Datasets:

ClarusC64
/

clinical-compensation-vs-recovery-v0.2

scenario_id string	heart_rate int64	map int64	lactate float64	urine_output int64	oxygen_requirement int64	vasopressor_dose float64	reserve_capacity string	support_dependency_trend string	organ_function_trend string	label int64
train_001	84	78	1.3	65	0	0	high	falling	improving	1
train_002	84	78	1.3	65	0	0.24	low	rising	stable	0
train_003	92	74	2	52	1	0	medium	falling	improving	1
train_004	92	74	2	52	1	0.2	low	rising	stable	0
train_005	88	76	1.8	58	0	0	high	stable	improving	1
train_006	88	76	1.8	58	0	0.18	low	rising	worsening	0
train_007	81	82	1.1	72	0	0	high	falling	improving	1
train_008	81	82	1.1	72	0	0.16	medium	rising	stable	0
train_009	96	72	2.4	45	1	0	medium	falling	improving	1
train_010	96	72	2.4	45	1	0.28	low	rising	worsening	0
train_011	86	79	1.5	62	0	0	high	stable	improving	1
train_012	86	79	1.5	62	0	0.22	low	rising	stable	0
train_013	90	75	1.9	55	1	0	medium	falling	stable	1
train_014	90	75	1.9	55	1	0.19	low	rising	worsening	0
train_015	83	80	1.4	68	0	0	high	falling	improving	1
train_016	83	80	1.4	68	0	0.25	low	stable	worsening	0
train_017	94	73	2.2	50	1	0	medium	falling	improving	1
train_018	94	73	2.2	50	1	0.27	low	rising	stable	0
train_019	87	77	1.7	60	0	0	high	stable	improving	1
train_020	87	77	1.7	60	0	0.21	low	rising	worsening	0
train_021	89	78	1.6	61	0	0	high	falling	improving	1
train_022	91	76	1.9	54	1	0.12	medium	falling	improving	1
train_023	85	80	1.3	66	0	0	high	stable	improving	1
train_024	93	74	2.1	51	1	0.08	medium	falling	stable	1
train_025	82	81	1.2	70	0	0	high	falling	improving	1
train_026	97	72	2.5	44	1	0.1	medium	falling	improving	1
train_027	88	79	1.6	64	0	0	high	stable	stable	1
train_028	95	73	2.3	48	1	0.09	medium	falling	improving	1
train_029	80	83	1.1	74	0	0	high	falling	improving	1
train_030	90	76	1.8	56	1	0.05	medium	falling	improving	1
train_031	89	78	1.6	61	0	0.23	low	rising	stable	0
train_032	91	76	1.9	54	1	0.26	low	rising	worsening	0
train_033	85	80	1.3	66	0	0.18	medium	rising	stable	0
train_034	93	74	2.1	51	1	0.3	low	rising	worsening	0
train_035	82	81	1.2	70	0	0.2	medium	rising	stable	0
train_036	97	72	2.5	44	1	0.32	low	rising	worsening	0
train_037	88	79	1.6	64	0	0.17	low	stable	worsening	0
train_038	95	73	2.3	48	1	0.29	low	rising	stable	0
train_039	80	83	1.1	74	0	0.15	medium	rising	stable	0
train_040	90	76	1.8	56	1	0.24	low	rising	worsening	0

What this dataset does

This dataset tests whether a model can distinguish genuine physiological recovery from compensated stability.

A patient may show normal blood pressure, heart rate, and lactate while still depending on external support.

The task is to classify whether the patient is genuinely recovered or merely compensated.

What changed in v0.2

v0.2 adds counterfactual twin cases.

Many rows are intentionally similar across visible observations.

The key distinction is whether stability is self-sustaining or externally maintained.

v0.2 also adds reserve and support-dependency signals.

These make the task harder than v0.1.

Core stability idea

Observable stability is not the same as recovery.

A patient can look stable because the underlying system has recovered.

A patient can also look stable because support is masking continued instability.

Correct classification requires reasoning across support dependence, reserve capacity, and organ function trends rather than current observations alone.

Prediction target

The label column is binary.

Label 1 means genuine recovery.

Label 0 means compensated state.

Row structure

Each row contains:

scenario_id
heart_rate
map
lactate
urine_output
oxygen_requirement
vasopressor_dose
reserve_capacity
support_dependency_trend
organ_function_trend
label

oxygen_requirement uses:

0 = room air or minimal support
1 = low-flow oxygen

reserve_capacity uses:

high
medium
low

support_dependency_trend uses:

falling
stable
rising

organ_function_trend uses:

improving
stable
worsening

Evaluation

Submissions must contain:

scenario_id,prediction
test_001,1
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 twin cases designed to prevent shortcut learning from individual variables.

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

The goal is to evaluate whether models can distinguish visible stability from true recovery.

License

MIT

Downloads last month: 27

Number of rows:

Total file size:

11.5 kB