Datasets:

ClarusC64
/

clinical-escalation-discipline-v0.2

scenario_id string	news_score int64	heart_rate int64	resp_rate int64	map int64	lactate float64	oxygen_requirement int64	urine_output int64	current_severity string	lactate_trend string	urine_output_trend string	treatment_response string	reserve_capacity string	label int64
train_001	3	88	20	77	1.5	0	62	low	stable	stable	improving	high	0
train_002	5	101	24	72	2.2	1	48	moderate	falling	improving	improving	medium	0
train_003	6	108	26	70	2.8	2	40	high	falling	improving	partial	medium	0
train_004	4	94	22	75	1.8	1	55	moderate	rising	worsening	poor	medium	1
train_005	5	99	23	73	2.1	1	50	moderate	rising	worsening	poor	low	1
train_006	3	90	20	76	1.6	0	60	low	rising	worsening	poor	medium	1
train_007	6	110	27	69	3	2	38	high	falling	improving	partial	medium	0
train_008	4	95	22	74	1.9	1	54	moderate	falling	stable	partial	medium	0
train_009	5	100	24	72	2.4	1	46	moderate	rising	worsening	poor	low	1
train_010	6	106	25	70	2.7	2	42	high	rising	worsening	poor	low	1
train_011	3	87	20	78	1.4	0	64	low	stable	stable	improving	high	0
train_012	5	102	24	72	2.3	1	47	moderate	falling	improving	partial	medium	0
train_013	4	93	22	75	1.7	1	56	moderate	rising	worsening	poor	medium	1
train_014	6	109	26	70	2.9	2	39	high	falling	improving	improving	medium	0
train_015	5	98	23	73	2	1	51	moderate	rising	worsening	poor	low	1
train_016	7	115	29	67	3.5	2	32	high	rising	worsening	none	low	1
train_017	4	92	21	76	1.6	1	58	moderate	stable	stable	partial	medium	0
train_018	5	103	24	71	2.5	1	45	moderate	rising	worsening	poor	low	1
train_019	6	111	27	69	3.1	2	37	high	falling	improving	partial	medium	0
train_020	3	89	20	77	1.5	0	61	low	rising	worsening	poor	medium	1
train_021	4	96	22	74	1.9	1	53	moderate	falling	improving	partial	medium	0
train_022	5	100	23	73	2.2	1	49	moderate	rising	worsening	poor	low	1
train_023	6	107	26	70	2.8	2	41	high	falling	improving	partial	medium	0
train_024	4	94	22	75	1.8	1	55	moderate	rising	worsening	none	medium	1
train_025	5	101	24	72	2.4	1	46	moderate	falling	improving	partial	medium	0
train_026	6	110	27	69	3	2	38	high	rising	worsening	poor	low	1
train_027	3	86	19	79	1.3	0	66	low	stable	stable	improving	high	0
train_028	5	99	23	73	2.1	1	50	moderate	rising	worsening	none	low	1
train_029	7	116	29	66	3.6	2	31	high	falling	improving	partial	medium	0
train_030	4	95	22	74	1.9	1	54	moderate	rising	worsening	poor	medium	1
train_031	5	102	24	72	2.3	1	48	moderate	falling	improving	improving	medium	0
train_032	3	90	20	76	1.6	0	60	low	rising	worsening	none	medium	1
train_033	6	108	26	70	2.9	2	40	high	falling	improving	partial	medium	0
train_034	5	100	24	72	2.2	1	49	moderate	rising	worsening	poor	low	1
train_035	4	93	21	76	1.7	1	57	moderate	stable	stable	partial	medium	0
train_036	7	114	28	67	3.4	2	33	high	rising	worsening	none	low	1
train_037	5	101	24	72	2.3	1	47	moderate	falling	improving	partial	medium	0
train_038	4	94	22	75	1.8	1	55	moderate	rising	worsening	poor	medium	1
train_039	6	109	26	70	3	2	39	high	falling	improving	improving	medium	0
train_040	5	98	23	73	2	1	52	moderate	rising	worsening	poor	low	1

What this dataset does

This dataset tests whether a model can decide when a patient should be escalated rather than monitored.

The task is not to identify the sickest-looking patient.

The task is to determine whether the current pattern requires escalation.

What changed in v0.2

v0.2 adds adversarial cases where the same NEWS score can have different labels.

Some high-score patients are improving and should be monitored.

Some moderate or low-score patients are deteriorating and should be escalated.

This makes the task harder than v0.1.

Core stability idea

Escalation depends on trajectory, reserve, and treatment response.

Visible severity alone is not enough.

A patient with a high score may be improving under current treatment.

A patient with a lower score may require escalation if treatment response is poor and reserve is falling.

Prediction target

The label column is binary.

Label 1 means escalate.

Label 0 means monitor.

Row structure

Each row contains:

scenario_id
news_score
heart_rate
resp_rate
map
lactate
oxygen_requirement
urine_output
current_severity
lactate_trend
urine_output_trend
treatment_response
reserve_capacity
label

oxygen_requirement uses:

0 = room air or minimal support
1 = low oxygen requirement
2 = high oxygen requirement

current_severity uses:

low
moderate
high

trend fields use:

stable
rising
falling
improving
worsening

treatment_response uses:

improving
partial
poor
none

reserve_capacity uses:

high
medium
low

Evaluation

Submissions must contain:

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

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 adversarial cases designed to prevent shortcut learning from NEWS score or visible severity.

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

The goal is to evaluate whether models can apply escalation discipline under clinical uncertainty.

License

MIT

Downloads last month: 20

Number of rows:

Total file size:

12.5 kB