export_2017-08-01_2017-08-31_1511297617.26078.json
37	healthkit-export_2017-09-01_2017-09-30_1511297617.26078.json
38	healthkit-export_2017-10-01_2017-10-31_1511297617.26078.json
39	healthkit-export_2017-11-01_2017-11-21_1511297617.26078.json
40	healthkit-export_2017-11-21_2018-01-03_1515014347.27459.json
42	taxonomy-1.json
44	taxonomy-2.json
45	rescuetime.json
46	healthkit-export_2018-01-03_2018-02-28_1526679272.60943.json
47	healthkit-export_2018-03-01_2018-03-31_1526679272.60943.json
48	healthkit-export_2018-04-01_2018-04-30_1526679272.60943.json
49	healthkit-export_2018-05-01_2018-05-18_1526679272.60943.json
50	ftdna-u7u9yzcj.csv
51	fitbit-data.json
52	example_data1.txt
53	23andMe-genotyping.txt
54	23andMe-genotyping.vcf.bz2
55	moves-storyline-data.json
57	Consent-text.pdf
58	rescuetime.json
59	moves-storyline-data.json

Let's iterate again and grab ouy fitbit data

Getting Fitbit data

In [5]:
for f in user['data']:
    if f['source'] == "direct-sharing-102":
        fitbit_download_url = f['download_url']
if not fitbit_download_url: 
    print("NO FITBIT DATA FOUND!")

To download the data we can use requests and subsequently convert the downloaded content into a JSON object

In [6]:
fitbit_temp = requests.get(fitbit_download_url).content
fitbit = json.loads(fitbit_temp)
dict_keys(['activities-overview', 'tracker-minutes-very-active', 'minutes-to-sleep', 'awake-minutes', 'sleep-efficiency', 'sleep-awakenings', 'heart', 'tracker-minutes-fairly-active', 'sleep-minutes-after-wakeup', 'tracker-minutes-lightly-active', 'sleep-minutes', 'tracker-steps', 'weight', 'tracker-activity-calories', 'tracker-elevation', 'tracker-calories', 'tracker-minutes-sedentary', 'tracker-distance', 'time-in-bed', 'weight-log', 'tracker-floors', 'profile', 'sleep-start-time'])

The keys give us an idea what data is in our fitbit store. Let's have a quick look at the step counts. To do this we convert our JSON file into a pandas dataframe which can then be used for plotting.

Visualizing Fitbit steps

In [7]:
dates = []
steps = []
for year in fitbit['tracker-steps']:
    for entry in fitbit['tracker-steps'][year]['activities-tracker-steps']:
        dates.append(datetime.datetime.strptime(entry['dateTime'], '%Y-%m-%d'))
fitbit_df = pd.DataFrame(data = {'date': dates, 
                          'steps': steps})
fitbit_df.index = fitbit_df["date"]

fitbit_df = fitbit_df.groupby(
fitbit_df.index = pd.to_datetime(fitbit_df.index)
fitbit_dfr = fitbit_df.rolling('60d').mean()
fitbit_pt = fitbit_dfr.plot(y=['steps'],figsize=(15,10),fontsize=14)
<matplotlib.legend.Legend at 0x7f00864dea90>

Let's get our HealthKit data

In a first step let's get all healthkit file URLs. As the HealthKit importer uploads a number of JSON files we want to save all of them as a list:

In [8]:
healthkit_urls = []
for f in user['data']:
    if f['source'] == "direct-sharing-14":

Now let's quickly grab all the steps collected and sum them up for a single day. For this we make a for loop over all the download URLs, download the content with requests and then import the content with json.loads(). Ultimately we save all daily steps in a dictionary called healthkit_steps

In [9]:
from collections import defaultdict
healthkit_steps = defaultdict(int)
for url in healthkit_urls:
    healthkit_content = requests.get(url).content
        healthkit_json = json.loads(healthkit_content)
        for entry in healthkit_json['HKQuantityTypeIdentifierStepCount']:
            date = entry['sdate'][:10]
            steps = entry['value']
            healthkit_steps[date] += steps
    except json.JSONDecodeError: 

