Fitbit

[carissalow]

Originally posted by @Meng6

For heartrate sensor or sleep sensor, we extracted two files from the Fitbit raw data: summary data and intraday data. For step sensor, we only extracted intraday data.

1. Heartrate Sensor

Data Pre-processing

Summary data (fitbit_heartrate_summary_with_datetime.csv) contains daily features extracted by Fitbit directly: "restinghr", "caloriesoutofrange", "caloriesfatburn", "caloriescardio", "caloriespeak".

Intraday data (fitbit_heartrate_intraday_with_datetime.csv) contains the heartrate column and heartrate_zone column during every time interval (e.g. 15s, 1min). According to Fitbit doc, we assign a heartrate_zone (outofrange, fatburn, cardio, peak) for each row (time period).

Feature Extraction

We extract features for different day segments (epochs) from intraday data: "maxhr", "minhr", "avghr", "medianhr", "modehr", "stdhr", "diffmaxmodehr", "diffminmodehr", "entropyhr", "minutesonoutofrangezone", "minutesonfatburnzone", "minutesoncardiozone", “minutesonpeakzone".

2. Sleep Sensor

Data Pre-processing

There are two versions of Fitbit’s sleep API (version 1 and version 1.2), and each provides raw sleep data in a different format. We extract summary data (fitbit_sleep_summary_with_datetime.csv) and Intraday data (fitbit_sleep_intraday_with_datetime.csv) according to the following rules:

• Sleep level. In v1, sleep level is an integer with three possible values (1, 2, 3) while in v1.2 is a string. We convert integer levels to strings, asleep, restless or awake respectively.
• Count summaries. For Summary data, v1 contains count_awake, duration_awake, count_awakenings, count_restless, and duration_restless fields for every sleep record while v1.2 does not.
• Types of sleep records. v1.2 has two types of sleep records: classic and stages. The classic type contains three sleep levels: awake, restless and asleep. The stages type contains four sleep levels: wake, deep, light, and rem. Sleep records from v1 will have the same sleep levels as v1.2 classic type; therefore we set their type to classic.
• Unified level of sleep. For intraday data, we unify sleep levels of each sleep record with a column named unified_level. Based on this Fitbit forum post , we merge levels into two categories:
  • For the classic type unified_level is one of {0, 1} where 0 means awake and groups awake + restless, while 1 means asleep and groups asleep.
  • For the stages type, unified_level is one of {0, 1} where 0 means awake and groups wake while 1 means asleep and groups deep + light + rem.
• Short Data. In v1.2, records of type stages contain shortData in addition to data. We merge both to extract intraday data.
  • data contains sleep stages and any wake periods > 3 minutes (180 seconds).
  • shortData contains short wake periods representing physiological awakenings that are <= 3 minutes (180 seconds).
• The following columns of Summary data are not computed by RAPIDS but taken directly from columns with a similar name provided by Fitbit's API: efficiency, minutes_after_wakeup, minutes_asleep, minutes_awake, minutes_to_fall_asleep, minutes_in_bed, is_main_sleep and type
• The following columns of Intraday data are not computed by RAPIDS but taken directly from columns with a similar name provided by Fitbit's API: original_level, is_main_sleep and type. We compute unified_level as explained above.

Here are the columns (except time related columns) we’ve got:
Summary data: each row is one sleep record

• v1.2: "efficiency", "minutes_after_wakeup", "minutes_asleep", "minutes_awake", "minutes_to_fall_asleep", "minutes_in_bed", "is_main_sleep", "type"
• v1: v1.2’s columns + ["count_awake", "duration_awake", "count_awakenings", "count_restless", "duration_restless"]

Intraday data: "original_level", "unified_level", "is_main_sleep", "type"

• classic type: sleep information of every 60s
• stages type: sleep information of every 30s

Feature Extraction

Currently, RAPIDS only extracts daily features from summary data: ["sumdurationafterwakeup", "sumdurationasleep", "sumdurationawake", "sumdurationtofallasleep", "sumdurationinbed", "avgefficiency", “countepisode"] for main/nap/all sleep periods.

3. Step Sensor

Data Pre-processing

Only intraday data (fitbit_step_intraday_with_datetime.csv) is parsed from step raw data. It contains the steps (step count) column during every time interval (e.g. 1min).

Feature Extraction

We extract features for different epochs from intraday data according to the following rules:

• Active and sedentary bouts. If the step count per minute is smaller than THRESHOLD_ACTIVE_BOUT (default value is 10), that minute is labelled as sedentary, otherwise, is labelled as active. Active and sedentary bouts are periods of consecutive minutes labelled as active or sedentary.
• validsensedminutes feature is not available for Step sensor as we cannot determine the valid minutes based on the raw Fitbit step data.
• We can exclude the time periods when the participant is asleep. There are two methods: (1) exclude the FIXED time periods of each day, you can define the start time and end time in config.yaml. (2) exclude the FITBIT_BASED time periods according to sleep summary data. (is_main_sleep != "main")

Here is the list of features:

• ALL_STEPS: ["sumallsteps", "maxallsteps", "minallsteps", "avgallsteps", "stdallsteps"]
• SEDENTARY_BOUT: ["countepisode", "sumduration", "maxduration", "minduration", "avgduration", "stdduration"]
• ACTIVE_BOUT: ["countepisode", "sumduration", "maxduration", "minduration", "avgduration", “stdduration"]