LIFETRACE Analysis Code

Overview

• Description: Python modules for data preprocessing, feature engineering, machine learning modeling, statistical analysis, and visualization of the LIFETRACE dataset.
• Purpose: Provides reusable, documented code for reproducing the analyses and figures presented in the LIFETRACE dataset paper.
• Language: Python 3.10+Let's

Modules in This Directory

• modeling.py: Feature engineering pipeline and XGBoost modeling utilities for predicting wellbeing metrics from multimodal daily signals. Includes lag/rolling feature generation, leave-one-group-out cross-validation, SHAP feature importance analysis, and baseline comparisons.
• plotting.py: Consolidated plotting module for generating all publication figures — EMA completion distributions, valence trajectories, walking protocol adherence, blood biomarker trends, and anthropometric summaries.
• ttest_analysis.py: T-test statistical analysis for discretized outcome comparisons. Discretises future targets (e.g., sleep hours, boredom) into low/high classes and performs independent t-tests with Bonferroni correction to identify significant feature associations.

All modules are organised around explicit responsibilities, keep side effects (file I/O, logging) transparent, and are safe to import as libraries or execute as standalone scripts.

Directory Structure

share/
├── data/              # Original dataset files (CSV)
├── img/               # Generated plots and figures (no subfolders)
├── out/               # Analysis results and model outputs
│   ├── results.csv    # Model evaluation summary from modeling.py
│   ├── feature_importances/  # SHAP plots from modeling.py
│   └── discrete_outcomes/    # T-test results from ttest_analysis.py
└── src/               # Analysis code (this directory)
    ├── modeling.py
    ├── plotting.py
    ├── ttest_analysis.py
    └── README.md

Dataset Access

The LIFETRACE dataset is required to run this code. Dataset files must be placed in the ../data/ directory.

The dataset is available at:

• Zenodo: https://doi.org/10.5281/zenodo.17243647

For access instructions and requirements, see the README.md file on Zenodo directory.

Installation

Requirements

pip install numpy pandas polars scikit-learn xgboost scipy matplotlib seaborn

Optional Dependencies

• SHAP: For feature importance visualisations in modeling.py

  pip install shap

• statsmodels: For Bonferroni correction in ttest_analysis.py

  pip install statsmodels

Usage

Running the Complete Modeling Pipeline

cd share/src
python modeling.py

Output:

• out/results.csv: Model performance metrics (RMSE, MAE) for all targets and horizons
• out/feature_importances/: SHAP beeswarm plots for each target-metric combination

What it does:

1. Loads and merges EMA, wearable, and contextual data
2. Generates lag features (1–3 days) and rolling statistics (7-day windows)
3. Trains XGBoost regressors with GridSearchCV and leave-one-group-out CV
4. Evaluates models on holdout period (days 29–49)
5. Compares against global mean, per-user mean, and lag-1 baselines

Generating All Figures

python plotting.py

Output:

• img/boxplot_49days.pdf, img/cdf_42days.pdf: EMA completion distributions
• img/positive_valance_boxplot.pdf, img/negative_valance_boxplot.pdf: Valence trajectories
• img/walking_consistency*.pdf: Walking protocol adherence
• img/steps_and_sport_time*.pdf: Activity trends and availability
• img/blood_metrics*.pdf: Blood biomarker distributions
• img/age_distribution.pdf, img/sex_*.pdf: Anthropometric summaries
• img/*_city_vs_park.pdf: City vs. park group comparisons

What it does: Aggregates plotting routines originally scattered across multiple scripts into a single, import-friendly module. Each function can be called individually or orchestrated via generate_all_plots().

