MissCP

Anonymous research code for missingness-aware conformal prediction.

This repository contains the implementation used for the MissCP experiments: split conformal prediction, missingness-defined Mondrian calibration, model-swap diagnostics, simulation sweeps, and clinical/non-clinical experiment runners. The repository intentionally excludes raw datasets, generated outputs, local experiment logs, paper drafts, and environment-specific paths.

Installation

uv sync

Optional external baselines:

uv sync --extra external-baselines

This extra installs conditionalconformal==0.0.5.

Optional non-clinical ACS dependency:

uv sync --extra non-clinical

Repository Layout

• src/sepsis_mcp/: implementation and experiment entry points.
• tests/: unit tests and smoke tests.
• VERSION: repository version marker.
• pyproject.toml: package metadata and dependencies.

Data

No raw data are included. Public datasets should be downloaded separately from their official sources and passed to commands through CLI arguments.

Common local layout examples:

data/physionet2019/training/
data/gossis/
data/tableshift_cache/
data/home_credit/application_train.csv

Example Commands

PhysioNet-style smoke run:

uv run python -m sepsis_mcp.cli run \
  --data-root data/physionet2019/training \
  --train-hospital A \
  --test-mode both \
  --model-type sklearn_gbdt \
  --train-patients 128 \
  --calibration-patients 64 \
  --test-patients 128 \
  --lookback-hours 6 \
  --horizon-hours 6 \
  --alpha 0.1 \
  --output-dir outputs/mvp

GOSSIS hospital-disjoint smoke run:

uv run python -m sepsis_mcp.gossis_experiment \
  --data-root data/gossis \
  --model-type xgboost \
  --random-state 0 \
  --output-dir outputs/gossis-smoke

Non-clinical validation smoke run:

uv run python -m sepsis_mcp.non_clinical_experiments \
  --dataset airquality \
  --seeds 2 \
  --output-dir outputs/non_clinical/airquality-smoke

Tests

uv run pytest -q

Anonymization Notes

This export is prepared for anonymous review:

• author metadata has been removed;
• raw data, generated outputs, manuscript files, and local logs are excluded;
• absolute local filesystem paths have been replaced with relative examples;
• no Hugging Face token or credential file is included.