README.md

---
license: cc-by-4.0
pretty_name: ActionAudit Public Data Companion
task_categories:
- tabular-classification
tags:
- calibration
- selective-labels
- survival-analysis
- synthetic-data
---

# ActionAudit Public Data Companion

This repository contains independently generated demonstration data for the
ActionAudit reproducibility package.

It contains no actual SUPPORT2 or MIMIC-IV patient rows.

## Directories

### `synthetic/`

The project's core Phase 0 synthetic benchmark contained
5,000 generated records.

This public file contains 500 independently generated
demonstration rows.

The sample illustrates the prediction, outcome, observation, action, and timing
interfaces. It is not the frozen row-level cohort used to calculate the paper's
reported numerical estimates.

### `support2_interface_mock/`

The SUPPORT2 source contains 9,105 patient records.

The frozen ActionAudit evaluation split contained 1,821
patients and used a 60-day mortality horizon.

This directory contains 182 independently generated mock
rows. The row count is approximately 10% of the frozen evaluation split.

Actual SUPPORT2 patient records shared here: **0**.

Researchers seeking empirical reproduction should obtain SUPPORT2 from its
authorized provider and run the public preprocessing scripts.

### `mimic_iv_interface_mock/`

The frozen ActionAudit MIMIC-IV cohort contained 155,059
admissions. The test split contained 31,012 admissions and used a
30-day mortality horizon.

This directory contains 3,101 independently generated mock
rows. The row count is approximately 10% of the frozen test split.

Actual MIMIC-IV admissions shared here: **0**.

Researchers seeking empirical reproduction must obtain credentialed MIMIC-IV
access through PhysioNet and accept the applicable data-use agreement.

## What the interface samples can do

The samples can be used to:

- inspect expected column names;
- test file loading;
- develop data adapters;
- test simple analysis code;
- understand the ActionAudit data interface.

They cannot reproduce the clinical estimates reported in the paper.

## Important distinction

These files are not:

- subsamples of the clinical datasets;
- anonymized clinical rows;
- transformed clinical values;
- patient-level predictions from the paper;
- statistical replicas intended to preserve real patient distributions.

They are newly generated demonstration records.