Now we can convert this dictionary into a nicely formatted pandas.dataframe. First we iterate over the date and steps in our dictionary and convert them into two lists. From there we can create the pd.DateFrame. After doing this we can start plotting our steps as measured by HealthKit

In [10]:
healthkit_dates = []
healthkit_data = []
for date,steps in healthkit_steps.items():
    healthkit_dates.append(datetime.datetime.strptime(date, '%Y-%m-%d'))
hk_df = pd.DataFrame(data = {'date': healthkit_dates, 
                          'hk_steps': healthkit_data})
hk_df.index = hk_df["date"]

hk_df = hk_df.groupby(
hk_df.index = pd.to_datetime(hk_df.index)
hk_dfr = hk_df.rolling('60d').mean()
hk_pt = hk_dfr.plot(y=['hk_steps'],figsize=(15,10),fontsize=14)
hk_pt.legend(["Healthkit Steps"])
<matplotlib.legend.Legend at 0x7fdbf3d5a358>

Comparing HealthKit & Fitbit data

In the next step we merge the two data sources, from fitbit_df and hk_df into a single dataframe using the dates as the keys. To remove some of the noise we also remove all the fitbit_df step entries that are 0, as most of these represent days in which the tracker wasn't worn and not actually not moving at all. Afterwards we can plot our merged data to see how they compare to each other.

In [11]:
df_merged = pd.merge(fitbit_df.reset_index(),hk_df.reset_index(),on="index", how='outer')
In [12]:
df_merged.index = df_merged["index"]
df_merged = df_merged.replace(0, np.nan)
hk_pt = df_merged.plot(y=['steps','hk_steps'],figsize=(15,10),fontsize=14)
hk_pt.legend(["Fitbit Steps","Healthkit Steps"])
<matplotlib.legend.Legend at 0x7fdbf3cd00b8>

This looks weird. For the overlaps at the end of 2015 and starting of end of 2710 we see that HealthKit data seems to yield much higher values than the Fitbit data. But does this represent a consistent signal? Let's switch to a scatterplot. With the small function add_identity we can also plot a diagonal that represents the case where fitbit == healthkit.

In [13]:
import numpy as np

def add_identity(axes, *line_args, **line_kwargs):
    identity, = axes.plot([], [], *line_args, **line_kwargs)
    def callback(axes):
        low_x, high_x = axes.get_xlim()
        low_y, high_y = axes.get_ylim()
        low = max(low_x, low_y)
        high = min(high_x, high_y)
        identity.set_data([low, high], [low, high])
    axes.callbacks.connect('xlim_changed', callback)
    axes.callbacks.connect('ylim_changed', callback)
    return axes

scatter_pt = df_merged.plot.scatter("steps", "hk_steps",figsize=(15,10),fontsize=14)
scatter_pt.set_xlabel("Fitbit Steps")
scatter_pt.set_ylabel("Healthkit Steps")
add_identity(scatter_pt, color='r', ls='--')
<matplotlib.axes._subplots.AxesSubplot at 0x7fdbed6ed940>

And sure enough, we do find a systematic error. Healthkit almost consistently reports ~2x as many steps as Fitbit does. So it looks like you should always trust HealthKit if you easily want to make your 10,000 steps!

But what's really going on here?!

I think the error has a rather simple reason: In my personal case HealthKit is tracking steps using two different methods:

  1. My iPhone and its internal accelerometer/gyroscope are feeding data into HealthKit when I'm carrying it around and they detect movement
  2. I'm also wearing an Apple Watch, which in just the same way records steps over the day (along with heart rate measurements etc.). And the data collected by the Watch end up in HealthKit too.

Which yields us roughly 2x as many recorded steps as compared to Fitbit, which only stores the steps recorded by its own hardware tracker.

On their end Apple cleverly manage this double recording of data: In cases where two data sources record steps at roughly the same time it will give the data recorded by the Apple Watch the lead and ignore the data recorded by the phone. Our complete export of the data done by the Open Humans HealthKit integration on the other hand does not differentiate between the two data sources and instead grabs all the step data it can find. Unfortunately, at this point it does not also export the data source, so there is no easy way to correct for this doubling of data.