Datasets:

ClarusC64
/

clinical-boundary-distance-v0.2

scenario_id stringlengths 9 9	oxygen_requirement int64 0 3	resp_rate int64 18 30	map int64 64 83	vasopressor_requirement stringclasses 4 values	lactate float64 1.1 4.9	urine_output int64 14 74	mental_status stringclasses 3 values	support_escalation_trend stringclasses 3 values	reserve_loss_trend stringclasses 4 values	buffer_capacity stringclasses 4 values	boundary_velocity stringclasses 4 values	label int64 0 2
train_001	0	18	82	none	1.2	70	baseline	stable	none	high	stable	0
train_002	1	20	78	none	1.6	60	baseline	stable	mild	high	stable	0
train_003	1	22	75	low	2.1	52	baseline	rising	mild	medium	slow	1
train_004	1	22	75	low	2.1	52	baseline	rising	severe	low	fast	2
train_005	2	24	71	low	2.8	38	mild_confusion	rising	moderate	medium	slow	1
train_006	2	24	71	low	2.8	38	mild_confusion	rising	severe	low	fast	2
train_007	2	25	70	moderate	3	34	mild_confusion	stable	moderate	medium	stable	1
train_008	2	25	70	moderate	3	34	mild_confusion	rising	severe	low	fast	2
train_009	3	28	67	moderate	3.8	26	confused	stable	moderate	medium	slow	1
train_010	3	28	67	moderate	3.8	26	confused	rising	severe	low	fast	2
train_011	0	18	81	low	1.5	64	baseline	stable	mild	high	stable	0
train_012	0	18	81	low	1.5	64	baseline	rising	moderate	medium	slow	1
train_013	1	21	76	none	1.8	56	baseline	stable	mild	high	stable	0
train_014	1	21	76	none	1.8	56	baseline	rising	moderate	medium	slow	1
train_015	1	23	74	moderate	2.5	44	mild_confusion	stable	moderate	medium	slow	1
train_016	1	23	74	moderate	2.5	44	mild_confusion	rising	severe	low	fast	2
train_017	2	26	69	moderate	3.4	30	confused	rising	severe	low	fast	2
train_018	2	26	69	moderate	3.4	30	confused	stable	moderate	medium	slow	1
train_019	3	30	64	high	4.9	14	confused	rising	severe	very_low	fast	2
train_020	3	30	64	high	4.9	14	confused	stable	severe	low	slow	2
train_021	0	18	83	none	1.1	74	baseline	falling	none	high	improving	0
train_022	1	20	79	low	1.7	58	baseline	falling	mild	high	improving	0
train_023	2	24	72	low	2.6	40	baseline	falling	moderate	medium	slow	1
train_024	2	24	72	low	2.6	40	mild_confusion	rising	severe	low	fast	2
train_025	2	25	70	none	2.3	50	baseline	stable	mild	medium	stable	1
train_026	2	25	70	none	2.3	50	baseline	rising	moderate	medium	slow	1
train_027	1	22	76	high	2.2	48	mild_confusion	rising	severe	low	fast	2
train_028	1	22	76	high	2.2	48	mild_confusion	stable	moderate	medium	slow	1
train_029	3	29	66	low	3.6	28	confused	stable	moderate	medium	slow	1
train_030	3	29	66	low	3.6	28	confused	rising	severe	low	fast	2
train_031	0	19	80	none	1.4	66	baseline	stable	none	high	stable	0
train_032	1	21	77	low	1.9	54	baseline	stable	mild	high	stable	0
train_033	2	24	71	moderate	3.1	36	mild_confusion	falling	moderate	medium	slow	1
train_034	2	24	71	moderate	3.1	36	mild_confusion	rising	severe	low	fast	2
train_035	3	27	68	moderate	3.7	25	confused	falling	moderate	medium	slow	1
train_036	1	23	74	low	2.4	45	baseline	rising	severe	low	fast	2
train_037	0	18	82	low	1.5	62	baseline	rising	mild	medium	slow	1
train_038	2	25	70	none	2.2	52	baseline	stable	mild	high	stable	0
train_039	3	30	65	high	4.7	16	confused	rising	severe	very_low	fast	2
train_040	1	21	78	none	1.7	57	baseline	falling	mild	high	improving	0

What this dataset does

This dataset tests whether a model can estimate how close a patient is to a clinical collapse boundary.

The task is not to identify whether collapse has already occurred.

The task is to classify remaining margin before ordinary monitoring or support may no longer be enough.

What changed in v0.2

v0.2 adds counterfactual and adversarial cases.

Some rows have the same MAP, oxygen requirement, respiratory rate, and lactate but different boundary distance.

Some high-looking cases have preserved buffer and stable boundary velocity.

Some moderate-looking cases are critically close to a boundary because support escalation is rising and reserve loss is fast.

This makes the task harder than v0.1.

Core stability idea

Collapse is often preceded by margin loss.

Boundary distance depends on more than current severity.

A patient may look severe but retain some buffer if boundary velocity is slow.

A patient may look only moderately unwell but be close to collapse if buffer is low and boundary velocity is fast.

Correct classification requires reasoning across current state, support escalation, reserve loss, buffer capacity, and direction of movement.

Prediction target

The label column has three classes.

Label 0 means safe margin.

Label 1 means narrowing margin.

Label 2 means critical boundary proximity.

Row structure

Each row contains:

scenario_id
oxygen_requirement
resp_rate
map
vasopressor_requirement
lactate
urine_output
mental_status
support_escalation_trend
reserve_loss_trend
buffer_capacity
boundary_velocity
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

mental_status uses:

baseline
mild_confusion
confused

support_escalation_trend uses:

falling
stable
rising

reserve_loss_trend uses:

none
mild
moderate
severe

buffer_capacity uses:

high
medium
low
very_low

boundary_velocity uses:

improving
stable
slow
fast

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 MAP, oxygen requirement, lactate, or current severity.

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

The goal is to evaluate whether models can detect margin loss before collapse becomes visible.

License

MIT

Downloads last month: 20

Number of rows:

Total file size:

13.3 kB