jwkoh94/bridge-trajectories

BRIDGE Adherence Trajectory Archetypes (aggregate)

Aggregate, group-level adherence trajectory archetypes for lipid-lowering therapy, accompanying the BRIDGE adherence-trajectory scorer. This dataset contains ONLY aggregate group-mean curves and aggregate counts. It contains NO individual patient records: no patient identifiers, no individual rows, no per-individual dates, and no free text that could identify a person.

What this is

The four adherence trajectory groups (G1 gradual decline, G2 early discontinuation, G3 rapid decline, G4 persistent adherence) are the production group-based trajectory model (GBTM) classes. Both the archetype curves and the group prevalences in this dataset are computed on a single consistent source: the BRIDGE development cohort of 50,857 patients. For each group we publish the group-mean monthly adherence (proportion of days covered, PDC), a smoothed display curve, and the group prevalence.

Files

• data/archetype_curves_smooth.csv long form group, label, month, smooth_pdc. The smoothed group archetype curve per group, on a monthly grid (months 1 to 50). smooth_pdc is the group-mean adherence, smoothed with a low-order spline for display.
• data/archetype_empirical_points.csv long form group, label, month, empirical_pdc. The raw empirical group-mean monthly PDC on the same cohort (the points the smooth curve is fitted to).
• data/group_prevalence.csv group, label, prevalence, n_patients. Aggregate group sizes and prevalences. Total cohort N = 50,857; group sizes G1 5,300 (10.4%), G2 20,633 (40.6%), G3 14,654 (28.8%), G4 10,270 (20.2%).
• data/trajectory_archetypes.json the full aggregate bundle: curves, empirical points, prevalences, group definition, seed and provenance, in one JSON.

Archetype shapes

• G1 gradual decline: high early coverage that holds for years, then erodes in the later windows.
• G2 early discontinuation: coverage collapses within the first year and does not recover.
• G3 rapid decline: sustained coverage for roughly two years, then a steep fall to near zero.
• G4 persistent adherence: high coverage maintained across the whole follow-up.

Privacy

Aggregate group-level data only. There are no patient identifiers, no individual rows, no individual dates, and no free text tied to a person. The only counts are aggregate group sizes. This dataset is safe to publish openly.

Provenance and reproducibility

Built by build_hf_dataset.py, which reads the aggregate curves served by the BRIDGE Space (themselves computed on the BRIDGE development cohort) and asserts the output is patient-data-free before writing, including a hard check that the group sizes sum to the development cohort N = 50,857. seed 42. Real data only; no fabricated values.

Push to the Hugging Face Hub

You run these (this machine has no Hugging Face auth). Replace <HF_USER>.

huggingface-cli login   # paste a write token

# Create a public dataset repo
huggingface-cli repo create bridge-adherence-archetypes --type dataset

# Push this directory (from hf_dataset/)
git clone https://huggingface.co/datasets/<HF_USER>/bridge-adherence-archetypes
cd bridge-adherence-archetypes
cp -r /path/to/hf_dataset/{README.md,data,build_hf_dataset.py} .
git add -A
git commit -m "BRIDGE adherence trajectory archetypes (aggregate, dev cohort N=50857)"
git push

Alternatively, upload the folder directly:

huggingface-cli upload <HF_USER>/bridge-adherence-archetypes \
  /path/to/hf_dataset . --repo-type dataset