digital-markers

digital-markers is an open-source Python library from the CASAS Lab at Washington State University. The package computes marker families from wearable sensor data: exertion, gait, location, and behavioral routine. Additional marker families will be added soon.

Supported markers

Exertion

compute_exertion_markers computes daily motion-only and heart-rate-plus-motion exertion markers, with participant-level summaries.

Gait

compute_gait_markers computes daily wrist-derived walking-bout markers, including step counts, cadence, step intervals, gait frequency, and signal measures.

Location

compute_location_markers returns participant-level and daily location measurements, plus home-location and staypoint tables in details.

Behavioral routine

compute_routine_markers computes model-free routine markers from 10 Hz motion data using fixed acceleration thresholds. Its details include five-minute interval, state-balance, region, and diagnostic tables.

Preparing participant data

The compute_*_markers functions accept a ParticipantData instance and return a MarkerResult. Prepare one DataFrame for each sensor stream and use the canonical column names below. Sensor timestamps must be parseable as datetimes. Exertion requires the motion timestamp and, when heart rate is provided, the heart-rate timestamp to have a Pandas datetime dtype. Gait and routine require a datetime motion timestamp.

The package uses Pandas readers, so inputs can come from CSV or Parquet. This CSV example assumes timestamps have no timezone or UTC offset and already represent local wall-clock time:

import pandas as pd
from digital_markers import ParticipantData

motion = pd.read_csv("motion.csv")
motion["timestamp"] = pd.to_datetime(motion["timestamp"])

heart_rate = pd.read_csv("heart_rate.csv")
heart_rate["timestamp"] = pd.to_datetime(heart_rate["timestamp"])

location = pd.read_csv("location.csv")
location["timestamp"] = pd.to_datetime(location["timestamp"])

participant = ParticipantData(
    participant_id="P001",
    timezone="America/Los_Angeles",
    motion=motion,
    heart_rate=heart_rate,
    location=location,
)

For Parquet, use pd.read_parquet("motion.parquet") (and likewise for the other streams; Pandas needs an optional Parquet engine such as pyarrow). Parquet timestamps are often already datetimes; if a timestamp column was stored as text, convert it with pd.to_datetime before calling the marker functions. Do not infer a timezone from a naive timestamp: establish from the source whether it represents local time or UTC. A naive local timestamp parsed with utc=True is interpreted as UTC, not local time.

Use these column names:

Use numeric values for sensor measurements. Exertion requires all three rotation-rate columns. Gait can run without rotation-rate columns; when provided, supply all three. Routine can run without them, but if any rotation column is present, all three are required. Location requires the latitude and longitude columns, but their values may be missing on rows without a GPS fix. Keep those rows and their timestamps: location uses them to estimate observed recording time. Routine uses location coordinates when supplied, or motion coordinates when present. Its motion input is expected at 10 Hz.

Timestamp and local-time guidance

Use the participant's IANA timezone (for example, America/Los_Angeles) as the source of truth for local time. Keep original timestamps and their source timezone/UTC-offset information in your stored data whenever possible. A timezone-naive timestamp does not identify a timezone: treat it as local only when the source confirms it is already local. If naive timestamps represent UTC, localize them as UTC before converting; do not parse them as local time. For example, for a CSV column known to contain UTC timestamps:

frame = pd.read_csv("motion.csv")
local_timezone = "America/Los_Angeles"
timestamps_utc = pd.to_datetime(frame["timestamp"], utc=True)
frame["timestamp"] = timestamps_utc.dt.tz_convert(local_timezone)

utc=True is appropriate in that example only because the source is known to be UTC (or contains explicit offsets). For naive local timestamps, use pd.to_datetime without utc=True, as in the earlier example. If localizing naive timestamps to a named timezone, choose an explicit policy for daylight saving transitions: a fall-back clock change can make a local time occur twice, and a spring-forward change can make some local times nonexistent. Pandas defaults to raising on ambiguous or nonexistent times; resolve them using source metadata rather than guessing.

