---
language:
- en
license: mit
task_categories:
- text-classification
tags:
- clinical
- clinical-reasoning
- intervention-selection
- competing-interventions
- treatment-choice
- stabilization-pathway
- counterfactual-twins
- stability
- evidence-fidelity
size_categories:
- n<1K
pretty_name: Clinical Competing Interventions v0.2
---

# What this dataset does

This dataset tests whether a model can choose the best stabilising intervention from competing plausible actions.

The task is not diagnosis.

The task is intervention selection under clinical uncertainty.

Each row presents a patient state and three possible intervention options.

The model must predict which intervention is most likely to improve the stabilization trajectory.

# What changed in v0.2

v0.2 adds counterfactual and adversarial cases.

Some rows have similar patient states but different best interventions because response profile, renal risk, ventilation risk, or support needs differ.

Some interventions that look plausible are harmful or insufficient in the given context.

This makes the task harder than simple escalation or severity classification.

# Core stability idea

The best intervention is not always the most aggressive intervention.

A patient may need fluids if fluid responsiveness is good and ventilation risk is low.

A patient may need vasopressors if fluid response is poor and circulatory support is required.

A patient may need oxygen escalation if respiratory load is the primary instability.

A patient may need dialysis if renal failure is the limiting recovery pathway.

A patient may need observation if the system is stable and intervention would add unnecessary burden.

Correct classification requires comparing competing stabilization pathways, not selecting treatment by current severity alone.

# Prediction target

The label column has three classes.

Label 0 means intervention_a is best.

Label 1 means intervention_b is best.

Label 2 means intervention_c is best.

# Row structure

Each row contains:

- scenario_id
- current_severity
- map
- lactate
- urine_output
- oxygen_requirement
- vasopressor_requirement
- fluid_responsiveness
- renal_risk
- ventilation_risk
- intervention_a
- intervention_b
- intervention_c
- label

current_severity uses:

- low
- moderate
- high

oxygen_requirement uses:

- 0 = room air or minimal support
- 1 = low oxygen requirement
- 2 = high oxygen requirement
- 3 = near respiratory boundary

vasopressor_requirement uses:

- none
- low
- moderate
- high

fluid_responsiveness uses:

- good
- poor

renal_risk uses:

- low
- medium
- high

ventilation_risk uses:

- low
- medium
- high

intervention fields may include:

- fluids
- vasopressors
- oxygen_escalation
- dialysis
- observe

# Evaluation

Submissions must contain:

```csv
scenario_id,prediction
test_001,0
test_002,1
test_003,2

Run:

python scorer.py predictions.csv

Optional truth path:

python scorer.py predictions.csv data/test.csv

The scorer reports:

Accuracy
Macro precision
Macro recall
Macro F1
Confusion matrix
Structural Note

This benchmark contains counterfactual and adversarial cases designed to prevent shortcut learning from current severity or a single intervention preference.

The dataset does not expose the hidden rationale behind each label.

The goal is to evaluate whether models can compare competing stabilization pathways under uncertainty.

License

MIT