ActiRhythm Toolkit

ActiRhythm Toolkit is a command-line workflow for analysing accelerometer activity data and characterising behavioural and circadian rhythms. It provides preprocessing, hidden Markov model state estimation, downstream rhythm analysis, and optional machine-learning and GLMM workflows.

This repository contains source code, notebooks and published figures only. No subject data is distributed: the maned wolf biologger recordings analysed in the accompanying thesis belong to the data owners and are not redistributable. Supply your own input files under data/ to run the workflow.

Repository layout

src/         importable package (actirhythm CLI and analysis modules)
scripts/     pipeline entry points, figure generators and reporting helpers
notebooks/   exploratory notebooks (outputs stripped)
tests/       unit tests
figures/     figures published in the thesis
docs/         pipeline description and methodological notes
data/        input data (git-ignored, placeholders only)
outputs/     run artefacts (git-ignored)

Installation

Install from PyPI (recommended)

python -m pip install actirhythm-toolkit

Optional extras from PyPI:

python -m pip install "actirhythm-toolkit[ml]"
python -m pip install "actirhythm-toolkit[glmm]"
python -m pip install "actirhythm-toolkit[notebooks]"

PyPI project page:

https://pypi.org/project/actirhythm-toolkit/

Optional capabilities

python -m pip install "actirhythm-toolkit[ml]"
python -m pip install "actirhythm-toolkit[glmm]"
python -m pip install "actirhythm-toolkit[notebooks]"

The glmm extra also requires a local R installation and compatible R packages. Check your environment with:

actirhythm glmm-doctor

Quick Start

Run the complete workflow from a project directory containing your input data and optional config.yaml:

actirhythm

Inspect the planned inputs and outputs before running:

actirhythm --dry-run

Run a named analysis version:

actirhythm full --run-version v1

Commands

actirhythm full
actirhythm preprocess
actirhythm analytics
actirhythm glmm-doctor

Run actirhythm --help for all configuration, data-path, output-path, and resume options.

Input Data

The workflow accepts accelerometer CSV data with a timestamp, subject identifier, activity measure, and optional axis or posture measurements. Use --data-revised-dir to specify an input directory and --raw-data-file to provide a fallback CSV file.

Choosing an activity signal

Set the analysis signal with --activity-signal (default ActMindata).

Some loggers export only the average of each accelerometer axis per epoch. The magnitude of a mean acceleration vector is dominated by gravity, so it describes body orientation, not movement intensity: it is longest when the animal holds a single posture and shortens as orientation varies within the epoch. Such a magnitude can therefore be negatively correlated with real activity and must not be used as an intensity index.

The toolkit still computes activity_xy and activity_xyz from the axes, but treats them as posture descriptors. A genuine intensity metric (ODBA/VeDBA) requires within-epoch variance, which averaged exports do not retain. Where a firmware activity count such as ActMindata is available, prefer it.

Posture descriptors remain useful for validating decoded states: resting states should show a higher and less variable posture magnitude than active ones. The preprocessing stage writes this check to posture_state_validation.csv.

Deployment windows

Records collected after a device is removed cannot be identified from the signal alone — a retrieved logger still registers varying posture and occasional activity while it is being transported, which is indistinguishable from genuine low-activity behaviour. Documented removal times are required.

Because removal times are study metadata that identify individual animals, none are shipped with this repository. Place your own deployment_windows.csv in the project directory and call qc.load_deployment_windows():

subject,end_timestamp,reason
A,2024-04-02 00:00,battery failure before explant
B,2023-07-07 10:15,documented device removal

Alternatively, pass a mapping directly to qc.apply_deployment_windows(). Without a window file, post-removal records are treated as valid observations.

Subject anonymisation

qc.assign_subject_codes() replaces subject names with codes (S01, S02, …) ordered by first observation, and the preprocessing stage writes the mapping to subject_code_mapping.csv. Treat that file as confidential — it is the re-identification key. It is not required to reproduce any result.

Dwell-time filtering

--min-dwell sets the minimum bout length in bins for the post-decoding filter (default 2; at 15-minute sampling that is 30 minutes).

This filter suppresses decoder flicker but also alters the decoded sequence, so transition probabilities and dwell-time summaries describe the filtered series. Choose the threshold against your own data rather than by convention: a value above the unfiltered median bout length discards a large share of genuine bouts. Each run writes dwell_time_sensitivity.csv comparing 15-, 30- and 60-minute thresholds so the choice can be justified.

Development

1. Clone the repository:

git clone https://github.com/nerminjukan/actirhythm-toolkit.git
cd actirhythm-toolkit

2. Create a virtual environment:

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install in editable mode:

python -m pip install -e .

Development extras:

python -m pip install -e .[dev]
python -m pip install -e .[ml]
python -m pip install -e .[glmm]
python -m pip install -e .[notebooks]

Source

Source code and issue tracking: https://github.com/nerminjukan/actirhythm-toolkit

License

GNU General Public License, version 3 or later. See LICENSE.