The current marker functions do not apply one common local-time conversion:

• Routine converts timezone-aware motion and location timestamps to ParticipantData.timezone, or UTC when timezone is omitted. Naive timestamps are left unchanged and treated as local clock values.
• Exertion, gait, and location do not use ParticipantData.timezone to convert inputs. Their daily grouping follows the timestamps as processed by each implementation; timezone-aware inputs may therefore not align with routine's local-day grouping.

Set ParticipantData.timezone for routine and ensure that motion and heart-rate timestamps use compatible conventions for exertion. If local-day alignment across all four marker families is essential, convert the streams consistently before calling them and validate results around midnight and daylight-saving transitions. Converting timezone-aware values to local time and then removing timezone information can align ordinary local-day boundaries, but it loses the distinction between the two occurrences of a repeated fall-back time; routine removes duplicate motion timestamps. Retain the original timezone-aware timestamps so that this ambiguity can be detected and handled for the study's data.

Calling the marker functions

Import each function from its marker package. All functions take the same participant object:

from digital_markers.exertion import compute_exertion_markers
from digital_markers.gait import compute_gait_markers
from digital_markers.location import compute_location_markers
from digital_markers.routine import compute_routine_markers

exertion = compute_exertion_markers(participant)
gait = compute_gait_markers(participant)
location_result = compute_location_markers(participant)
routine = compute_routine_markers(participant)

Exertion requires motion; heart_rate is optional and can be omitted from ParticipantData or set to None. Gait and routine require motion. Location requires location; motion is optional and contributes timestamps to its recorded-time calculations. Routine can use location data but does not require it. Location and routine currently take no marker-specific keyword arguments; routine configuration uses fixed thresholds.

Optional exertion arguments

Exertion defaults are suitable for typical use. The keyword-only arguments let you customize windowing, signal filtering, and HR availability:

For example:

exertion = compute_exertion_markers(
    participant,
    window_seconds=30,
    filter_signals=True,
    min_hr_values_for_markers=30,
)

Optional gait arguments

Gait accepts keyword-only parameters for sampling rate and bout detection. The defaults are tuned for 10 Hz wrist motion; adjust them to match the sampling rate and study protocol:

Reading results

MarkerResult has markers, daily, details, and metadata attributes:

print(exertion.markers.head())  # participant-level summary
print(exertion.daily.head())    # one row per day, when provided by the marker

print(gait.daily[["date", "num_steps", "num_walking_bouts"]])
print(location_result.details["home_locations"])
print(routine.details["state_balance"])

for name, table in routine.details.items():
    print(name, table.shape)

markers contains participant-level values; daily contains daily measurements when that marker provides them. Routine does not provide a daily table: its five-minute records and summaries are in details. Use details.keys() to see which additional tables are available, and inspect metadata for marker-specific information such as participant ID and sampling rate.

The tables are marker-specific. Exertion daily rows include motion-only and HR-plus-motion exertion, peak HR, HR/motion relationships, and wear time; participant summaries report statistics across daily measurements. Gait daily rows include step and bout counts, walking duration, cadence, step-interval and signal measures; participant summaries include means, standard deviations, trends, and valid/recorded day counts. In summary column names, mean_ denotes the average across days, std_ the spread, slope_ the linear trend over time, and ac1_ lag-one autocorrelation when consecutive-day data is available; exact calculations vary by marker. Location returns daily and participant-level movement summaries, with details["home_locations"] and details["staypoints"] for home and staypoint tables. Routine includes details["intervals"], details["state_balance"], details["regions"], details["motion_diagnostics"], and details["location_diagnostics"].

Missing values are meaningful: a value may be unavailable because the stream was not supplied or there were too few valid observations. In particular, exertion's HR-derived metrics are unavailable when fewer than min_hr_values_for_markers valid HR samples exist; motion-only results remain available. Location can return empty summary tables when there are too few GPS fixes or no recorded minutes. Check the relevant counts and detail tables instead of replacing missing values with zero.

Installation

python -m pip install digital-markers

For development:

git clone https://github.com/WSU-CASAS/digital-markers.git
cd digital-markers
python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e ".[dev]"
pytest