Datasets:

ClarusC64
/

clinical-recovery-geometry-v0.2

scenario_id string	current_severity string	heart_rate_trend string	map_trend string	lactate_trend string	urine_output_trend string	oxygen_requirement_trend string	vasopressor_trend string	mental_status_trend string	organ_function_trend string	reserve_capacity string	support_withdrawal_success string	label int64
train_001	moderate	improving	improving	falling	improving	falling	falling	improving	improving	medium	yes	1
train_002	moderate	improving	stable	falling	improving	falling	stable	stable	improving	medium	yes	1
train_003	high	improving	improving	falling	improving	falling	falling	improving	stable	medium	partial	1
train_004	low	stable	stable	falling	worsening	rising	rising	stable	worsening	low	no	0
train_005	moderate	improving	improving	falling	worsening	rising	rising	stable	worsening	low	no	0
train_006	low	improving	stable	falling	stable	stable	stable	stable	improving	high	yes	1
train_007	moderate	stable	improving	falling	improving	falling	falling	stable	improving	medium	yes	1
train_008	high	improving	stable	falling	improving	falling	falling	improving	improving	medium	partial	1
train_009	moderate	improving	improving	falling	stable	rising	rising	stable	worsening	low	no	0
train_010	low	stable	improving	stable	worsening	rising	stable	worsening	worsening	low	no	0
train_011	moderate	stable	stable	rising	worsening	rising	rising	worsening	worsening	low	no	0
train_012	high	improving	improving	falling	improving	falling	falling	stable	improving	medium	partial	1
train_013	low	improving	stable	falling	improving	stable	falling	stable	improving	high	yes	1
train_014	moderate	improving	stable	falling	worsening	rising	stable	stable	worsening	low	no	0
train_015	high	stable	improving	falling	improving	falling	falling	improving	stable	medium	partial	1
train_016	low	improving	improving	falling	stable	falling	stable	stable	improving	high	yes	1
train_017	moderate	stable	stable	falling	worsening	rising	rising	stable	worsening	low	no	0
train_018	high	improving	improving	falling	worsening	rising	rising	stable	worsening	low	no	0
train_019	moderate	improving	improving	falling	improving	falling	stable	stable	improving	medium	yes	1
train_020	low	stable	stable	stable	worsening	rising	stable	worsening	worsening	low	no	0
train_021	high	improving	stable	falling	improving	falling	falling	improving	improving	medium	partial	1
train_022	moderate	improving	improving	falling	worsening	stable	rising	stable	worsening	low	no	0
train_023	low	stable	improving	falling	improving	falling	stable	stable	improving	high	yes	1
train_024	moderate	stable	improving	stable	worsening	rising	rising	worsening	worsening	low	no	0
train_025	high	improving	improving	falling	improving	falling	falling	improving	improving	medium	partial	1
train_026	low	improving	stable	falling	improving	stable	stable	stable	improving	high	yes	1
train_027	moderate	stable	stable	rising	worsening	rising	rising	stable	worsening	low	no	0
train_028	high	improving	improving	falling	worsening	rising	rising	worsening	worsening	low	no	0
train_029	moderate	improving	stable	falling	improving	falling	falling	stable	improving	medium	yes	1
train_030	low	stable	stable	falling	worsening	rising	stable	stable	worsening	low	no	0
train_031	high	stable	improving	falling	improving	falling	falling	improving	stable	medium	partial	1
train_032	moderate	improving	improving	falling	stable	rising	rising	stable	worsening	low	no	0
train_033	low	improving	improving	falling	improving	falling	stable	stable	improving	high	yes	1
train_034	moderate	stable	improving	stable	worsening	rising	rising	worsening	worsening	low	no	0
train_035	high	improving	stable	falling	improving	falling	falling	improving	improving	medium	partial	1
train_036	low	stable	stable	stable	worsening	rising	stable	worsening	worsening	low	no	0
train_037	moderate	improving	improving	falling	improving	falling	falling	stable	improving	medium	yes	1
train_038	high	improving	improving	falling	worsening	rising	rising	stable	worsening	low	no	0
train_039	low	improving	stable	falling	improving	stable	falling	stable	improving	high	yes	1
train_040	moderate	stable	stable	rising	worsening	rising	rising	worsening	worsening	low	no	0

What this dataset does

This dataset tests whether a model can distinguish genuine recovery from superficial improvement.

The task is not to detect whether one marker improved.

The task is to decide whether the overall patient state is moving toward a healthier stability basin.

What changed in v0.2

v0.2 adds misleading improvement cases.

Some rows show improving heart rate, MAP, or lactate while urine output, oxygen requirement, vasopressor trend, or organ function worsens.

v0.2 also adds support withdrawal status.

This makes the task harder than v0.1.

Core stability idea

Recovery is structural, not cosmetic.

A patient may improve on visible markers while still losing reserve or becoming more dependent on support.

A patient may still look severe while moving coherently toward recovery.

Correct classification requires reasoning across trajectory, reserve, organ function, and whether support can be withdrawn without destabilization.

Prediction target

The label column is binary.

Label 1 means genuine recovery trajectory.

Label 0 means superficial improvement or non-recovery.

Row structure

Each row contains:

scenario_id
current_severity
heart_rate_trend
map_trend
lactate_trend
urine_output_trend
oxygen_requirement_trend
vasopressor_trend
mental_status_trend
organ_function_trend
reserve_capacity
support_withdrawal_success
label

current_severity uses:

low
moderate
high

trend fields use:

improving
stable
worsening
rising
falling

reserve_capacity uses:

high
medium
low

support_withdrawal_success uses:

yes
partial
no

Evaluation

Submissions must contain:

scenario_id,prediction
test_001,1
test_002,1
test_003,0
Run:

python scorer.py predictions.csv

Optional truth path:

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

The scorer reports:

Accuracy
Precision
Recall
F1
Confusion matrix
Structural Note

This benchmark contains counterfactual and misleading improvement cases designed to prevent shortcut learning from individual improving markers.

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

The goal is to evaluate whether models can detect genuine recovery geometry rather than superficial improvement.

License

MIT

Downloads last month: 34

Number of rows:

Total file size:

14.3 kB