metrics-adjuster

metrics-adjuster computes conventional and adjusted group-aware metrics for binary prediction models. It provides a typed Python API, a command-line interface, deterministic synthetic demos, self-contained HTML reports, and optional parquet artifacts for downstream inspection.

Research and software provenance

The underlying adjusted-risk-distribution research was originally led by Jinbo Chen and Sarah E. Hegarty. Hegarty developed the original R implementation, fairRisk, for the aTPR methodology.

Sae-Hwan Park later joined the research team and performed additional VA analyses with his Python rewrite of that workflow. The rewrite was validated internally and reviewed and approved by Hegarty for the team's analysis. It subsequently became the foundation of metrics-adjuster, which is maintained as a separate Python project and does not replace or supersede fairRisk.

For scientific provenance, users of the aTPR methodology should cite the underlying research article and acknowledge fairRisk as the original R implementation. Cite metrics-adjuster separately when using this Python software.

How metrics-adjuster extends the workflow

fairRisk remains the original native R implementation and the methodological reference for the adjusted error-rate work. Compared with the currently documented fairRisk v1.1.1.1 interface, metrics-adjuster exposes additional capabilities for Python-based analysis and reproducible delivery:

These additions broaden the operational workflow; they do not make metrics-adjuster a scientific replacement for fairRisk. Shared aTPR analyses should align the calibration model, density-ratio model, thresholds, reference group, and other estimator settings before comparing results across languages. Intentional differences in supported metrics, defaults, or output schemas should be documented rather than treated as implementation disagreement.

Installation

Install from PyPI (Python 3.11+):

python -m pip install metrics-adjuster

That one command installs both the metrics-adjuster CLI and the Python API. Parquet input and artifact output work out of the box (pyarrow is included).

Other setups:

• CLI only, without a project environment: pipx install metrics-adjuster
• uv project: uv add metrics-adjuster
• conda or mamba env: run the pip command above inside your activated environment
• pinned release: python -m pip install "metrics-adjuster==1.1.1"
• source checkout: see Contributing

Quick Example

metrics-adjuster demo \
  --output-dir demo_outputs \
  --report \
  --save-artifacts \
  --report-figures

For caller-provided CSV or Parquet data:

metrics-adjuster run \
  --input cohort.csv \
  --output-dir adjusted_metric_outputs \
  --group-col group \
  --ref-group ref \
  --response-col outcome \
  --risk-col risk \
  --id-col patient_id \
  --metrics aTPR,aPPV,aNB,aHR \
  --report \
  --save-artifacts \
  --report-figures

The CLI writes one CSV per adjusted metric. Optional outputs include:

• report.html when --report is enabled
• figure_1_calibrated_density.svg, figure_2_weight_ratio.svg, and figure_3_decision_curve.svg when --report-figures is also used
• calibration.parquet and weights.parquet when --save-artifacts is used

Python API

from pathlib import Path

from metrics_adjuster import ColumnSpec, MetricConfig, OutputConfig, adjusted_metrics
from metrics_adjuster.synthetic import generate_synthetic_metrics_data

frame = generate_synthetic_metrics_data(n=600, seed=2026)
config = MetricConfig(
  columns=ColumnSpec(group="group", response="outcome", risk="risk", id="patient_id"),
  ref_group="ref",
  output=OutputConfig(
    calibration_path=Path("artifacts/calibration.parquet"),
    density_ratio_path=Path("artifacts/weights.parquet"),
    include_intermediates=True,
  ),
  random_state=2026,
)

result = adjusted_metrics(frame, config)
print(result.metrics["aTPR"])
print(result.weighted[["cal_risk", "dens_ratio"]].head())

For reports and standalone figures, use adjusted_metrics_report(...) and write_report_figures(...). See the API manual.

Documentation

• Quickstart
• API manual
• CLI manual
• R CLI guide
• Contributing

Citation

If you use the aTPR methodology, cite the underlying research article and acknowledge Sarah E. Hegarty's fairRisk as the original R implementation. If you use metrics-adjuster, cite this Python software separately as:

APA:

Park, S. (2026). metrics-adjuster (Version 1.1.1) [Computer software]. GitHub. https://github.com/SaehwanPark/metrics-adjuster

AMA:

Park S. metrics-adjuster [computer program]. Version 1.1.1. Published 2026. Accessed August 20, 2026. https://github.com/SaehwanPark/metrics-adjuster

License

metrics-adjuster is distributed under the GNU General Public License v3.0.