Running T-Test Analysis

python ttest_analysis.py

Output:

• out/discrete_outcomes/discrete_analysis_*.csv: Feature-level t-test results with p-values, Cohen's d, and significance markers
• out/discrete_outcomes/discrete_analysis_*.txt: Metadata summaries (thresholds, sample sizes)

What it does:

1. Reuses the feature engineering pipeline from modeling.py
2. Discretises continuous targets (sleep hours, boredom, loneliness) into low/high groups (40th/60th percentiles)
3. Performs independent t-tests between groups for each feature
4. Applies Bonferroni correction for multiple comparisons
5. Reports effect sizes (Cohen's d) and significance levels

Using as a Library

from modeling import DatasetPaths, ModelingConfig, prepare_daily_dataframe
from plotting import PlotConfig, plot_valence_boxplots
from ttest_analysis import run_analysis

# Load preprocessed data
paths = DatasetPaths()
config = ModelingConfig()
daily_df = prepare_daily_dataframe(paths, config)

# Generate a single plot
plot_config = PlotConfig()
plot_config.ensure_output()
plot_valence_boxplots(plot_config)

# Run custom t-test analysis
outputs = run_analysis()
print(outputs)  # Dict of written file paths

Configuration

Modeling Configuration

Edit ModelingConfig in modeling.py:

• validation_cutoff: Day threshold for train/validation split (default: 28)
• lag_days: Number of lag features to generate (default: 3)
• rolling_window: Window size for rolling statistics (default: 7)
• future_days: Prediction horizons in days (default: (1, 7))
• shap_max_display: Number of features in SHAP plots (default: 15)
• param_grid: XGBoost hyperparameter grid

Plotting Configuration

Edit PlotConfig in plotting.py:

• font_scale: Seaborn font scaling factor (default: 1.5)

T-Test Configuration

Edit DiscreteOutcomeConfig in ttest_analysis.py:

• targets: Tuple of target column names (default: ("sleep_hours", "boredom", "loneliness"))
• future_days: Prediction horizons to analyse (default: (1, 7))
• low_percentile, high_percentile: Discretisation thresholds (default: 40.0, 60.0)
• alpha: Significance threshold after correction (default: 0.05)

Reproducibility Notes

• Random seed: Not explicitly set — XGBoost early stopping introduces minor non-determinism. For exact reproduction, set random_state in XGBRegressor initialisation.
• Z-score normalisation: All percentage-based and continuous features are z-scored within each participant using statistics from days 1–28 (training period).
• Data interpolation: Training data uses Polars .interpolate() to fill missing values; validation data retains raw missingness to reflect real-world deployment.
• SHAP computation: TreeExplainer runs on training data only; plots may take several minutes for large feature sets.

Extending the Code

Adding a New Target Variable

1. Ensure the column exists in the merged daily dataframe
2. Add column name to select_feature_columns() exclusion list if it's a metadata field
3. Rerun modeling.py or ttest_analysis.py

Adding a New Plot

1. Define a new function in plotting.py following the existing pattern:

   def plot_my_new_figure(config: PlotConfig) -> None:
       """Brief description."""
       df = _load_csv(config, "my_data.csv")
       fig, ax = plt.subplots(figsize=(8, 6))
       # ... plotting code ...
       _save_figure(fig, config.output_dir / "my_figure.pdf")

2. Add function call to generate_all_plots()

Customising the Feature Engineering Pipeline

Modify feature_engineering_pipeline() in modeling.py to add new feature transformations (e.g., polynomial features, interaction terms, different window sizes).

Citation

When using this code, please cite the associated dataset and paper:

Paper:

Francesco Bombassei De Bona, Ioana Andreea Câmpanu, Marc Langheinrich, Martin Gjoreski, and Georgiana Juravle. 2025. LIFETRACE: A Longitudinal Multimodal Dataset on Daily Physical Activity, Well-Being, and Habits. Proc. ACM Interact. Mob. Wearable Ubiquitous Technol. 9, 4, Article 161 (December 2025), 33 pages. https://doi.org/10.1145/3770676

Dataset:

Francesco Bombassei De Bona, Ioana Andreea Câmpanu, Marc Langheinrich, Martin Gjoreski, and Georgiana Juravle. 2025. Data from "LIFETRACE: A Longitudinal Multimodal Dataset on Daily Physical Activity, Well-Being, and Habits" [Data set]. Zenodo. https://doi.org/10.5281/zenodo.17249917

Contact

For questions about the code or dataset access:

• Author: Francesco Bombassei De Bona
• Email: francesco.bombassei.de.bona@usi.ch