ClarusC64/clinical-constraint-pressure-v0.2

What this dataset does

This dataset tests whether a model can estimate clinical constraint pressure.

The task is not to identify current illness severity.

The task is to classify how much pressure the patient system is under relative to available reserve.

What changed in v0.2

v0.2 adds counterfactual and adversarial cases.

Some rows have the same oxygen requirement or vasopressor requirement but different reserve states.

Some high-looking cases have preserved reserve and improving trends.

Some medium-looking cases are actually high pressure because reserve is falling and support needs are rising.

This makes the task harder than v0.1.

Core stability idea

Constraint pressure is not the same as visible severity.

Pressure rises when support needs increase and reserve capacity falls.

A patient with moderate support needs may be high pressure if reserve is collapsing.

A patient with high support needs may be medium pressure if reserve is preserved and support needs are falling.

Correct classification requires reasoning across support level, reserve capacity, organ function, and direction of change.

Prediction target

The label column has three classes.

Label 0 means low constraint pressure.

Label 1 means medium constraint pressure.

Label 2 means high constraint pressure.

Row structure

Each row contains:

• —scenario_id
• —oxygen_requirement
• —vasopressor_requirement
• —renal_function
• —urine_output
• —mental_status
• —respiratory_reserve
• —circulatory_reserve
• —renal_reserve
• —support_trend
• —reserve_trend
• —label

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

renal_function uses:

• —normal
• —mild_impairment
• —moderate_impairment
• —severe_impairment

urine_output uses:

• —normal
• —reduced
• —oliguria
• —anuria

mental_status uses:

• —baseline
• —mild_confusion
• —confused

reserve fields use:

• —high
• —medium
• —low
• —very_low

support_trend uses:

• —falling
• —stable
• —rising

reserve_trend uses:

• —improving
• —stable
• —worsening

Evaluation

Submissions must contain:

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 oxygen requirement, vasopressor requirement, or current illness severity.

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

The goal is to evaluate whether models can detect pressure building before collapse becomes obvious.

License

MIT